1/*
2 * CDDL HEADER START
3 *
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.
7 *
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.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2011 by Delphix. All rights reserved.
25 */
26
27/*
28 * DVA-based Adjustable Replacement Cache
29 *
30 * While much of the theory of operation used here is
31 * based on the self-tuning, low overhead replacement cache
32 * presented by Megiddo and Modha at FAST 2003, there are some
33 * significant differences:
34 *
35 * 1. The Megiddo and Modha model assumes any page is evictable.
36 * Pages in its cache cannot be "locked" into memory. This makes
37 * the eviction algorithm simple: evict the last page in the list.
38 * This also make the performance characteristics easy to reason
39 * about. Our cache is not so simple. At any given moment, some
40 * subset of the blocks in the cache are un-evictable because we
41 * have handed out a reference to them. Blocks are only evictable
42 * when there are no external references active. This makes
43 * eviction far more problematic: we choose to evict the evictable
44 * blocks that are the "lowest" in the list.
45 *
46 * There are times when it is not possible to evict the requested
47 * space. In these circumstances we are unable to adjust the cache
48 * size. To prevent the cache growing unbounded at these times we
49 * implement a "cache throttle" that slows the flow of new data
50 * into the cache until we can make space available.
51 *
52 * 2. The Megiddo and Modha model assumes a fixed cache size.
53 * Pages are evicted when the cache is full and there is a cache
54 * miss. Our model has a variable sized cache. It grows with
55 * high use, but also tries to react to memory pressure from the
56 * operating system: decreasing its size when system memory is
57 * tight.
58 *
59 * 3. The Megiddo and Modha model assumes a fixed page size. All
60 * elements of the cache are therefor exactly the same size. So
61 * when adjusting the cache size following a cache miss, its simply
62 * a matter of choosing a single page to evict. In our model, we
63 * have variable sized cache blocks (rangeing from 512 bytes to
64 * 128K bytes). We therefor choose a set of blocks to evict to make
65 * space for a cache miss that approximates as closely as possible
66 * the space used by the new block.
67 *
68 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
69 * by N. Megiddo & D. Modha, FAST 2003
70 */
71
72/*
73 * The locking model:
74 *
75 * A new reference to a cache buffer can be obtained in two
76 * ways: 1) via a hash table lookup using the DVA as a key,
77 * or 2) via one of the ARC lists. The arc_read() interface
78 * uses method 1, while the internal arc algorithms for
79 * adjusting the cache use method 2. We therefor provide two
80 * types of locks: 1) the hash table lock array, and 2) the
81 * arc list locks.
82 *
83 * Buffers do not have their own mutexs, rather they rely on the
84 * hash table mutexs for the bulk of their protection (i.e. most
85 * fields in the arc_buf_hdr_t are protected by these mutexs).
86 *
87 * buf_hash_find() returns the appropriate mutex (held) when it
88 * locates the requested buffer in the hash table. It returns
89 * NULL for the mutex if the buffer was not in the table.
90 *
91 * buf_hash_remove() expects the appropriate hash mutex to be
92 * already held before it is invoked.
93 *
94 * Each arc state also has a mutex which is used to protect the
95 * buffer list associated with the state. When attempting to
96 * obtain a hash table lock while holding an arc list lock you
97 * must use: mutex_tryenter() to avoid deadlock. Also note that
98 * the active state mutex must be held before the ghost state mutex.
99 *
100 * Arc buffers may have an associated eviction callback function.
101 * This function will be invoked prior to removing the buffer (e.g.
102 * in arc_do_user_evicts()). Note however that the data associated
103 * with the buffer may be evicted prior to the callback. The callback
104 * must be made with *no locks held* (to prevent deadlock). Additionally,
105 * the users of callbacks must ensure that their private data is
106 * protected from simultaneous callbacks from arc_buf_evict()
107 * and arc_do_user_evicts().
108 *
109 * Note that the majority of the performance stats are manipulated
110 * with atomic operations.
111 *
112 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
113 *
114 * - L2ARC buflist creation
115 * - L2ARC buflist eviction
116 * - L2ARC write completion, which walks L2ARC buflists
117 * - ARC header destruction, as it removes from L2ARC buflists
118 * - ARC header release, as it removes from L2ARC buflists
119 */
120
121#include <sys/spa.h>
122#include <sys/zio.h>
123#include <sys/zfs_context.h>
124#include <sys/arc.h>
125#include <sys/refcount.h>
126#include <sys/vdev.h>
127#include <sys/vdev_impl.h>
128#ifdef _KERNEL
129#include <sys/dnlc.h>
130#endif
131#include <sys/callb.h>
132#include <sys/kstat.h>
133#include <zfs_fletcher.h>
134#include <sys/sdt.h>
135
136#include <vm/vm_pageout.h>
137
138#ifdef illumos
139#ifndef _KERNEL
140/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
141boolean_t arc_watch = B_FALSE;
142int arc_procfd;
143#endif
144#endif /* illumos */
145
146static kmutex_t arc_reclaim_thr_lock;
147static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
148static uint8_t arc_thread_exit;
149
150extern int zfs_write_limit_shift;
151extern uint64_t zfs_write_limit_max;
152extern kmutex_t zfs_write_limit_lock;
153
154#define ARC_REDUCE_DNLC_PERCENT 3
155uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
156
157typedef enum arc_reclaim_strategy {
158 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
159 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
160} arc_reclaim_strategy_t;
161
162/* number of seconds before growing cache again */
163static int arc_grow_retry = 60;
164
165/* shift of arc_c for calculating both min and max arc_p */
166static int arc_p_min_shift = 4;
167
168/* log2(fraction of arc to reclaim) */
169static int arc_shrink_shift = 5;
170
171/*
172 * minimum lifespan of a prefetch block in clock ticks
173 * (initialized in arc_init())
174 */
175static int arc_min_prefetch_lifespan;
176
177static int arc_dead;
178extern int zfs_prefetch_disable;
179
180/*
181 * The arc has filled available memory and has now warmed up.
182 */
183static boolean_t arc_warm;
184
185/*
186 * These tunables are for performance analysis.
187 */
188uint64_t zfs_arc_max;
189uint64_t zfs_arc_min;
190uint64_t zfs_arc_meta_limit = 0;
191int zfs_arc_grow_retry = 0;
192int zfs_arc_shrink_shift = 0;
193int zfs_arc_p_min_shift = 0;
194int zfs_disable_dup_eviction = 0;
195
196TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
197TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
198TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
199SYSCTL_DECL(_vfs_zfs);
200SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0,
201 "Maximum ARC size");
202SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0,
203 "Minimum ARC size");
204
205/*
206 * Note that buffers can be in one of 6 states:
207 * ARC_anon - anonymous (discussed below)
208 * ARC_mru - recently used, currently cached
209 * ARC_mru_ghost - recentely used, no longer in cache
210 * ARC_mfu - frequently used, currently cached
211 * ARC_mfu_ghost - frequently used, no longer in cache
212 * ARC_l2c_only - exists in L2ARC but not other states
213 * When there are no active references to the buffer, they are
214 * are linked onto a list in one of these arc states. These are
215 * the only buffers that can be evicted or deleted. Within each
216 * state there are multiple lists, one for meta-data and one for
217 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
218 * etc.) is tracked separately so that it can be managed more
219 * explicitly: favored over data, limited explicitly.
220 *
221 * Anonymous buffers are buffers that are not associated with
222 * a DVA. These are buffers that hold dirty block copies
223 * before they are written to stable storage. By definition,
224 * they are "ref'd" and are considered part of arc_mru
225 * that cannot be freed. Generally, they will aquire a DVA
226 * as they are written and migrate onto the arc_mru list.
227 *
228 * The ARC_l2c_only state is for buffers that are in the second
229 * level ARC but no longer in any of the ARC_m* lists. The second
230 * level ARC itself may also contain buffers that are in any of
231 * the ARC_m* states - meaning that a buffer can exist in two
232 * places. The reason for the ARC_l2c_only state is to keep the
233 * buffer header in the hash table, so that reads that hit the
234 * second level ARC benefit from these fast lookups.
235 */
236
237#define ARCS_LOCK_PAD CACHE_LINE_SIZE
238struct arcs_lock {
239 kmutex_t arcs_lock;
240#ifdef _KERNEL
241 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))];
242#endif
243};
244
245/*
246 * must be power of two for mask use to work
247 *
248 */
249#define ARC_BUFC_NUMDATALISTS 16
250#define ARC_BUFC_NUMMETADATALISTS 16
251#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS)
252
253typedef struct arc_state {
254 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
255 uint64_t arcs_size; /* total amount of data in this state */
256 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */
257 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE);
258} arc_state_t;
259
260#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock))
261
262/* The 6 states: */
263static arc_state_t ARC_anon;
264static arc_state_t ARC_mru;
265static arc_state_t ARC_mru_ghost;
266static arc_state_t ARC_mfu;
267static arc_state_t ARC_mfu_ghost;
268static arc_state_t ARC_l2c_only;
269
270typedef struct arc_stats {
271 kstat_named_t arcstat_hits;
272 kstat_named_t arcstat_misses;
273 kstat_named_t arcstat_demand_data_hits;
274 kstat_named_t arcstat_demand_data_misses;
275 kstat_named_t arcstat_demand_metadata_hits;
276 kstat_named_t arcstat_demand_metadata_misses;
277 kstat_named_t arcstat_prefetch_data_hits;
278 kstat_named_t arcstat_prefetch_data_misses;
279 kstat_named_t arcstat_prefetch_metadata_hits;
280 kstat_named_t arcstat_prefetch_metadata_misses;
281 kstat_named_t arcstat_mru_hits;
282 kstat_named_t arcstat_mru_ghost_hits;
283 kstat_named_t arcstat_mfu_hits;
284 kstat_named_t arcstat_mfu_ghost_hits;
285 kstat_named_t arcstat_allocated;
286 kstat_named_t arcstat_deleted;
287 kstat_named_t arcstat_stolen;
288 kstat_named_t arcstat_recycle_miss;
289 kstat_named_t arcstat_mutex_miss;
290 kstat_named_t arcstat_evict_skip;
291 kstat_named_t arcstat_evict_l2_cached;
292 kstat_named_t arcstat_evict_l2_eligible;
293 kstat_named_t arcstat_evict_l2_ineligible;
294 kstat_named_t arcstat_hash_elements;
295 kstat_named_t arcstat_hash_elements_max;
296 kstat_named_t arcstat_hash_collisions;
297 kstat_named_t arcstat_hash_chains;
298 kstat_named_t arcstat_hash_chain_max;
299 kstat_named_t arcstat_p;
300 kstat_named_t arcstat_c;
301 kstat_named_t arcstat_c_min;
302 kstat_named_t arcstat_c_max;
303 kstat_named_t arcstat_size;
304 kstat_named_t arcstat_hdr_size;
305 kstat_named_t arcstat_data_size;
306 kstat_named_t arcstat_other_size;
307 kstat_named_t arcstat_l2_hits;
308 kstat_named_t arcstat_l2_misses;
309 kstat_named_t arcstat_l2_feeds;
310 kstat_named_t arcstat_l2_rw_clash;
311 kstat_named_t arcstat_l2_read_bytes;
312 kstat_named_t arcstat_l2_write_bytes;
313 kstat_named_t arcstat_l2_writes_sent;
314 kstat_named_t arcstat_l2_writes_done;
315 kstat_named_t arcstat_l2_writes_error;
316 kstat_named_t arcstat_l2_writes_hdr_miss;
317 kstat_named_t arcstat_l2_evict_lock_retry;
318 kstat_named_t arcstat_l2_evict_reading;
319 kstat_named_t arcstat_l2_free_on_write;
320 kstat_named_t arcstat_l2_abort_lowmem;
321 kstat_named_t arcstat_l2_cksum_bad;
322 kstat_named_t arcstat_l2_io_error;
323 kstat_named_t arcstat_l2_size;
324 kstat_named_t arcstat_l2_hdr_size;
325 kstat_named_t arcstat_l2_write_trylock_fail;
326 kstat_named_t arcstat_l2_write_passed_headroom;
327 kstat_named_t arcstat_l2_write_spa_mismatch;
328 kstat_named_t arcstat_l2_write_in_l2;
329 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
330 kstat_named_t arcstat_l2_write_not_cacheable;
331 kstat_named_t arcstat_l2_write_full;
332 kstat_named_t arcstat_l2_write_buffer_iter;
333 kstat_named_t arcstat_l2_write_pios;
334 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
335 kstat_named_t arcstat_l2_write_buffer_list_iter;
336 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
337 kstat_named_t arcstat_memory_throttle_count;
338 kstat_named_t arcstat_duplicate_buffers;
339 kstat_named_t arcstat_duplicate_buffers_size;
340 kstat_named_t arcstat_duplicate_reads;
341} arc_stats_t;
342
343static arc_stats_t arc_stats = {
344 { "hits", KSTAT_DATA_UINT64 },
345 { "misses", KSTAT_DATA_UINT64 },
346 { "demand_data_hits", KSTAT_DATA_UINT64 },
347 { "demand_data_misses", KSTAT_DATA_UINT64 },
348 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
349 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
350 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
351 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
352 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
353 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
354 { "mru_hits", KSTAT_DATA_UINT64 },
355 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
356 { "mfu_hits", KSTAT_DATA_UINT64 },
357 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
358 { "allocated", KSTAT_DATA_UINT64 },
359 { "deleted", KSTAT_DATA_UINT64 },
360 { "stolen", KSTAT_DATA_UINT64 },
361 { "recycle_miss", KSTAT_DATA_UINT64 },
362 { "mutex_miss", KSTAT_DATA_UINT64 },
363 { "evict_skip", KSTAT_DATA_UINT64 },
364 { "evict_l2_cached", KSTAT_DATA_UINT64 },
365 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
366 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
367 { "hash_elements", KSTAT_DATA_UINT64 },
368 { "hash_elements_max", KSTAT_DATA_UINT64 },
369 { "hash_collisions", KSTAT_DATA_UINT64 },
370 { "hash_chains", KSTAT_DATA_UINT64 },
371 { "hash_chain_max", KSTAT_DATA_UINT64 },
372 { "p", KSTAT_DATA_UINT64 },
373 { "c", KSTAT_DATA_UINT64 },
374 { "c_min", KSTAT_DATA_UINT64 },
375 { "c_max", KSTAT_DATA_UINT64 },
376 { "size", KSTAT_DATA_UINT64 },
377 { "hdr_size", KSTAT_DATA_UINT64 },
378 { "data_size", KSTAT_DATA_UINT64 },
379 { "other_size", KSTAT_DATA_UINT64 },
380 { "l2_hits", KSTAT_DATA_UINT64 },
381 { "l2_misses", KSTAT_DATA_UINT64 },
382 { "l2_feeds", KSTAT_DATA_UINT64 },
383 { "l2_rw_clash", KSTAT_DATA_UINT64 },
384 { "l2_read_bytes", KSTAT_DATA_UINT64 },
385 { "l2_write_bytes", KSTAT_DATA_UINT64 },
386 { "l2_writes_sent", KSTAT_DATA_UINT64 },
387 { "l2_writes_done", KSTAT_DATA_UINT64 },
388 { "l2_writes_error", KSTAT_DATA_UINT64 },
389 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
390 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
391 { "l2_evict_reading", KSTAT_DATA_UINT64 },
392 { "l2_free_on_write", KSTAT_DATA_UINT64 },
393 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
394 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
395 { "l2_io_error", KSTAT_DATA_UINT64 },
396 { "l2_size", KSTAT_DATA_UINT64 },
397 { "l2_hdr_size", KSTAT_DATA_UINT64 },
398 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
399 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
400 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
401 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
402 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
403 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
404 { "l2_write_full", KSTAT_DATA_UINT64 },
405 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
406 { "l2_write_pios", KSTAT_DATA_UINT64 },
407 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
408 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
409 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
410 { "memory_throttle_count", KSTAT_DATA_UINT64 },
411 { "duplicate_buffers", KSTAT_DATA_UINT64 },
412 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
413 { "duplicate_reads", KSTAT_DATA_UINT64 }
414};
415
416#define ARCSTAT(stat) (arc_stats.stat.value.ui64)
417
418#define ARCSTAT_INCR(stat, val) \
419 atomic_add_64(&arc_stats.stat.value.ui64, (val));
420
421#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
422#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
423
424#define ARCSTAT_MAX(stat, val) { \
425 uint64_t m; \
426 while ((val) > (m = arc_stats.stat.value.ui64) && \
427 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
428 continue; \
429}
430
431#define ARCSTAT_MAXSTAT(stat) \
432 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
433
434/*
435 * We define a macro to allow ARC hits/misses to be easily broken down by
436 * two separate conditions, giving a total of four different subtypes for
437 * each of hits and misses (so eight statistics total).
438 */
439#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
440 if (cond1) { \
441 if (cond2) { \
442 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
443 } else { \
444 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
445 } \
446 } else { \
447 if (cond2) { \
448 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
449 } else { \
450 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
451 } \
452 }
453
454kstat_t *arc_ksp;
455static arc_state_t *arc_anon;
456static arc_state_t *arc_mru;
457static arc_state_t *arc_mru_ghost;
458static arc_state_t *arc_mfu;
459static arc_state_t *arc_mfu_ghost;
460static arc_state_t *arc_l2c_only;
461
462/*
463 * There are several ARC variables that are critical to export as kstats --
464 * but we don't want to have to grovel around in the kstat whenever we wish to
465 * manipulate them. For these variables, we therefore define them to be in
466 * terms of the statistic variable. This assures that we are not introducing
467 * the possibility of inconsistency by having shadow copies of the variables,
468 * while still allowing the code to be readable.
469 */
470#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
471#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
472#define arc_c ARCSTAT(arcstat_c) /* target size of cache */
473#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
474#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
475
476static int arc_no_grow; /* Don't try to grow cache size */
477static uint64_t arc_tempreserve;
478static uint64_t arc_loaned_bytes;
479static uint64_t arc_meta_used;
480static uint64_t arc_meta_limit;
481static uint64_t arc_meta_max = 0;
482SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0,
483 "ARC metadata used");
484SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0,
485 "ARC metadata limit");
486
487typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
488
489typedef struct arc_callback arc_callback_t;
490
491struct arc_callback {
492 void *acb_private;
493 arc_done_func_t *acb_done;
494 arc_buf_t *acb_buf;
495 zio_t *acb_zio_dummy;
496 arc_callback_t *acb_next;
497};
498
499typedef struct arc_write_callback arc_write_callback_t;
500
501struct arc_write_callback {
502 void *awcb_private;
503 arc_done_func_t *awcb_ready;
504 arc_done_func_t *awcb_done;
505 arc_buf_t *awcb_buf;
506};
507
508struct arc_buf_hdr {
509 /* protected by hash lock */
510 dva_t b_dva;
511 uint64_t b_birth;
512 uint64_t b_cksum0;
513
514 kmutex_t b_freeze_lock;
515 zio_cksum_t *b_freeze_cksum;
516 void *b_thawed;
517
518 arc_buf_hdr_t *b_hash_next;
519 arc_buf_t *b_buf;
520 uint32_t b_flags;
521 uint32_t b_datacnt;
522
523 arc_callback_t *b_acb;
524 kcondvar_t b_cv;
525
526 /* immutable */
527 arc_buf_contents_t b_type;
528 uint64_t b_size;
529 uint64_t b_spa;
530
531 /* protected by arc state mutex */
532 arc_state_t *b_state;
533 list_node_t b_arc_node;
534
535 /* updated atomically */
536 clock_t b_arc_access;
537
538 /* self protecting */
539 refcount_t b_refcnt;
540
541 l2arc_buf_hdr_t *b_l2hdr;
542 list_node_t b_l2node;
543};
544
545static arc_buf_t *arc_eviction_list;
546static kmutex_t arc_eviction_mtx;
547static arc_buf_hdr_t arc_eviction_hdr;
548static void arc_get_data_buf(arc_buf_t *buf);
549static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
550static int arc_evict_needed(arc_buf_contents_t type);
551static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
552#ifdef illumos
553static void arc_buf_watch(arc_buf_t *buf);
554#endif /* illumos */
555
556static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
557
558#define GHOST_STATE(state) \
559 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
560 (state) == arc_l2c_only)
561
562/*
563 * Private ARC flags. These flags are private ARC only flags that will show up
564 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
565 * be passed in as arc_flags in things like arc_read. However, these flags
566 * should never be passed and should only be set by ARC code. When adding new
567 * public flags, make sure not to smash the private ones.
568 */
569
570#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
571#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
572#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
573#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
574#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
575#define ARC_INDIRECT (1 << 14) /* this is an indirect block */
576#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
577#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
578#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
579#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
580
581#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
582#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
583#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
584#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
585#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
586#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
587#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
588#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
589#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
590 (hdr)->b_l2hdr != NULL)
591#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
592#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
593#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
594
595/*
596 * Other sizes
597 */
598
599#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
600#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
601
602/*
603 * Hash table routines
604 */
605
606#define HT_LOCK_PAD CACHE_LINE_SIZE
607
608struct ht_lock {
609 kmutex_t ht_lock;
610#ifdef _KERNEL
611 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
612#endif
613};
614
615#define BUF_LOCKS 256
616typedef struct buf_hash_table {
617 uint64_t ht_mask;
618 arc_buf_hdr_t **ht_table;
619 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
620} buf_hash_table_t;
621
622static buf_hash_table_t buf_hash_table;
623
624#define BUF_HASH_INDEX(spa, dva, birth) \
625 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
626#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
627#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
628#define HDR_LOCK(hdr) \
629 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
630
631uint64_t zfs_crc64_table[256];
632
633/*
634 * Level 2 ARC
635 */
636
637#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
638#define L2ARC_HEADROOM 2 /* num of writes */
639#define L2ARC_FEED_SECS 1 /* caching interval secs */
640#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
641
642#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
643#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
644
645/*
646 * L2ARC Performance Tunables
647 */
648uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
649uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
650uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
651uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
652uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
653boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
654boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
655boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
656
657SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
658 &l2arc_write_max, 0, "max write size");
659SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
660 &l2arc_write_boost, 0, "extra write during warmup");
661SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
662 &l2arc_headroom, 0, "number of dev writes");
663SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
664 &l2arc_feed_secs, 0, "interval seconds");
665SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
666 &l2arc_feed_min_ms, 0, "min interval milliseconds");
667
668SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
669 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
670SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
671 &l2arc_feed_again, 0, "turbo warmup");
672SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
673 &l2arc_norw, 0, "no reads during writes");
674
675SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
676 &ARC_anon.arcs_size, 0, "size of anonymous state");
677SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
678 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
679SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
680 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
681
682SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
683 &ARC_mru.arcs_size, 0, "size of mru state");
684SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
685 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
686SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
687 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
688
689SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
690 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state");
691SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
692 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
693 "size of metadata in mru ghost state");
694SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
695 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
696 "size of data in mru ghost state");
697
698SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
699 &ARC_mfu.arcs_size, 0, "size of mfu state");
700SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
701 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
702SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
703 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
704
705SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
706 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state");
707SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
708 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
709 "size of metadata in mfu ghost state");
710SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
711 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
712 "size of data in mfu ghost state");
713
714SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
715 &ARC_l2c_only.arcs_size, 0, "size of mru state");
716
717/*
718 * L2ARC Internals
719 */
720typedef struct l2arc_dev {
721 vdev_t *l2ad_vdev; /* vdev */
722 spa_t *l2ad_spa; /* spa */
723 uint64_t l2ad_hand; /* next write location */
724 uint64_t l2ad_write; /* desired write size, bytes */
725 uint64_t l2ad_boost; /* warmup write boost, bytes */
726 uint64_t l2ad_start; /* first addr on device */
727 uint64_t l2ad_end; /* last addr on device */
728 uint64_t l2ad_evict; /* last addr eviction reached */
729 boolean_t l2ad_first; /* first sweep through */
730 boolean_t l2ad_writing; /* currently writing */
731 list_t *l2ad_buflist; /* buffer list */
732 list_node_t l2ad_node; /* device list node */
733} l2arc_dev_t;
734
735static list_t L2ARC_dev_list; /* device list */
736static list_t *l2arc_dev_list; /* device list pointer */
737static kmutex_t l2arc_dev_mtx; /* device list mutex */
738static l2arc_dev_t *l2arc_dev_last; /* last device used */
739static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
740static list_t L2ARC_free_on_write; /* free after write buf list */
741static list_t *l2arc_free_on_write; /* free after write list ptr */
742static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
743static uint64_t l2arc_ndev; /* number of devices */
744
745typedef struct l2arc_read_callback {
746 arc_buf_t *l2rcb_buf; /* read buffer */
747 spa_t *l2rcb_spa; /* spa */
748 blkptr_t l2rcb_bp; /* original blkptr */
749 zbookmark_t l2rcb_zb; /* original bookmark */
750 int l2rcb_flags; /* original flags */
751} l2arc_read_callback_t;
752
753typedef struct l2arc_write_callback {
754 l2arc_dev_t *l2wcb_dev; /* device info */
755 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
756} l2arc_write_callback_t;
757
758struct l2arc_buf_hdr {
759 /* protected by arc_buf_hdr mutex */
760 l2arc_dev_t *b_dev; /* L2ARC device */
761 uint64_t b_daddr; /* disk address, offset byte */
762};
763
764typedef struct l2arc_data_free {
765 /* protected by l2arc_free_on_write_mtx */
766 void *l2df_data;
767 size_t l2df_size;
768 void (*l2df_func)(void *, size_t);
769 list_node_t l2df_list_node;
770} l2arc_data_free_t;
771
772static kmutex_t l2arc_feed_thr_lock;
773static kcondvar_t l2arc_feed_thr_cv;
774static uint8_t l2arc_thread_exit;
775
776static void l2arc_read_done(zio_t *zio);
777static void l2arc_hdr_stat_add(void);
778static void l2arc_hdr_stat_remove(void);
779
780static uint64_t
781buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
782{
783 uint8_t *vdva = (uint8_t *)dva;
784 uint64_t crc = -1ULL;
785 int i;
786
787 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
788
789 for (i = 0; i < sizeof (dva_t); i++)
790 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
791
792 crc ^= (spa>>8) ^ birth;
793
794 return (crc);
795}
796
797#define BUF_EMPTY(buf) \
798 ((buf)->b_dva.dva_word[0] == 0 && \
799 (buf)->b_dva.dva_word[1] == 0 && \
800 (buf)->b_birth == 0)
801
802#define BUF_EQUAL(spa, dva, birth, buf) \
803 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
804 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
805 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
806
807static void
808buf_discard_identity(arc_buf_hdr_t *hdr)
809{
810 hdr->b_dva.dva_word[0] = 0;
811 hdr->b_dva.dva_word[1] = 0;
812 hdr->b_birth = 0;
813 hdr->b_cksum0 = 0;
814}
815
816static arc_buf_hdr_t *
817buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
818{
819 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
820 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
821 arc_buf_hdr_t *buf;
822
823 mutex_enter(hash_lock);
824 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
825 buf = buf->b_hash_next) {
826 if (BUF_EQUAL(spa, dva, birth, buf)) {
827 *lockp = hash_lock;
828 return (buf);
829 }
830 }
831 mutex_exit(hash_lock);
832 *lockp = NULL;
833 return (NULL);
834}
835
836/*
837 * Insert an entry into the hash table. If there is already an element
838 * equal to elem in the hash table, then the already existing element
839 * will be returned and the new element will not be inserted.
840 * Otherwise returns NULL.
841 */
842static arc_buf_hdr_t *
843buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
844{
845 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
846 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
847 arc_buf_hdr_t *fbuf;
848 uint32_t i;
849
850 ASSERT(!HDR_IN_HASH_TABLE(buf));
851 *lockp = hash_lock;
852 mutex_enter(hash_lock);
853 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
854 fbuf = fbuf->b_hash_next, i++) {
855 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
856 return (fbuf);
857 }
858
859 buf->b_hash_next = buf_hash_table.ht_table[idx];
860 buf_hash_table.ht_table[idx] = buf;
861 buf->b_flags |= ARC_IN_HASH_TABLE;
862
863 /* collect some hash table performance data */
864 if (i > 0) {
865 ARCSTAT_BUMP(arcstat_hash_collisions);
866 if (i == 1)
867 ARCSTAT_BUMP(arcstat_hash_chains);
868
869 ARCSTAT_MAX(arcstat_hash_chain_max, i);
870 }
871
872 ARCSTAT_BUMP(arcstat_hash_elements);
873 ARCSTAT_MAXSTAT(arcstat_hash_elements);
874
875 return (NULL);
876}
877
878static void
879buf_hash_remove(arc_buf_hdr_t *buf)
880{
881 arc_buf_hdr_t *fbuf, **bufp;
882 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
883
884 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
885 ASSERT(HDR_IN_HASH_TABLE(buf));
886
887 bufp = &buf_hash_table.ht_table[idx];
888 while ((fbuf = *bufp) != buf) {
889 ASSERT(fbuf != NULL);
890 bufp = &fbuf->b_hash_next;
891 }
892 *bufp = buf->b_hash_next;
893 buf->b_hash_next = NULL;
894 buf->b_flags &= ~ARC_IN_HASH_TABLE;
895
896 /* collect some hash table performance data */
897 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
898
899 if (buf_hash_table.ht_table[idx] &&
900 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
901 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
902}
903
904/*
905 * Global data structures and functions for the buf kmem cache.
906 */
907static kmem_cache_t *hdr_cache;
908static kmem_cache_t *buf_cache;
909
910static void
911buf_fini(void)
912{
913 int i;
914
915 kmem_free(buf_hash_table.ht_table,
916 (buf_hash_table.ht_mask + 1) * sizeof (void *));
917 for (i = 0; i < BUF_LOCKS; i++)
918 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
919 kmem_cache_destroy(hdr_cache);
920 kmem_cache_destroy(buf_cache);
921}
922
923/*
924 * Constructor callback - called when the cache is empty
925 * and a new buf is requested.
926 */
927/* ARGSUSED */
928static int
929hdr_cons(void *vbuf, void *unused, int kmflag)
930{
931 arc_buf_hdr_t *buf = vbuf;
932
933 bzero(buf, sizeof (arc_buf_hdr_t));
934 refcount_create(&buf->b_refcnt);
935 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
936 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
937 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
938
939 return (0);
940}
941
942/* ARGSUSED */
943static int
944buf_cons(void *vbuf, void *unused, int kmflag)
945{
946 arc_buf_t *buf = vbuf;
947
948 bzero(buf, sizeof (arc_buf_t));
949 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
950 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
951
952 return (0);
953}
954
955/*
956 * Destructor callback - called when a cached buf is
957 * no longer required.
958 */
959/* ARGSUSED */
960static void
961hdr_dest(void *vbuf, void *unused)
962{
963 arc_buf_hdr_t *buf = vbuf;
964
965 ASSERT(BUF_EMPTY(buf));
966 refcount_destroy(&buf->b_refcnt);
967 cv_destroy(&buf->b_cv);
968 mutex_destroy(&buf->b_freeze_lock);
969 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
970}
971
972/* ARGSUSED */
973static void
974buf_dest(void *vbuf, void *unused)
975{
976 arc_buf_t *buf = vbuf;
977
978 mutex_destroy(&buf->b_evict_lock);
979 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
980}
981
982/*
983 * Reclaim callback -- invoked when memory is low.
984 */
985/* ARGSUSED */
986static void
987hdr_recl(void *unused)
988{
989 dprintf("hdr_recl called\n");
990 /*
991 * umem calls the reclaim func when we destroy the buf cache,
992 * which is after we do arc_fini().
