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) 2013 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 mutexes, rather they rely on the
84 * hash table mutexes for the bulk of their protection (i.e. most
85 * fields in the arc_buf_hdr_t are protected by these mutexes).
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/vmsystm.h>
130 #include <vm/anon.h>
131 #include <sys/fs/swapnode.h>
132 #include <sys/dnlc.h>
133 #endif
134 #include <sys/callb.h>
135 #include <sys/kstat.h>
136 #include <zfs_fletcher.h>
137
138 #ifndef _KERNEL
139 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
140 boolean_t arc_watch = B_FALSE;
141 int arc_procfd;
142 #endif
143
144 static kmutex_t arc_reclaim_thr_lock;
145 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
146 static uint8_t arc_thread_exit;
147
148 extern int zfs_write_limit_shift;
149 extern uint64_t zfs_write_limit_max;
150 extern kmutex_t zfs_write_limit_lock;
151
152 #define ARC_REDUCE_DNLC_PERCENT 3
153 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
154
155 typedef enum arc_reclaim_strategy {
156 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
157 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
158 } arc_reclaim_strategy_t;
159
160 /* number of seconds before growing cache again */
161 static int arc_grow_retry = 60;
162
163 /* shift of arc_c for calculating both min and max arc_p */
164 static int arc_p_min_shift = 4;
165
166 /* log2(fraction of arc to reclaim) */
167 static int arc_shrink_shift = 5;
168
169 /*
170 * minimum lifespan of a prefetch block in clock ticks
171 * (initialized in arc_init())
172 */
173 static int arc_min_prefetch_lifespan;
174
175 static int arc_dead;
176
177 /*
178 * The arc has filled available memory and has now warmed up.
179 */
180 static boolean_t arc_warm;
181
182 /*
183 * These tunables are for performance analysis.
184 */
185 uint64_t zfs_arc_max;
186 uint64_t zfs_arc_min;
187 uint64_t zfs_arc_meta_limit = 0;
188 int zfs_arc_grow_retry = 0;
189 int zfs_arc_shrink_shift = 0;
190 int zfs_arc_p_min_shift = 0;
191 int zfs_disable_dup_eviction = 0;
192
193 /*
194 * Note that buffers can be in one of 6 states:
195 * ARC_anon - anonymous (discussed below)
196 * ARC_mru - recently used, currently cached
197 * ARC_mru_ghost - recentely used, no longer in cache
198 * ARC_mfu - frequently used, currently cached
199 * ARC_mfu_ghost - frequently used, no longer in cache
200 * ARC_l2c_only - exists in L2ARC but not other states
201 * When there are no active references to the buffer, they are
202 * are linked onto a list in one of these arc states. These are
203 * the only buffers that can be evicted or deleted. Within each
204 * state there are multiple lists, one for meta-data and one for
205 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
206 * etc.) is tracked separately so that it can be managed more
207 * explicitly: favored over data, limited explicitly.
208 *
209 * Anonymous buffers are buffers that are not associated with
210 * a DVA. These are buffers that hold dirty block copies
211 * before they are written to stable storage. By definition,
212 * they are "ref'd" and are considered part of arc_mru
213 * that cannot be freed. Generally, they will aquire a DVA
214 * as they are written and migrate onto the arc_mru list.
215 *
216 * The ARC_l2c_only state is for buffers that are in the second
217 * level ARC but no longer in any of the ARC_m* lists. The second
218 * level ARC itself may also contain buffers that are in any of
219 * the ARC_m* states - meaning that a buffer can exist in two
220 * places. The reason for the ARC_l2c_only state is to keep the
221 * buffer header in the hash table, so that reads that hit the
222 * second level ARC benefit from these fast lookups.