993 */
994 if (!arc_dead)
995 cv_signal(&arc_reclaim_thr_cv);
996}
997
998static void
999buf_init(void)
1000{
1001 uint64_t *ct;
1002 uint64_t hsize = 1ULL << 12;
1003 int i, j;
1004
1005 /*
1006 * The hash table is big enough to fill all of physical memory
1007 * with an average 64K block size. The table will take up
1008 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
1009 */
1010 while (hsize * 65536 < (uint64_t)physmem * PAGESIZE)
1011 hsize <<= 1;
1012retry:
1013 buf_hash_table.ht_mask = hsize - 1;
1014 buf_hash_table.ht_table =
1015 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1016 if (buf_hash_table.ht_table == NULL) {
1017 ASSERT(hsize > (1ULL << 8));
1018 hsize >>= 1;
1019 goto retry;
1020 }
1021
1022 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
1023 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
1024 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1025 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1026
1027 for (i = 0; i < 256; i++)
1028 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1029 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1030
1031 for (i = 0; i < BUF_LOCKS; i++) {
1032 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1033 NULL, MUTEX_DEFAULT, NULL);
1034 }
1035}
1036
1037#define ARC_MINTIME (hz>>4) /* 62 ms */
1038
1039static void
1040arc_cksum_verify(arc_buf_t *buf)
1041{
1042 zio_cksum_t zc;
1043
1044 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1045 return;
1046
1047 mutex_enter(&buf->b_hdr->b_freeze_lock);
1048 if (buf->b_hdr->b_freeze_cksum == NULL ||
1049 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
1050 mutex_exit(&buf->b_hdr->b_freeze_lock);
1051 return;
1052 }
1053 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1054 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1055 panic("buffer modified while frozen!");
1056 mutex_exit(&buf->b_hdr->b_freeze_lock);
1057}
1058
1059static int
1060arc_cksum_equal(arc_buf_t *buf)
1061{
1062 zio_cksum_t zc;
1063 int equal;
1064
1065 mutex_enter(&buf->b_hdr->b_freeze_lock);
1066 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1067 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1068 mutex_exit(&buf->b_hdr->b_freeze_lock);
1069
1070 return (equal);
1071}
1072
1073static void
1074arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1075{
1076 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1077 return;
1078
1079 mutex_enter(&buf->b_hdr->b_freeze_lock);
1080 if (buf->b_hdr->b_freeze_cksum != NULL) {
1081 mutex_exit(&buf->b_hdr->b_freeze_lock);
1082 return;
1083 }
1084 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1085 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1086 buf->b_hdr->b_freeze_cksum);
1087 mutex_exit(&buf->b_hdr->b_freeze_lock);
1088#ifdef illumos
1089 arc_buf_watch(buf);
1090#endif /* illumos */
1091}
1092
1093#ifdef illumos
1094#ifndef _KERNEL
1095typedef struct procctl {
1096 long cmd;
1097 prwatch_t prwatch;
1098} procctl_t;
1099#endif
1100
1101/* ARGSUSED */
1102static void
1103arc_buf_unwatch(arc_buf_t *buf)
1104{
1105#ifndef _KERNEL
1106 if (arc_watch) {
1107 int result;
1108 procctl_t ctl;
1109 ctl.cmd = PCWATCH;
1110 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1111 ctl.prwatch.pr_size = 0;
1112 ctl.prwatch.pr_wflags = 0;
1113 result = write(arc_procfd, &ctl, sizeof (ctl));
1114 ASSERT3U(result, ==, sizeof (ctl));
1115 }
1116#endif
1117}
1118
1119/* ARGSUSED */
1120static void
1121arc_buf_watch(arc_buf_t *buf)
1122{
1123#ifndef _KERNEL
1124 if (arc_watch) {
1125 int result;
1126 procctl_t ctl;
1127 ctl.cmd = PCWATCH;
1128 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1129 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1130 ctl.prwatch.pr_wflags = WA_WRITE;
1131 result = write(arc_procfd, &ctl, sizeof (ctl));
1132 ASSERT3U(result, ==, sizeof (ctl));
1133 }
1134#endif
1135}
1136#endif /* illumos */
1137
1138void
1139arc_buf_thaw(arc_buf_t *buf)
1140{
1141 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1142 if (buf->b_hdr->b_state != arc_anon)
1143 panic("modifying non-anon buffer!");
1144 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1145 panic("modifying buffer while i/o in progress!");
1146 arc_cksum_verify(buf);
1147 }
1148
1149 mutex_enter(&buf->b_hdr->b_freeze_lock);
1150 if (buf->b_hdr->b_freeze_cksum != NULL) {
1151 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1152 buf->b_hdr->b_freeze_cksum = NULL;
1153 }
1154
1155 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1156 if (buf->b_hdr->b_thawed)
1157 kmem_free(buf->b_hdr->b_thawed, 1);
1158 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1159 }
1160
1161 mutex_exit(&buf->b_hdr->b_freeze_lock);
1162
1163#ifdef illumos
1164 arc_buf_unwatch(buf);
1165#endif /* illumos */
1166}
1167
1168void
1169arc_buf_freeze(arc_buf_t *buf)
1170{
1171 kmutex_t *hash_lock;
1172
1173 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1174 return;
1175
1176 hash_lock = HDR_LOCK(buf->b_hdr);
1177 mutex_enter(hash_lock);
1178
1179 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1180 buf->b_hdr->b_state == arc_anon);
1181 arc_cksum_compute(buf, B_FALSE);
1182 mutex_exit(hash_lock);
1183
1184}
1185
1186static void
1187get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock)
1188{
1189 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth);
1190
1191 if (ab->b_type == ARC_BUFC_METADATA)
1192 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1);
1193 else {
1194 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1);
1195 buf_hashid += ARC_BUFC_NUMMETADATALISTS;
1196 }
1197
1198 *list = &state->arcs_lists[buf_hashid];
1199 *lock = ARCS_LOCK(state, buf_hashid);
1200}
1201
1202
1203static void
1204add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1205{
1206 ASSERT(MUTEX_HELD(hash_lock));
1207
1208 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1209 (ab->b_state != arc_anon)) {
1210 uint64_t delta = ab->b_size * ab->b_datacnt;
1211 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1212 list_t *list;
1213 kmutex_t *lock;
1214
1215 get_buf_info(ab, ab->b_state, &list, &lock);
1216 ASSERT(!MUTEX_HELD(lock));
1217 mutex_enter(lock);
1218 ASSERT(list_link_active(&ab->b_arc_node));
1219 list_remove(list, ab);
1220 if (GHOST_STATE(ab->b_state)) {
1221 ASSERT0(ab->b_datacnt);
1222 ASSERT3P(ab->b_buf, ==, NULL);
1223 delta = ab->b_size;
1224 }
1225 ASSERT(delta > 0);
1226 ASSERT3U(*size, >=, delta);
1227 atomic_add_64(size, -delta);
1228 mutex_exit(lock);
1229 /* remove the prefetch flag if we get a reference */
1230 if (ab->b_flags & ARC_PREFETCH)
1231 ab->b_flags &= ~ARC_PREFETCH;
1232 }
1233}
1234
1235static int
1236remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1237{
1238 int cnt;
1239 arc_state_t *state = ab->b_state;
1240
1241 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1242 ASSERT(!GHOST_STATE(state));
1243
1244 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1245 (state != arc_anon)) {
1246 uint64_t *size = &state->arcs_lsize[ab->b_type];
1247 list_t *list;
1248 kmutex_t *lock;
1249
1250 get_buf_info(ab, state, &list, &lock);
1251 ASSERT(!MUTEX_HELD(lock));
1252 mutex_enter(lock);
1253 ASSERT(!list_link_active(&ab->b_arc_node));
1254 list_insert_head(list, ab);
1255 ASSERT(ab->b_datacnt > 0);
1256 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1257 mutex_exit(lock);
1258 }
1259 return (cnt);
1260}
1261
1262/*
1263 * Move the supplied buffer to the indicated state. The mutex
1264 * for the buffer must be held by the caller.
1265 */
1266static void
1267arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1268{
1269 arc_state_t *old_state = ab->b_state;
1270 int64_t refcnt = refcount_count(&ab->b_refcnt);
1271 uint64_t from_delta, to_delta;
1272 list_t *list;
1273 kmutex_t *lock;
1274
1275 ASSERT(MUTEX_HELD(hash_lock));
1276 ASSERT(new_state != old_state);
1277 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1278 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1279 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1280
1281 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1282
1283 /*
1284 * If this buffer is evictable, transfer it from the
1285 * old state list to the new state list.
1286 */
1287 if (refcnt == 0) {
1288 if (old_state != arc_anon) {
1289 int use_mutex;
1290 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1291
1292 get_buf_info(ab, old_state, &list, &lock);
1293 use_mutex = !MUTEX_HELD(lock);
1294 if (use_mutex)
1295 mutex_enter(lock);
1296
1297 ASSERT(list_link_active(&ab->b_arc_node));
1298 list_remove(list, ab);
1299
1300 /*
1301 * If prefetching out of the ghost cache,
1302 * we will have a non-zero datacnt.
1303 */
1304 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1305 /* ghost elements have a ghost size */
1306 ASSERT(ab->b_buf == NULL);
1307 from_delta = ab->b_size;
1308 }
1309 ASSERT3U(*size, >=, from_delta);
1310 atomic_add_64(size, -from_delta);
1311
1312 if (use_mutex)
1313 mutex_exit(lock);
1314 }
1315 if (new_state != arc_anon) {
1316 int use_mutex;
1317 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1318
1319 get_buf_info(ab, new_state, &list, &lock);
1320 use_mutex = !MUTEX_HELD(lock);
1321 if (use_mutex)
1322 mutex_enter(lock);
1323
1324 list_insert_head(list, ab);
1325
1326 /* ghost elements have a ghost size */
1327 if (GHOST_STATE(new_state)) {
1328 ASSERT(ab->b_datacnt == 0);
1329 ASSERT(ab->b_buf == NULL);
1330 to_delta = ab->b_size;
1331 }
1332 atomic_add_64(size, to_delta);
1333
1334 if (use_mutex)
1335 mutex_exit(lock);
1336 }
1337 }
1338
1339 ASSERT(!BUF_EMPTY(ab));
1340 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1341 buf_hash_remove(ab);
1342
1343 /* adjust state sizes */
1344 if (to_delta)
1345 atomic_add_64(&new_state->arcs_size, to_delta);
1346 if (from_delta) {
1347 ASSERT3U(old_state->arcs_size, >=, from_delta);
1348 atomic_add_64(&old_state->arcs_size, -from_delta);
1349 }
1350 ab->b_state = new_state;
1351
1352 /* adjust l2arc hdr stats */
1353 if (new_state == arc_l2c_only)
1354 l2arc_hdr_stat_add();
1355 else if (old_state == arc_l2c_only)
1356 l2arc_hdr_stat_remove();
1357}
1358
1359void
1360arc_space_consume(uint64_t space, arc_space_type_t type)
1361{
1362 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1363
1364 switch (type) {
1365 case ARC_SPACE_DATA:
1366 ARCSTAT_INCR(arcstat_data_size, space);
1367 break;
1368 case ARC_SPACE_OTHER:
1369 ARCSTAT_INCR(arcstat_other_size, space);
1370 break;
1371 case ARC_SPACE_HDRS:
1372 ARCSTAT_INCR(arcstat_hdr_size, space);
1373 break;
1374 case ARC_SPACE_L2HDRS:
1375 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1376 break;
1377 }
1378
1379 atomic_add_64(&arc_meta_used, space);
1380 atomic_add_64(&arc_size, space);
1381}
1382
1383void
1384arc_space_return(uint64_t space, arc_space_type_t type)
1385{
1386 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1387
1388 switch (type) {
1389 case ARC_SPACE_DATA:
1390 ARCSTAT_INCR(arcstat_data_size, -space);
1391 break;
1392 case ARC_SPACE_OTHER:
1393 ARCSTAT_INCR(arcstat_other_size, -space);
1394 break;
1395 case ARC_SPACE_HDRS:
1396 ARCSTAT_INCR(arcstat_hdr_size, -space);
1397 break;
1398 case ARC_SPACE_L2HDRS:
1399 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1400 break;
1401 }
1402
1403 ASSERT(arc_meta_used >= space);
1404 if (arc_meta_max < arc_meta_used)
1405 arc_meta_max = arc_meta_used;
1406 atomic_add_64(&arc_meta_used, -space);
1407 ASSERT(arc_size >= space);
1408 atomic_add_64(&arc_size, -space);
1409}
1410
1411void *
1412arc_data_buf_alloc(uint64_t size)
1413{
1414 if (arc_evict_needed(ARC_BUFC_DATA))
1415 cv_signal(&arc_reclaim_thr_cv);
1416 atomic_add_64(&arc_size, size);
1417 return (zio_data_buf_alloc(size));
1418}
1419
1420void
1421arc_data_buf_free(void *buf, uint64_t size)
1422{
1423 zio_data_buf_free(buf, size);
1424 ASSERT(arc_size >= size);
1425 atomic_add_64(&arc_size, -size);
1426}
1427
1428arc_buf_t *
1429arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1430{
1431 arc_buf_hdr_t *hdr;
1432 arc_buf_t *buf;
1433
1434 ASSERT3U(size, >, 0);
1435 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1436 ASSERT(BUF_EMPTY(hdr));
1437 hdr->b_size = size;
1438 hdr->b_type = type;
1439 hdr->b_spa = spa_load_guid(spa);
1440 hdr->b_state = arc_anon;
1441 hdr->b_arc_access = 0;
1442 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1443 buf->b_hdr = hdr;
1444 buf->b_data = NULL;
1445 buf->b_efunc = NULL;
1446 buf->b_private = NULL;
1447 buf->b_next = NULL;
1448 hdr->b_buf = buf;
1449 arc_get_data_buf(buf);
1450 hdr->b_datacnt = 1;
1451 hdr->b_flags = 0;
1452 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1453 (void) refcount_add(&hdr->b_refcnt, tag);
1454
1455 return (buf);
1456}
1457
1458static char *arc_onloan_tag = "onloan";
1459
1460/*
1461 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1462 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1463 * buffers must be returned to the arc before they can be used by the DMU or
1464 * freed.
1465 */
1466arc_buf_t *
1467arc_loan_buf(spa_t *spa, int size)
1468{
1469 arc_buf_t *buf;
1470
1471 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1472
1473 atomic_add_64(&arc_loaned_bytes, size);
1474 return (buf);
1475}
1476
1477/*
1478 * Return a loaned arc buffer to the arc.
1479 */
1480void
1481arc_return_buf(arc_buf_t *buf, void *tag)
1482{
1483 arc_buf_hdr_t *hdr = buf->b_hdr;
1484
1485 ASSERT(buf->b_data != NULL);
1486 (void) refcount_add(&hdr->b_refcnt, tag);
1487 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1488
1489 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1490}
1491
1492/* Detach an arc_buf from a dbuf (tag) */
1493void
1494arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1495{
1496 arc_buf_hdr_t *hdr;
1497
1498 ASSERT(buf->b_data != NULL);
1499 hdr = buf->b_hdr;
1500 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1501 (void) refcount_remove(&hdr->b_refcnt, tag);
1502 buf->b_efunc = NULL;
1503 buf->b_private = NULL;
1504
1505 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1506}
1507
1508static arc_buf_t *
1509arc_buf_clone(arc_buf_t *from)
1510{
1511 arc_buf_t *buf;
1512 arc_buf_hdr_t *hdr = from->b_hdr;
1513 uint64_t size = hdr->b_size;
1514
1515 ASSERT(hdr->b_state != arc_anon);
1516
1517 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1518 buf->b_hdr = hdr;
1519 buf->b_data = NULL;
1520 buf->b_efunc = NULL;
1521 buf->b_private = NULL;
1522 buf->b_next = hdr->b_buf;
1523 hdr->b_buf = buf;
1524 arc_get_data_buf(buf);
1525 bcopy(from->b_data, buf->b_data, size);
1526
1527 /*
1528 * This buffer already exists in the arc so create a duplicate
1529 * copy for the caller. If the buffer is associated with user data
1530 * then track the size and number of duplicates. These stats will be
1531 * updated as duplicate buffers are created and destroyed.
1532 */
1533 if (hdr->b_type == ARC_BUFC_DATA) {
1534 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1535 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1536 }
1537 hdr->b_datacnt += 1;
1538 return (buf);
1539}
1540
1541void
1542arc_buf_add_ref(arc_buf_t *buf, void* tag)
1543{
1544 arc_buf_hdr_t *hdr;
1545 kmutex_t *hash_lock;
1546
1547 /*
1548 * Check to see if this buffer is evicted. Callers
1549 * must verify b_data != NULL to know if the add_ref
1550 * was successful.
1551 */
1552 mutex_enter(&buf->b_evict_lock);
1553 if (buf->b_data == NULL) {
1554 mutex_exit(&buf->b_evict_lock);
1555 return;
1556 }
1557 hash_lock = HDR_LOCK(buf->b_hdr);
1558 mutex_enter(hash_lock);
1559 hdr = buf->b_hdr;
1560 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1561 mutex_exit(&buf->b_evict_lock);
1562
1563 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1564 add_reference(hdr, hash_lock, tag);
1565 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1566 arc_access(hdr, hash_lock);
1567 mutex_exit(hash_lock);
1568 ARCSTAT_BUMP(arcstat_hits);
1569 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1570 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1571 data, metadata, hits);
1572}
1573
1574/*
1575 * Free the arc data buffer. If it is an l2arc write in progress,
1576 * the buffer is placed on l2arc_free_on_write to be freed later.
1577 */
1578static void
1579arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1580{
1581 arc_buf_hdr_t *hdr = buf->b_hdr;
1582
1583 if (HDR_L2_WRITING(hdr)) {
1584 l2arc_data_free_t *df;
1585 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1586 df->l2df_data = buf->b_data;
1587 df->l2df_size = hdr->b_size;
1588 df->l2df_func = free_func;
1589 mutex_enter(&l2arc_free_on_write_mtx);
1590 list_insert_head(l2arc_free_on_write, df);
1591 mutex_exit(&l2arc_free_on_write_mtx);
1592 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1593 } else {
1594 free_func(buf->b_data, hdr->b_size);
1595 }
1596}
1597
1598static void
1599arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1600{
1601 arc_buf_t **bufp;
1602
1603 /* free up data associated with the buf */
1604 if (buf->b_data) {
1605 arc_state_t *state = buf->b_hdr->b_state;
1606 uint64_t size = buf->b_hdr->b_size;
1607 arc_buf_contents_t type = buf->b_hdr->b_type;
1608
1609 arc_cksum_verify(buf);
1610#ifdef illumos
1611 arc_buf_unwatch(buf);
1612#endif /* illumos */
1613
1614 if (!recycle) {
1615 if (type == ARC_BUFC_METADATA) {
1616 arc_buf_data_free(buf, zio_buf_free);
1617 arc_space_return(size, ARC_SPACE_DATA);
1618 } else {
1619 ASSERT(type == ARC_BUFC_DATA);
1620 arc_buf_data_free(buf, zio_data_buf_free);
1621 ARCSTAT_INCR(arcstat_data_size, -size);
1622 atomic_add_64(&arc_size, -size);
1623 }
1624 }
1625 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1626 uint64_t *cnt = &state->arcs_lsize[type];
1627
1628 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1629 ASSERT(state != arc_anon);
1630
1631 ASSERT3U(*cnt, >=, size);
1632 atomic_add_64(cnt, -size);
1633 }
1634 ASSERT3U(state->arcs_size, >=, size);
1635 atomic_add_64(&state->arcs_size, -size);
1636 buf->b_data = NULL;
1637
1638 /*
1639 * If we're destroying a duplicate buffer make sure
1640 * that the appropriate statistics are updated.
1641 */
1642 if (buf->b_hdr->b_datacnt > 1 &&
1643 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1644 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1645 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1646 }
1647 ASSERT(buf->b_hdr->b_datacnt > 0);
1648 buf->b_hdr->b_datacnt -= 1;
1649 }
1650
1651 /* only remove the buf if requested */
1652 if (!all)
1653 return;
1654
1655 /* remove the buf from the hdr list */
1656 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1657 continue;
1658 *bufp = buf->b_next;
1659 buf->b_next = NULL;
1660
1661 ASSERT(buf->b_efunc == NULL);
1662
1663 /* clean up the buf */
1664 buf->b_hdr = NULL;
1665 kmem_cache_free(buf_cache, buf);
1666}
1667
1668static void
1669arc_hdr_destroy(arc_buf_hdr_t *hdr)
1670{
1671 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1672 ASSERT3P(hdr->b_state, ==, arc_anon);
1673 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1674 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1675
1676 if (l2hdr != NULL) {
1677 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1678 /*
1679 * To prevent arc_free() and l2arc_evict() from
1680 * attempting to free the same buffer at the same time,
1681 * a FREE_IN_PROGRESS flag is given to arc_free() to
1682 * give it priority. l2arc_evict() can't destroy this
1683 * header while we are waiting on l2arc_buflist_mtx.
1684 *
1685 * The hdr may be removed from l2ad_buflist before we
1686 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1687 */
1688 if (!buflist_held) {
1689 mutex_enter(&l2arc_buflist_mtx);
1690 l2hdr = hdr->b_l2hdr;
1691 }
1692
1693 if (l2hdr != NULL) {
1694 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1695 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1696 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1697 if (hdr->b_state == arc_l2c_only)
1698 l2arc_hdr_stat_remove();
1699 hdr->b_l2hdr = NULL;
1700 }
1701
1702 if (!buflist_held)
1703 mutex_exit(&l2arc_buflist_mtx);
1704 }
1705
1706 if (!BUF_EMPTY(hdr)) {
1707 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1708 buf_discard_identity(hdr);
1709 }
1710 while (hdr->b_buf) {
1711 arc_buf_t *buf = hdr->b_buf;
1712
1713 if (buf->b_efunc) {
1714 mutex_enter(&arc_eviction_mtx);
1715 mutex_enter(&buf->b_evict_lock);
1716 ASSERT(buf->b_hdr != NULL);
1717 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1718 hdr->b_buf = buf->b_next;
1719 buf->b_hdr = &arc_eviction_hdr;
1720 buf->b_next = arc_eviction_list;
1721 arc_eviction_list = buf;
1722 mutex_exit(&buf->b_evict_lock);
1723 mutex_exit(&arc_eviction_mtx);
1724 } else {
1725 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1726 }
1727 }
1728 if (hdr->b_freeze_cksum != NULL) {
1729 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1730 hdr->b_freeze_cksum = NULL;
1731 }
1732 if (hdr->b_thawed) {
1733 kmem_free(hdr->b_thawed, 1);
1734 hdr->b_thawed = NULL;
1735 }
1736
1737 ASSERT(!list_link_active(&hdr->b_arc_node));
1738 ASSERT3P(hdr->b_hash_next, ==, NULL);
1739 ASSERT3P(hdr->b_acb, ==, NULL);
1740 kmem_cache_free(hdr_cache, hdr);
1741}
1742
1743void
1744arc_buf_free(arc_buf_t *buf, void *tag)
1745{
1746 arc_buf_hdr_t *hdr = buf->b_hdr;
1747 int hashed = hdr->b_state != arc_anon;
1748
1749 ASSERT(buf->b_efunc == NULL);
1750 ASSERT(buf->b_data != NULL);
1751
1752 if (hashed) {
1753 kmutex_t *hash_lock = HDR_LOCK(hdr);
1754
1755 mutex_enter(hash_lock);
1756 hdr = buf->b_hdr;
1757 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1758
1759 (void) remove_reference(hdr, hash_lock, tag);
1760 if (hdr->b_datacnt > 1) {
1761 arc_buf_destroy(buf, FALSE, TRUE);
1762 } else {
1763 ASSERT(buf == hdr->b_buf);
1764 ASSERT(buf->b_efunc == NULL);
1765 hdr->b_flags |= ARC_BUF_AVAILABLE;
1766 }
1767 mutex_exit(hash_lock);
1768 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1769 int destroy_hdr;
1770 /*
1771 * We are in the middle of an async write. Don't destroy
1772 * this buffer unless the write completes before we finish
1773 * decrementing the reference count.
1774 */
1775 mutex_enter(&arc_eviction_mtx);
1776 (void) remove_reference(hdr, NULL, tag);
1777 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1778 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1779 mutex_exit(&arc_eviction_mtx);
1780 if (destroy_hdr)
1781 arc_hdr_destroy(hdr);
1782 } else {
1783 if (remove_reference(hdr, NULL, tag) > 0)
1784 arc_buf_destroy(buf, FALSE, TRUE);
1785 else
1786 arc_hdr_destroy(hdr);
1787 }
1788}
1789
1790int
1791arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1792{
1793 arc_buf_hdr_t *hdr = buf->b_hdr;
1794 kmutex_t *hash_lock = HDR_LOCK(hdr);
1795 int no_callback = (buf->b_efunc == NULL);
1796
1797 if (hdr->b_state == arc_anon) {
1798 ASSERT(hdr->b_datacnt == 1);
1799 arc_buf_free(buf, tag);
1800 return (no_callback);
1801 }
1802
1803 mutex_enter(hash_lock);
1804 hdr = buf->b_hdr;
1805 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1806 ASSERT(hdr->b_state != arc_anon);
1807 ASSERT(buf->b_data != NULL);
1808
1809 (void) remove_reference(hdr, hash_lock, tag);
1810 if (hdr->b_datacnt > 1) {
1811 if (no_callback)
1812 arc_buf_destroy(buf, FALSE, TRUE);
1813 } else if (no_callback) {
1814 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1815 ASSERT(buf->b_efunc == NULL);
1816 hdr->b_flags |= ARC_BUF_AVAILABLE;
1817 }
1818 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1819 refcount_is_zero(&hdr->b_refcnt));
1820 mutex_exit(hash_lock);
1821 return (no_callback);
1822}
1823
1824int
1825arc_buf_size(arc_buf_t *buf)
1826{
1827 return (buf->b_hdr->b_size);
1828}
1829
1830/*
1831 * Called from the DMU to determine if the current buffer should be
1832 * evicted. In order to ensure proper locking, the eviction must be initiated
1833 * from the DMU. Return true if the buffer is associated with user data and
1834 * duplicate buffers still exist.
1835 */
1836boolean_t
1837arc_buf_eviction_needed(arc_buf_t *buf)
1838{
1839 arc_buf_hdr_t *hdr;
1840 boolean_t evict_needed = B_FALSE;
1841
1842 if (zfs_disable_dup_eviction)
1843 return (B_FALSE);
1844
1845 mutex_enter(&buf->b_evict_lock);
1846 hdr = buf->b_hdr;
1847 if (hdr == NULL) {
1848 /*
1849 * We are in arc_do_user_evicts(); let that function
1850 * perform the eviction.
1851 */
1852 ASSERT(buf->b_data == NULL);
1853 mutex_exit(&buf->b_evict_lock);
1854 return (B_FALSE);
1855 } else if (buf->b_data == NULL) {
1856 /*
1857 * We have already been added to the arc eviction list;
1858 * recommend eviction.
1859 */
1860 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1861 mutex_exit(&buf->b_evict_lock);
1862 return (B_TRUE);
1863 }
1864
1865 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1866 evict_needed = B_TRUE;
1867
1868 mutex_exit(&buf->b_evict_lock);
1869 return (evict_needed);
1870}
1871
1872/*
1873 * Evict buffers from list until we've removed the specified number of
1874 * bytes. Move the removed buffers to the appropriate evict state.
1875 * If the recycle flag is set, then attempt to "recycle" a buffer:
1876 * - look for a buffer to evict that is `bytes' long.
1877 * - return the data block from this buffer rather than freeing it.
1878 * This flag is used by callers that are trying to make space for a
1879 * new buffer in a full arc cache.
1880 *
1881 * This function makes a "best effort". It skips over any buffers
1882 * it can't get a hash_lock on, and so may not catch all candidates.
1883 * It may also return without evicting as much space as requested.
1884 */
1885static void *
1886arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1887 arc_buf_contents_t type)
1888{
1889 arc_state_t *evicted_state;
1890 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1891 int64_t bytes_remaining;
1892 arc_buf_hdr_t *ab, *ab_prev = NULL;
1893 list_t *evicted_list, *list, *evicted_list_start, *list_start;
1894 kmutex_t *lock, *evicted_lock;
1895 kmutex_t *hash_lock;
1896 boolean_t have_lock;
1897 void *stolen = NULL;
1898 static int evict_metadata_offset, evict_data_offset;
1899 int i, idx, offset, list_count, count;
1900
1901 ASSERT(state == arc_mru || state == arc_mfu);
1902
1903 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1904
1905 if (type == ARC_BUFC_METADATA) {
1906 offset = 0;
1907 list_count = ARC_BUFC_NUMMETADATALISTS;
1908 list_start = &state->arcs_lists[0];
1909 evicted_list_start = &evicted_state->arcs_lists[0];
1910 idx = evict_metadata_offset;
1911 } else {
1912 offset = ARC_BUFC_NUMMETADATALISTS;
1913 list_start = &state->arcs_lists[offset];
1914 evicted_list_start = &evicted_state->arcs_lists[offset];
1915 list_count = ARC_BUFC_NUMDATALISTS;
1916 idx = evict_data_offset;
1917 }
1918 bytes_remaining = evicted_state->arcs_lsize[type];
1919 count = 0;
1920
1921evict_start:
1922 list = &list_start[idx];
1923 evicted_list = &evicted_list_start[idx];
1924 lock = ARCS_LOCK(state, (offset + idx));
1925 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx));
1926
1927 mutex_enter(lock);
1928 mutex_enter(evicted_lock);
1929
1930 for (ab = list_tail(list); ab; ab = ab_prev) {
1931 ab_prev = list_prev(list, ab);
1932 bytes_remaining -= (ab->b_size * ab->b_datacnt);
1933 /* prefetch buffers have a minimum lifespan */
1934 if (HDR_IO_IN_PROGRESS(ab) ||
1935 (spa && ab->b_spa != spa) ||
1936 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1937 ddi_get_lbolt() - ab->b_arc_access <
1938 arc_min_prefetch_lifespan)) {
1939 skipped++;
1940 continue;
1941 }
1942 /* "lookahead" for better eviction candidate */
1943 if (recycle && ab->b_size != bytes &&
1944 ab_prev && ab_prev->b_size == bytes)
1945 continue;
1946 hash_lock = HDR_LOCK(ab);
1947 have_lock = MUTEX_HELD(hash_lock);
1948 if (have_lock || mutex_tryenter(hash_lock)) {
1949 ASSERT0(refcount_count(&ab->b_refcnt));
1950 ASSERT(ab->b_datacnt > 0);
1951 while (ab->b_buf) {
1952 arc_buf_t *buf = ab->b_buf;
1953 if (!mutex_tryenter(&buf->b_evict_lock)) {
1954 missed += 1;
1955 break;
1956 }
1957 if (buf->b_data) {
1958 bytes_evicted += ab->b_size;
1959 if (recycle && ab->b_type == type &&
1960 ab->b_size == bytes &&
1961 !HDR_L2_WRITING(ab)) {
1962 stolen = buf->b_data;
1963 recycle = FALSE;
1964 }
1965 }
1966 if (buf->b_efunc) {
1967 mutex_enter(&arc_eviction_mtx);
1968 arc_buf_destroy(buf,
1969 buf->b_data == stolen, FALSE);
1970 ab->b_buf = buf->b_next;
1971 buf->b_hdr = &arc_eviction_hdr;
1972 buf->b_next = arc_eviction_list;
1973 arc_eviction_list = buf;
1974 mutex_exit(&arc_eviction_mtx);
1975 mutex_exit(&buf->b_evict_lock);
1976 } else {
1977 mutex_exit(&buf->b_evict_lock);
1978 arc_buf_destroy(buf,
1979 buf->b_data == stolen, TRUE);
1980 }
1981 }
1982
1983 if (ab->b_l2hdr) {
1984 ARCSTAT_INCR(arcstat_evict_l2_cached,
1985 ab->b_size);
1986 } else {
1987 if (l2arc_write_eligible(ab->b_spa, ab)) {
1988 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1989 ab->b_size);
1990 } else {
1991 ARCSTAT_INCR(
1992 arcstat_evict_l2_ineligible,
1993 ab->b_size);
1994 }
1995 }
1996
1997 if (ab->b_datacnt == 0) {
1998 arc_change_state(evicted_state, ab, hash_lock);
1999 ASSERT(HDR_IN_HASH_TABLE(ab));
2000 ab->b_flags |= ARC_IN_HASH_TABLE;
2001 ab->b_flags &= ~ARC_BUF_AVAILABLE;
2002 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
2003 }
2004 if (!have_lock)
2005 mutex_exit(hash_lock);
2006 if (bytes >= 0 && bytes_evicted >= bytes)
2007 break;
2008 if (bytes_remaining > 0) {
2009 mutex_exit(evicted_lock);
2010 mutex_exit(lock);
2011 idx = ((idx + 1) & (list_count - 1));
2012 count++;
2013 goto evict_start;
2014 }
2015 } else {
2016 missed += 1;
2017 }
2018 }
2019
2020 mutex_exit(evicted_lock);
2021 mutex_exit(lock);
2022
2023 idx = ((idx + 1) & (list_count - 1));
2024 count++;
2025
2026 if (bytes_evicted < bytes) {
2027 if (count < list_count)
2028 goto evict_start;
2029 else
2030 dprintf("only evicted %lld bytes from %x",
2031 (longlong_t)bytes_evicted, state);
2032 }
2033 if (type == ARC_BUFC_METADATA)
2034 evict_metadata_offset = idx;
2035 else
2036 evict_data_offset = idx;
2037
2038 if (skipped)
2039 ARCSTAT_INCR(arcstat_evict_skip, skipped);
2040
2041 if (missed)
2042 ARCSTAT_INCR(arcstat_mutex_miss, missed);
2043
2044 /*
2045 * We have just evicted some date into the ghost state, make
2046 * sure we also adjust the ghost state size if necessary.
2047 */
2048 if (arc_no_grow &&
2049 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
2050 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
2051 arc_mru_ghost->arcs_size - arc_c;
2052
2053 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
2054 int64_t todelete =
2055 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
2056 arc_evict_ghost(arc_mru_ghost, 0, todelete);
2057 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
2058 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
2059 arc_mru_ghost->arcs_size +
2060 arc_mfu_ghost->arcs_size - arc_c);
2061 arc_evict_ghost(arc_mfu_ghost, 0, todelete);
2062 }
2063 }
2064 if (stolen)
2065 ARCSTAT_BUMP(arcstat_stolen);
2066
2067 return (stolen);
2068}
2069
2070/*
2071 * Remove buffers from list until we've removed the specified number of
2072 * bytes. Destroy the buffers that are removed.
2073 */
2074static void
2075arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
2076{
2077 arc_buf_hdr_t *ab, *ab_prev;
2078 arc_buf_hdr_t marker = { 0 };
2079 list_t *list, *list_start;
2080 kmutex_t *hash_lock, *lock;
2081 uint64_t bytes_deleted = 0;
2082 uint64_t bufs_skipped = 0;
2083 static int evict_offset;
2084 int list_count, idx = evict_offset;
2085 int offset, count = 0;
2086
2087 ASSERT(GHOST_STATE(state));
2088
2089 /*
2090 * data lists come after metadata lists
2091 */
2092 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS];
2093 list_count = ARC_BUFC_NUMDATALISTS;
2094 offset = ARC_BUFC_NUMMETADATALISTS;
2095
2096evict_start:
2097 list = &list_start[idx];
2098 lock = ARCS_LOCK(state, idx + offset);
2099
2100 mutex_enter(lock);
2101 for (ab = list_tail(list); ab; ab = ab_prev) {
2102 ab_prev = list_prev(list, ab);
2103 if (spa && ab->b_spa != spa)
2104 continue;
2105
2106 /* ignore markers */
2107 if (ab->b_spa == 0)
2108 continue;
2109
2110 hash_lock = HDR_LOCK(ab);
2111 /* caller may be trying to modify this buffer, skip it */
2112 if (MUTEX_HELD(hash_lock))
2113 continue;
2114 if (mutex_tryenter(hash_lock)) {
2115 ASSERT(!HDR_IO_IN_PROGRESS(ab));
2116 ASSERT(ab->b_buf == NULL);
2117 ARCSTAT_BUMP(arcstat_deleted);
2118 bytes_deleted += ab->b_size;
2119
2120 if (ab->b_l2hdr != NULL) {
2121 /*
2122 * This buffer is cached on the 2nd Level ARC;
2123 * don't destroy the header.
2124 */
2125 arc_change_state(arc_l2c_only, ab, hash_lock);
2126 mutex_exit(hash_lock);
2127 } else {
2128 arc_change_state(arc_anon, ab, hash_lock);
2129 mutex_exit(hash_lock);
2130 arc_hdr_destroy(ab);
2131 }
2132
2133 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
2134 if (bytes >= 0 && bytes_deleted >= bytes)
2135 break;
2136 } else if (bytes < 0) {
2137 /*
2138 * Insert a list marker and then wait for the
2139 * hash lock to become available. Once its
2140 * available, restart from where we left off.
2141 */
2142 list_insert_after(list, ab, &marker);
2143 mutex_exit(lock);
2144 mutex_enter(hash_lock);
2145 mutex_exit(hash_lock);
2146 mutex_enter(lock);
2147 ab_prev = list_prev(list, &marker);
2148 list_remove(list, &marker);
2149 } else
2150 bufs_skipped += 1;
2151 }
2152 mutex_exit(lock);
2153 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1));
2154 count++;
2155
2156 if (count < list_count)
2157 goto evict_start;
2158
2159 evict_offset = idx;
2160 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] &&
2161 (bytes < 0 || bytes_deleted < bytes)) {
2162 list_start = &state->arcs_lists[0];
2163 list_count = ARC_BUFC_NUMMETADATALISTS;
2164 offset = count = 0;
2165 goto evict_start;
2166 }
2167
2168 if (bufs_skipped) {
2169 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2170 ASSERT(bytes >= 0);
2171 }
2172
2173 if (bytes_deleted < bytes)
2174 dprintf("only deleted %lld bytes from %p",
2175 (longlong_t)bytes_deleted, state);
2176}
2177
2178static void
2179arc_adjust(void)
2180{
2181 int64_t adjustment, delta;
2182
2183 /*
2184 * Adjust MRU size
2185 */
2186
2187 adjustment = MIN((int64_t)(arc_size - arc_c),
2188 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2189 arc_p));
2190
2191 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2192 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2193 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
2194 adjustment -= delta;
2195 }
2196
2197 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2198 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2199 (void) arc_evict(arc_mru, 0, delta, FALSE,
2200 ARC_BUFC_METADATA);
2201 }
2202
2203 /*
2204 * Adjust MFU size
2205 */
2206
2207 adjustment = arc_size - arc_c;
2208
2209 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2210 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2211 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
2212 adjustment -= delta;
2213 }
2214
2215 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2216 int64_t delta = MIN(adjustment,
2217 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2218 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2219 ARC_BUFC_METADATA);
2220 }
2221
2222 /*
2223 * Adjust ghost lists
2224 */
2225
2226 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2227
2228 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2229 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2230 arc_evict_ghost(arc_mru_ghost, 0, delta);
2231 }
2232
2233 adjustment =
2234 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2235
2236 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2237 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2238 arc_evict_ghost(arc_mfu_ghost, 0, delta);
2239 }
2240}
2241
2242static void
2243arc_do_user_evicts(void)
2244{
2245 static arc_buf_t *tmp_arc_eviction_list;
2246
2247 /*
2248 * Move list over to avoid LOR
2249 */
2250restart:
2251 mutex_enter(&arc_eviction_mtx);
2252 tmp_arc_eviction_list = arc_eviction_list;
2253 arc_eviction_list = NULL;
2254 mutex_exit(&arc_eviction_mtx);
2255
2256 while (tmp_arc_eviction_list != NULL) {
2257 arc_buf_t *buf = tmp_arc_eviction_list;
2258 tmp_arc_eviction_list = buf->b_next;
2259 mutex_enter(&buf->b_evict_lock);
2260 buf->b_hdr = NULL;
2261 mutex_exit(&buf->b_evict_lock);
2262
2263 if (buf->b_efunc != NULL)
2264 VERIFY(buf->b_efunc(buf) == 0);
2265
2266 buf->b_efunc = NULL;
2267 buf->b_private = NULL;
2268 kmem_cache_free(buf_cache, buf);
2269 }
2270
2271 if (arc_eviction_list != NULL)
2272 goto restart;
2273}
2274
2275/*
2276 * Flush all *evictable* data from the cache for the given spa.
2277 * NOTE: this will not touch "active" (i.e. referenced) data.
2278 */
2279void
2280arc_flush(spa_t *spa)
2281{
2282 uint64_t guid = 0;
2283
2284 if (spa)
2285 guid = spa_load_guid(spa);
2286
2287 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) {
2288 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2289 if (spa)
2290 break;
2291 }
2292 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) {
2293 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2294 if (spa)
2295 break;
2296 }
2297 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) {
2298 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2299 if (spa)
2300 break;
2301 }
2302 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) {
2303 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2304 if (spa)
2305 break;
2306 }
2307
2308 arc_evict_ghost(arc_mru_ghost, guid, -1);
2309 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2310
2311 mutex_enter(&arc_reclaim_thr_lock);
2312 arc_do_user_evicts();
2313 mutex_exit(&arc_reclaim_thr_lock);
2314 ASSERT(spa || arc_eviction_list == NULL);
2315}
2316
2317void
2318arc_shrink(void)
2319{
2320 if (arc_c > arc_c_min) {
2321 uint64_t to_free;
2322
2323#ifdef _KERNEL
2324 to_free = arc_c >> arc_shrink_shift;
2325#else
2326 to_free = arc_c >> arc_shrink_shift;
2327#endif
2328 if (arc_c > arc_c_min + to_free)
2329 atomic_add_64(&arc_c, -to_free);
2330 else
2331 arc_c = arc_c_min;
2332
2333 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2334 if (arc_c > arc_size)
2335 arc_c = MAX(arc_size, arc_c_min);
2336 if (arc_p > arc_c)
2337 arc_p = (arc_c >> 1);
2338 ASSERT(arc_c >= arc_c_min);
2339 ASSERT((int64_t)arc_p >= 0);
2340 }
2341
2342 if (arc_size > arc_c)
2343 arc_adjust();
2344}
2345
2346static int needfree = 0;
2347
2348static int
2349arc_reclaim_needed(void)
2350{
2351
2352#ifdef _KERNEL
2353
2354 if (needfree)
2355 return (1);
2356
2357 /*
2358 * Cooperate with pagedaemon when it's time for it to scan
2359 * and reclaim some pages.
2360 */
2361 if (vm_paging_needed())
2362 return (1);
2363
2364#ifdef sun
2365 /*
2366 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2367 */
2368 extra = desfree;
2369
2370 /*
2371 * check that we're out of range of the pageout scanner. It starts to
2372 * schedule paging if freemem is less than lotsfree and needfree.
2373 * lotsfree is the high-water mark for pageout, and needfree is the
2374 * number of needed free pages. We add extra pages here to make sure
2375 * the scanner doesn't start up while we're freeing memory.
2376 */
2377 if (freemem < lotsfree + needfree + extra)
2378 return (1);
2379
2380 /*
2381 * check to make sure that swapfs has enough space so that anon
2382 * reservations can still succeed. anon_resvmem() checks that the
2383 * availrmem is greater than swapfs_minfree, and the number of reserved
2384 * swap pages. We also add a bit of extra here just to prevent
2385 * circumstances from getting really dire.
2386 */
2387 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2388 return (1);
2389
2390#if defined(__i386)
2391 /*
2392 * If we're on an i386 platform, it's possible that we'll exhaust the
2393 * kernel heap space before we ever run out of available physical
2394 * memory. Most checks of the size of the heap_area compare against
2395 * tune.t_minarmem, which is the minimum available real memory that we
2396 * can have in the system. However, this is generally fixed at 25 pages
2397 * which is so low that it's useless. In this comparison, we seek to
2398 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2399 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2400 * free)
2401 */
2402 if (btop(vmem_size(heap_arena, VMEM_FREE)) <
2403 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
2404 return (1);
2405#endif
2406#else /* !sun */
2407 if (kmem_used() > (kmem_size() * 3) / 4)
2408 return (1);
2409#endif /* sun */
2410
2411#else
2412 if (spa_get_random(100) == 0)
2413 return (1);
2414#endif
2415 return (0);
2416}
2417
2418extern kmem_cache_t *zio_buf_cache[];
2419extern kmem_cache_t *zio_data_buf_cache[];
2420
2421static void
2422arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2423{
2424 size_t i;
2425 kmem_cache_t *prev_cache = NULL;
2426 kmem_cache_t *prev_data_cache = NULL;
2427
2428#ifdef _KERNEL
2429 if (arc_meta_used >= arc_meta_limit) {
2430 /*
2431 * We are exceeding our meta-data cache limit.
2432 * Purge some DNLC entries to release holds on meta-data.
2433 */
2434 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2435 }
2436#if defined(__i386)
2437 /*
2438 * Reclaim unused memory from all kmem caches.
2439 */
2440 kmem_reap();
2441#endif
2442#endif
2443
2444 /*
2445 * An aggressive reclamation will shrink the cache size as well as
2446 * reap free buffers from the arc kmem caches.
2447 */
2448 if (strat == ARC_RECLAIM_AGGR)
2449 arc_shrink();
2450
2451 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2452 if (zio_buf_cache[i] != prev_cache) {
2453 prev_cache = zio_buf_cache[i];
2454 kmem_cache_reap_now(zio_buf_cache[i]);
2455 }
2456 if (zio_data_buf_cache[i] != prev_data_cache) {
2457 prev_data_cache = zio_data_buf_cache[i];
2458 kmem_cache_reap_now(zio_data_buf_cache[i]);
2459 }
2460 }
2461 kmem_cache_reap_now(buf_cache);
2462 kmem_cache_reap_now(hdr_cache);
2463}
2464
2465static void
2466arc_reclaim_thread(void *dummy __unused)
2467{
2468 clock_t growtime = 0;
2469 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2470 callb_cpr_t cpr;
2471
2472 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2473
2474 mutex_enter(&arc_reclaim_thr_lock);
2475 while (arc_thread_exit == 0) {
2476 if (arc_reclaim_needed()) {
2477
2478 if (arc_no_grow) {
2479 if (last_reclaim == ARC_RECLAIM_CONS) {
2480 last_reclaim = ARC_RECLAIM_AGGR;
2481 } else {
2482 last_reclaim = ARC_RECLAIM_CONS;
2483 }
2484 } else {
2485 arc_no_grow = TRUE;
2486 last_reclaim = ARC_RECLAIM_AGGR;
2487 membar_producer();
2488 }
2489
2490 /* reset the growth delay for every reclaim */
2491 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2492
2493 if (needfree && last_reclaim == ARC_RECLAIM_CONS) {
2494 /*
2495 * If needfree is TRUE our vm_lowmem hook
2496 * was called and in that case we must free some
2497 * memory, so switch to aggressive mode.
2498 */
2499 arc_no_grow = TRUE;
2500 last_reclaim = ARC_RECLAIM_AGGR;
2501 }
2502 arc_kmem_reap_now(last_reclaim);
2503 arc_warm = B_TRUE;
2504
2505 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2506 arc_no_grow = FALSE;
2507 }
2508
2509 arc_adjust();
2510
2511 if (arc_eviction_list != NULL)
2512 arc_do_user_evicts();
2513
2514#ifdef _KERNEL
2515 if (needfree) {
2516 needfree = 0;
2517 wakeup(&needfree);
2518 }
2519#endif
2520
2521 /* block until needed, or one second, whichever is shorter */
2522 CALLB_CPR_SAFE_BEGIN(&cpr);
2523 (void) cv_timedwait(&arc_reclaim_thr_cv,
2524 &arc_reclaim_thr_lock, hz);
2525 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2526 }
2527
2528 arc_thread_exit = 0;
2529 cv_broadcast(&arc_reclaim_thr_cv);
2530 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2531 thread_exit();
2532}
2533
2534/*
2535 * Adapt arc info given the number of bytes we are trying to add and
2536 * the state that we are comming from. This function is only called
2537 * when we are adding new content to the cache.
2538 */
2539static void
2540arc_adapt(int bytes, arc_state_t *state)
2541{
2542 int mult;
2543 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2544
2545 if (state == arc_l2c_only)
2546 return;
2547
2548 ASSERT(bytes > 0);
2549 /*
2550 * Adapt the target size of the MRU list:
2551 * - if we just hit in the MRU ghost list, then increase
2552 * the target size of the MRU list.
2553 * - if we just hit in the MFU ghost list, then increase
2554 * the target size of the MFU list by decreasing the
2555 * target size of the MRU list.
2556 */
2557 if (state == arc_mru_ghost) {
2558 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2559 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2560 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2561
2562 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2563 } else if (state == arc_mfu_ghost) {
2564 uint64_t delta;
2565
2566 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2567 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2568 mult = MIN(mult, 10);
2569
2570 delta = MIN(bytes * mult, arc_p);
2571 arc_p = MAX(arc_p_min, arc_p - delta);
2572 }
2573 ASSERT((int64_t)arc_p >= 0);
2574
2575 if (arc_reclaim_needed()) {
2576 cv_signal(&arc_reclaim_thr_cv);
2577 return;
2578 }
2579
2580 if (arc_no_grow)
2581 return;
2582
2583 if (arc_c >= arc_c_max)
2584 return;
2585
2586 /*
2587 * If we're within (2 * maxblocksize) bytes of the target
2588 * cache size, increment the target cache size
2589 */
2590 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2591 atomic_add_64(&arc_c, (int64_t)bytes);
2592 if (arc_c > arc_c_max)
2593 arc_c = arc_c_max;
2594 else if (state == arc_anon)
2595 atomic_add_64(&arc_p, (int64_t)bytes);
2596 if (arc_p > arc_c)
2597 arc_p = arc_c;
2598 }
2599 ASSERT((int64_t)arc_p >= 0);
2600}
2601
2602/*
2603 * Check if the cache has reached its limits and eviction is required
2604 * prior to insert.
2605 */
2606static int
2607arc_evict_needed(arc_buf_contents_t type)
2608{
2609 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2610 return (1);
2611
2612#ifdef sun
2613#ifdef _KERNEL
2614 /*
2615 * If zio data pages are being allocated out of a separate heap segment,
2616 * then enforce that the size of available vmem for this area remains
2617 * above about 1/32nd free.
2618 */
2619 if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2620 vmem_size(zio_arena, VMEM_FREE) <
2621 (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2622 return (1);
2623#endif
2624#endif /* sun */
2625
2626 if (arc_reclaim_needed())
2627 return (1);
2628
2629 return (arc_size > arc_c);
2630}
2631
2632/*
2633 * The buffer, supplied as the first argument, needs a data block.
2634 * So, if we are at cache max, determine which cache should be victimized.
2635 * We have the following cases:
2636 *
2637 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2638 * In this situation if we're out of space, but the resident size of the MFU is
2639 * under the limit, victimize the MFU cache to satisfy this insertion request.
2640 *
2641 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2642 * Here, we've used up all of the available space for the MRU, so we need to
2643 * evict from our own cache instead. Evict from the set of resident MRU
2644 * entries.
2645 *
2646 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2647 * c minus p represents the MFU space in the cache, since p is the size of the
2648 * cache that is dedicated to the MRU. In this situation there's still space on
2649 * the MFU side, so the MRU side needs to be victimized.
2650 *
2651 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2652 * MFU's resident set is consuming more space than it has been allotted. In
2653 * this situation, we must victimize our own cache, the MFU, for this insertion.
2654 */
2655static void
2656arc_get_data_buf(arc_buf_t *buf)
2657{
2658 arc_state_t *state = buf->b_hdr->b_state;
2659 uint64_t size = buf->b_hdr->b_size;
2660 arc_buf_contents_t type = buf->b_hdr->b_type;
2661
2662 arc_adapt(size, state);
2663
2664 /*
2665 * We have not yet reached cache maximum size,
2666 * just allocate a new buffer.
2667 */
2668 if (!arc_evict_needed(type)) {
2669 if (type == ARC_BUFC_METADATA) {
2670 buf->b_data = zio_buf_alloc(size);
2671 arc_space_consume(size, ARC_SPACE_DATA);
2672 } else {
2673 ASSERT(type == ARC_BUFC_DATA);
2674 buf->b_data = zio_data_buf_alloc(size);
2675 ARCSTAT_INCR(arcstat_data_size, size);
2676 atomic_add_64(&arc_size, size);
2677 }
2678 goto out;
2679 }
2680
2681 /*
2682 * If we are prefetching from the mfu ghost list, this buffer
2683 * will end up on the mru list; so steal space from there.
2684 */
2685 if (state == arc_mfu_ghost)
2686 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2687 else if (state == arc_mru_ghost)
2688 state = arc_mru;
2689
2690 if (state == arc_mru || state == arc_anon) {
2691 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2692 state = (arc_mfu->arcs_lsize[type] >= size &&
2693 arc_p > mru_used) ? arc_mfu : arc_mru;
2694 } else {
2695 /* MFU cases */
2696 uint64_t mfu_space = arc_c - arc_p;
2697 state = (arc_mru->arcs_lsize[type] >= size &&
2698 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2699 }
2700 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2701 if (type == ARC_BUFC_METADATA) {
2702 buf->b_data = zio_buf_alloc(size);
2703 arc_space_consume(size, ARC_SPACE_DATA);
2704 } else {
2705 ASSERT(type == ARC_BUFC_DATA);
2706 buf->b_data = zio_data_buf_alloc(size);
2707 ARCSTAT_INCR(arcstat_data_size, size);
2708 atomic_add_64(&arc_size, size);
2709 }
2710 ARCSTAT_BUMP(arcstat_recycle_miss);
2711 }
2712 ASSERT(buf->b_data != NULL);
2713out:
2714 /*
2715 * Update the state size. Note that ghost states have a
2716 * "ghost size" and so don't need to be updated.
2717 */
2718 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2719 arc_buf_hdr_t *hdr = buf->b_hdr;
2720
2721 atomic_add_64(&hdr->b_state->arcs_size, size);
2722 if (list_link_active(&hdr->b_arc_node)) {
2723 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2724 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2725 }
2726 /*
2727 * If we are growing the cache, and we are adding anonymous
2728 * data, and we have outgrown arc_p, update arc_p
2729 */
2730 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2731 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2732 arc_p = MIN(arc_c, arc_p + size);
2733 }
2734 ARCSTAT_BUMP(arcstat_allocated);
2735}
2736
2737/*
2738 * This routine is called whenever a buffer is accessed.
2739 * NOTE: the hash lock is dropped in this function.
2740 */
2741static void
2742arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2743{
2744 clock_t now;
2745
2746 ASSERT(MUTEX_HELD(hash_lock));
2747
2748 if (buf->b_state == arc_anon) {
2749 /*
2750 * This buffer is not in the cache, and does not
2751 * appear in our "ghost" list. Add the new buffer
2752 * to the MRU state.
2753 */
2754
2755 ASSERT(buf->b_arc_access == 0);
2756 buf->b_arc_access = ddi_get_lbolt();
2757 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2758 arc_change_state(arc_mru, buf, hash_lock);
2759
2760 } else if (buf->b_state == arc_mru) {
2761 now = ddi_get_lbolt();
2762
2763 /*
2764 * If this buffer is here because of a prefetch, then either:
2765 * - clear the flag if this is a "referencing" read
2766 * (any subsequent access will bump this into the MFU state).
2767 * or
2768 * - move the buffer to the head of the list if this is
2769 * another prefetch (to make it less likely to be evicted).
2770 */
2771 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2772 if (refcount_count(&buf->b_refcnt) == 0) {
2773 ASSERT(list_link_active(&buf->b_arc_node));
2774 } else {
2775 buf->b_flags &= ~ARC_PREFETCH;
2776 ARCSTAT_BUMP(arcstat_mru_hits);
2777 }
2778 buf->b_arc_access = now;
2779 return;
2780 }
2781
2782 /*
2783 * This buffer has been "accessed" only once so far,
2784 * but it is still in the cache. Move it to the MFU
2785 * state.
2786 */
2787 if (now > buf->b_arc_access + ARC_MINTIME) {
2788 /*
2789 * More than 125ms have passed since we
2790 * instantiated this buffer. Move it to the
2791 * most frequently used state.
2792 */
2793 buf->b_arc_access = now;
2794 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2795 arc_change_state(arc_mfu, buf, hash_lock);
2796 }
2797 ARCSTAT_BUMP(arcstat_mru_hits);
2798 } else if (buf->b_state == arc_mru_ghost) {
2799 arc_state_t *new_state;
2800 /*
2801 * This buffer has been "accessed" recently, but
2802 * was evicted from the cache. Move it to the
2803 * MFU state.
2804 */
2805
2806 if (buf->b_flags & ARC_PREFETCH) {
2807 new_state = arc_mru;
2808 if (refcount_count(&buf->b_refcnt) > 0)
2809 buf->b_flags &= ~ARC_PREFETCH;
2810 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2811 } else {
2812 new_state = arc_mfu;
2813 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2814 }
2815
2816 buf->b_arc_access = ddi_get_lbolt();
2817 arc_change_state(new_state, buf, hash_lock);
2818
2819 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2820 } else if (buf->b_state == arc_mfu) {
2821 /*
2822 * This buffer has been accessed more than once and is
2823 * still in the cache. Keep it in the MFU state.
2824 *
2825 * NOTE: an add_reference() that occurred when we did
2826 * the arc_read() will have kicked this off the list.
2827 * If it was a prefetch, we will explicitly move it to
2828 * the head of the list now.
2829 */
2830 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2831 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2832 ASSERT(list_link_active(&buf->b_arc_node));
2833 }
2834 ARCSTAT_BUMP(arcstat_mfu_hits);
2835 buf->b_arc_access = ddi_get_lbolt();
2836 } else if (buf->b_state == arc_mfu_ghost) {
2837 arc_state_t *new_state = arc_mfu;
2838 /*
2839 * This buffer has been accessed more than once but has
2840 * been evicted from the cache. Move it back to the
2841 * MFU state.
2842 */
2843
2844 if (buf->b_flags & ARC_PREFETCH) {
2845 /*
2846 * This is a prefetch access...
2847 * move this block back to the MRU state.
2848 */
2849 ASSERT0(refcount_count(&buf->b_refcnt));
2850 new_state = arc_mru;
2851 }
2852
2853 buf->b_arc_access = ddi_get_lbolt();
2854 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2855 arc_change_state(new_state, buf, hash_lock);
2856
2857 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2858 } else if (buf->b_state == arc_l2c_only) {
2859 /*
2860 * This buffer is on the 2nd Level ARC.
2861 */
2862
2863 buf->b_arc_access = ddi_get_lbolt();
2864 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2865 arc_change_state(arc_mfu, buf, hash_lock);
2866 } else {
2867 ASSERT(!"invalid arc state");
2868 }
2869}
2870
2871/* a generic arc_done_func_t which you can use */
2872/* ARGSUSED */
2873void
2874arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2875{
2876 if (zio == NULL || zio->io_error == 0)
2877 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2878 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2879}
2880
2881/* a generic arc_done_func_t */
2882void
2883arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2884{
2885 arc_buf_t **bufp = arg;
2886 if (zio && zio->io_error) {
2887 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2888 *bufp = NULL;
2889 } else {
2890 *bufp = buf;
2891 ASSERT(buf->b_data);
2892 }
2893}
2894
2895static void
2896arc_read_done(zio_t *zio)
2897{
2898 arc_buf_hdr_t *hdr, *found;
2899 arc_buf_t *buf;
2900 arc_buf_t *abuf; /* buffer we're assigning to callback */
2901 kmutex_t *hash_lock;
2902 arc_callback_t *callback_list, *acb;
2903 int freeable = FALSE;
2904
2905 buf = zio->io_private;
2906 hdr = buf->b_hdr;
2907
2908 /*
2909 * The hdr was inserted into hash-table and removed from lists
2910 * prior to starting I/O. We should find this header, since
2911 * it's in the hash table, and it should be legit since it's
2912 * not possible to evict it during the I/O. The only possible
2913 * reason for it not to be found is if we were freed during the
2914 * read.
2915 */
2916 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2917 &hash_lock);
2918
2919 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2920 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2921 (found == hdr && HDR_L2_READING(hdr)));
2922
2923 hdr->b_flags &= ~ARC_L2_EVICTED;
2924 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2925 hdr->b_flags &= ~ARC_L2CACHE;
2926
2927 /* byteswap if necessary */
2928 callback_list = hdr->b_acb;
2929 ASSERT(callback_list != NULL);
2930 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2931 dmu_object_byteswap_t bswap =
2932 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2933 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2934 byteswap_uint64_array :
2935 dmu_ot_byteswap[bswap].ob_func;
2936 func(buf->b_data, hdr->b_size);
2937 }
2938
2939 arc_cksum_compute(buf, B_FALSE);
2940#ifdef illumos
2941 arc_buf_watch(buf);
2942#endif /* illumos */
2943
2944 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2945 /*
2946 * Only call arc_access on anonymous buffers. This is because
2947 * if we've issued an I/O for an evicted buffer, we've already
2948 * called arc_access (to prevent any simultaneous readers from
2949 * getting confused).
2950 */
2951 arc_access(hdr, hash_lock);
2952 }
2953
2954 /* create copies of the data buffer for the callers */
2955 abuf = buf;
2956 for (acb = callback_list; acb; acb = acb->acb_next) {
2957 if (acb->acb_done) {
2958 if (abuf == NULL) {
2959 ARCSTAT_BUMP(arcstat_duplicate_reads);
2960 abuf = arc_buf_clone(buf);
2961 }
2962 acb->acb_buf = abuf;
2963 abuf = NULL;
2964 }
2965 }
2966 hdr->b_acb = NULL;
2967 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2968 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2969 if (abuf == buf) {
2970 ASSERT(buf->b_efunc == NULL);
2971 ASSERT(hdr->b_datacnt == 1);
2972 hdr->b_flags |= ARC_BUF_AVAILABLE;
2973 }
2974
2975 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2976
2977 if (zio->io_error != 0) {
2978 hdr->b_flags |= ARC_IO_ERROR;
2979 if (hdr->b_state != arc_anon)
2980 arc_change_state(arc_anon, hdr, hash_lock);
2981 if (HDR_IN_HASH_TABLE(hdr))
2982 buf_hash_remove(hdr);
2983 freeable = refcount_is_zero(&hdr->b_refcnt);
2984 }
2985
2986 /*
2987 * Broadcast before we drop the hash_lock to avoid the possibility
2988 * that the hdr (and hence the cv) might be freed before we get to
2989 * the cv_broadcast().
2990 */
2991 cv_broadcast(&hdr->b_cv);
2992
2993 if (hash_lock) {
2994 mutex_exit(hash_lock);
2995 } else {
2996 /*
2997 * This block was freed while we waited for the read to
2998 * complete. It has been removed from the hash table and
2999 * moved to the anonymous state (so that it won't show up
3000 * in the cache).
3001 */
3002 ASSERT3P(hdr->b_state, ==, arc_anon);
3003 freeable = refcount_is_zero(&hdr->b_refcnt);
3004 }
3005
3006 /* execute each callback and free its structure */
3007 while ((acb = callback_list) != NULL) {
3008 if (acb->acb_done)
3009 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3010
3011 if (acb->acb_zio_dummy != NULL) {
3012 acb->acb_zio_dummy->io_error = zio->io_error;
3013 zio_nowait(acb->acb_zio_dummy);
3014 }
3015
3016 callback_list = acb->acb_next;
3017 kmem_free(acb, sizeof (arc_callback_t));
3018 }
3019
3020 if (freeable)
3021 arc_hdr_destroy(hdr);
3022}
3023
3024/*
3025 * "Read" the block block at the specified DVA (in bp) via the
3026 * cache. If the block is found in the cache, invoke the provided
3027 * callback immediately and return. Note that the `zio' parameter
3028 * in the callback will be NULL in this case, since no IO was
3029 * required. If the block is not in the cache pass the read request
3030 * on to the spa with a substitute callback function, so that the
3031 * requested block will be added to the cache.
3032 *
3033 * If a read request arrives for a block that has a read in-progress,
3034 * either wait for the in-progress read to complete (and return the
3035 * results); or, if this is a read with a "done" func, add a record
3036 * to the read to invoke the "done" func when the read completes,
3037 * and return; or just return.
3038 *
3039 * arc_read_done() will invoke all the requested "done" functions
3040 * for readers of this block.
3041 */
3042int
3043arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
3044 void *private, int priority, int zio_flags, uint32_t *arc_flags,
3045 const zbookmark_t *zb)
3046{
3047 arc_buf_hdr_t *hdr;
| 1/*
2 * CDDL HEADER START
3 *
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.
7 *
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.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21/*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2011 by Delphix. All rights reserved.
25 */
26
27/*
28 * DVA-based Adjustable Replacement Cache
29 *
30 * While much of the theory of operation used here is
31 * based on the self-tuning, low overhead replacement cache
32 * presented by Megiddo and Modha at FAST 2003, there are some
33 * significant differences:
34 *
35 * 1. The Megiddo and Modha model assumes any page is evictable.
36 * Pages in its cache cannot be "locked" into memory. This makes
37 * the eviction algorithm simple: evict the last page in the list.
38 * This also make the performance characteristics easy to reason
39 * about. Our cache is not so simple. At any given moment, some
40 * subset of the blocks in the cache are un-evictable because we
41 * have handed out a reference to them. Blocks are only evictable
42 * when there are no external references active. This makes
43 * eviction far more problematic: we choose to evict the evictable
44 * blocks that are the "lowest" in the list.
45 *
46 * There are times when it is not possible to evict the requested
47 * space. In these circumstances we are unable to adjust the cache
48 * size. To prevent the cache growing unbounded at these times we
49 * implement a "cache throttle" that slows the flow of new data
50 * into the cache until we can make space available.
51 *
52 * 2. The Megiddo and Modha model assumes a fixed cache size.
53 * Pages are evicted when the cache is full and there is a cache
54 * miss. Our model has a variable sized cache. It grows with
55 * high use, but also tries to react to memory pressure from the
56 * operating system: decreasing its size when system memory is
57 * tight.
58 *
59 * 3. The Megiddo and Modha model assumes a fixed page size. All
60 * elements of the cache are therefor exactly the same size. So
61 * when adjusting the cache size following a cache miss, its simply
62 * a matter of choosing a single page to evict. In our model, we
63 * have variable sized cache blocks (rangeing from 512 bytes to
64 * 128K bytes). We therefor choose a set of blocks to evict to make
65 * space for a cache miss that approximates as closely as possible
66 * the space used by the new block.
67 *
68 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
69 * by N. Megiddo & D. Modha, FAST 2003
70 */
71
72/*
73 * The locking model:
74 *
75 * A new reference to a cache buffer can be obtained in two
76 * ways: 1) via a hash table lookup using the DVA as a key,
77 * or 2) via one of the ARC lists. The arc_read() interface
78 * uses method 1, while the internal arc algorithms for
79 * adjusting the cache use method 2. We therefor provide two
80 * types of locks: 1) the hash table lock array, and 2) the
81 * arc list locks.
82 *
83 * Buffers do not have their own mutexs, rather they rely on the
84 * hash table mutexs for the bulk of their protection (i.e. most
85 * fields in the arc_buf_hdr_t are protected by these mutexs).
86 *
87 * buf_hash_find() returns the appropriate mutex (held) when it
88 * locates the requested buffer in the hash table. It returns
89 * NULL for the mutex if the buffer was not in the table.
90 *
91 * buf_hash_remove() expects the appropriate hash mutex to be
92 * already held before it is invoked.
93 *
94 * Each arc state also has a mutex which is used to protect the
95 * buffer list associated with the state. When attempting to
96 * obtain a hash table lock while holding an arc list lock you
97 * must use: mutex_tryenter() to avoid deadlock. Also note that
98 * the active state mutex must be held before the ghost state mutex.
99 *
100 * Arc buffers may have an associated eviction callback function.
101 * This function will be invoked prior to removing the buffer (e.g.
102 * in arc_do_user_evicts()). Note however that the data associated
103 * with the buffer may be evicted prior to the callback. The callback
104 * must be made with *no locks held* (to prevent deadlock). Additionally,
105 * the users of callbacks must ensure that their private data is
106 * protected from simultaneous callbacks from arc_buf_evict()
107 * and arc_do_user_evicts().
108 *
109 * Note that the majority of the performance stats are manipulated
110 * with atomic operations.
111 *
112 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
113 *
114 * - L2ARC buflist creation
115 * - L2ARC buflist eviction
116 * - L2ARC write completion, which walks L2ARC buflists
117 * - ARC header destruction, as it removes from L2ARC buflists
118 * - ARC header release, as it removes from L2ARC buflists
119 */
120
121#include <sys/spa.h>
122#include <sys/zio.h>
123#include <sys/zfs_context.h>
124#include <sys/arc.h>
125#include <sys/refcount.h>
126#include <sys/vdev.h>
127#include <sys/vdev_impl.h>
128#ifdef _KERNEL
129#include <sys/dnlc.h>
130#endif
131#include <sys/callb.h>
132#include <sys/kstat.h>
133#include <zfs_fletcher.h>
134#include <sys/sdt.h>
135
136#include <vm/vm_pageout.h>
137
138#ifdef illumos
139#ifndef _KERNEL
140/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
141boolean_t arc_watch = B_FALSE;
142int arc_procfd;
143#endif
144#endif /* illumos */
145
146static kmutex_t arc_reclaim_thr_lock;
147static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
148static uint8_t arc_thread_exit;
149
150extern int zfs_write_limit_shift;
151extern uint64_t zfs_write_limit_max;
152extern kmutex_t zfs_write_limit_lock;
153
154#define ARC_REDUCE_DNLC_PERCENT 3
155uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
156
157typedef enum arc_reclaim_strategy {
158 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
159 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
160} arc_reclaim_strategy_t;
161
162/* number of seconds before growing cache again */
163static int arc_grow_retry = 60;
164
165/* shift of arc_c for calculating both min and max arc_p */
166static int arc_p_min_shift = 4;
167
168/* log2(fraction of arc to reclaim) */
169static int arc_shrink_shift = 5;
170
171/*
172 * minimum lifespan of a prefetch block in clock ticks
173 * (initialized in arc_init())
174 */
175static int arc_min_prefetch_lifespan;
176
177static int arc_dead;
178extern int zfs_prefetch_disable;
179
180/*
181 * The arc has filled available memory and has now warmed up.
182 */
183static boolean_t arc_warm;
184
185/*
186 * These tunables are for performance analysis.
187 */
188uint64_t zfs_arc_max;
189uint64_t zfs_arc_min;
190uint64_t zfs_arc_meta_limit = 0;
191int zfs_arc_grow_retry = 0;
192int zfs_arc_shrink_shift = 0;
193int zfs_arc_p_min_shift = 0;
194int zfs_disable_dup_eviction = 0;
195
196TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
197TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
198TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
199SYSCTL_DECL(_vfs_zfs);
200SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0,
201 "Maximum ARC size");
202SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0,
203 "Minimum ARC size");
204
205/*
206 * Note that buffers can be in one of 6 states:
207 * ARC_anon - anonymous (discussed below)
208 * ARC_mru - recently used, currently cached
209 * ARC_mru_ghost - recentely used, no longer in cache
210 * ARC_mfu - frequently used, currently cached
211 * ARC_mfu_ghost - frequently used, no longer in cache
212 * ARC_l2c_only - exists in L2ARC but not other states
213 * When there are no active references to the buffer, they are
214 * are linked onto a list in one of these arc states. These are
215 * the only buffers that can be evicted or deleted. Within each
216 * state there are multiple lists, one for meta-data and one for
217 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
218 * etc.) is tracked separately so that it can be managed more
219 * explicitly: favored over data, limited explicitly.
220 *
221 * Anonymous buffers are buffers that are not associated with
222 * a DVA. These are buffers that hold dirty block copies
223 * before they are written to stable storage. By definition,
224 * they are "ref'd" and are considered part of arc_mru
225 * that cannot be freed. Generally, they will aquire a DVA
226 * as they are written and migrate onto the arc_mru list.
227 *
228 * The ARC_l2c_only state is for buffers that are in the second
229 * level ARC but no longer in any of the ARC_m* lists. The second
230 * level ARC itself may also contain buffers that are in any of
231 * the ARC_m* states - meaning that a buffer can exist in two
232 * places. The reason for the ARC_l2c_only state is to keep the
233 * buffer header in the hash table, so that reads that hit the
234 * second level ARC benefit from these fast lookups.
235 */
236
237#define ARCS_LOCK_PAD CACHE_LINE_SIZE
238struct arcs_lock {
239 kmutex_t arcs_lock;
240#ifdef _KERNEL
241 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))];
242#endif
243};
244
245/*
246 * must be power of two for mask use to work
247 *
248 */
249#define ARC_BUFC_NUMDATALISTS 16
250#define ARC_BUFC_NUMMETADATALISTS 16
251#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS)
252
253typedef struct arc_state {
254 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
255 uint64_t arcs_size; /* total amount of data in this state */
256 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */
257 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE);
258} arc_state_t;
259
260#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock))
261
262/* The 6 states: */
263static arc_state_t ARC_anon;
264static arc_state_t ARC_mru;
265static arc_state_t ARC_mru_ghost;
266static arc_state_t ARC_mfu;
267static arc_state_t ARC_mfu_ghost;
268static arc_state_t ARC_l2c_only;
269
270typedef struct arc_stats {
271 kstat_named_t arcstat_hits;
272 kstat_named_t arcstat_misses;
273 kstat_named_t arcstat_demand_data_hits;
274 kstat_named_t arcstat_demand_data_misses;
275 kstat_named_t arcstat_demand_metadata_hits;
276 kstat_named_t arcstat_demand_metadata_misses;
277 kstat_named_t arcstat_prefetch_data_hits;
278 kstat_named_t arcstat_prefetch_data_misses;
279 kstat_named_t arcstat_prefetch_metadata_hits;
280 kstat_named_t arcstat_prefetch_metadata_misses;
281 kstat_named_t arcstat_mru_hits;
282 kstat_named_t arcstat_mru_ghost_hits;
283 kstat_named_t arcstat_mfu_hits;
284 kstat_named_t arcstat_mfu_ghost_hits;
285 kstat_named_t arcstat_allocated;
286 kstat_named_t arcstat_deleted;
287 kstat_named_t arcstat_stolen;
288 kstat_named_t arcstat_recycle_miss;
289 kstat_named_t arcstat_mutex_miss;
290 kstat_named_t arcstat_evict_skip;
291 kstat_named_t arcstat_evict_l2_cached;
292 kstat_named_t arcstat_evict_l2_eligible;
293 kstat_named_t arcstat_evict_l2_ineligible;
294 kstat_named_t arcstat_hash_elements;
295 kstat_named_t arcstat_hash_elements_max;
296 kstat_named_t arcstat_hash_collisions;
297 kstat_named_t arcstat_hash_chains;
298 kstat_named_t arcstat_hash_chain_max;
299 kstat_named_t arcstat_p;
300 kstat_named_t arcstat_c;
301 kstat_named_t arcstat_c_min;
302 kstat_named_t arcstat_c_max;
303 kstat_named_t arcstat_size;
304 kstat_named_t arcstat_hdr_size;
305 kstat_named_t arcstat_data_size;
306 kstat_named_t arcstat_other_size;
307 kstat_named_t arcstat_l2_hits;
308 kstat_named_t arcstat_l2_misses;
309 kstat_named_t arcstat_l2_feeds;
310 kstat_named_t arcstat_l2_rw_clash;
311 kstat_named_t arcstat_l2_read_bytes;
312 kstat_named_t arcstat_l2_write_bytes;
313 kstat_named_t arcstat_l2_writes_sent;
314 kstat_named_t arcstat_l2_writes_done;
315 kstat_named_t arcstat_l2_writes_error;
316 kstat_named_t arcstat_l2_writes_hdr_miss;
317 kstat_named_t arcstat_l2_evict_lock_retry;
318 kstat_named_t arcstat_l2_evict_reading;
319 kstat_named_t arcstat_l2_free_on_write;
320 kstat_named_t arcstat_l2_abort_lowmem;
321 kstat_named_t arcstat_l2_cksum_bad;
322 kstat_named_t arcstat_l2_io_error;
323 kstat_named_t arcstat_l2_size;
324 kstat_named_t arcstat_l2_hdr_size;
325 kstat_named_t arcstat_l2_write_trylock_fail;
326 kstat_named_t arcstat_l2_write_passed_headroom;
327 kstat_named_t arcstat_l2_write_spa_mismatch;
328 kstat_named_t arcstat_l2_write_in_l2;
329 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
330 kstat_named_t arcstat_l2_write_not_cacheable;
331 kstat_named_t arcstat_l2_write_full;
332 kstat_named_t arcstat_l2_write_buffer_iter;
333 kstat_named_t arcstat_l2_write_pios;
334 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
335 kstat_named_t arcstat_l2_write_buffer_list_iter;
336 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
337 kstat_named_t arcstat_memory_throttle_count;
338 kstat_named_t arcstat_duplicate_buffers;
339 kstat_named_t arcstat_duplicate_buffers_size;
340 kstat_named_t arcstat_duplicate_reads;
341} arc_stats_t;
342
343static arc_stats_t arc_stats = {
344 { "hits", KSTAT_DATA_UINT64 },
345 { "misses", KSTAT_DATA_UINT64 },
346 { "demand_data_hits", KSTAT_DATA_UINT64 },
347 { "demand_data_misses", KSTAT_DATA_UINT64 },
348 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
349 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
350 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
351 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
352 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
353 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
354 { "mru_hits", KSTAT_DATA_UINT64 },
355 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
356 { "mfu_hits", KSTAT_DATA_UINT64 },
357 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
358 { "allocated", KSTAT_DATA_UINT64 },
359 { "deleted", KSTAT_DATA_UINT64 },
360 { "stolen", KSTAT_DATA_UINT64 },
361 { "recycle_miss", KSTAT_DATA_UINT64 },
362 { "mutex_miss", KSTAT_DATA_UINT64 },
363 { "evict_skip", KSTAT_DATA_UINT64 },
364 { "evict_l2_cached", KSTAT_DATA_UINT64 },
365 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
366 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
367 { "hash_elements", KSTAT_DATA_UINT64 },
368 { "hash_elements_max", KSTAT_DATA_UINT64 },
369 { "hash_collisions", KSTAT_DATA_UINT64 },
370 { "hash_chains", KSTAT_DATA_UINT64 },
371 { "hash_chain_max", KSTAT_DATA_UINT64 },
372 { "p", KSTAT_DATA_UINT64 },
373 { "c", KSTAT_DATA_UINT64 },
374 { "c_min", KSTAT_DATA_UINT64 },
375 { "c_max", KSTAT_DATA_UINT64 },
376 { "size", KSTAT_DATA_UINT64 },
377 { "hdr_size", KSTAT_DATA_UINT64 },
378 { "data_size", KSTAT_DATA_UINT64 },
379 { "other_size", KSTAT_DATA_UINT64 },
380 { "l2_hits", KSTAT_DATA_UINT64 },
381 { "l2_misses", KSTAT_DATA_UINT64 },
382 { "l2_feeds", KSTAT_DATA_UINT64 },
383 { "l2_rw_clash", KSTAT_DATA_UINT64 },
384 { "l2_read_bytes", KSTAT_DATA_UINT64 },
385 { "l2_write_bytes", KSTAT_DATA_UINT64 },
386 { "l2_writes_sent", KSTAT_DATA_UINT64 },
387 { "l2_writes_done", KSTAT_DATA_UINT64 },
388 { "l2_writes_error", KSTAT_DATA_UINT64 },
389 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
390 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
391 { "l2_evict_reading", KSTAT_DATA_UINT64 },
392 { "l2_free_on_write", KSTAT_DATA_UINT64 },
393 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
394 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
395 { "l2_io_error", KSTAT_DATA_UINT64 },
396 { "l2_size", KSTAT_DATA_UINT64 },
397 { "l2_hdr_size", KSTAT_DATA_UINT64 },
398 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
399 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
400 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
401 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
402 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
403 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
404 { "l2_write_full", KSTAT_DATA_UINT64 },
405 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
406 { "l2_write_pios", KSTAT_DATA_UINT64 },
407 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
408 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
409 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 },
410 { "memory_throttle_count", KSTAT_DATA_UINT64 },
411 { "duplicate_buffers", KSTAT_DATA_UINT64 },
412 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
413 { "duplicate_reads", KSTAT_DATA_UINT64 }
414};
415
416#define ARCSTAT(stat) (arc_stats.stat.value.ui64)
417
418#define ARCSTAT_INCR(stat, val) \
419 atomic_add_64(&arc_stats.stat.value.ui64, (val));
420
421#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
422#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
423
424#define ARCSTAT_MAX(stat, val) { \
425 uint64_t m; \
426 while ((val) > (m = arc_stats.stat.value.ui64) && \
427 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
428 continue; \
429}
430
431#define ARCSTAT_MAXSTAT(stat) \
432 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
433
434/*
435 * We define a macro to allow ARC hits/misses to be easily broken down by
436 * two separate conditions, giving a total of four different subtypes for
437 * each of hits and misses (so eight statistics total).
438 */
439#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
440 if (cond1) { \
441 if (cond2) { \
442 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
443 } else { \
444 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
445 } \
446 } else { \
447 if (cond2) { \
448 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
449 } else { \
450 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
451 } \
452 }
453
454kstat_t *arc_ksp;
455static arc_state_t *arc_anon;
456static arc_state_t *arc_mru;
457static arc_state_t *arc_mru_ghost;
458static arc_state_t *arc_mfu;
459static arc_state_t *arc_mfu_ghost;
460static arc_state_t *arc_l2c_only;
461
462/*
463 * There are several ARC variables that are critical to export as kstats --
464 * but we don't want to have to grovel around in the kstat whenever we wish to
465 * manipulate them. For these variables, we therefore define them to be in
466 * terms of the statistic variable. This assures that we are not introducing
467 * the possibility of inconsistency by having shadow copies of the variables,
468 * while still allowing the code to be readable.
469 */
470#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
471#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
472#define arc_c ARCSTAT(arcstat_c) /* target size of cache */
473#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
474#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
475
476static int arc_no_grow; /* Don't try to grow cache size */
477static uint64_t arc_tempreserve;
478static uint64_t arc_loaned_bytes;
479static uint64_t arc_meta_used;
480static uint64_t arc_meta_limit;
481static uint64_t arc_meta_max = 0;
482SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0,
483 "ARC metadata used");
484SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0,
485 "ARC metadata limit");
486
487typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
488
489typedef struct arc_callback arc_callback_t;
490
491struct arc_callback {
492 void *acb_private;
493 arc_done_func_t *acb_done;
494 arc_buf_t *acb_buf;
495 zio_t *acb_zio_dummy;
496 arc_callback_t *acb_next;
497};
498
499typedef struct arc_write_callback arc_write_callback_t;
500
501struct arc_write_callback {
502 void *awcb_private;
503 arc_done_func_t *awcb_ready;
504 arc_done_func_t *awcb_done;
505 arc_buf_t *awcb_buf;
506};
507
508struct arc_buf_hdr {
509 /* protected by hash lock */
510 dva_t b_dva;
511 uint64_t b_birth;
512 uint64_t b_cksum0;
513
514 kmutex_t b_freeze_lock;
515 zio_cksum_t *b_freeze_cksum;
516 void *b_thawed;
517
518 arc_buf_hdr_t *b_hash_next;
519 arc_buf_t *b_buf;
520 uint32_t b_flags;
521 uint32_t b_datacnt;
522
523 arc_callback_t *b_acb;
524 kcondvar_t b_cv;
525
526 /* immutable */
527 arc_buf_contents_t b_type;
528 uint64_t b_size;
529 uint64_t b_spa;
530
531 /* protected by arc state mutex */
532 arc_state_t *b_state;
533 list_node_t b_arc_node;
534
535 /* updated atomically */
536 clock_t b_arc_access;
537
538 /* self protecting */
539 refcount_t b_refcnt;
540
541 l2arc_buf_hdr_t *b_l2hdr;
542 list_node_t b_l2node;
543};
544
545static arc_buf_t *arc_eviction_list;
546static kmutex_t arc_eviction_mtx;
547static arc_buf_hdr_t arc_eviction_hdr;
548static void arc_get_data_buf(arc_buf_t *buf);
549static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
550static int arc_evict_needed(arc_buf_contents_t type);
551static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
552#ifdef illumos
553static void arc_buf_watch(arc_buf_t *buf);
554#endif /* illumos */
555
556static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
557
558#define GHOST_STATE(state) \
559 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
560 (state) == arc_l2c_only)
561
562/*
563 * Private ARC flags. These flags are private ARC only flags that will show up
564 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
565 * be passed in as arc_flags in things like arc_read. However, these flags
566 * should never be passed and should only be set by ARC code. When adding new
567 * public flags, make sure not to smash the private ones.
568 */
569
570#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
571#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
572#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
573#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
574#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
575#define ARC_INDIRECT (1 << 14) /* this is an indirect block */
576#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
577#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
578#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
579#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
580
581#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
582#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
583#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
584#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
585#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
586#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
587#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
588#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
589#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
590 (hdr)->b_l2hdr != NULL)
591#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
592#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
593#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
594
595/*
596 * Other sizes
597 */
598
599#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
600#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
601
602/*
603 * Hash table routines
604 */
605
606#define HT_LOCK_PAD CACHE_LINE_SIZE
607
608struct ht_lock {
609 kmutex_t ht_lock;
610#ifdef _KERNEL
611 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
612#endif
613};
614
615#define BUF_LOCKS 256
616typedef struct buf_hash_table {
617 uint64_t ht_mask;
618 arc_buf_hdr_t **ht_table;
619 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
620} buf_hash_table_t;
621
622static buf_hash_table_t buf_hash_table;
623
624#define BUF_HASH_INDEX(spa, dva, birth) \
625 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
626#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
627#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
628#define HDR_LOCK(hdr) \
629 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
630
631uint64_t zfs_crc64_table[256];
632
633/*
634 * Level 2 ARC
635 */
636
637#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
638#define L2ARC_HEADROOM 2 /* num of writes */
639#define L2ARC_FEED_SECS 1 /* caching interval secs */
640#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
641
642#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
643#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
644
645/*
646 * L2ARC Performance Tunables
647 */
648uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
649uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
650uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
651uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
652uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
653boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
654boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
655boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
656
657SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
658 &l2arc_write_max, 0, "max write size");
659SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
660 &l2arc_write_boost, 0, "extra write during warmup");
661SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
662 &l2arc_headroom, 0, "number of dev writes");
663SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
664 &l2arc_feed_secs, 0, "interval seconds");
665SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
666 &l2arc_feed_min_ms, 0, "min interval milliseconds");
667
668SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
669 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
670SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
671 &l2arc_feed_again, 0, "turbo warmup");
672SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
673 &l2arc_norw, 0, "no reads during writes");
674
675SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
676 &ARC_anon.arcs_size, 0, "size of anonymous state");
677SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
678 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
679SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
680 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
681
682SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
683 &ARC_mru.arcs_size, 0, "size of mru state");
684SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
685 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
686SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
687 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
688
689SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
690 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state");
691SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
692 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
693 "size of metadata in mru ghost state");
694SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
695 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
696 "size of data in mru ghost state");
697
698SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
699 &ARC_mfu.arcs_size, 0, "size of mfu state");
700SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
701 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
702SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
703 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
704
705SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
706 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state");
707SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
708 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
709 "size of metadata in mfu ghost state");
710SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
711 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
712 "size of data in mfu ghost state");
713
714SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
715 &ARC_l2c_only.arcs_size, 0, "size of mru state");
716
717/*
718 * L2ARC Internals
719 */
720typedef struct l2arc_dev {
721 vdev_t *l2ad_vdev; /* vdev */
722 spa_t *l2ad_spa; /* spa */
723 uint64_t l2ad_hand; /* next write location */
724 uint64_t l2ad_write; /* desired write size, bytes */
725 uint64_t l2ad_boost; /* warmup write boost, bytes */
726 uint64_t l2ad_start; /* first addr on device */
727 uint64_t l2ad_end; /* last addr on device */
728 uint64_t l2ad_evict; /* last addr eviction reached */
729 boolean_t l2ad_first; /* first sweep through */
730 boolean_t l2ad_writing; /* currently writing */
731 list_t *l2ad_buflist; /* buffer list */
732 list_node_t l2ad_node; /* device list node */
733} l2arc_dev_t;
734
735static list_t L2ARC_dev_list; /* device list */
736static list_t *l2arc_dev_list; /* device list pointer */
737static kmutex_t l2arc_dev_mtx; /* device list mutex */
738static l2arc_dev_t *l2arc_dev_last; /* last device used */
739static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
740static list_t L2ARC_free_on_write; /* free after write buf list */
741static list_t *l2arc_free_on_write; /* free after write list ptr */
742static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
743static uint64_t l2arc_ndev; /* number of devices */
744
745typedef struct l2arc_read_callback {
746 arc_buf_t *l2rcb_buf; /* read buffer */
747 spa_t *l2rcb_spa; /* spa */
748 blkptr_t l2rcb_bp; /* original blkptr */
749 zbookmark_t l2rcb_zb; /* original bookmark */
750 int l2rcb_flags; /* original flags */
751} l2arc_read_callback_t;
752
753typedef struct l2arc_write_callback {
754 l2arc_dev_t *l2wcb_dev; /* device info */
755 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
756} l2arc_write_callback_t;
757
758struct l2arc_buf_hdr {
759 /* protected by arc_buf_hdr mutex */
760 l2arc_dev_t *b_dev; /* L2ARC device */
761 uint64_t b_daddr; /* disk address, offset byte */
762};
763
764typedef struct l2arc_data_free {
765 /* protected by l2arc_free_on_write_mtx */
766 void *l2df_data;
767 size_t l2df_size;
768 void (*l2df_func)(void *, size_t);
769 list_node_t l2df_list_node;
770} l2arc_data_free_t;
771
772static kmutex_t l2arc_feed_thr_lock;
773static kcondvar_t l2arc_feed_thr_cv;
774static uint8_t l2arc_thread_exit;
775
776static void l2arc_read_done(zio_t *zio);
777static void l2arc_hdr_stat_add(void);
778static void l2arc_hdr_stat_remove(void);
779
780static uint64_t
781buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
782{
783 uint8_t *vdva = (uint8_t *)dva;
784 uint64_t crc = -1ULL;
785 int i;
786
787 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
788
789 for (i = 0; i < sizeof (dva_t); i++)
790 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
791
792 crc ^= (spa>>8) ^ birth;
793
794 return (crc);
795}
796
797#define BUF_EMPTY(buf) \
798 ((buf)->b_dva.dva_word[0] == 0 && \
799 (buf)->b_dva.dva_word[1] == 0 && \
800 (buf)->b_birth == 0)
801
802#define BUF_EQUAL(spa, dva, birth, buf) \
803 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
804 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
805 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
806
807static void
808buf_discard_identity(arc_buf_hdr_t *hdr)
809{
810 hdr->b_dva.dva_word[0] = 0;
811 hdr->b_dva.dva_word[1] = 0;
812 hdr->b_birth = 0;
813 hdr->b_cksum0 = 0;
814}
815
816static arc_buf_hdr_t *
817buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
818{
819 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
820 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
821 arc_buf_hdr_t *buf;
822
823 mutex_enter(hash_lock);
824 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
825 buf = buf->b_hash_next) {
826 if (BUF_EQUAL(spa, dva, birth, buf)) {
827 *lockp = hash_lock;
828 return (buf);
829 }
830 }
831 mutex_exit(hash_lock);
832 *lockp = NULL;
833 return (NULL);
834}
835
836/*
837 * Insert an entry into the hash table. If there is already an element
838 * equal to elem in the hash table, then the already existing element
839 * will be returned and the new element will not be inserted.
840 * Otherwise returns NULL.
841 */
842static arc_buf_hdr_t *
843buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
844{
845 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
846 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
847 arc_buf_hdr_t *fbuf;
848 uint32_t i;
849
850 ASSERT(!HDR_IN_HASH_TABLE(buf));
851 *lockp = hash_lock;
852 mutex_enter(hash_lock);
853 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
854 fbuf = fbuf->b_hash_next, i++) {
855 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
856 return (fbuf);
857 }
858
859 buf->b_hash_next = buf_hash_table.ht_table[idx];
860 buf_hash_table.ht_table[idx] = buf;
861 buf->b_flags |= ARC_IN_HASH_TABLE;
862
863 /* collect some hash table performance data */
864 if (i > 0) {
865 ARCSTAT_BUMP(arcstat_hash_collisions);
866 if (i == 1)
867 ARCSTAT_BUMP(arcstat_hash_chains);
868
869 ARCSTAT_MAX(arcstat_hash_chain_max, i);
870 }
871
872 ARCSTAT_BUMP(arcstat_hash_elements);
873 ARCSTAT_MAXSTAT(arcstat_hash_elements);
874
875 return (NULL);
876}
877
878static void
879buf_hash_remove(arc_buf_hdr_t *buf)
880{
881 arc_buf_hdr_t *fbuf, **bufp;
882 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
883
884 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
885 ASSERT(HDR_IN_HASH_TABLE(buf));
886
887 bufp = &buf_hash_table.ht_table[idx];
888 while ((fbuf = *bufp) != buf) {
889 ASSERT(fbuf != NULL);
890 bufp = &fbuf->b_hash_next;
891 }
892 *bufp = buf->b_hash_next;
893 buf->b_hash_next = NULL;
894 buf->b_flags &= ~ARC_IN_HASH_TABLE;
895
896 /* collect some hash table performance data */
897 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
898
899 if (buf_hash_table.ht_table[idx] &&
900 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
901 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
902}
903
904/*
905 * Global data structures and functions for the buf kmem cache.
906 */
907static kmem_cache_t *hdr_cache;
908static kmem_cache_t *buf_cache;
909
910static void
911buf_fini(void)
912{
913 int i;
914
915 kmem_free(buf_hash_table.ht_table,
916 (buf_hash_table.ht_mask + 1) * sizeof (void *));
917 for (i = 0; i < BUF_LOCKS; i++)
918 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
919 kmem_cache_destroy(hdr_cache);
920 kmem_cache_destroy(buf_cache);
921}
922
923/*
924 * Constructor callback - called when the cache is empty
925 * and a new buf is requested.
926 */
927/* ARGSUSED */
928static int
929hdr_cons(void *vbuf, void *unused, int kmflag)
930{
931 arc_buf_hdr_t *buf = vbuf;
932
933 bzero(buf, sizeof (arc_buf_hdr_t));
934 refcount_create(&buf->b_refcnt);
935 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
936 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
937 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
938
939 return (0);
940}
941
942/* ARGSUSED */
943static int
944buf_cons(void *vbuf, void *unused, int kmflag)
945{
946 arc_buf_t *buf = vbuf;
947
948 bzero(buf, sizeof (arc_buf_t));
949 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
950 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
951
952 return (0);
953}
954
955/*
956 * Destructor callback - called when a cached buf is
957 * no longer required.
958 */
959/* ARGSUSED */
960static void
961hdr_dest(void *vbuf, void *unused)
962{
963 arc_buf_hdr_t *buf = vbuf;
964
965 ASSERT(BUF_EMPTY(buf));
966 refcount_destroy(&buf->b_refcnt);
967 cv_destroy(&buf->b_cv);
968 mutex_destroy(&buf->b_freeze_lock);
969 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
970}
971
972/* ARGSUSED */
973static void
974buf_dest(void *vbuf, void *unused)
975{
976 arc_buf_t *buf = vbuf;
977
978 mutex_destroy(&buf->b_evict_lock);
979 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
980}
981
982/*
983 * Reclaim callback -- invoked when memory is low.
984 */
985/* ARGSUSED */
986static void
987hdr_recl(void *unused)
988{
989 dprintf("hdr_recl called\n");
990 /*
991 * umem calls the reclaim func when we destroy the buf cache,
992 * which is after we do arc_fini().
993 */
994 if (!arc_dead)
995 cv_signal(&arc_reclaim_thr_cv);
996}
997
998static void
999buf_init(void)
1000{
1001 uint64_t *ct;
1002 uint64_t hsize = 1ULL << 12;
1003 int i, j;
1004
1005 /*
1006 * The hash table is big enough to fill all of physical memory
1007 * with an average 64K block size. The table will take up
1008 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
1009 */
1010 while (hsize * 65536 < (uint64_t)physmem * PAGESIZE)
1011 hsize <<= 1;
1012retry:
1013 buf_hash_table.ht_mask = hsize - 1;
1014 buf_hash_table.ht_table =
1015 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1016 if (buf_hash_table.ht_table == NULL) {
1017 ASSERT(hsize > (1ULL << 8));
1018 hsize >>= 1;
1019 goto retry;
1020 }
1021
1022 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
1023 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
1024 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1025 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1026
1027 for (i = 0; i < 256; i++)
1028 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1029 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1030
1031 for (i = 0; i < BUF_LOCKS; i++) {
1032 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1033 NULL, MUTEX_DEFAULT, NULL);
1034 }
1035}
1036
1037#define ARC_MINTIME (hz>>4) /* 62 ms */
1038
1039static void
1040arc_cksum_verify(arc_buf_t *buf)
1041{
1042 zio_cksum_t zc;
1043
1044 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1045 return;
1046
1047 mutex_enter(&buf->b_hdr->b_freeze_lock);
1048 if (buf->b_hdr->b_freeze_cksum == NULL ||
1049 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
1050 mutex_exit(&buf->b_hdr->b_freeze_lock);
1051 return;
1052 }
1053 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1054 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1055 panic("buffer modified while frozen!");
1056 mutex_exit(&buf->b_hdr->b_freeze_lock);
1057}
1058
1059static int
1060arc_cksum_equal(arc_buf_t *buf)
1061{
1062 zio_cksum_t zc;
1063 int equal;
1064
1065 mutex_enter(&buf->b_hdr->b_freeze_lock);
1066 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1067 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1068 mutex_exit(&buf->b_hdr->b_freeze_lock);
1069
1070 return (equal);
1071}
1072
1073static void
1074arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1075{
1076 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1077 return;
1078
1079 mutex_enter(&buf->b_hdr->b_freeze_lock);
1080 if (buf->b_hdr->b_freeze_cksum != NULL) {
1081 mutex_exit(&buf->b_hdr->b_freeze_lock);
1082 return;
1083 }
1084 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1085 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1086 buf->b_hdr->b_freeze_cksum);
1087 mutex_exit(&buf->b_hdr->b_freeze_lock);
1088#ifdef illumos
1089 arc_buf_watch(buf);
1090#endif /* illumos */
1091}
1092
1093#ifdef illumos
1094#ifndef _KERNEL
1095typedef struct procctl {
1096 long cmd;
1097 prwatch_t prwatch;
1098} procctl_t;
1099#endif
1100
1101/* ARGSUSED */
1102static void
1103arc_buf_unwatch(arc_buf_t *buf)
1104{
1105#ifndef _KERNEL
1106 if (arc_watch) {
1107 int result;
1108 procctl_t ctl;
1109 ctl.cmd = PCWATCH;
1110 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1111 ctl.prwatch.pr_size = 0;
1112 ctl.prwatch.pr_wflags = 0;
1113 result = write(arc_procfd, &ctl, sizeof (ctl));
1114 ASSERT3U(result, ==, sizeof (ctl));
1115 }
1116#endif
1117}
1118
1119/* ARGSUSED */
1120static void
1121arc_buf_watch(arc_buf_t *buf)
1122{
1123#ifndef _KERNEL
1124 if (arc_watch) {
1125 int result;
1126 procctl_t ctl;
1127 ctl.cmd = PCWATCH;
1128 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1129 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1130 ctl.prwatch.pr_wflags = WA_WRITE;
1131 result = write(arc_procfd, &ctl, sizeof (ctl));
1132 ASSERT3U(result, ==, sizeof (ctl));
1133 }
1134#endif
1135}
1136#endif /* illumos */
1137
1138void
1139arc_buf_thaw(arc_buf_t *buf)
1140{
1141 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1142 if (buf->b_hdr->b_state != arc_anon)
1143 panic("modifying non-anon buffer!");
1144 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1145 panic("modifying buffer while i/o in progress!");
1146 arc_cksum_verify(buf);
1147 }
1148
1149 mutex_enter(&buf->b_hdr->b_freeze_lock);
1150 if (buf->b_hdr->b_freeze_cksum != NULL) {
1151 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1152 buf->b_hdr->b_freeze_cksum = NULL;
1153 }
1154
1155 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1156 if (buf->b_hdr->b_thawed)
1157 kmem_free(buf->b_hdr->b_thawed, 1);
1158 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1159 }
1160
1161 mutex_exit(&buf->b_hdr->b_freeze_lock);
1162
1163#ifdef illumos
1164 arc_buf_unwatch(buf);
1165#endif /* illumos */
1166}
1167
1168void
1169arc_buf_freeze(arc_buf_t *buf)
1170{
1171 kmutex_t *hash_lock;
1172
1173 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1174 return;
1175
1176 hash_lock = HDR_LOCK(buf->b_hdr);
1177 mutex_enter(hash_lock);
1178
1179 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1180 buf->b_hdr->b_state == arc_anon);
1181 arc_cksum_compute(buf, B_FALSE);
1182 mutex_exit(hash_lock);
1183
1184}
1185
1186static void
1187get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock)
1188{
1189 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth);
1190
1191 if (ab->b_type == ARC_BUFC_METADATA)
1192 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1);
1193 else {
1194 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1);
1195 buf_hashid += ARC_BUFC_NUMMETADATALISTS;
1196 }
1197
1198 *list = &state->arcs_lists[buf_hashid];
1199 *lock = ARCS_LOCK(state, buf_hashid);
1200}
1201
1202
1203static void
1204add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1205{
1206 ASSERT(MUTEX_HELD(hash_lock));
1207
1208 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1209 (ab->b_state != arc_anon)) {
1210 uint64_t delta = ab->b_size * ab->b_datacnt;
1211 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1212 list_t *list;
1213 kmutex_t *lock;
1214
1215 get_buf_info(ab, ab->b_state, &list, &lock);
1216 ASSERT(!MUTEX_HELD(lock));
1217 mutex_enter(lock);
1218 ASSERT(list_link_active(&ab->b_arc_node));
1219 list_remove(list, ab);
1220 if (GHOST_STATE(ab->b_state)) {
1221 ASSERT0(ab->b_datacnt);
1222 ASSERT3P(ab->b_buf, ==, NULL);
1223 delta = ab->b_size;
1224 }
1225 ASSERT(delta > 0);
1226 ASSERT3U(*size, >=, delta);
1227 atomic_add_64(size, -delta);
1228 mutex_exit(lock);
1229 /* remove the prefetch flag if we get a reference */
1230 if (ab->b_flags & ARC_PREFETCH)
1231 ab->b_flags &= ~ARC_PREFETCH;
1232 }
1233}
1234
1235static int
1236remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1237{
1238 int cnt;
1239 arc_state_t *state = ab->b_state;
1240
1241 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1242 ASSERT(!GHOST_STATE(state));
1243
1244 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1245 (state != arc_anon)) {
1246 uint64_t *size = &state->arcs_lsize[ab->b_type];
1247 list_t *list;
1248 kmutex_t *lock;
1249
1250 get_buf_info(ab, state, &list, &lock);
1251 ASSERT(!MUTEX_HELD(lock));
1252 mutex_enter(lock);
1253 ASSERT(!list_link_active(&ab->b_arc_node));
1254 list_insert_head(list, ab);
1255 ASSERT(ab->b_datacnt > 0);
1256 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1257 mutex_exit(lock);
1258 }
1259 return (cnt);
1260}
1261
1262/*
1263 * Move the supplied buffer to the indicated state. The mutex
1264 * for the buffer must be held by the caller.
1265 */
1266static void
1267arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1268{
1269 arc_state_t *old_state = ab->b_state;
1270 int64_t refcnt = refcount_count(&ab->b_refcnt);
1271 uint64_t from_delta, to_delta;
1272 list_t *list;
1273 kmutex_t *lock;
1274
1275 ASSERT(MUTEX_HELD(hash_lock));
1276 ASSERT(new_state != old_state);
1277 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1278 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1279 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1280
1281 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1282
1283 /*
1284 * If this buffer is evictable, transfer it from the
1285 * old state list to the new state list.
1286 */
1287 if (refcnt == 0) {
1288 if (old_state != arc_anon) {
1289 int use_mutex;
1290 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1291
1292 get_buf_info(ab, old_state, &list, &lock);
1293 use_mutex = !MUTEX_HELD(lock);
1294 if (use_mutex)
1295 mutex_enter(lock);
1296
1297 ASSERT(list_link_active(&ab->b_arc_node));
1298 list_remove(list, ab);
1299
1300 /*
1301 * If prefetching out of the ghost cache,
1302 * we will have a non-zero datacnt.
1303 */
1304 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1305 /* ghost elements have a ghost size */
1306 ASSERT(ab->b_buf == NULL);
1307 from_delta = ab->b_size;
1308 }
1309 ASSERT3U(*size, >=, from_delta);
1310 atomic_add_64(size, -from_delta);
1311
1312 if (use_mutex)
1313 mutex_exit(lock);
1314 }
1315 if (new_state != arc_anon) {
1316 int use_mutex;
1317 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1318
1319 get_buf_info(ab, new_state, &list, &lock);
1320 use_mutex = !MUTEX_HELD(lock);
1321 if (use_mutex)
1322 mutex_enter(lock);
1323
1324 list_insert_head(list, ab);
1325
1326 /* ghost elements have a ghost size */
1327 if (GHOST_STATE(new_state)) {
1328 ASSERT(ab->b_datacnt == 0);
1329 ASSERT(ab->b_buf == NULL);
1330 to_delta = ab->b_size;
1331 }
1332 atomic_add_64(size, to_delta);
1333
1334 if (use_mutex)
1335 mutex_exit(lock);
1336 }
1337 }
1338
1339 ASSERT(!BUF_EMPTY(ab));
1340 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1341 buf_hash_remove(ab);
1342
1343 /* adjust state sizes */
1344 if (to_delta)
1345 atomic_add_64(&new_state->arcs_size, to_delta);
1346 if (from_delta) {
1347 ASSERT3U(old_state->arcs_size, >=, from_delta);
1348 atomic_add_64(&old_state->arcs_size, -from_delta);
1349 }
1350 ab->b_state = new_state;
1351
1352 /* adjust l2arc hdr stats */
1353 if (new_state == arc_l2c_only)
1354 l2arc_hdr_stat_add();
1355 else if (old_state == arc_l2c_only)
1356 l2arc_hdr_stat_remove();
1357}
1358
1359void
1360arc_space_consume(uint64_t space, arc_space_type_t type)
1361{
1362 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1363
1364 switch (type) {
1365 case ARC_SPACE_DATA:
1366 ARCSTAT_INCR(arcstat_data_size, space);
1367 break;
1368 case ARC_SPACE_OTHER:
1369 ARCSTAT_INCR(arcstat_other_size, space);
1370 break;
1371 case ARC_SPACE_HDRS:
1372 ARCSTAT_INCR(arcstat_hdr_size, space);
1373 break;
1374 case ARC_SPACE_L2HDRS:
1375 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1376 break;
1377 }
1378
1379 atomic_add_64(&arc_meta_used, space);
1380 atomic_add_64(&arc_size, space);
1381}
1382
1383void
1384arc_space_return(uint64_t space, arc_space_type_t type)
1385{
1386 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1387
1388 switch (type) {
1389 case ARC_SPACE_DATA:
1390 ARCSTAT_INCR(arcstat_data_size, -space);
1391 break;
1392 case ARC_SPACE_OTHER:
1393 ARCSTAT_INCR(arcstat_other_size, -space);
1394 break;
1395 case ARC_SPACE_HDRS:
1396 ARCSTAT_INCR(arcstat_hdr_size, -space);
1397 break;
1398 case ARC_SPACE_L2HDRS:
1399 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1400 break;
1401 }
1402
1403 ASSERT(arc_meta_used >= space);
1404 if (arc_meta_max < arc_meta_used)
1405 arc_meta_max = arc_meta_used;
1406 atomic_add_64(&arc_meta_used, -space);
1407 ASSERT(arc_size >= space);
1408 atomic_add_64(&arc_size, -space);
1409}
1410
1411void *
1412arc_data_buf_alloc(uint64_t size)
1413{
1414 if (arc_evict_needed(ARC_BUFC_DATA))
1415 cv_signal(&arc_reclaim_thr_cv);
1416 atomic_add_64(&arc_size, size);
1417 return (zio_data_buf_alloc(size));
1418}
1419
1420void
1421arc_data_buf_free(void *buf, uint64_t size)
1422{
1423 zio_data_buf_free(buf, size);
1424 ASSERT(arc_size >= size);
1425 atomic_add_64(&arc_size, -size);
1426}
1427
1428arc_buf_t *
1429arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1430{
1431 arc_buf_hdr_t *hdr;
1432 arc_buf_t *buf;
1433
1434 ASSERT3U(size, >, 0);
1435 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1436 ASSERT(BUF_EMPTY(hdr));
1437 hdr->b_size = size;
1438 hdr->b_type = type;
1439 hdr->b_spa = spa_load_guid(spa);
1440 hdr->b_state = arc_anon;
1441 hdr->b_arc_access = 0;
1442 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1443 buf->b_hdr = hdr;
1444 buf->b_data = NULL;
1445 buf->b_efunc = NULL;
1446 buf->b_private = NULL;
1447 buf->b_next = NULL;
1448 hdr->b_buf = buf;
1449 arc_get_data_buf(buf);
1450 hdr->b_datacnt = 1;
1451 hdr->b_flags = 0;
1452 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1453 (void) refcount_add(&hdr->b_refcnt, tag);
1454
1455 return (buf);
1456}
1457
1458static char *arc_onloan_tag = "onloan";
1459
1460/*
1461 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1462 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1463 * buffers must be returned to the arc before they can be used by the DMU or
1464 * freed.
1465 */
1466arc_buf_t *
1467arc_loan_buf(spa_t *spa, int size)
1468{
1469 arc_buf_t *buf;
1470
1471 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1472
1473 atomic_add_64(&arc_loaned_bytes, size);
1474 return (buf);
1475}
1476
1477/*
1478 * Return a loaned arc buffer to the arc.
1479 */
1480void
1481arc_return_buf(arc_buf_t *buf, void *tag)
1482{
1483 arc_buf_hdr_t *hdr = buf->b_hdr;
1484
1485 ASSERT(buf->b_data != NULL);
1486 (void) refcount_add(&hdr->b_refcnt, tag);
1487 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1488
1489 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1490}
1491
1492/* Detach an arc_buf from a dbuf (tag) */
1493void
1494arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1495{
1496 arc_buf_hdr_t *hdr;
1497
1498 ASSERT(buf->b_data != NULL);
1499 hdr = buf->b_hdr;
1500 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1501 (void) refcount_remove(&hdr->b_refcnt, tag);
1502 buf->b_efunc = NULL;
1503 buf->b_private = NULL;
1504
1505 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1506}
1507
1508static arc_buf_t *
1509arc_buf_clone(arc_buf_t *from)
1510{
1511 arc_buf_t *buf;
1512 arc_buf_hdr_t *hdr = from->b_hdr;
1513 uint64_t size = hdr->b_size;
1514
1515 ASSERT(hdr->b_state != arc_anon);
1516
1517 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1518 buf->b_hdr = hdr;
1519 buf->b_data = NULL;
1520 buf->b_efunc = NULL;
1521 buf->b_private = NULL;
1522 buf->b_next = hdr->b_buf;
1523 hdr->b_buf = buf;
1524 arc_get_data_buf(buf);
1525 bcopy(from->b_data, buf->b_data, size);
1526
1527 /*
1528 * This buffer already exists in the arc so create a duplicate
1529 * copy for the caller. If the buffer is associated with user data
1530 * then track the size and number of duplicates. These stats will be
1531 * updated as duplicate buffers are created and destroyed.
1532 */
1533 if (hdr->b_type == ARC_BUFC_DATA) {
1534 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1535 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1536 }
1537 hdr->b_datacnt += 1;
1538 return (buf);
1539}
1540
1541void
1542arc_buf_add_ref(arc_buf_t *buf, void* tag)
1543{
1544 arc_buf_hdr_t *hdr;
1545 kmutex_t *hash_lock;
1546
1547 /*
1548 * Check to see if this buffer is evicted. Callers
1549 * must verify b_data != NULL to know if the add_ref
1550 * was successful.
1551 */
1552 mutex_enter(&buf->b_evict_lock);
1553 if (buf->b_data == NULL) {
1554 mutex_exit(&buf->b_evict_lock);
1555 return;
1556 }
1557 hash_lock = HDR_LOCK(buf->b_hdr);
1558 mutex_enter(hash_lock);
1559 hdr = buf->b_hdr;
1560 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1561 mutex_exit(&buf->b_evict_lock);
1562
1563 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1564 add_reference(hdr, hash_lock, tag);
1565 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1566 arc_access(hdr, hash_lock);
1567 mutex_exit(hash_lock);
1568 ARCSTAT_BUMP(arcstat_hits);
1569 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1570 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1571 data, metadata, hits);
1572}
1573
1574/*
1575 * Free the arc data buffer. If it is an l2arc write in progress,
1576 * the buffer is placed on l2arc_free_on_write to be freed later.
1577 */
1578static void
1579arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1580{
1581 arc_buf_hdr_t *hdr = buf->b_hdr;
1582
1583 if (HDR_L2_WRITING(hdr)) {
1584 l2arc_data_free_t *df;
1585 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1586 df->l2df_data = buf->b_data;
1587 df->l2df_size = hdr->b_size;
1588 df->l2df_func = free_func;
1589 mutex_enter(&l2arc_free_on_write_mtx);
1590 list_insert_head(l2arc_free_on_write, df);
1591 mutex_exit(&l2arc_free_on_write_mtx);
1592 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1593 } else {
1594 free_func(buf->b_data, hdr->b_size);
1595 }
1596}
1597
1598static void
1599arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1600{
1601 arc_buf_t **bufp;
1602
1603 /* free up data associated with the buf */
1604 if (buf->b_data) {
1605 arc_state_t *state = buf->b_hdr->b_state;
1606 uint64_t size = buf->b_hdr->b_size;
1607 arc_buf_contents_t type = buf->b_hdr->b_type;
1608
1609 arc_cksum_verify(buf);
1610#ifdef illumos
1611 arc_buf_unwatch(buf);
1612#endif /* illumos */
1613
1614 if (!recycle) {
1615 if (type == ARC_BUFC_METADATA) {
1616 arc_buf_data_free(buf, zio_buf_free);
1617 arc_space_return(size, ARC_SPACE_DATA);
1618 } else {
1619 ASSERT(type == ARC_BUFC_DATA);
1620 arc_buf_data_free(buf, zio_data_buf_free);
1621 ARCSTAT_INCR(arcstat_data_size, -size);
1622 atomic_add_64(&arc_size, -size);
1623 }
1624 }
1625 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1626 uint64_t *cnt = &state->arcs_lsize[type];
1627
1628 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1629 ASSERT(state != arc_anon);
1630
1631 ASSERT3U(*cnt, >=, size);
1632 atomic_add_64(cnt, -size);
1633 }
1634 ASSERT3U(state->arcs_size, >=, size);
1635 atomic_add_64(&state->arcs_size, -size);
1636 buf->b_data = NULL;
1637
1638 /*
1639 * If we're destroying a duplicate buffer make sure
1640 * that the appropriate statistics are updated.
1641 */
1642 if (buf->b_hdr->b_datacnt > 1 &&
1643 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1644 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1645 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1646 }
1647 ASSERT(buf->b_hdr->b_datacnt > 0);
1648 buf->b_hdr->b_datacnt -= 1;
1649 }
1650
1651 /* only remove the buf if requested */
1652 if (!all)
1653 return;
1654
1655 /* remove the buf from the hdr list */
1656 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1657 continue;
1658 *bufp = buf->b_next;
1659 buf->b_next = NULL;
1660
1661 ASSERT(buf->b_efunc == NULL);
1662
1663 /* clean up the buf */
1664 buf->b_hdr = NULL;
1665 kmem_cache_free(buf_cache, buf);
1666}
1667
1668static void
1669arc_hdr_destroy(arc_buf_hdr_t *hdr)
1670{
1671 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1672 ASSERT3P(hdr->b_state, ==, arc_anon);
1673 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1674 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1675
1676 if (l2hdr != NULL) {
1677 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1678 /*
1679 * To prevent arc_free() and l2arc_evict() from
1680 * attempting to free the same buffer at the same time,
1681 * a FREE_IN_PROGRESS flag is given to arc_free() to
1682 * give it priority. l2arc_evict() can't destroy this
1683 * header while we are waiting on l2arc_buflist_mtx.
1684 *
1685 * The hdr may be removed from l2ad_buflist before we
1686 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1687 */
1688 if (!buflist_held) {
1689 mutex_enter(&l2arc_buflist_mtx);
1690 l2hdr = hdr->b_l2hdr;
1691 }
1692
1693 if (l2hdr != NULL) {
1694 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1695 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1696 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1697 if (hdr->b_state == arc_l2c_only)
1698 l2arc_hdr_stat_remove();
1699 hdr->b_l2hdr = NULL;
1700 }
1701
1702 if (!buflist_held)
1703 mutex_exit(&l2arc_buflist_mtx);
1704 }
1705
1706 if (!BUF_EMPTY(hdr)) {
1707 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1708 buf_discard_identity(hdr);
1709 }
1710 while (hdr->b_buf) {
1711 arc_buf_t *buf = hdr->b_buf;
1712
1713 if (buf->b_efunc) {
1714 mutex_enter(&arc_eviction_mtx);
1715 mutex_enter(&buf->b_evict_lock);
1716 ASSERT(buf->b_hdr != NULL);
1717 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1718 hdr->b_buf = buf->b_next;
1719 buf->b_hdr = &arc_eviction_hdr;
1720 buf->b_next = arc_eviction_list;
1721 arc_eviction_list = buf;
1722 mutex_exit(&buf->b_evict_lock);
1723 mutex_exit(&arc_eviction_mtx);
1724 } else {
1725 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1726 }
1727 }
1728 if (hdr->b_freeze_cksum != NULL) {
1729 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1730 hdr->b_freeze_cksum = NULL;
1731 }
1732 if (hdr->b_thawed) {
1733 kmem_free(hdr->b_thawed, 1);
1734 hdr->b_thawed = NULL;
1735 }
1736
1737 ASSERT(!list_link_active(&hdr->b_arc_node));
1738 ASSERT3P(hdr->b_hash_next, ==, NULL);
1739 ASSERT3P(hdr->b_acb, ==, NULL);
1740 kmem_cache_free(hdr_cache, hdr);
1741}
1742
1743void
1744arc_buf_free(arc_buf_t *buf, void *tag)
1745{
1746 arc_buf_hdr_t *hdr = buf->b_hdr;
1747 int hashed = hdr->b_state != arc_anon;
1748
1749 ASSERT(buf->b_efunc == NULL);
1750 ASSERT(buf->b_data != NULL);
1751
1752 if (hashed) {
1753 kmutex_t *hash_lock = HDR_LOCK(hdr);
1754
1755 mutex_enter(hash_lock);
1756 hdr = buf->b_hdr;
1757 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1758
1759 (void) remove_reference(hdr, hash_lock, tag);
1760 if (hdr->b_datacnt > 1) {
1761 arc_buf_destroy(buf, FALSE, TRUE);
1762 } else {
1763 ASSERT(buf == hdr->b_buf);
1764 ASSERT(buf->b_efunc == NULL);
1765 hdr->b_flags |= ARC_BUF_AVAILABLE;
1766 }
1767 mutex_exit(hash_lock);
1768 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1769 int destroy_hdr;
1770 /*
1771 * We are in the middle of an async write. Don't destroy
1772 * this buffer unless the write completes before we finish
1773 * decrementing the reference count.
1774 */
1775 mutex_enter(&arc_eviction_mtx);
1776 (void) remove_reference(hdr, NULL, tag);
1777 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1778 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1779 mutex_exit(&arc_eviction_mtx);
1780 if (destroy_hdr)
1781 arc_hdr_destroy(hdr);
1782 } else {
1783 if (remove_reference(hdr, NULL, tag) > 0)
1784 arc_buf_destroy(buf, FALSE, TRUE);
1785 else
1786 arc_hdr_destroy(hdr);
1787 }
1788}
1789
1790int
1791arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1792{
1793 arc_buf_hdr_t *hdr = buf->b_hdr;
1794 kmutex_t *hash_lock = HDR_LOCK(hdr);
1795 int no_callback = (buf->b_efunc == NULL);
1796
1797 if (hdr->b_state == arc_anon) {
1798 ASSERT(hdr->b_datacnt == 1);
1799 arc_buf_free(buf, tag);
1800 return (no_callback);
1801 }
1802
1803 mutex_enter(hash_lock);
1804 hdr = buf->b_hdr;
1805 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1806 ASSERT(hdr->b_state != arc_anon);
1807 ASSERT(buf->b_data != NULL);
1808
1809 (void) remove_reference(hdr, hash_lock, tag);
1810 if (hdr->b_datacnt > 1) {
1811 if (no_callback)
1812 arc_buf_destroy(buf, FALSE, TRUE);
1813 } else if (no_callback) {
1814 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1815 ASSERT(buf->b_efunc == NULL);
1816 hdr->b_flags |= ARC_BUF_AVAILABLE;
1817 }
1818 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1819 refcount_is_zero(&hdr->b_refcnt));
1820 mutex_exit(hash_lock);
1821 return (no_callback);
1822}
1823
1824int
1825arc_buf_size(arc_buf_t *buf)
1826{
1827 return (buf->b_hdr->b_size);
1828}
1829
1830/*
1831 * Called from the DMU to determine if the current buffer should be
1832 * evicted. In order to ensure proper locking, the eviction must be initiated
1833 * from the DMU. Return true if the buffer is associated with user data and
1834 * duplicate buffers still exist.
1835 */
1836boolean_t
1837arc_buf_eviction_needed(arc_buf_t *buf)
1838{
1839 arc_buf_hdr_t *hdr;
1840 boolean_t evict_needed = B_FALSE;
1841
1842 if (zfs_disable_dup_eviction)
1843 return (B_FALSE);
1844
1845 mutex_enter(&buf->b_evict_lock);
1846 hdr = buf->b_hdr;
1847 if (hdr == NULL) {
1848 /*
1849 * We are in arc_do_user_evicts(); let that function
1850 * perform the eviction.
1851 */
1852 ASSERT(buf->b_data == NULL);
1853 mutex_exit(&buf->b_evict_lock);
1854 return (B_FALSE);
1855 } else if (buf->b_data == NULL) {
1856 /*
1857 * We have already been added to the arc eviction list;
1858 * recommend eviction.
1859 */
1860 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1861 mutex_exit(&buf->b_evict_lock);
1862 return (B_TRUE);
1863 }
1864
1865 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1866 evict_needed = B_TRUE;
1867
1868 mutex_exit(&buf->b_evict_lock);
1869 return (evict_needed);
1870}
1871
1872/*
1873 * Evict buffers from list until we've removed the specified number of
1874 * bytes. Move the removed buffers to the appropriate evict state.
1875 * If the recycle flag is set, then attempt to "recycle" a buffer:
1876 * - look for a buffer to evict that is `bytes' long.
1877 * - return the data block from this buffer rather than freeing it.
1878 * This flag is used by callers that are trying to make space for a
1879 * new buffer in a full arc cache.
1880 *
1881 * This function makes a "best effort". It skips over any buffers
1882 * it can't get a hash_lock on, and so may not catch all candidates.
1883 * It may also return without evicting as much space as requested.
1884 */
1885static void *
1886arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1887 arc_buf_contents_t type)
1888{
1889 arc_state_t *evicted_state;
1890 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1891 int64_t bytes_remaining;
1892 arc_buf_hdr_t *ab, *ab_prev = NULL;
1893 list_t *evicted_list, *list, *evicted_list_start, *list_start;
1894 kmutex_t *lock, *evicted_lock;
1895 kmutex_t *hash_lock;
1896 boolean_t have_lock;
1897 void *stolen = NULL;
1898 static int evict_metadata_offset, evict_data_offset;
1899 int i, idx, offset, list_count, count;
1900
1901 ASSERT(state == arc_mru || state == arc_mfu);
1902
1903 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1904
1905 if (type == ARC_BUFC_METADATA) {
1906 offset = 0;
1907 list_count = ARC_BUFC_NUMMETADATALISTS;
1908 list_start = &state->arcs_lists[0];
1909 evicted_list_start = &evicted_state->arcs_lists[0];
1910 idx = evict_metadata_offset;
1911 } else {
1912 offset = ARC_BUFC_NUMMETADATALISTS;
1913 list_start = &state->arcs_lists[offset];
1914 evicted_list_start = &evicted_state->arcs_lists[offset];
1915 list_count = ARC_BUFC_NUMDATALISTS;
1916 idx = evict_data_offset;
1917 }
1918 bytes_remaining = evicted_state->arcs_lsize[type];
1919 count = 0;
1920
1921evict_start:
1922 list = &list_start[idx];
1923 evicted_list = &evicted_list_start[idx];
1924 lock = ARCS_LOCK(state, (offset + idx));
1925 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx));
1926
1927 mutex_enter(lock);
1928 mutex_enter(evicted_lock);
1929
1930 for (ab = list_tail(list); ab; ab = ab_prev) {
1931 ab_prev = list_prev(list, ab);
1932 bytes_remaining -= (ab->b_size * ab->b_datacnt);
1933 /* prefetch buffers have a minimum lifespan */
1934 if (HDR_IO_IN_PROGRESS(ab) ||
1935 (spa && ab->b_spa != spa) ||
1936 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1937 ddi_get_lbolt() - ab->b_arc_access <
1938 arc_min_prefetch_lifespan)) {
1939 skipped++;
1940 continue;
1941 }
1942 /* "lookahead" for better eviction candidate */
1943 if (recycle && ab->b_size != bytes &&
1944 ab_prev && ab_prev->b_size == bytes)
1945 continue;
1946 hash_lock = HDR_LOCK(ab);
1947 have_lock = MUTEX_HELD(hash_lock);
1948 if (have_lock || mutex_tryenter(hash_lock)) {
1949 ASSERT0(refcount_count(&ab->b_refcnt));
1950 ASSERT(ab->b_datacnt > 0);
1951 while (ab->b_buf) {
1952 arc_buf_t *buf = ab->b_buf;
1953 if (!mutex_tryenter(&buf->b_evict_lock)) {
1954 missed += 1;
1955 break;
1956 }
1957 if (buf->b_data) {
1958 bytes_evicted += ab->b_size;
1959 if (recycle && ab->b_type == type &&
1960 ab->b_size == bytes &&
1961 !HDR_L2_WRITING(ab)) {
1962 stolen = buf->b_data;
1963 recycle = FALSE;
1964 }
1965 }
1966 if (buf->b_efunc) {
1967 mutex_enter(&arc_eviction_mtx);
1968 arc_buf_destroy(buf,
1969 buf->b_data == stolen, FALSE);
1970 ab->b_buf = buf->b_next;
1971 buf->b_hdr = &arc_eviction_hdr;
1972 buf->b_next = arc_eviction_list;
1973 arc_eviction_list = buf;
1974 mutex_exit(&arc_eviction_mtx);
1975 mutex_exit(&buf->b_evict_lock);
1976 } else {
1977 mutex_exit(&buf->b_evict_lock);
1978 arc_buf_destroy(buf,
1979 buf->b_data == stolen, TRUE);
1980 }
1981 }
1982
1983 if (ab->b_l2hdr) {
1984 ARCSTAT_INCR(arcstat_evict_l2_cached,
1985 ab->b_size);
1986 } else {
1987 if (l2arc_write_eligible(ab->b_spa, ab)) {
1988 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1989 ab->b_size);
1990 } else {
1991 ARCSTAT_INCR(
1992 arcstat_evict_l2_ineligible,
1993 ab->b_size);
1994 }
1995 }
1996
1997 if (ab->b_datacnt == 0) {
1998 arc_change_state(evicted_state, ab, hash_lock);
1999 ASSERT(HDR_IN_HASH_TABLE(ab));
2000 ab->b_flags |= ARC_IN_HASH_TABLE;
2001 ab->b_flags &= ~ARC_BUF_AVAILABLE;
2002 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
2003 }
2004 if (!have_lock)
2005 mutex_exit(hash_lock);
2006 if (bytes >= 0 && bytes_evicted >= bytes)
2007 break;
2008 if (bytes_remaining > 0) {
2009 mutex_exit(evicted_lock);
2010 mutex_exit(lock);
2011 idx = ((idx + 1) & (list_count - 1));
2012 count++;
2013 goto evict_start;
2014 }
2015 } else {
2016 missed += 1;
2017 }
2018 }
2019
2020 mutex_exit(evicted_lock);
2021 mutex_exit(lock);
2022
2023 idx = ((idx + 1) & (list_count - 1));
2024 count++;
2025
2026 if (bytes_evicted < bytes) {
2027 if (count < list_count)
2028 goto evict_start;
2029 else
2030 dprintf("only evicted %lld bytes from %x",
2031 (longlong_t)bytes_evicted, state);
2032 }
2033 if (type == ARC_BUFC_METADATA)
2034 evict_metadata_offset = idx;
2035 else
2036 evict_data_offset = idx;
2037
2038 if (skipped)
2039 ARCSTAT_INCR(arcstat_evict_skip, skipped);
2040
2041 if (missed)
2042 ARCSTAT_INCR(arcstat_mutex_miss, missed);
2043
2044 /*
2045 * We have just evicted some date into the ghost state, make
2046 * sure we also adjust the ghost state size if necessary.
2047 */
2048 if (arc_no_grow &&
2049 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
2050 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
2051 arc_mru_ghost->arcs_size - arc_c;
2052
2053 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
2054 int64_t todelete =
2055 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
2056 arc_evict_ghost(arc_mru_ghost, 0, todelete);
2057 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
2058 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
2059 arc_mru_ghost->arcs_size +
2060 arc_mfu_ghost->arcs_size - arc_c);
2061 arc_evict_ghost(arc_mfu_ghost, 0, todelete);
2062 }
2063 }
2064 if (stolen)
2065 ARCSTAT_BUMP(arcstat_stolen);
2066
2067 return (stolen);
2068}
2069
2070/*
2071 * Remove buffers from list until we've removed the specified number of
2072 * bytes. Destroy the buffers that are removed.
2073 */
2074static void
2075arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
2076{
2077 arc_buf_hdr_t *ab, *ab_prev;
2078 arc_buf_hdr_t marker = { 0 };
2079 list_t *list, *list_start;
2080 kmutex_t *hash_lock, *lock;
2081 uint64_t bytes_deleted = 0;
2082 uint64_t bufs_skipped = 0;
2083 static int evict_offset;
2084 int list_count, idx = evict_offset;
2085 int offset, count = 0;
2086
2087 ASSERT(GHOST_STATE(state));
2088
2089 /*
2090 * data lists come after metadata lists
2091 */
2092 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS];
2093 list_count = ARC_BUFC_NUMDATALISTS;
2094 offset = ARC_BUFC_NUMMETADATALISTS;
2095
2096evict_start:
2097 list = &list_start[idx];
2098 lock = ARCS_LOCK(state, idx + offset);
2099
2100 mutex_enter(lock);
2101 for (ab = list_tail(list); ab; ab = ab_prev) {
2102 ab_prev = list_prev(list, ab);
2103 if (spa && ab->b_spa != spa)
2104 continue;
2105
2106 /* ignore markers */
2107 if (ab->b_spa == 0)
2108 continue;
2109
2110 hash_lock = HDR_LOCK(ab);
2111 /* caller may be trying to modify this buffer, skip it */
2112 if (MUTEX_HELD(hash_lock))
2113 continue;
2114 if (mutex_tryenter(hash_lock)) {
2115 ASSERT(!HDR_IO_IN_PROGRESS(ab));
2116 ASSERT(ab->b_buf == NULL);
2117 ARCSTAT_BUMP(arcstat_deleted);
2118 bytes_deleted += ab->b_size;
2119
2120 if (ab->b_l2hdr != NULL) {
2121 /*
2122 * This buffer is cached on the 2nd Level ARC;
2123 * don't destroy the header.
2124 */
2125 arc_change_state(arc_l2c_only, ab, hash_lock);
2126 mutex_exit(hash_lock);
2127 } else {
2128 arc_change_state(arc_anon, ab, hash_lock);
2129 mutex_exit(hash_lock);
2130 arc_hdr_destroy(ab);
2131 }
2132
2133 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
2134 if (bytes >= 0 && bytes_deleted >= bytes)
2135 break;
2136 } else if (bytes < 0) {
2137 /*
2138 * Insert a list marker and then wait for the
2139 * hash lock to become available. Once its
2140 * available, restart from where we left off.
2141 */
2142 list_insert_after(list, ab, &marker);
2143 mutex_exit(lock);
2144 mutex_enter(hash_lock);
2145 mutex_exit(hash_lock);
2146 mutex_enter(lock);
2147 ab_prev = list_prev(list, &marker);
2148 list_remove(list, &marker);
2149 } else
2150 bufs_skipped += 1;
2151 }
2152 mutex_exit(lock);
2153 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1));
2154 count++;
2155
2156 if (count < list_count)
2157 goto evict_start;
2158
2159 evict_offset = idx;
2160 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] &&
2161 (bytes < 0 || bytes_deleted < bytes)) {
2162 list_start = &state->arcs_lists[0];
2163 list_count = ARC_BUFC_NUMMETADATALISTS;
2164 offset = count = 0;
2165 goto evict_start;
2166 }
2167
2168 if (bufs_skipped) {
2169 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2170 ASSERT(bytes >= 0);
2171 }
2172
2173 if (bytes_deleted < bytes)
2174 dprintf("only deleted %lld bytes from %p",
2175 (longlong_t)bytes_deleted, state);
2176}
2177
2178static void
2179arc_adjust(void)
2180{
2181 int64_t adjustment, delta;
2182
2183 /*
2184 * Adjust MRU size
2185 */
2186
2187 adjustment = MIN((int64_t)(arc_size - arc_c),
2188 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2189 arc_p));
2190
2191 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2192 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2193 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA);
2194 adjustment -= delta;
2195 }
2196
2197 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2198 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2199 (void) arc_evict(arc_mru, 0, delta, FALSE,
2200 ARC_BUFC_METADATA);
2201 }
2202
2203 /*
2204 * Adjust MFU size
2205 */
2206
2207 adjustment = arc_size - arc_c;
2208
2209 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2210 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2211 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA);
2212 adjustment -= delta;
2213 }
2214
2215 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2216 int64_t delta = MIN(adjustment,
2217 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2218 (void) arc_evict(arc_mfu, 0, delta, FALSE,
2219 ARC_BUFC_METADATA);
2220 }
2221
2222 /*
2223 * Adjust ghost lists
2224 */
2225
2226 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2227
2228 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2229 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2230 arc_evict_ghost(arc_mru_ghost, 0, delta);
2231 }
2232
2233 adjustment =
2234 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2235
2236 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2237 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2238 arc_evict_ghost(arc_mfu_ghost, 0, delta);
2239 }
2240}
2241
2242static void
2243arc_do_user_evicts(void)
2244{
2245 static arc_buf_t *tmp_arc_eviction_list;
2246
2247 /*
2248 * Move list over to avoid LOR
2249 */
2250restart:
2251 mutex_enter(&arc_eviction_mtx);
2252 tmp_arc_eviction_list = arc_eviction_list;
2253 arc_eviction_list = NULL;
2254 mutex_exit(&arc_eviction_mtx);
2255
2256 while (tmp_arc_eviction_list != NULL) {
2257 arc_buf_t *buf = tmp_arc_eviction_list;
2258 tmp_arc_eviction_list = buf->b_next;
2259 mutex_enter(&buf->b_evict_lock);
2260 buf->b_hdr = NULL;
2261 mutex_exit(&buf->b_evict_lock);
2262
2263 if (buf->b_efunc != NULL)
2264 VERIFY(buf->b_efunc(buf) == 0);
2265
2266 buf->b_efunc = NULL;
2267 buf->b_private = NULL;
2268 kmem_cache_free(buf_cache, buf);
2269 }
2270
2271 if (arc_eviction_list != NULL)
2272 goto restart;
2273}
2274
2275/*
2276 * Flush all *evictable* data from the cache for the given spa.
2277 * NOTE: this will not touch "active" (i.e. referenced) data.
2278 */
2279void
2280arc_flush(spa_t *spa)
2281{
2282 uint64_t guid = 0;
2283
2284 if (spa)
2285 guid = spa_load_guid(spa);
2286
2287 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) {
2288 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2289 if (spa)
2290 break;
2291 }
2292 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) {
2293 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2294 if (spa)
2295 break;
2296 }
2297 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) {
2298 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2299 if (spa)
2300 break;
2301 }
2302 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) {
2303 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2304 if (spa)
2305 break;
2306 }
2307
2308 arc_evict_ghost(arc_mru_ghost, guid, -1);
2309 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2310
2311 mutex_enter(&arc_reclaim_thr_lock);
2312 arc_do_user_evicts();
2313 mutex_exit(&arc_reclaim_thr_lock);
2314 ASSERT(spa || arc_eviction_list == NULL);
2315}
2316
2317void
2318arc_shrink(void)
2319{
2320 if (arc_c > arc_c_min) {
2321 uint64_t to_free;
2322
2323#ifdef _KERNEL
2324 to_free = arc_c >> arc_shrink_shift;
2325#else
2326 to_free = arc_c >> arc_shrink_shift;
2327#endif
2328 if (arc_c > arc_c_min + to_free)
2329 atomic_add_64(&arc_c, -to_free);
2330 else
2331 arc_c = arc_c_min;
2332
2333 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2334 if (arc_c > arc_size)
2335 arc_c = MAX(arc_size, arc_c_min);
2336 if (arc_p > arc_c)
2337 arc_p = (arc_c >> 1);
2338 ASSERT(arc_c >= arc_c_min);
2339 ASSERT((int64_t)arc_p >= 0);
2340 }
2341
2342 if (arc_size > arc_c)
2343 arc_adjust();
2344}
2345
2346static int needfree = 0;
2347
2348static int
2349arc_reclaim_needed(void)
2350{
2351
2352#ifdef _KERNEL
2353
2354 if (needfree)
2355 return (1);
2356
2357 /*
2358 * Cooperate with pagedaemon when it's time for it to scan
2359 * and reclaim some pages.
2360 */
2361 if (vm_paging_needed())
2362 return (1);
2363
2364#ifdef sun
2365 /*
2366 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2367 */
2368 extra = desfree;
2369
2370 /*
2371 * check that we're out of range of the pageout scanner. It starts to
2372 * schedule paging if freemem is less than lotsfree and needfree.
2373 * lotsfree is the high-water mark for pageout, and needfree is the
2374 * number of needed free pages. We add extra pages here to make sure
2375 * the scanner doesn't start up while we're freeing memory.
2376 */
2377 if (freemem < lotsfree + needfree + extra)
2378 return (1);
2379
2380 /*
2381 * check to make sure that swapfs has enough space so that anon
2382 * reservations can still succeed. anon_resvmem() checks that the
2383 * availrmem is greater than swapfs_minfree, and the number of reserved
2384 * swap pages. We also add a bit of extra here just to prevent
2385 * circumstances from getting really dire.
2386 */
2387 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2388 return (1);
2389
2390#if defined(__i386)
2391 /*
2392 * If we're on an i386 platform, it's possible that we'll exhaust the
2393 * kernel heap space before we ever run out of available physical
2394 * memory. Most checks of the size of the heap_area compare against
2395 * tune.t_minarmem, which is the minimum available real memory that we
2396 * can have in the system. However, this is generally fixed at 25 pages
2397 * which is so low that it's useless. In this comparison, we seek to
2398 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2399 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2400 * free)
2401 */
2402 if (btop(vmem_size(heap_arena, VMEM_FREE)) <
2403 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
2404 return (1);
2405#endif
2406#else /* !sun */
2407 if (kmem_used() > (kmem_size() * 3) / 4)
2408 return (1);
2409#endif /* sun */
2410
2411#else
2412 if (spa_get_random(100) == 0)
2413 return (1);
2414#endif
2415 return (0);
2416}
2417
2418extern kmem_cache_t *zio_buf_cache[];
2419extern kmem_cache_t *zio_data_buf_cache[];
2420
2421static void
2422arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2423{
2424 size_t i;
2425 kmem_cache_t *prev_cache = NULL;
2426 kmem_cache_t *prev_data_cache = NULL;
2427
2428#ifdef _KERNEL
2429 if (arc_meta_used >= arc_meta_limit) {
2430 /*
2431 * We are exceeding our meta-data cache limit.
2432 * Purge some DNLC entries to release holds on meta-data.
2433 */
2434 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2435 }
2436#if defined(__i386)
2437 /*
2438 * Reclaim unused memory from all kmem caches.
2439 */
2440 kmem_reap();
2441#endif
2442#endif
2443
2444 /*
2445 * An aggressive reclamation will shrink the cache size as well as
2446 * reap free buffers from the arc kmem caches.
2447 */
2448 if (strat == ARC_RECLAIM_AGGR)
2449 arc_shrink();
2450
2451 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2452 if (zio_buf_cache[i] != prev_cache) {
2453 prev_cache = zio_buf_cache[i];
2454 kmem_cache_reap_now(zio_buf_cache[i]);
2455 }
2456 if (zio_data_buf_cache[i] != prev_data_cache) {
2457 prev_data_cache = zio_data_buf_cache[i];
2458 kmem_cache_reap_now(zio_data_buf_cache[i]);
2459 }
2460 }
2461 kmem_cache_reap_now(buf_cache);
2462 kmem_cache_reap_now(hdr_cache);
2463}
2464
2465static void
2466arc_reclaim_thread(void *dummy __unused)
2467{
2468 clock_t growtime = 0;
2469 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2470 callb_cpr_t cpr;
2471
2472 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2473
2474 mutex_enter(&arc_reclaim_thr_lock);
2475 while (arc_thread_exit == 0) {
2476 if (arc_reclaim_needed()) {
2477
2478 if (arc_no_grow) {
2479 if (last_reclaim == ARC_RECLAIM_CONS) {
2480 last_reclaim = ARC_RECLAIM_AGGR;
2481 } else {
2482 last_reclaim = ARC_RECLAIM_CONS;
2483 }
2484 } else {
2485 arc_no_grow = TRUE;
2486 last_reclaim = ARC_RECLAIM_AGGR;
2487 membar_producer();
2488 }
2489
2490 /* reset the growth delay for every reclaim */
2491 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2492
2493 if (needfree && last_reclaim == ARC_RECLAIM_CONS) {
2494 /*
2495 * If needfree is TRUE our vm_lowmem hook
2496 * was called and in that case we must free some
2497 * memory, so switch to aggressive mode.
2498 */
2499 arc_no_grow = TRUE;
2500 last_reclaim = ARC_RECLAIM_AGGR;
2501 }
2502 arc_kmem_reap_now(last_reclaim);
2503 arc_warm = B_TRUE;
2504
2505 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2506 arc_no_grow = FALSE;
2507 }
2508
2509 arc_adjust();
2510
2511 if (arc_eviction_list != NULL)
2512 arc_do_user_evicts();
2513
2514#ifdef _KERNEL
2515 if (needfree) {
2516 needfree = 0;
2517 wakeup(&needfree);
2518 }
2519#endif
2520
2521 /* block until needed, or one second, whichever is shorter */
2522 CALLB_CPR_SAFE_BEGIN(&cpr);
2523 (void) cv_timedwait(&arc_reclaim_thr_cv,
2524 &arc_reclaim_thr_lock, hz);
2525 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2526 }
2527
2528 arc_thread_exit = 0;
2529 cv_broadcast(&arc_reclaim_thr_cv);
2530 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2531 thread_exit();
2532}
2533
2534/*
2535 * Adapt arc info given the number of bytes we are trying to add and
2536 * the state that we are comming from. This function is only called
2537 * when we are adding new content to the cache.
2538 */
2539static void
2540arc_adapt(int bytes, arc_state_t *state)
2541{
2542 int mult;
2543 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2544
2545 if (state == arc_l2c_only)
2546 return;
2547
2548 ASSERT(bytes > 0);
2549 /*
2550 * Adapt the target size of the MRU list:
2551 * - if we just hit in the MRU ghost list, then increase
2552 * the target size of the MRU list.
2553 * - if we just hit in the MFU ghost list, then increase
2554 * the target size of the MFU list by decreasing the
2555 * target size of the MRU list.
2556 */
2557 if (state == arc_mru_ghost) {
2558 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2559 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2560 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2561
2562 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2563 } else if (state == arc_mfu_ghost) {
2564 uint64_t delta;
2565
2566 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2567 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2568 mult = MIN(mult, 10);
2569
2570 delta = MIN(bytes * mult, arc_p);
2571 arc_p = MAX(arc_p_min, arc_p - delta);
2572 }
2573 ASSERT((int64_t)arc_p >= 0);
2574
2575 if (arc_reclaim_needed()) {
2576 cv_signal(&arc_reclaim_thr_cv);
2577 return;
2578 }
2579
2580 if (arc_no_grow)
2581 return;
2582
2583 if (arc_c >= arc_c_max)
2584 return;
2585
2586 /*
2587 * If we're within (2 * maxblocksize) bytes of the target
2588 * cache size, increment the target cache size
2589 */
2590 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2591 atomic_add_64(&arc_c, (int64_t)bytes);
2592 if (arc_c > arc_c_max)
2593 arc_c = arc_c_max;
2594 else if (state == arc_anon)
2595 atomic_add_64(&arc_p, (int64_t)bytes);
2596 if (arc_p > arc_c)
2597 arc_p = arc_c;
2598 }
2599 ASSERT((int64_t)arc_p >= 0);
2600}
2601
2602/*
2603 * Check if the cache has reached its limits and eviction is required
2604 * prior to insert.
2605 */
2606static int
2607arc_evict_needed(arc_buf_contents_t type)
2608{
2609 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2610 return (1);
2611
2612#ifdef sun
2613#ifdef _KERNEL
2614 /*
2615 * If zio data pages are being allocated out of a separate heap segment,
2616 * then enforce that the size of available vmem for this area remains
2617 * above about 1/32nd free.
2618 */
2619 if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2620 vmem_size(zio_arena, VMEM_FREE) <
2621 (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2622 return (1);
2623#endif
2624#endif /* sun */
2625
2626 if (arc_reclaim_needed())
2627 return (1);
2628
2629 return (arc_size > arc_c);
2630}
2631
2632/*
2633 * The buffer, supplied as the first argument, needs a data block.
2634 * So, if we are at cache max, determine which cache should be victimized.
2635 * We have the following cases:
2636 *
2637 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2638 * In this situation if we're out of space, but the resident size of the MFU is
2639 * under the limit, victimize the MFU cache to satisfy this insertion request.
2640 *
2641 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2642 * Here, we've used up all of the available space for the MRU, so we need to
2643 * evict from our own cache instead. Evict from the set of resident MRU
2644 * entries.
2645 *
2646 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2647 * c minus p represents the MFU space in the cache, since p is the size of the
2648 * cache that is dedicated to the MRU. In this situation there's still space on
2649 * the MFU side, so the MRU side needs to be victimized.
2650 *
2651 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2652 * MFU's resident set is consuming more space than it has been allotted. In
2653 * this situation, we must victimize our own cache, the MFU, for this insertion.
2654 */
2655static void
2656arc_get_data_buf(arc_buf_t *buf)
2657{
2658 arc_state_t *state = buf->b_hdr->b_state;
2659 uint64_t size = buf->b_hdr->b_size;
2660 arc_buf_contents_t type = buf->b_hdr->b_type;
2661
2662 arc_adapt(size, state);
2663
2664 /*
2665 * We have not yet reached cache maximum size,
2666 * just allocate a new buffer.
2667 */
2668 if (!arc_evict_needed(type)) {
2669 if (type == ARC_BUFC_METADATA) {
2670 buf->b_data = zio_buf_alloc(size);
2671 arc_space_consume(size, ARC_SPACE_DATA);
2672 } else {
2673 ASSERT(type == ARC_BUFC_DATA);
2674 buf->b_data = zio_data_buf_alloc(size);
2675 ARCSTAT_INCR(arcstat_data_size, size);
2676 atomic_add_64(&arc_size, size);
2677 }
2678 goto out;
2679 }
2680
2681 /*
2682 * If we are prefetching from the mfu ghost list, this buffer
2683 * will end up on the mru list; so steal space from there.
2684 */
2685 if (state == arc_mfu_ghost)
2686 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2687 else if (state == arc_mru_ghost)
2688 state = arc_mru;
2689
2690 if (state == arc_mru || state == arc_anon) {
2691 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2692 state = (arc_mfu->arcs_lsize[type] >= size &&
2693 arc_p > mru_used) ? arc_mfu : arc_mru;
2694 } else {
2695 /* MFU cases */
2696 uint64_t mfu_space = arc_c - arc_p;
2697 state = (arc_mru->arcs_lsize[type] >= size &&
2698 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2699 }
2700 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) {
2701 if (type == ARC_BUFC_METADATA) {
2702 buf->b_data = zio_buf_alloc(size);
2703 arc_space_consume(size, ARC_SPACE_DATA);
2704 } else {
2705 ASSERT(type == ARC_BUFC_DATA);
2706 buf->b_data = zio_data_buf_alloc(size);
2707 ARCSTAT_INCR(arcstat_data_size, size);
2708 atomic_add_64(&arc_size, size);
2709 }
2710 ARCSTAT_BUMP(arcstat_recycle_miss);
2711 }
2712 ASSERT(buf->b_data != NULL);
2713out:
2714 /*
2715 * Update the state size. Note that ghost states have a
2716 * "ghost size" and so don't need to be updated.
2717 */
2718 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2719 arc_buf_hdr_t *hdr = buf->b_hdr;
2720
2721 atomic_add_64(&hdr->b_state->arcs_size, size);
2722 if (list_link_active(&hdr->b_arc_node)) {
2723 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2724 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2725 }
2726 /*
2727 * If we are growing the cache, and we are adding anonymous
2728 * data, and we have outgrown arc_p, update arc_p
2729 */
2730 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2731 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2732 arc_p = MIN(arc_c, arc_p + size);
2733 }
2734 ARCSTAT_BUMP(arcstat_allocated);
2735}
2736
2737/*
2738 * This routine is called whenever a buffer is accessed.
2739 * NOTE: the hash lock is dropped in this function.
2740 */
2741static void
2742arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2743{
2744 clock_t now;
2745
2746 ASSERT(MUTEX_HELD(hash_lock));
2747
2748 if (buf->b_state == arc_anon) {
2749 /*
2750 * This buffer is not in the cache, and does not
2751 * appear in our "ghost" list. Add the new buffer
2752 * to the MRU state.
2753 */
2754
2755 ASSERT(buf->b_arc_access == 0);
2756 buf->b_arc_access = ddi_get_lbolt();
2757 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2758 arc_change_state(arc_mru, buf, hash_lock);
2759
2760 } else if (buf->b_state == arc_mru) {
2761 now = ddi_get_lbolt();
2762
2763 /*
2764 * If this buffer is here because of a prefetch, then either:
2765 * - clear the flag if this is a "referencing" read
2766 * (any subsequent access will bump this into the MFU state).
2767 * or
2768 * - move the buffer to the head of the list if this is
2769 * another prefetch (to make it less likely to be evicted).
2770 */
2771 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2772 if (refcount_count(&buf->b_refcnt) == 0) {
2773 ASSERT(list_link_active(&buf->b_arc_node));
2774 } else {
2775 buf->b_flags &= ~ARC_PREFETCH;
2776 ARCSTAT_BUMP(arcstat_mru_hits);
2777 }
2778 buf->b_arc_access = now;
2779 return;
2780 }
2781
2782 /*
2783 * This buffer has been "accessed" only once so far,
2784 * but it is still in the cache. Move it to the MFU
2785 * state.
2786 */
2787 if (now > buf->b_arc_access + ARC_MINTIME) {
2788 /*
2789 * More than 125ms have passed since we
2790 * instantiated this buffer. Move it to the
2791 * most frequently used state.
2792 */
2793 buf->b_arc_access = now;
2794 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2795 arc_change_state(arc_mfu, buf, hash_lock);
2796 }
2797 ARCSTAT_BUMP(arcstat_mru_hits);
2798 } else if (buf->b_state == arc_mru_ghost) {
2799 arc_state_t *new_state;
2800 /*
2801 * This buffer has been "accessed" recently, but
2802 * was evicted from the cache. Move it to the
2803 * MFU state.
2804 */
2805
2806 if (buf->b_flags & ARC_PREFETCH) {
2807 new_state = arc_mru;
2808 if (refcount_count(&buf->b_refcnt) > 0)
2809 buf->b_flags &= ~ARC_PREFETCH;
2810 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2811 } else {
2812 new_state = arc_mfu;
2813 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2814 }
2815
2816 buf->b_arc_access = ddi_get_lbolt();
2817 arc_change_state(new_state, buf, hash_lock);
2818
2819 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2820 } else if (buf->b_state == arc_mfu) {
2821 /*
2822 * This buffer has been accessed more than once and is
2823 * still in the cache. Keep it in the MFU state.
2824 *
2825 * NOTE: an add_reference() that occurred when we did
2826 * the arc_read() will have kicked this off the list.
2827 * If it was a prefetch, we will explicitly move it to
2828 * the head of the list now.
2829 */
2830 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2831 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2832 ASSERT(list_link_active(&buf->b_arc_node));
2833 }
2834 ARCSTAT_BUMP(arcstat_mfu_hits);
2835 buf->b_arc_access = ddi_get_lbolt();
2836 } else if (buf->b_state == arc_mfu_ghost) {
2837 arc_state_t *new_state = arc_mfu;
2838 /*
2839 * This buffer has been accessed more than once but has
2840 * been evicted from the cache. Move it back to the
2841 * MFU state.
2842 */
2843
2844 if (buf->b_flags & ARC_PREFETCH) {
2845 /*
2846 * This is a prefetch access...
2847 * move this block back to the MRU state.
2848 */
2849 ASSERT0(refcount_count(&buf->b_refcnt));
2850 new_state = arc_mru;
2851 }
2852
2853 buf->b_arc_access = ddi_get_lbolt();
2854 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2855 arc_change_state(new_state, buf, hash_lock);
2856
2857 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2858 } else if (buf->b_state == arc_l2c_only) {
2859 /*
2860 * This buffer is on the 2nd Level ARC.
2861 */
2862
2863 buf->b_arc_access = ddi_get_lbolt();
2864 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2865 arc_change_state(arc_mfu, buf, hash_lock);
2866 } else {
2867 ASSERT(!"invalid arc state");
2868 }
2869}
2870
2871/* a generic arc_done_func_t which you can use */
2872/* ARGSUSED */
2873void
2874arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2875{
2876 if (zio == NULL || zio->io_error == 0)
2877 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2878 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2879}
2880
2881/* a generic arc_done_func_t */
2882void
2883arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2884{
2885 arc_buf_t **bufp = arg;
2886 if (zio && zio->io_error) {
2887 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2888 *bufp = NULL;
2889 } else {
2890 *bufp = buf;
2891 ASSERT(buf->b_data);
2892 }
2893}
2894
2895static void
2896arc_read_done(zio_t *zio)
2897{
2898 arc_buf_hdr_t *hdr, *found;
2899 arc_buf_t *buf;
2900 arc_buf_t *abuf; /* buffer we're assigning to callback */
2901 kmutex_t *hash_lock;
2902 arc_callback_t *callback_list, *acb;
2903 int freeable = FALSE;
2904
2905 buf = zio->io_private;
2906 hdr = buf->b_hdr;
2907
2908 /*
2909 * The hdr was inserted into hash-table and removed from lists
2910 * prior to starting I/O. We should find this header, since
2911 * it's in the hash table, and it should be legit since it's
2912 * not possible to evict it during the I/O. The only possible
2913 * reason for it not to be found is if we were freed during the
2914 * read.
2915 */
2916 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2917 &hash_lock);
2918
2919 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2920 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2921 (found == hdr && HDR_L2_READING(hdr)));
2922
2923 hdr->b_flags &= ~ARC_L2_EVICTED;
2924 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2925 hdr->b_flags &= ~ARC_L2CACHE;
2926
2927 /* byteswap if necessary */
2928 callback_list = hdr->b_acb;
2929 ASSERT(callback_list != NULL);
2930 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2931 dmu_object_byteswap_t bswap =
2932 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2933 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2934 byteswap_uint64_array :
2935 dmu_ot_byteswap[bswap].ob_func;
2936 func(buf->b_data, hdr->b_size);
2937 }
2938
2939 arc_cksum_compute(buf, B_FALSE);
2940#ifdef illumos
2941 arc_buf_watch(buf);
2942#endif /* illumos */
2943
2944 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2945 /*
2946 * Only call arc_access on anonymous buffers. This is because
2947 * if we've issued an I/O for an evicted buffer, we've already
2948 * called arc_access (to prevent any simultaneous readers from
2949 * getting confused).
2950 */
2951 arc_access(hdr, hash_lock);
2952 }
2953
2954 /* create copies of the data buffer for the callers */
2955 abuf = buf;
2956 for (acb = callback_list; acb; acb = acb->acb_next) {
2957 if (acb->acb_done) {
2958 if (abuf == NULL) {
2959 ARCSTAT_BUMP(arcstat_duplicate_reads);
2960 abuf = arc_buf_clone(buf);
2961 }
2962 acb->acb_buf = abuf;
2963 abuf = NULL;
2964 }
2965 }
2966 hdr->b_acb = NULL;
2967 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2968 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2969 if (abuf == buf) {
2970 ASSERT(buf->b_efunc == NULL);
2971 ASSERT(hdr->b_datacnt == 1);
2972 hdr->b_flags |= ARC_BUF_AVAILABLE;
2973 }
2974
2975 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2976
2977 if (zio->io_error != 0) {
2978 hdr->b_flags |= ARC_IO_ERROR;
2979 if (hdr->b_state != arc_anon)
2980 arc_change_state(arc_anon, hdr, hash_lock);
2981 if (HDR_IN_HASH_TABLE(hdr))
2982 buf_hash_remove(hdr);
2983 freeable = refcount_is_zero(&hdr->b_refcnt);
2984 }
2985
2986 /*
2987 * Broadcast before we drop the hash_lock to avoid the possibility
2988 * that the hdr (and hence the cv) might be freed before we get to
2989 * the cv_broadcast().
2990 */
2991 cv_broadcast(&hdr->b_cv);
2992
2993 if (hash_lock) {
2994 mutex_exit(hash_lock);
2995 } else {
2996 /*
2997 * This block was freed while we waited for the read to
2998 * complete. It has been removed from the hash table and
2999 * moved to the anonymous state (so that it won't show up
3000 * in the cache).
3001 */
3002 ASSERT3P(hdr->b_state, ==, arc_anon);
3003 freeable = refcount_is_zero(&hdr->b_refcnt);
3004 }
3005
3006 /* execute each callback and free its structure */
3007 while ((acb = callback_list) != NULL) {
3008 if (acb->acb_done)
3009 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3010
3011 if (acb->acb_zio_dummy != NULL) {
3012 acb->acb_zio_dummy->io_error = zio->io_error;
3013 zio_nowait(acb->acb_zio_dummy);
3014 }
3015
3016 callback_list = acb->acb_next;
3017 kmem_free(acb, sizeof (arc_callback_t));
3018 }
3019
3020 if (freeable)
3021 arc_hdr_destroy(hdr);
3022}
3023
3024/*
3025 * "Read" the block block at the specified DVA (in bp) via the
3026 * cache. If the block is found in the cache, invoke the provided
3027 * callback immediately and return. Note that the `zio' parameter
3028 * in the callback will be NULL in this case, since no IO was
3029 * required. If the block is not in the cache pass the read request
3030 * on to the spa with a substitute callback function, so that the
3031 * requested block will be added to the cache.
3032 *
3033 * If a read request arrives for a block that has a read in-progress,
3034 * either wait for the in-progress read to complete (and return the
3035 * results); or, if this is a read with a "done" func, add a record
3036 * to the read to invoke the "done" func when the read completes,
3037 * and return; or just return.
3038 *
3039 * arc_read_done() will invoke all the requested "done" functions
3040 * for readers of this block.
3041 */
3042int
3043arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
3044 void *private, int priority, int zio_flags, uint32_t *arc_flags,
3045 const zbookmark_t *zb)
3046{
3047 arc_buf_hdr_t *hdr;
|
3445
3446 /*
3447 * Do we have more than one buf?
3448 */
3449 if (hdr->b_datacnt > 1) {
3450 arc_buf_hdr_t *nhdr;
3451 arc_buf_t **bufp;
3452 uint64_t blksz = hdr->b_size;
3453 uint64_t spa = hdr->b_spa;
3454 arc_buf_contents_t type = hdr->b_type;
3455 uint32_t flags = hdr->b_flags;
3456
3457 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3458 /*
3459 * Pull the data off of this hdr and attach it to
3460 * a new anonymous hdr.
3461 */
3462 (void) remove_reference(hdr, hash_lock, tag);
3463 bufp = &hdr->b_buf;
3464 while (*bufp != buf)
3465 bufp = &(*bufp)->b_next;
3466 *bufp = buf->b_next;
3467 buf->b_next = NULL;
3468
3469 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3470 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3471 if (refcount_is_zero(&hdr->b_refcnt)) {
3472 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3473 ASSERT3U(*size, >=, hdr->b_size);
3474 atomic_add_64(size, -hdr->b_size);
3475 }
3476
3477 /*
3478 * We're releasing a duplicate user data buffer, update
3479 * our statistics accordingly.
3480 */
3481 if (hdr->b_type == ARC_BUFC_DATA) {
3482 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3483 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3484 -hdr->b_size);
3485 }
3486 hdr->b_datacnt -= 1;
3487 arc_cksum_verify(buf);
3488#ifdef illumos
3489 arc_buf_unwatch(buf);
3490#endif /* illumos */
3491
3492 mutex_exit(hash_lock);
3493
3494 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3495 nhdr->b_size = blksz;
3496 nhdr->b_spa = spa;
3497 nhdr->b_type = type;
3498 nhdr->b_buf = buf;
3499 nhdr->b_state = arc_anon;
3500 nhdr->b_arc_access = 0;
3501 nhdr->b_flags = flags & ARC_L2_WRITING;
3502 nhdr->b_l2hdr = NULL;
3503 nhdr->b_datacnt = 1;
3504 nhdr->b_freeze_cksum = NULL;
3505 (void) refcount_add(&nhdr->b_refcnt, tag);
3506 buf->b_hdr = nhdr;
3507 mutex_exit(&buf->b_evict_lock);
3508 atomic_add_64(&arc_anon->arcs_size, blksz);
3509 } else {
3510 mutex_exit(&buf->b_evict_lock);
3511 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3512 ASSERT(!list_link_active(&hdr->b_arc_node));
3513 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3514 if (hdr->b_state != arc_anon)
3515 arc_change_state(arc_anon, hdr, hash_lock);
3516 hdr->b_arc_access = 0;
3517 if (hash_lock)
3518 mutex_exit(hash_lock);
3519
3520 buf_discard_identity(hdr);
3521 arc_buf_thaw(buf);
3522 }
3523 buf->b_efunc = NULL;
3524 buf->b_private = NULL;
3525
3526 if (l2hdr) {
3527 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3528 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3529 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3530 mutex_exit(&l2arc_buflist_mtx);
3531 }
3532}
3533
3534int
3535arc_released(arc_buf_t *buf)
3536{
3537 int released;
3538
3539 mutex_enter(&buf->b_evict_lock);
3540 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3541 mutex_exit(&buf->b_evict_lock);
3542 return (released);
3543}
3544
3545int
3546arc_has_callback(arc_buf_t *buf)
3547{
3548 int callback;
3549
3550 mutex_enter(&buf->b_evict_lock);
3551 callback = (buf->b_efunc != NULL);
3552 mutex_exit(&buf->b_evict_lock);
3553 return (callback);
3554}
3555
3556#ifdef ZFS_DEBUG
3557int
3558arc_referenced(arc_buf_t *buf)
3559{
3560 int referenced;
3561
3562 mutex_enter(&buf->b_evict_lock);
3563 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3564 mutex_exit(&buf->b_evict_lock);
3565 return (referenced);
3566}
3567#endif
3568
3569static void
3570arc_write_ready(zio_t *zio)
3571{
3572 arc_write_callback_t *callback = zio->io_private;
3573 arc_buf_t *buf = callback->awcb_buf;
3574 arc_buf_hdr_t *hdr = buf->b_hdr;
3575
3576 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3577 callback->awcb_ready(zio, buf, callback->awcb_private);
3578
3579 /*
3580 * If the IO is already in progress, then this is a re-write
3581 * attempt, so we need to thaw and re-compute the cksum.
3582 * It is the responsibility of the callback to handle the
3583 * accounting for any re-write attempt.
3584 */
3585 if (HDR_IO_IN_PROGRESS(hdr)) {
3586 mutex_enter(&hdr->b_freeze_lock);
3587 if (hdr->b_freeze_cksum != NULL) {
3588 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3589 hdr->b_freeze_cksum = NULL;
3590 }
3591 mutex_exit(&hdr->b_freeze_lock);
3592 }
3593 arc_cksum_compute(buf, B_FALSE);
3594 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3595}
3596
3597static void
3598arc_write_done(zio_t *zio)
3599{
3600 arc_write_callback_t *callback = zio->io_private;
3601 arc_buf_t *buf = callback->awcb_buf;
3602 arc_buf_hdr_t *hdr = buf->b_hdr;
3603
3604 ASSERT(hdr->b_acb == NULL);
3605
3606 if (zio->io_error == 0) {
3607 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3608 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3609 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3610 } else {
3611 ASSERT(BUF_EMPTY(hdr));
3612 }
3613
3614 /*
3615 * If the block to be written was all-zero, we may have
3616 * compressed it away. In this case no write was performed
3617 * so there will be no dva/birth/checksum. The buffer must
3618 * therefore remain anonymous (and uncached).
3619 */
3620 if (!BUF_EMPTY(hdr)) {
3621 arc_buf_hdr_t *exists;
3622 kmutex_t *hash_lock;
3623
3624 ASSERT(zio->io_error == 0);
3625
3626 arc_cksum_verify(buf);
3627
3628 exists = buf_hash_insert(hdr, &hash_lock);
3629 if (exists) {
3630 /*
3631 * This can only happen if we overwrite for
3632 * sync-to-convergence, because we remove
3633 * buffers from the hash table when we arc_free().
3634 */
3635 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3636 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3637 panic("bad overwrite, hdr=%p exists=%p",
3638 (void *)hdr, (void *)exists);
3639 ASSERT(refcount_is_zero(&exists->b_refcnt));
3640 arc_change_state(arc_anon, exists, hash_lock);
3641 mutex_exit(hash_lock);
3642 arc_hdr_destroy(exists);
3643 exists = buf_hash_insert(hdr, &hash_lock);
3644 ASSERT3P(exists, ==, NULL);
3645 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3646 /* nopwrite */
3647 ASSERT(zio->io_prop.zp_nopwrite);
3648 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3649 panic("bad nopwrite, hdr=%p exists=%p",
3650 (void *)hdr, (void *)exists);
3651 } else {
3652 /* Dedup */
3653 ASSERT(hdr->b_datacnt == 1);
3654 ASSERT(hdr->b_state == arc_anon);
3655 ASSERT(BP_GET_DEDUP(zio->io_bp));
3656 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3657 }
3658 }
3659 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3660 /* if it's not anon, we are doing a scrub */
3661 if (!exists && hdr->b_state == arc_anon)
3662 arc_access(hdr, hash_lock);
3663 mutex_exit(hash_lock);
3664 } else {
3665 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3666 }
3667
3668 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3669 callback->awcb_done(zio, buf, callback->awcb_private);
3670
3671 kmem_free(callback, sizeof (arc_write_callback_t));
3672}
3673
3674zio_t *
3675arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3676 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
3677 arc_done_func_t *ready, arc_done_func_t *done, void *private,
3678 int priority, int zio_flags, const zbookmark_t *zb)
3679{
3680 arc_buf_hdr_t *hdr = buf->b_hdr;
3681 arc_write_callback_t *callback;
3682 zio_t *zio;
3683
3684 ASSERT(ready != NULL);
3685 ASSERT(done != NULL);
3686 ASSERT(!HDR_IO_ERROR(hdr));
3687 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3688 ASSERT(hdr->b_acb == NULL);
3689 if (l2arc)
3690 hdr->b_flags |= ARC_L2CACHE;
3691 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3692 callback->awcb_ready = ready;
3693 callback->awcb_done = done;
3694 callback->awcb_private = private;
3695 callback->awcb_buf = buf;
3696
3697 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3698 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3699
3700 return (zio);
3701}
3702
3703static int
3704arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3705{
3706#ifdef _KERNEL
3707 uint64_t available_memory =
3708 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count);
3709 static uint64_t page_load = 0;
3710 static uint64_t last_txg = 0;
3711
3712#ifdef sun
3713#if defined(__i386)
3714 available_memory =
3715 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3716#endif
3717#endif /* sun */
3718 if (available_memory >= zfs_write_limit_max)
3719 return (0);
3720
3721 if (txg > last_txg) {
3722 last_txg = txg;
3723 page_load = 0;
3724 }
3725 /*
3726 * If we are in pageout, we know that memory is already tight,
3727 * the arc is already going to be evicting, so we just want to
3728 * continue to let page writes occur as quickly as possible.
3729 */
3730 if (curproc == pageproc) {
3731 if (page_load > available_memory / 4)
3732 return (ERESTART);
3733 /* Note: reserve is inflated, so we deflate */
3734 page_load += reserve / 8;
3735 return (0);
3736 } else if (page_load > 0 && arc_reclaim_needed()) {
3737 /* memory is low, delay before restarting */
3738 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3739 return (EAGAIN);
3740 }
3741 page_load = 0;
3742
3743 if (arc_size > arc_c_min) {
3744 uint64_t evictable_memory =
3745 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3746 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3747 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3748 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3749 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3750 }
3751
3752 if (inflight_data > available_memory / 4) {
3753 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3754 return (ERESTART);
3755 }
3756#endif
3757 return (0);
3758}
3759
3760void
3761arc_tempreserve_clear(uint64_t reserve)
3762{
3763 atomic_add_64(&arc_tempreserve, -reserve);
3764 ASSERT((int64_t)arc_tempreserve >= 0);
3765}
3766
3767int
3768arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3769{
3770 int error;
3771 uint64_t anon_size;
3772
3773#ifdef ZFS_DEBUG
3774 /*
3775 * Once in a while, fail for no reason. Everything should cope.
3776 */
3777 if (spa_get_random(10000) == 0) {
3778 dprintf("forcing random failure\n");
3779 return (ERESTART);
3780 }
3781#endif
3782 if (reserve > arc_c/4 && !arc_no_grow)
3783 arc_c = MIN(arc_c_max, reserve * 4);
3784 if (reserve > arc_c)
3785 return (ENOMEM);
3786
3787 /*
3788 * Don't count loaned bufs as in flight dirty data to prevent long
3789 * network delays from blocking transactions that are ready to be
3790 * assigned to a txg.
3791 */
3792 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3793
3794 /*
3795 * Writes will, almost always, require additional memory allocations
3796 * in order to compress/encrypt/etc the data. We therefor need to
3797 * make sure that there is sufficient available memory for this.
3798 */
3799 if (error = arc_memory_throttle(reserve, anon_size, txg))
3800 return (error);
3801
3802 /*
3803 * Throttle writes when the amount of dirty data in the cache
3804 * gets too large. We try to keep the cache less than half full
3805 * of dirty blocks so that our sync times don't grow too large.
3806 * Note: if two requests come in concurrently, we might let them
3807 * both succeed, when one of them should fail. Not a huge deal.
3808 */
3809
3810 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3811 anon_size > arc_c / 4) {
3812 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3813 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3814 arc_tempreserve>>10,
3815 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3816 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3817 reserve>>10, arc_c>>10);
3818 return (ERESTART);
3819 }
3820 atomic_add_64(&arc_tempreserve, reserve);
3821 return (0);
3822}
3823
3824static kmutex_t arc_lowmem_lock;
3825#ifdef _KERNEL
3826static eventhandler_tag arc_event_lowmem = NULL;
3827
3828static void
3829arc_lowmem(void *arg __unused, int howto __unused)
3830{
3831
3832 /* Serialize access via arc_lowmem_lock. */
3833 mutex_enter(&arc_lowmem_lock);
3834 mutex_enter(&arc_reclaim_thr_lock);
3835 needfree = 1;
3836 cv_signal(&arc_reclaim_thr_cv);
3837
3838 /*
3839 * It is unsafe to block here in arbitrary threads, because we can come
3840 * here from ARC itself and may hold ARC locks and thus risk a deadlock
3841 * with ARC reclaim thread.
3842 */
3843 if (curproc == pageproc) {
3844 while (needfree)
3845 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0);
3846 }
3847 mutex_exit(&arc_reclaim_thr_lock);
3848 mutex_exit(&arc_lowmem_lock);
3849}
3850#endif
3851
3852void
3853arc_init(void)
3854{
3855 int i, prefetch_tunable_set = 0;
3856
3857 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3858 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3859 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL);
3860
3861 /* Convert seconds to clock ticks */
3862 arc_min_prefetch_lifespan = 1 * hz;
3863
3864 /* Start out with 1/8 of all memory */
3865 arc_c = kmem_size() / 8;
3866
3867#ifdef sun
3868#ifdef _KERNEL
3869 /*
3870 * On architectures where the physical memory can be larger
3871 * than the addressable space (intel in 32-bit mode), we may
3872 * need to limit the cache to 1/8 of VM size.
3873 */
3874 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3875#endif
3876#endif /* sun */
3877 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */
3878 arc_c_min = MAX(arc_c / 4, 64<<18);
3879 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
3880 if (arc_c * 8 >= 1<<30)
3881 arc_c_max = (arc_c * 8) - (1<<30);
3882 else
3883 arc_c_max = arc_c_min;
3884 arc_c_max = MAX(arc_c * 5, arc_c_max);
3885
3886#ifdef _KERNEL
3887 /*
3888 * Allow the tunables to override our calculations if they are
3889 * reasonable (ie. over 16MB)
3890 */
3891 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size())
3892 arc_c_max = zfs_arc_max;
3893 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max)
3894 arc_c_min = zfs_arc_min;
3895#endif
3896
3897 arc_c = arc_c_max;
3898 arc_p = (arc_c >> 1);
3899
3900 /* limit meta-data to 1/4 of the arc capacity */
3901 arc_meta_limit = arc_c_max / 4;
3902
3903 /* Allow the tunable to override if it is reasonable */
3904 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3905 arc_meta_limit = zfs_arc_meta_limit;
3906
3907 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3908 arc_c_min = arc_meta_limit / 2;
3909
3910 if (zfs_arc_grow_retry > 0)
3911 arc_grow_retry = zfs_arc_grow_retry;
3912
3913 if (zfs_arc_shrink_shift > 0)
3914 arc_shrink_shift = zfs_arc_shrink_shift;
3915
3916 if (zfs_arc_p_min_shift > 0)
3917 arc_p_min_shift = zfs_arc_p_min_shift;
3918
3919 /* if kmem_flags are set, lets try to use less memory */
3920 if (kmem_debugging())
3921 arc_c = arc_c / 2;
3922 if (arc_c < arc_c_min)
3923 arc_c = arc_c_min;
3924
3925 zfs_arc_min = arc_c_min;
3926 zfs_arc_max = arc_c_max;
3927
3928 arc_anon = &ARC_anon;
3929 arc_mru = &ARC_mru;
3930 arc_mru_ghost = &ARC_mru_ghost;
3931 arc_mfu = &ARC_mfu;
3932 arc_mfu_ghost = &ARC_mfu_ghost;
3933 arc_l2c_only = &ARC_l2c_only;
3934 arc_size = 0;
3935
3936 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
3937 mutex_init(&arc_anon->arcs_locks[i].arcs_lock,
3938 NULL, MUTEX_DEFAULT, NULL);
3939 mutex_init(&arc_mru->arcs_locks[i].arcs_lock,
3940 NULL, MUTEX_DEFAULT, NULL);
3941 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock,
3942 NULL, MUTEX_DEFAULT, NULL);
3943 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock,
3944 NULL, MUTEX_DEFAULT, NULL);
3945 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock,
3946 NULL, MUTEX_DEFAULT, NULL);
3947 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock,
3948 NULL, MUTEX_DEFAULT, NULL);
3949
3950 list_create(&arc_mru->arcs_lists[i],
3951 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3952 list_create(&arc_mru_ghost->arcs_lists[i],
3953 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3954 list_create(&arc_mfu->arcs_lists[i],
3955 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3956 list_create(&arc_mfu_ghost->arcs_lists[i],
3957 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3958 list_create(&arc_mfu_ghost->arcs_lists[i],
3959 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3960 list_create(&arc_l2c_only->arcs_lists[i],
3961 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3962 }
3963
3964 buf_init();
3965
3966 arc_thread_exit = 0;
3967 arc_eviction_list = NULL;
3968 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3969 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3970
3971 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3972 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3973
3974 if (arc_ksp != NULL) {
3975 arc_ksp->ks_data = &arc_stats;
3976 kstat_install(arc_ksp);
3977 }
3978
3979 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3980 TS_RUN, minclsyspri);
3981
3982#ifdef _KERNEL
3983 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
3984 EVENTHANDLER_PRI_FIRST);
3985#endif
3986
3987 arc_dead = FALSE;
3988 arc_warm = B_FALSE;
3989
3990 if (zfs_write_limit_max == 0)
3991 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3992 else
3993 zfs_write_limit_shift = 0;
3994 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3995
3996#ifdef _KERNEL
3997 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
3998 prefetch_tunable_set = 1;
3999
4000#ifdef __i386__
4001 if (prefetch_tunable_set == 0) {
4002 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
4003 "-- to enable,\n");
4004 printf(" add \"vfs.zfs.prefetch_disable=0\" "
4005 "to /boot/loader.conf.\n");
4006 zfs_prefetch_disable = 1;
4007 }
4008#else
4009 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
4010 prefetch_tunable_set == 0) {
4011 printf("ZFS NOTICE: Prefetch is disabled by default if less "
4012 "than 4GB of RAM is present;\n"
4013 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
4014 "to /boot/loader.conf.\n");
4015 zfs_prefetch_disable = 1;
4016 }
4017#endif
4018 /* Warn about ZFS memory and address space requirements. */
4019 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
4020 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
4021 "expect unstable behavior.\n");
4022 }
4023 if (kmem_size() < 512 * (1 << 20)) {
4024 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
4025 "expect unstable behavior.\n");
4026 printf(" Consider tuning vm.kmem_size and "
4027 "vm.kmem_size_max\n");
4028 printf(" in /boot/loader.conf.\n");
4029 }
4030#endif
4031}
4032
4033void
4034arc_fini(void)
4035{
4036 int i;
4037
4038 mutex_enter(&arc_reclaim_thr_lock);
4039 arc_thread_exit = 1;
4040 cv_signal(&arc_reclaim_thr_cv);
4041 while (arc_thread_exit != 0)
4042 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
4043 mutex_exit(&arc_reclaim_thr_lock);
4044
4045 arc_flush(NULL);
4046
4047 arc_dead = TRUE;
4048
4049 if (arc_ksp != NULL) {
4050 kstat_delete(arc_ksp);
4051 arc_ksp = NULL;
4052 }
4053
4054 mutex_destroy(&arc_eviction_mtx);
4055 mutex_destroy(&arc_reclaim_thr_lock);
4056 cv_destroy(&arc_reclaim_thr_cv);
4057
4058 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
4059 list_destroy(&arc_mru->arcs_lists[i]);
4060 list_destroy(&arc_mru_ghost->arcs_lists[i]);
4061 list_destroy(&arc_mfu->arcs_lists[i]);
4062 list_destroy(&arc_mfu_ghost->arcs_lists[i]);
4063 list_destroy(&arc_l2c_only->arcs_lists[i]);
4064
4065 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock);
4066 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock);
4067 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock);
4068 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock);
4069 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock);
4070 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock);
4071 }
4072
4073 mutex_destroy(&zfs_write_limit_lock);
4074
4075 buf_fini();
4076
4077 ASSERT(arc_loaned_bytes == 0);
4078
4079 mutex_destroy(&arc_lowmem_lock);
4080#ifdef _KERNEL
4081 if (arc_event_lowmem != NULL)
4082 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
4083#endif
4084}
4085
4086/*
4087 * Level 2 ARC
4088 *
4089 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
4090 * It uses dedicated storage devices to hold cached data, which are populated
4091 * using large infrequent writes. The main role of this cache is to boost
4092 * the performance of random read workloads. The intended L2ARC devices
4093 * include short-stroked disks, solid state disks, and other media with
4094 * substantially faster read latency than disk.
4095 *
4096 * +-----------------------+
4097 * | ARC |
4098 * +-----------------------+
4099 * | ^ ^
4100 * | | |
4101 * l2arc_feed_thread() arc_read()
4102 * | | |
4103 * | l2arc read |
4104 * V | |
4105 * +---------------+ |
4106 * | L2ARC | |
4107 * +---------------+ |
4108 * | ^ |
4109 * l2arc_write() | |
4110 * | | |
4111 * V | |
4112 * +-------+ +-------+
4113 * | vdev | | vdev |
4114 * | cache | | cache |
4115 * +-------+ +-------+
4116 * +=========+ .-----.
4117 * : L2ARC : |-_____-|
4118 * : devices : | Disks |
4119 * +=========+ `-_____-'
4120 *
4121 * Read requests are satisfied from the following sources, in order:
4122 *
4123 * 1) ARC
4124 * 2) vdev cache of L2ARC devices
4125 * 3) L2ARC devices
4126 * 4) vdev cache of disks
4127 * 5) disks
4128 *
4129 * Some L2ARC device types exhibit extremely slow write performance.
4130 * To accommodate for this there are some significant differences between
4131 * the L2ARC and traditional cache design:
4132 *
4133 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4134 * the ARC behave as usual, freeing buffers and placing headers on ghost
4135 * lists. The ARC does not send buffers to the L2ARC during eviction as
4136 * this would add inflated write latencies for all ARC memory pressure.
4137 *
4138 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4139 * It does this by periodically scanning buffers from the eviction-end of
4140 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4141 * not already there. It scans until a headroom of buffers is satisfied,
4142 * which itself is a buffer for ARC eviction. The thread that does this is
4143 * l2arc_feed_thread(), illustrated below; example sizes are included to
4144 * provide a better sense of ratio than this diagram:
4145 *
4146 * head --> tail
4147 * +---------------------+----------+
4148 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4149 * +---------------------+----------+ | o L2ARC eligible
4150 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4151 * +---------------------+----------+ |
4152 * 15.9 Gbytes ^ 32 Mbytes |
4153 * headroom |
4154 * l2arc_feed_thread()
4155 * |
4156 * l2arc write hand <--[oooo]--'
4157 * | 8 Mbyte
4158 * | write max
4159 * V
4160 * +==============================+
4161 * L2ARC dev |####|#|###|###| |####| ... |
4162 * +==============================+
4163 * 32 Gbytes
4164 *
4165 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4166 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4167 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4168 * safe to say that this is an uncommon case, since buffers at the end of
4169 * the ARC lists have moved there due to inactivity.
4170 *
4171 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4172 * then the L2ARC simply misses copying some buffers. This serves as a
4173 * pressure valve to prevent heavy read workloads from both stalling the ARC
4174 * with waits and clogging the L2ARC with writes. This also helps prevent
4175 * the potential for the L2ARC to churn if it attempts to cache content too
4176 * quickly, such as during backups of the entire pool.
4177 *
4178 * 5. After system boot and before the ARC has filled main memory, there are
4179 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4180 * lists can remain mostly static. Instead of searching from tail of these
4181 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4182 * for eligible buffers, greatly increasing its chance of finding them.
4183 *
4184 * The L2ARC device write speed is also boosted during this time so that
4185 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4186 * there are no L2ARC reads, and no fear of degrading read performance
4187 * through increased writes.
4188 *
4189 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4190 * the vdev queue can aggregate them into larger and fewer writes. Each
4191 * device is written to in a rotor fashion, sweeping writes through
4192 * available space then repeating.
4193 *
4194 * 7. The L2ARC does not store dirty content. It never needs to flush
4195 * write buffers back to disk based storage.
4196 *
4197 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4198 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4199 *
4200 * The performance of the L2ARC can be tweaked by a number of tunables, which
4201 * may be necessary for different workloads:
4202 *
4203 * l2arc_write_max max write bytes per interval
4204 * l2arc_write_boost extra write bytes during device warmup
4205 * l2arc_noprefetch skip caching prefetched buffers
4206 * l2arc_headroom number of max device writes to precache
4207 * l2arc_feed_secs seconds between L2ARC writing
4208 *
4209 * Tunables may be removed or added as future performance improvements are
4210 * integrated, and also may become zpool properties.
4211 *
4212 * There are three key functions that control how the L2ARC warms up:
4213 *
4214 * l2arc_write_eligible() check if a buffer is eligible to cache
4215 * l2arc_write_size() calculate how much to write
4216 * l2arc_write_interval() calculate sleep delay between writes
4217 *
4218 * These three functions determine what to write, how much, and how quickly
4219 * to send writes.
4220 */
4221
4222static boolean_t
4223l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4224{
4225 /*
4226 * A buffer is *not* eligible for the L2ARC if it:
4227 * 1. belongs to a different spa.
4228 * 2. is already cached on the L2ARC.
4229 * 3. has an I/O in progress (it may be an incomplete read).
4230 * 4. is flagged not eligible (zfs property).
4231 */
4232 if (ab->b_spa != spa_guid) {
4233 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
4234 return (B_FALSE);
4235 }
4236 if (ab->b_l2hdr != NULL) {
4237 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
4238 return (B_FALSE);
4239 }
4240 if (HDR_IO_IN_PROGRESS(ab)) {
4241 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
4242 return (B_FALSE);
4243 }
4244 if (!HDR_L2CACHE(ab)) {
4245 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
4246 return (B_FALSE);
4247 }
4248
4249 return (B_TRUE);
4250}
4251
4252static uint64_t
4253l2arc_write_size(l2arc_dev_t *dev)
4254{
4255 uint64_t size;
4256
4257 size = dev->l2ad_write;
4258
4259 if (arc_warm == B_FALSE)
4260 size += dev->l2ad_boost;
4261
4262 return (size);
4263
4264}
4265
4266static clock_t
4267l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4268{
4269 clock_t interval, next, now;
4270
4271 /*
4272 * If the ARC lists are busy, increase our write rate; if the
4273 * lists are stale, idle back. This is achieved by checking
4274 * how much we previously wrote - if it was more than half of
4275 * what we wanted, schedule the next write much sooner.
4276 */
4277 if (l2arc_feed_again && wrote > (wanted / 2))
4278 interval = (hz * l2arc_feed_min_ms) / 1000;
4279 else
4280 interval = hz * l2arc_feed_secs;
4281
4282 now = ddi_get_lbolt();
4283 next = MAX(now, MIN(now + interval, began + interval));
4284
4285 return (next);
4286}
4287
4288static void
4289l2arc_hdr_stat_add(void)
4290{
4291 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4292 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4293}
4294
4295static void
4296l2arc_hdr_stat_remove(void)
4297{
4298 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4299 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4300}
4301
4302/*
4303 * Cycle through L2ARC devices. This is how L2ARC load balances.
4304 * If a device is returned, this also returns holding the spa config lock.
4305 */
4306static l2arc_dev_t *
4307l2arc_dev_get_next(void)
4308{
4309 l2arc_dev_t *first, *next = NULL;
4310
4311 /*
4312 * Lock out the removal of spas (spa_namespace_lock), then removal
4313 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4314 * both locks will be dropped and a spa config lock held instead.
4315 */
4316 mutex_enter(&spa_namespace_lock);
4317 mutex_enter(&l2arc_dev_mtx);
4318
4319 /* if there are no vdevs, there is nothing to do */
4320 if (l2arc_ndev == 0)
4321 goto out;
4322
4323 first = NULL;
4324 next = l2arc_dev_last;
4325 do {
4326 /* loop around the list looking for a non-faulted vdev */
4327 if (next == NULL) {
4328 next = list_head(l2arc_dev_list);
4329 } else {
4330 next = list_next(l2arc_dev_list, next);
4331 if (next == NULL)
4332 next = list_head(l2arc_dev_list);
4333 }
4334
4335 /* if we have come back to the start, bail out */
4336 if (first == NULL)
4337 first = next;
4338 else if (next == first)
4339 break;
4340
4341 } while (vdev_is_dead(next->l2ad_vdev));
4342
4343 /* if we were unable to find any usable vdevs, return NULL */
4344 if (vdev_is_dead(next->l2ad_vdev))
4345 next = NULL;
4346
4347 l2arc_dev_last = next;
4348
4349out:
4350 mutex_exit(&l2arc_dev_mtx);
4351
4352 /*
4353 * Grab the config lock to prevent the 'next' device from being
4354 * removed while we are writing to it.
4355 */
4356 if (next != NULL)
4357 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4358 mutex_exit(&spa_namespace_lock);
4359
4360 return (next);
4361}
4362
4363/*
4364 * Free buffers that were tagged for destruction.
4365 */
4366static void
4367l2arc_do_free_on_write()
4368{
4369 list_t *buflist;
4370 l2arc_data_free_t *df, *df_prev;
4371
4372 mutex_enter(&l2arc_free_on_write_mtx);
4373 buflist = l2arc_free_on_write;
4374
4375 for (df = list_tail(buflist); df; df = df_prev) {
4376 df_prev = list_prev(buflist, df);
4377 ASSERT(df->l2df_data != NULL);
4378 ASSERT(df->l2df_func != NULL);
4379 df->l2df_func(df->l2df_data, df->l2df_size);
4380 list_remove(buflist, df);
4381 kmem_free(df, sizeof (l2arc_data_free_t));
4382 }
4383
4384 mutex_exit(&l2arc_free_on_write_mtx);
4385}
4386
4387/*
4388 * A write to a cache device has completed. Update all headers to allow
4389 * reads from these buffers to begin.
4390 */
4391static void
4392l2arc_write_done(zio_t *zio)
4393{
4394 l2arc_write_callback_t *cb;
4395 l2arc_dev_t *dev;
4396 list_t *buflist;
4397 arc_buf_hdr_t *head, *ab, *ab_prev;
4398 l2arc_buf_hdr_t *abl2;
4399 kmutex_t *hash_lock;
4400
4401 cb = zio->io_private;
4402 ASSERT(cb != NULL);
4403 dev = cb->l2wcb_dev;
4404 ASSERT(dev != NULL);
4405 head = cb->l2wcb_head;
4406 ASSERT(head != NULL);
4407 buflist = dev->l2ad_buflist;
4408 ASSERT(buflist != NULL);
4409 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4410 l2arc_write_callback_t *, cb);
4411
4412 if (zio->io_error != 0)
4413 ARCSTAT_BUMP(arcstat_l2_writes_error);
4414
4415 mutex_enter(&l2arc_buflist_mtx);
4416
4417 /*
4418 * All writes completed, or an error was hit.
4419 */
4420 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4421 ab_prev = list_prev(buflist, ab);
4422
4423 hash_lock = HDR_LOCK(ab);
4424 if (!mutex_tryenter(hash_lock)) {
4425 /*
4426 * This buffer misses out. It may be in a stage
4427 * of eviction. Its ARC_L2_WRITING flag will be
4428 * left set, denying reads to this buffer.
4429 */
4430 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4431 continue;
4432 }
4433
4434 if (zio->io_error != 0) {
4435 /*
4436 * Error - drop L2ARC entry.
4437 */
4438 list_remove(buflist, ab);
4439 abl2 = ab->b_l2hdr;
4440 ab->b_l2hdr = NULL;
4441 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4442 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4443 }
4444
4445 /*
4446 * Allow ARC to begin reads to this L2ARC entry.
4447 */
4448 ab->b_flags &= ~ARC_L2_WRITING;
4449
4450 mutex_exit(hash_lock);
4451 }
4452
4453 atomic_inc_64(&l2arc_writes_done);
4454 list_remove(buflist, head);
4455 kmem_cache_free(hdr_cache, head);
4456 mutex_exit(&l2arc_buflist_mtx);
4457
4458 l2arc_do_free_on_write();
4459
4460 kmem_free(cb, sizeof (l2arc_write_callback_t));
4461}
4462
4463/*
4464 * A read to a cache device completed. Validate buffer contents before
4465 * handing over to the regular ARC routines.
4466 */
4467static void
4468l2arc_read_done(zio_t *zio)
4469{
4470 l2arc_read_callback_t *cb;
4471 arc_buf_hdr_t *hdr;
4472 arc_buf_t *buf;
4473 kmutex_t *hash_lock;
4474 int equal;
4475
4476 ASSERT(zio->io_vd != NULL);
4477 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4478
4479 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4480
4481 cb = zio->io_private;
4482 ASSERT(cb != NULL);
4483 buf = cb->l2rcb_buf;
4484 ASSERT(buf != NULL);
4485
4486 hash_lock = HDR_LOCK(buf->b_hdr);
4487 mutex_enter(hash_lock);
4488 hdr = buf->b_hdr;
4489 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4490
4491 /*
4492 * Check this survived the L2ARC journey.
4493 */
4494 equal = arc_cksum_equal(buf);
4495 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4496 mutex_exit(hash_lock);
4497 zio->io_private = buf;
4498 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4499 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4500 arc_read_done(zio);
4501 } else {
4502 mutex_exit(hash_lock);
4503 /*
4504 * Buffer didn't survive caching. Increment stats and
4505 * reissue to the original storage device.
4506 */
4507 if (zio->io_error != 0) {
4508 ARCSTAT_BUMP(arcstat_l2_io_error);
4509 } else {
4510 zio->io_error = EIO;
4511 }
4512 if (!equal)
4513 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4514
4515 /*
4516 * If there's no waiter, issue an async i/o to the primary
4517 * storage now. If there *is* a waiter, the caller must
4518 * issue the i/o in a context where it's OK to block.
4519 */
4520 if (zio->io_waiter == NULL) {
4521 zio_t *pio = zio_unique_parent(zio);
4522
4523 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4524
4525 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4526 buf->b_data, zio->io_size, arc_read_done, buf,
4527 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4528 }
4529 }
4530
4531 kmem_free(cb, sizeof (l2arc_read_callback_t));
4532}
4533
4534/*
4535 * This is the list priority from which the L2ARC will search for pages to
4536 * cache. This is used within loops (0..3) to cycle through lists in the
4537 * desired order. This order can have a significant effect on cache
4538 * performance.
4539 *
4540 * Currently the metadata lists are hit first, MFU then MRU, followed by
4541 * the data lists. This function returns a locked list, and also returns
4542 * the lock pointer.
4543 */
4544static list_t *
4545l2arc_list_locked(int list_num, kmutex_t **lock)
4546{
| 3449
3450 /*
3451 * Do we have more than one buf?
3452 */
3453 if (hdr->b_datacnt > 1) {
3454 arc_buf_hdr_t *nhdr;
3455 arc_buf_t **bufp;
3456 uint64_t blksz = hdr->b_size;
3457 uint64_t spa = hdr->b_spa;
3458 arc_buf_contents_t type = hdr->b_type;
3459 uint32_t flags = hdr->b_flags;
3460
3461 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3462 /*
3463 * Pull the data off of this hdr and attach it to
3464 * a new anonymous hdr.
3465 */
3466 (void) remove_reference(hdr, hash_lock, tag);
3467 bufp = &hdr->b_buf;
3468 while (*bufp != buf)
3469 bufp = &(*bufp)->b_next;
3470 *bufp = buf->b_next;
3471 buf->b_next = NULL;
3472
3473 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3474 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3475 if (refcount_is_zero(&hdr->b_refcnt)) {
3476 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3477 ASSERT3U(*size, >=, hdr->b_size);
3478 atomic_add_64(size, -hdr->b_size);
3479 }
3480
3481 /*
3482 * We're releasing a duplicate user data buffer, update
3483 * our statistics accordingly.
3484 */
3485 if (hdr->b_type == ARC_BUFC_DATA) {
3486 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3487 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3488 -hdr->b_size);
3489 }
3490 hdr->b_datacnt -= 1;
3491 arc_cksum_verify(buf);
3492#ifdef illumos
3493 arc_buf_unwatch(buf);
3494#endif /* illumos */
3495
3496 mutex_exit(hash_lock);
3497
3498 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3499 nhdr->b_size = blksz;
3500 nhdr->b_spa = spa;
3501 nhdr->b_type = type;
3502 nhdr->b_buf = buf;
3503 nhdr->b_state = arc_anon;
3504 nhdr->b_arc_access = 0;
3505 nhdr->b_flags = flags & ARC_L2_WRITING;
3506 nhdr->b_l2hdr = NULL;
3507 nhdr->b_datacnt = 1;
3508 nhdr->b_freeze_cksum = NULL;
3509 (void) refcount_add(&nhdr->b_refcnt, tag);
3510 buf->b_hdr = nhdr;
3511 mutex_exit(&buf->b_evict_lock);
3512 atomic_add_64(&arc_anon->arcs_size, blksz);
3513 } else {
3514 mutex_exit(&buf->b_evict_lock);
3515 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3516 ASSERT(!list_link_active(&hdr->b_arc_node));
3517 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3518 if (hdr->b_state != arc_anon)
3519 arc_change_state(arc_anon, hdr, hash_lock);
3520 hdr->b_arc_access = 0;
3521 if (hash_lock)
3522 mutex_exit(hash_lock);
3523
3524 buf_discard_identity(hdr);
3525 arc_buf_thaw(buf);
3526 }
3527 buf->b_efunc = NULL;
3528 buf->b_private = NULL;
3529
3530 if (l2hdr) {
3531 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3532 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3533 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3534 mutex_exit(&l2arc_buflist_mtx);
3535 }
3536}
3537
3538int
3539arc_released(arc_buf_t *buf)
3540{
3541 int released;
3542
3543 mutex_enter(&buf->b_evict_lock);
3544 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3545 mutex_exit(&buf->b_evict_lock);
3546 return (released);
3547}
3548
3549int
3550arc_has_callback(arc_buf_t *buf)
3551{
3552 int callback;
3553
3554 mutex_enter(&buf->b_evict_lock);
3555 callback = (buf->b_efunc != NULL);
3556 mutex_exit(&buf->b_evict_lock);
3557 return (callback);
3558}
3559
3560#ifdef ZFS_DEBUG
3561int
3562arc_referenced(arc_buf_t *buf)
3563{
3564 int referenced;
3565
3566 mutex_enter(&buf->b_evict_lock);
3567 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3568 mutex_exit(&buf->b_evict_lock);
3569 return (referenced);
3570}
3571#endif
3572
3573static void
3574arc_write_ready(zio_t *zio)
3575{
3576 arc_write_callback_t *callback = zio->io_private;
3577 arc_buf_t *buf = callback->awcb_buf;
3578 arc_buf_hdr_t *hdr = buf->b_hdr;
3579
3580 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3581 callback->awcb_ready(zio, buf, callback->awcb_private);
3582
3583 /*
3584 * If the IO is already in progress, then this is a re-write
3585 * attempt, so we need to thaw and re-compute the cksum.
3586 * It is the responsibility of the callback to handle the
3587 * accounting for any re-write attempt.
3588 */
3589 if (HDR_IO_IN_PROGRESS(hdr)) {
3590 mutex_enter(&hdr->b_freeze_lock);
3591 if (hdr->b_freeze_cksum != NULL) {
3592 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3593 hdr->b_freeze_cksum = NULL;
3594 }
3595 mutex_exit(&hdr->b_freeze_lock);
3596 }
3597 arc_cksum_compute(buf, B_FALSE);
3598 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3599}
3600
3601static void
3602arc_write_done(zio_t *zio)
3603{
3604 arc_write_callback_t *callback = zio->io_private;
3605 arc_buf_t *buf = callback->awcb_buf;
3606 arc_buf_hdr_t *hdr = buf->b_hdr;
3607
3608 ASSERT(hdr->b_acb == NULL);
3609
3610 if (zio->io_error == 0) {
3611 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3612 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3613 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3614 } else {
3615 ASSERT(BUF_EMPTY(hdr));
3616 }
3617
3618 /*
3619 * If the block to be written was all-zero, we may have
3620 * compressed it away. In this case no write was performed
3621 * so there will be no dva/birth/checksum. The buffer must
3622 * therefore remain anonymous (and uncached).
3623 */
3624 if (!BUF_EMPTY(hdr)) {
3625 arc_buf_hdr_t *exists;
3626 kmutex_t *hash_lock;
3627
3628 ASSERT(zio->io_error == 0);
3629
3630 arc_cksum_verify(buf);
3631
3632 exists = buf_hash_insert(hdr, &hash_lock);
3633 if (exists) {
3634 /*
3635 * This can only happen if we overwrite for
3636 * sync-to-convergence, because we remove
3637 * buffers from the hash table when we arc_free().
3638 */
3639 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3640 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3641 panic("bad overwrite, hdr=%p exists=%p",
3642 (void *)hdr, (void *)exists);
3643 ASSERT(refcount_is_zero(&exists->b_refcnt));
3644 arc_change_state(arc_anon, exists, hash_lock);
3645 mutex_exit(hash_lock);
3646 arc_hdr_destroy(exists);
3647 exists = buf_hash_insert(hdr, &hash_lock);
3648 ASSERT3P(exists, ==, NULL);
3649 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3650 /* nopwrite */
3651 ASSERT(zio->io_prop.zp_nopwrite);
3652 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3653 panic("bad nopwrite, hdr=%p exists=%p",
3654 (void *)hdr, (void *)exists);
3655 } else {
3656 /* Dedup */
3657 ASSERT(hdr->b_datacnt == 1);
3658 ASSERT(hdr->b_state == arc_anon);
3659 ASSERT(BP_GET_DEDUP(zio->io_bp));
3660 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3661 }
3662 }
3663 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3664 /* if it's not anon, we are doing a scrub */
3665 if (!exists && hdr->b_state == arc_anon)
3666 arc_access(hdr, hash_lock);
3667 mutex_exit(hash_lock);
3668 } else {
3669 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3670 }
3671
3672 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3673 callback->awcb_done(zio, buf, callback->awcb_private);
3674
3675 kmem_free(callback, sizeof (arc_write_callback_t));
3676}
3677
3678zio_t *
3679arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3680 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
3681 arc_done_func_t *ready, arc_done_func_t *done, void *private,
3682 int priority, int zio_flags, const zbookmark_t *zb)
3683{
3684 arc_buf_hdr_t *hdr = buf->b_hdr;
3685 arc_write_callback_t *callback;
3686 zio_t *zio;
3687
3688 ASSERT(ready != NULL);
3689 ASSERT(done != NULL);
3690 ASSERT(!HDR_IO_ERROR(hdr));
3691 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3692 ASSERT(hdr->b_acb == NULL);
3693 if (l2arc)
3694 hdr->b_flags |= ARC_L2CACHE;
3695 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3696 callback->awcb_ready = ready;
3697 callback->awcb_done = done;
3698 callback->awcb_private = private;
3699 callback->awcb_buf = buf;
3700
3701 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3702 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3703
3704 return (zio);
3705}
3706
3707static int
3708arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3709{
3710#ifdef _KERNEL
3711 uint64_t available_memory =
3712 ptoa((uintmax_t)cnt.v_free_count + cnt.v_cache_count);
3713 static uint64_t page_load = 0;
3714 static uint64_t last_txg = 0;
3715
3716#ifdef sun
3717#if defined(__i386)
3718 available_memory =
3719 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3720#endif
3721#endif /* sun */
3722 if (available_memory >= zfs_write_limit_max)
3723 return (0);
3724
3725 if (txg > last_txg) {
3726 last_txg = txg;
3727 page_load = 0;
3728 }
3729 /*
3730 * If we are in pageout, we know that memory is already tight,
3731 * the arc is already going to be evicting, so we just want to
3732 * continue to let page writes occur as quickly as possible.
3733 */
3734 if (curproc == pageproc) {
3735 if (page_load > available_memory / 4)
3736 return (ERESTART);
3737 /* Note: reserve is inflated, so we deflate */
3738 page_load += reserve / 8;
3739 return (0);
3740 } else if (page_load > 0 && arc_reclaim_needed()) {
3741 /* memory is low, delay before restarting */
3742 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3743 return (EAGAIN);
3744 }
3745 page_load = 0;
3746
3747 if (arc_size > arc_c_min) {
3748 uint64_t evictable_memory =
3749 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3750 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3751 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3752 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3753 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3754 }
3755
3756 if (inflight_data > available_memory / 4) {
3757 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3758 return (ERESTART);
3759 }
3760#endif
3761 return (0);
3762}
3763
3764void
3765arc_tempreserve_clear(uint64_t reserve)
3766{
3767 atomic_add_64(&arc_tempreserve, -reserve);
3768 ASSERT((int64_t)arc_tempreserve >= 0);
3769}
3770
3771int
3772arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3773{
3774 int error;
3775 uint64_t anon_size;
3776
3777#ifdef ZFS_DEBUG
3778 /*
3779 * Once in a while, fail for no reason. Everything should cope.
3780 */
3781 if (spa_get_random(10000) == 0) {
3782 dprintf("forcing random failure\n");
3783 return (ERESTART);
3784 }
3785#endif
3786 if (reserve > arc_c/4 && !arc_no_grow)
3787 arc_c = MIN(arc_c_max, reserve * 4);
3788 if (reserve > arc_c)
3789 return (ENOMEM);
3790
3791 /*
3792 * Don't count loaned bufs as in flight dirty data to prevent long
3793 * network delays from blocking transactions that are ready to be
3794 * assigned to a txg.
3795 */
3796 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3797
3798 /*
3799 * Writes will, almost always, require additional memory allocations
3800 * in order to compress/encrypt/etc the data. We therefor need to
3801 * make sure that there is sufficient available memory for this.
3802 */
3803 if (error = arc_memory_throttle(reserve, anon_size, txg))
3804 return (error);
3805
3806 /*
3807 * Throttle writes when the amount of dirty data in the cache
3808 * gets too large. We try to keep the cache less than half full
3809 * of dirty blocks so that our sync times don't grow too large.
3810 * Note: if two requests come in concurrently, we might let them
3811 * both succeed, when one of them should fail. Not a huge deal.
3812 */
3813
3814 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3815 anon_size > arc_c / 4) {
3816 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3817 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3818 arc_tempreserve>>10,
3819 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3820 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3821 reserve>>10, arc_c>>10);
3822 return (ERESTART);
3823 }
3824 atomic_add_64(&arc_tempreserve, reserve);
3825 return (0);
3826}
3827
3828static kmutex_t arc_lowmem_lock;
3829#ifdef _KERNEL
3830static eventhandler_tag arc_event_lowmem = NULL;
3831
3832static void
3833arc_lowmem(void *arg __unused, int howto __unused)
3834{
3835
3836 /* Serialize access via arc_lowmem_lock. */
3837 mutex_enter(&arc_lowmem_lock);
3838 mutex_enter(&arc_reclaim_thr_lock);
3839 needfree = 1;
3840 cv_signal(&arc_reclaim_thr_cv);
3841
3842 /*
3843 * It is unsafe to block here in arbitrary threads, because we can come
3844 * here from ARC itself and may hold ARC locks and thus risk a deadlock
3845 * with ARC reclaim thread.
3846 */
3847 if (curproc == pageproc) {
3848 while (needfree)
3849 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0);
3850 }
3851 mutex_exit(&arc_reclaim_thr_lock);
3852 mutex_exit(&arc_lowmem_lock);
3853}
3854#endif
3855
3856void
3857arc_init(void)
3858{
3859 int i, prefetch_tunable_set = 0;
3860
3861 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3862 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3863 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL);
3864
3865 /* Convert seconds to clock ticks */
3866 arc_min_prefetch_lifespan = 1 * hz;
3867
3868 /* Start out with 1/8 of all memory */
3869 arc_c = kmem_size() / 8;
3870
3871#ifdef sun
3872#ifdef _KERNEL
3873 /*
3874 * On architectures where the physical memory can be larger
3875 * than the addressable space (intel in 32-bit mode), we may
3876 * need to limit the cache to 1/8 of VM size.
3877 */
3878 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3879#endif
3880#endif /* sun */
3881 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */
3882 arc_c_min = MAX(arc_c / 4, 64<<18);
3883 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
3884 if (arc_c * 8 >= 1<<30)
3885 arc_c_max = (arc_c * 8) - (1<<30);
3886 else
3887 arc_c_max = arc_c_min;
3888 arc_c_max = MAX(arc_c * 5, arc_c_max);
3889
3890#ifdef _KERNEL
3891 /*
3892 * Allow the tunables to override our calculations if they are
3893 * reasonable (ie. over 16MB)
3894 */
3895 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size())
3896 arc_c_max = zfs_arc_max;
3897 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max)
3898 arc_c_min = zfs_arc_min;
3899#endif
3900
3901 arc_c = arc_c_max;
3902 arc_p = (arc_c >> 1);
3903
3904 /* limit meta-data to 1/4 of the arc capacity */
3905 arc_meta_limit = arc_c_max / 4;
3906
3907 /* Allow the tunable to override if it is reasonable */
3908 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3909 arc_meta_limit = zfs_arc_meta_limit;
3910
3911 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3912 arc_c_min = arc_meta_limit / 2;
3913
3914 if (zfs_arc_grow_retry > 0)
3915 arc_grow_retry = zfs_arc_grow_retry;
3916
3917 if (zfs_arc_shrink_shift > 0)
3918 arc_shrink_shift = zfs_arc_shrink_shift;
3919
3920 if (zfs_arc_p_min_shift > 0)
3921 arc_p_min_shift = zfs_arc_p_min_shift;
3922
3923 /* if kmem_flags are set, lets try to use less memory */
3924 if (kmem_debugging())
3925 arc_c = arc_c / 2;
3926 if (arc_c < arc_c_min)
3927 arc_c = arc_c_min;
3928
3929 zfs_arc_min = arc_c_min;
3930 zfs_arc_max = arc_c_max;
3931
3932 arc_anon = &ARC_anon;
3933 arc_mru = &ARC_mru;
3934 arc_mru_ghost = &ARC_mru_ghost;
3935 arc_mfu = &ARC_mfu;
3936 arc_mfu_ghost = &ARC_mfu_ghost;
3937 arc_l2c_only = &ARC_l2c_only;
3938 arc_size = 0;
3939
3940 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
3941 mutex_init(&arc_anon->arcs_locks[i].arcs_lock,
3942 NULL, MUTEX_DEFAULT, NULL);
3943 mutex_init(&arc_mru->arcs_locks[i].arcs_lock,
3944 NULL, MUTEX_DEFAULT, NULL);
3945 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock,
3946 NULL, MUTEX_DEFAULT, NULL);
3947 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock,
3948 NULL, MUTEX_DEFAULT, NULL);
3949 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock,
3950 NULL, MUTEX_DEFAULT, NULL);
3951 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock,
3952 NULL, MUTEX_DEFAULT, NULL);
3953
3954 list_create(&arc_mru->arcs_lists[i],
3955 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3956 list_create(&arc_mru_ghost->arcs_lists[i],
3957 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3958 list_create(&arc_mfu->arcs_lists[i],
3959 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3960 list_create(&arc_mfu_ghost->arcs_lists[i],
3961 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3962 list_create(&arc_mfu_ghost->arcs_lists[i],
3963 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3964 list_create(&arc_l2c_only->arcs_lists[i],
3965 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3966 }
3967
3968 buf_init();
3969
3970 arc_thread_exit = 0;
3971 arc_eviction_list = NULL;
3972 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3973 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3974
3975 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3976 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3977
3978 if (arc_ksp != NULL) {
3979 arc_ksp->ks_data = &arc_stats;
3980 kstat_install(arc_ksp);
3981 }
3982
3983 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3984 TS_RUN, minclsyspri);
3985
3986#ifdef _KERNEL
3987 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
3988 EVENTHANDLER_PRI_FIRST);
3989#endif
3990
3991 arc_dead = FALSE;
3992 arc_warm = B_FALSE;
3993
3994 if (zfs_write_limit_max == 0)
3995 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3996 else
3997 zfs_write_limit_shift = 0;
3998 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3999
4000#ifdef _KERNEL
4001 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
4002 prefetch_tunable_set = 1;
4003
4004#ifdef __i386__
4005 if (prefetch_tunable_set == 0) {
4006 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
4007 "-- to enable,\n");
4008 printf(" add \"vfs.zfs.prefetch_disable=0\" "
4009 "to /boot/loader.conf.\n");
4010 zfs_prefetch_disable = 1;
4011 }
4012#else
4013 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
4014 prefetch_tunable_set == 0) {
4015 printf("ZFS NOTICE: Prefetch is disabled by default if less "
4016 "than 4GB of RAM is present;\n"
4017 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
4018 "to /boot/loader.conf.\n");
4019 zfs_prefetch_disable = 1;
4020 }
4021#endif
4022 /* Warn about ZFS memory and address space requirements. */
4023 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
4024 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
4025 "expect unstable behavior.\n");
4026 }
4027 if (kmem_size() < 512 * (1 << 20)) {
4028 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
4029 "expect unstable behavior.\n");
4030 printf(" Consider tuning vm.kmem_size and "
4031 "vm.kmem_size_max\n");
4032 printf(" in /boot/loader.conf.\n");
4033 }
4034#endif
4035}
4036
4037void
4038arc_fini(void)
4039{
4040 int i;
4041
4042 mutex_enter(&arc_reclaim_thr_lock);
4043 arc_thread_exit = 1;
4044 cv_signal(&arc_reclaim_thr_cv);
4045 while (arc_thread_exit != 0)
4046 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
4047 mutex_exit(&arc_reclaim_thr_lock);
4048
4049 arc_flush(NULL);
4050
4051 arc_dead = TRUE;
4052
4053 if (arc_ksp != NULL) {
4054 kstat_delete(arc_ksp);
4055 arc_ksp = NULL;
4056 }
4057
4058 mutex_destroy(&arc_eviction_mtx);
4059 mutex_destroy(&arc_reclaim_thr_lock);
4060 cv_destroy(&arc_reclaim_thr_cv);
4061
4062 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
4063 list_destroy(&arc_mru->arcs_lists[i]);
4064 list_destroy(&arc_mru_ghost->arcs_lists[i]);
4065 list_destroy(&arc_mfu->arcs_lists[i]);
4066 list_destroy(&arc_mfu_ghost->arcs_lists[i]);
4067 list_destroy(&arc_l2c_only->arcs_lists[i]);
4068
4069 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock);
4070 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock);
4071 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock);
4072 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock);
4073 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock);
4074 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock);
4075 }
4076
4077 mutex_destroy(&zfs_write_limit_lock);
4078
4079 buf_fini();
4080
4081 ASSERT(arc_loaned_bytes == 0);
4082
4083 mutex_destroy(&arc_lowmem_lock);
4084#ifdef _KERNEL
4085 if (arc_event_lowmem != NULL)
4086 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
4087#endif
4088}
4089
4090/*
4091 * Level 2 ARC
4092 *
4093 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
4094 * It uses dedicated storage devices to hold cached data, which are populated
4095 * using large infrequent writes. The main role of this cache is to boost
4096 * the performance of random read workloads. The intended L2ARC devices
4097 * include short-stroked disks, solid state disks, and other media with
4098 * substantially faster read latency than disk.
4099 *
4100 * +-----------------------+
4101 * | ARC |
4102 * +-----------------------+
4103 * | ^ ^
4104 * | | |
4105 * l2arc_feed_thread() arc_read()
4106 * | | |
4107 * | l2arc read |
4108 * V | |
4109 * +---------------+ |
4110 * | L2ARC | |
4111 * +---------------+ |
4112 * | ^ |
4113 * l2arc_write() | |
4114 * | | |
4115 * V | |
4116 * +-------+ +-------+
4117 * | vdev | | vdev |
4118 * | cache | | cache |
4119 * +-------+ +-------+
4120 * +=========+ .-----.
4121 * : L2ARC : |-_____-|
4122 * : devices : | Disks |
4123 * +=========+ `-_____-'
4124 *
4125 * Read requests are satisfied from the following sources, in order:
4126 *
4127 * 1) ARC
4128 * 2) vdev cache of L2ARC devices
4129 * 3) L2ARC devices
4130 * 4) vdev cache of disks
4131 * 5) disks
4132 *
4133 * Some L2ARC device types exhibit extremely slow write performance.
4134 * To accommodate for this there are some significant differences between
4135 * the L2ARC and traditional cache design:
4136 *
4137 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4138 * the ARC behave as usual, freeing buffers and placing headers on ghost
4139 * lists. The ARC does not send buffers to the L2ARC during eviction as
4140 * this would add inflated write latencies for all ARC memory pressure.
4141 *
4142 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4143 * It does this by periodically scanning buffers from the eviction-end of
4144 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4145 * not already there. It scans until a headroom of buffers is satisfied,
4146 * which itself is a buffer for ARC eviction. The thread that does this is
4147 * l2arc_feed_thread(), illustrated below; example sizes are included to
4148 * provide a better sense of ratio than this diagram:
4149 *
4150 * head --> tail
4151 * +---------------------+----------+
4152 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4153 * +---------------------+----------+ | o L2ARC eligible
4154 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4155 * +---------------------+----------+ |
4156 * 15.9 Gbytes ^ 32 Mbytes |
4157 * headroom |
4158 * l2arc_feed_thread()
4159 * |
4160 * l2arc write hand <--[oooo]--'
4161 * | 8 Mbyte
4162 * | write max
4163 * V
4164 * +==============================+
4165 * L2ARC dev |####|#|###|###| |####| ... |
4166 * +==============================+
4167 * 32 Gbytes
4168 *
4169 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4170 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4171 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4172 * safe to say that this is an uncommon case, since buffers at the end of
4173 * the ARC lists have moved there due to inactivity.
4174 *
4175 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4176 * then the L2ARC simply misses copying some buffers. This serves as a
4177 * pressure valve to prevent heavy read workloads from both stalling the ARC
4178 * with waits and clogging the L2ARC with writes. This also helps prevent
4179 * the potential for the L2ARC to churn if it attempts to cache content too
4180 * quickly, such as during backups of the entire pool.
4181 *
4182 * 5. After system boot and before the ARC has filled main memory, there are
4183 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4184 * lists can remain mostly static. Instead of searching from tail of these
4185 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4186 * for eligible buffers, greatly increasing its chance of finding them.
4187 *
4188 * The L2ARC device write speed is also boosted during this time so that
4189 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4190 * there are no L2ARC reads, and no fear of degrading read performance
4191 * through increased writes.
4192 *
4193 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4194 * the vdev queue can aggregate them into larger and fewer writes. Each
4195 * device is written to in a rotor fashion, sweeping writes through
4196 * available space then repeating.
4197 *
4198 * 7. The L2ARC does not store dirty content. It never needs to flush
4199 * write buffers back to disk based storage.
4200 *
4201 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4202 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4203 *
4204 * The performance of the L2ARC can be tweaked by a number of tunables, which
4205 * may be necessary for different workloads:
4206 *
4207 * l2arc_write_max max write bytes per interval
4208 * l2arc_write_boost extra write bytes during device warmup
4209 * l2arc_noprefetch skip caching prefetched buffers
4210 * l2arc_headroom number of max device writes to precache
4211 * l2arc_feed_secs seconds between L2ARC writing
4212 *
4213 * Tunables may be removed or added as future performance improvements are
4214 * integrated, and also may become zpool properties.
4215 *
4216 * There are three key functions that control how the L2ARC warms up:
4217 *
4218 * l2arc_write_eligible() check if a buffer is eligible to cache
4219 * l2arc_write_size() calculate how much to write
4220 * l2arc_write_interval() calculate sleep delay between writes
4221 *
4222 * These three functions determine what to write, how much, and how quickly
4223 * to send writes.
4224 */
4225
4226static boolean_t
4227l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4228{
4229 /*
4230 * A buffer is *not* eligible for the L2ARC if it:
4231 * 1. belongs to a different spa.
4232 * 2. is already cached on the L2ARC.
4233 * 3. has an I/O in progress (it may be an incomplete read).
4234 * 4. is flagged not eligible (zfs property).
4235 */
4236 if (ab->b_spa != spa_guid) {
4237 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
4238 return (B_FALSE);
4239 }
4240 if (ab->b_l2hdr != NULL) {
4241 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
4242 return (B_FALSE);
4243 }
4244 if (HDR_IO_IN_PROGRESS(ab)) {
4245 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
4246 return (B_FALSE);
4247 }
4248 if (!HDR_L2CACHE(ab)) {
4249 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
4250 return (B_FALSE);
4251 }
4252
4253 return (B_TRUE);
4254}
4255
4256static uint64_t
4257l2arc_write_size(l2arc_dev_t *dev)
4258{
4259 uint64_t size;
4260
4261 size = dev->l2ad_write;
4262
4263 if (arc_warm == B_FALSE)
4264 size += dev->l2ad_boost;
4265
4266 return (size);
4267
4268}
4269
4270static clock_t
4271l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4272{
4273 clock_t interval, next, now;
4274
4275 /*
4276 * If the ARC lists are busy, increase our write rate; if the
4277 * lists are stale, idle back. This is achieved by checking
4278 * how much we previously wrote - if it was more than half of
4279 * what we wanted, schedule the next write much sooner.
4280 */
4281 if (l2arc_feed_again && wrote > (wanted / 2))
4282 interval = (hz * l2arc_feed_min_ms) / 1000;
4283 else
4284 interval = hz * l2arc_feed_secs;
4285
4286 now = ddi_get_lbolt();
4287 next = MAX(now, MIN(now + interval, began + interval));
4288
4289 return (next);
4290}
4291
4292static void
4293l2arc_hdr_stat_add(void)
4294{
4295 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4296 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4297}
4298
4299static void
4300l2arc_hdr_stat_remove(void)
4301{
4302 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4303 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4304}
4305
4306/*
4307 * Cycle through L2ARC devices. This is how L2ARC load balances.
4308 * If a device is returned, this also returns holding the spa config lock.
4309 */
4310static l2arc_dev_t *
4311l2arc_dev_get_next(void)
4312{
4313 l2arc_dev_t *first, *next = NULL;
4314
4315 /*
4316 * Lock out the removal of spas (spa_namespace_lock), then removal
4317 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4318 * both locks will be dropped and a spa config lock held instead.
4319 */
4320 mutex_enter(&spa_namespace_lock);
4321 mutex_enter(&l2arc_dev_mtx);
4322
4323 /* if there are no vdevs, there is nothing to do */
4324 if (l2arc_ndev == 0)
4325 goto out;
4326
4327 first = NULL;
4328 next = l2arc_dev_last;
4329 do {
4330 /* loop around the list looking for a non-faulted vdev */
4331 if (next == NULL) {
4332 next = list_head(l2arc_dev_list);
4333 } else {
4334 next = list_next(l2arc_dev_list, next);
4335 if (next == NULL)
4336 next = list_head(l2arc_dev_list);
4337 }
4338
4339 /* if we have come back to the start, bail out */
4340 if (first == NULL)
4341 first = next;
4342 else if (next == first)
4343 break;
4344
4345 } while (vdev_is_dead(next->l2ad_vdev));
4346
4347 /* if we were unable to find any usable vdevs, return NULL */
4348 if (vdev_is_dead(next->l2ad_vdev))
4349 next = NULL;
4350
4351 l2arc_dev_last = next;
4352
4353out:
4354 mutex_exit(&l2arc_dev_mtx);
4355
4356 /*
4357 * Grab the config lock to prevent the 'next' device from being
4358 * removed while we are writing to it.
4359 */
4360 if (next != NULL)
4361 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4362 mutex_exit(&spa_namespace_lock);
4363
4364 return (next);
4365}
4366
4367/*
4368 * Free buffers that were tagged for destruction.
4369 */
4370static void
4371l2arc_do_free_on_write()
4372{
4373 list_t *buflist;
4374 l2arc_data_free_t *df, *df_prev;
4375
4376 mutex_enter(&l2arc_free_on_write_mtx);
4377 buflist = l2arc_free_on_write;
4378
4379 for (df = list_tail(buflist); df; df = df_prev) {
4380 df_prev = list_prev(buflist, df);
4381 ASSERT(df->l2df_data != NULL);
4382 ASSERT(df->l2df_func != NULL);
4383 df->l2df_func(df->l2df_data, df->l2df_size);
4384 list_remove(buflist, df);
4385 kmem_free(df, sizeof (l2arc_data_free_t));
4386 }
4387
4388 mutex_exit(&l2arc_free_on_write_mtx);
4389}
4390
4391/*
4392 * A write to a cache device has completed. Update all headers to allow
4393 * reads from these buffers to begin.
4394 */
4395static void
4396l2arc_write_done(zio_t *zio)
4397{
4398 l2arc_write_callback_t *cb;
4399 l2arc_dev_t *dev;
4400 list_t *buflist;
4401 arc_buf_hdr_t *head, *ab, *ab_prev;
4402 l2arc_buf_hdr_t *abl2;
4403 kmutex_t *hash_lock;
4404
4405 cb = zio->io_private;
4406 ASSERT(cb != NULL);
4407 dev = cb->l2wcb_dev;
4408 ASSERT(dev != NULL);
4409 head = cb->l2wcb_head;
4410 ASSERT(head != NULL);
4411 buflist = dev->l2ad_buflist;
4412 ASSERT(buflist != NULL);
4413 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4414 l2arc_write_callback_t *, cb);
4415
4416 if (zio->io_error != 0)
4417 ARCSTAT_BUMP(arcstat_l2_writes_error);
4418
4419 mutex_enter(&l2arc_buflist_mtx);
4420
4421 /*
4422 * All writes completed, or an error was hit.
4423 */
4424 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4425 ab_prev = list_prev(buflist, ab);
4426
4427 hash_lock = HDR_LOCK(ab);
4428 if (!mutex_tryenter(hash_lock)) {
4429 /*
4430 * This buffer misses out. It may be in a stage
4431 * of eviction. Its ARC_L2_WRITING flag will be
4432 * left set, denying reads to this buffer.
4433 */
4434 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4435 continue;
4436 }
4437
4438 if (zio->io_error != 0) {
4439 /*
4440 * Error - drop L2ARC entry.
4441 */
4442 list_remove(buflist, ab);
4443 abl2 = ab->b_l2hdr;
4444 ab->b_l2hdr = NULL;
4445 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4446 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4447 }
4448
4449 /*
4450 * Allow ARC to begin reads to this L2ARC entry.
4451 */
4452 ab->b_flags &= ~ARC_L2_WRITING;
4453
4454 mutex_exit(hash_lock);
4455 }
4456
4457 atomic_inc_64(&l2arc_writes_done);
4458 list_remove(buflist, head);
4459 kmem_cache_free(hdr_cache, head);
4460 mutex_exit(&l2arc_buflist_mtx);
4461
4462 l2arc_do_free_on_write();
4463
4464 kmem_free(cb, sizeof (l2arc_write_callback_t));
4465}
4466
4467/*
4468 * A read to a cache device completed. Validate buffer contents before
4469 * handing over to the regular ARC routines.
4470 */
4471static void
4472l2arc_read_done(zio_t *zio)
4473{
4474 l2arc_read_callback_t *cb;
4475 arc_buf_hdr_t *hdr;
4476 arc_buf_t *buf;
4477 kmutex_t *hash_lock;
4478 int equal;
4479
4480 ASSERT(zio->io_vd != NULL);
4481 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4482
4483 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4484
4485 cb = zio->io_private;
4486 ASSERT(cb != NULL);
4487 buf = cb->l2rcb_buf;
4488 ASSERT(buf != NULL);
4489
4490 hash_lock = HDR_LOCK(buf->b_hdr);
4491 mutex_enter(hash_lock);
4492 hdr = buf->b_hdr;
4493 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4494
4495 /*
4496 * Check this survived the L2ARC journey.
4497 */
4498 equal = arc_cksum_equal(buf);
4499 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4500 mutex_exit(hash_lock);
4501 zio->io_private = buf;
4502 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4503 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4504 arc_read_done(zio);
4505 } else {
4506 mutex_exit(hash_lock);
4507 /*
4508 * Buffer didn't survive caching. Increment stats and
4509 * reissue to the original storage device.
4510 */
4511 if (zio->io_error != 0) {
4512 ARCSTAT_BUMP(arcstat_l2_io_error);
4513 } else {
4514 zio->io_error = EIO;
4515 }
4516 if (!equal)
4517 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4518
4519 /*
4520 * If there's no waiter, issue an async i/o to the primary
4521 * storage now. If there *is* a waiter, the caller must
4522 * issue the i/o in a context where it's OK to block.
4523 */
4524 if (zio->io_waiter == NULL) {
4525 zio_t *pio = zio_unique_parent(zio);
4526
4527 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4528
4529 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4530 buf->b_data, zio->io_size, arc_read_done, buf,
4531 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4532 }
4533 }
4534
4535 kmem_free(cb, sizeof (l2arc_read_callback_t));
4536}
4537
4538/*
4539 * This is the list priority from which the L2ARC will search for pages to
4540 * cache. This is used within loops (0..3) to cycle through lists in the
4541 * desired order. This order can have a significant effect on cache
4542 * performance.
4543 *
4544 * Currently the metadata lists are hit first, MFU then MRU, followed by
4545 * the data lists. This function returns a locked list, and also returns
4546 * the lock pointer.
4547 */
4548static list_t *
4549l2arc_list_locked(int list_num, kmutex_t **lock)
4550{
|