223 */
224
225 typedef struct arc_state {
226 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
227 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
228 uint64_t arcs_size; /* total amount of data in this state */
229 kmutex_t arcs_mtx;
230 } arc_state_t;
231
232 /* The 6 states: */
233 static arc_state_t ARC_anon;
234 static arc_state_t ARC_mru;
235 static arc_state_t ARC_mru_ghost;
236 static arc_state_t ARC_mfu;
237 static arc_state_t ARC_mfu_ghost;
238 static arc_state_t ARC_l2c_only;
239
240 typedef struct arc_stats {
241 kstat_named_t arcstat_hits;
242 kstat_named_t arcstat_misses;
243 kstat_named_t arcstat_demand_data_hits;
244 kstat_named_t arcstat_demand_data_misses;
245 kstat_named_t arcstat_demand_metadata_hits;
246 kstat_named_t arcstat_demand_metadata_misses;
247 kstat_named_t arcstat_prefetch_data_hits;
248 kstat_named_t arcstat_prefetch_data_misses;
249 kstat_named_t arcstat_prefetch_metadata_hits;
250 kstat_named_t arcstat_prefetch_metadata_misses;
251 kstat_named_t arcstat_mru_hits;
252 kstat_named_t arcstat_mru_ghost_hits;
253 kstat_named_t arcstat_mfu_hits;
254 kstat_named_t arcstat_mfu_ghost_hits;
255 kstat_named_t arcstat_deleted;
256 kstat_named_t arcstat_recycle_miss;
257 kstat_named_t arcstat_mutex_miss;
258 kstat_named_t arcstat_evict_skip;
259 kstat_named_t arcstat_evict_l2_cached;
260 kstat_named_t arcstat_evict_l2_eligible;
261 kstat_named_t arcstat_evict_l2_ineligible;
262 kstat_named_t arcstat_hash_elements;
263 kstat_named_t arcstat_hash_elements_max;
264 kstat_named_t arcstat_hash_collisions;
265 kstat_named_t arcstat_hash_chains;
266 kstat_named_t arcstat_hash_chain_max;
267 kstat_named_t arcstat_p;
268 kstat_named_t arcstat_c;
269 kstat_named_t arcstat_c_min;
270 kstat_named_t arcstat_c_max;
271 kstat_named_t arcstat_size;
272 kstat_named_t arcstat_hdr_size;
273 kstat_named_t arcstat_data_size;
274 kstat_named_t arcstat_other_size;
275 kstat_named_t arcstat_l2_hits;
276 kstat_named_t arcstat_l2_misses;
277 kstat_named_t arcstat_l2_feeds;
278 kstat_named_t arcstat_l2_rw_clash;
279 kstat_named_t arcstat_l2_read_bytes;
280 kstat_named_t arcstat_l2_write_bytes;
281 kstat_named_t arcstat_l2_writes_sent;
282 kstat_named_t arcstat_l2_writes_done;
283 kstat_named_t arcstat_l2_writes_error;
284 kstat_named_t arcstat_l2_writes_hdr_miss;
285 kstat_named_t arcstat_l2_evict_lock_retry;
286 kstat_named_t arcstat_l2_evict_reading;
287 kstat_named_t arcstat_l2_free_on_write;
288 kstat_named_t arcstat_l2_abort_lowmem;
289 kstat_named_t arcstat_l2_cksum_bad;
290 kstat_named_t arcstat_l2_io_error;
291 kstat_named_t arcstat_l2_size;
292 kstat_named_t arcstat_l2_hdr_size;
293 kstat_named_t arcstat_memory_throttle_count;
294 kstat_named_t arcstat_duplicate_buffers;
295 kstat_named_t arcstat_duplicate_buffers_size;
296 kstat_named_t arcstat_duplicate_reads;
297 kstat_named_t arcstat_meta_used;
298 kstat_named_t arcstat_meta_limit;
299 kstat_named_t arcstat_meta_max;
300 } arc_stats_t;
301
302 static arc_stats_t arc_stats = {
303 { "hits", KSTAT_DATA_UINT64 },
304 { "misses", KSTAT_DATA_UINT64 },
305 { "demand_data_hits", KSTAT_DATA_UINT64 },
306 { "demand_data_misses", KSTAT_DATA_UINT64 },
307 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
308 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
309 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
310 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
311 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
312 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
313 { "mru_hits", KSTAT_DATA_UINT64 },
314 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
315 { "mfu_hits", KSTAT_DATA_UINT64 },
316 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
317 { "deleted", KSTAT_DATA_UINT64 },
318 { "recycle_miss", KSTAT_DATA_UINT64 },
319 { "mutex_miss", KSTAT_DATA_UINT64 },
320 { "evict_skip", KSTAT_DATA_UINT64 },
321 { "evict_l2_cached", KSTAT_DATA_UINT64 },
322 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
323 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
324 { "hash_elements", KSTAT_DATA_UINT64 },
325 { "hash_elements_max", KSTAT_DATA_UINT64 },
326 { "hash_collisions", KSTAT_DATA_UINT64 },
327 { "hash_chains", KSTAT_DATA_UINT64 },
328 { "hash_chain_max", KSTAT_DATA_UINT64 },
329 { "p", KSTAT_DATA_UINT64 },
330 { "c", KSTAT_DATA_UINT64 },
331 { "c_min", KSTAT_DATA_UINT64 },
332 { "c_max", KSTAT_DATA_UINT64 },
333 { "size", KSTAT_DATA_UINT64 },
334 { "hdr_size", KSTAT_DATA_UINT64 },
335 { "data_size", KSTAT_DATA_UINT64 },
336 { "other_size", KSTAT_DATA_UINT64 },
337 { "l2_hits", KSTAT_DATA_UINT64 },
338 { "l2_misses", KSTAT_DATA_UINT64 },
339 { "l2_feeds", KSTAT_DATA_UINT64 },
340 { "l2_rw_clash", KSTAT_DATA_UINT64 },
341 { "l2_read_bytes", KSTAT_DATA_UINT64 },
342 { "l2_write_bytes", KSTAT_DATA_UINT64 },
343 { "l2_writes_sent", KSTAT_DATA_UINT64 },
344 { "l2_writes_done", KSTAT_DATA_UINT64 },
345 { "l2_writes_error", KSTAT_DATA_UINT64 },
346 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
347 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
348 { "l2_evict_reading", KSTAT_DATA_UINT64 },
349 { "l2_free_on_write", KSTAT_DATA_UINT64 },
350 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
351 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
352 { "l2_io_error", KSTAT_DATA_UINT64 },
353 { "l2_size", KSTAT_DATA_UINT64 },
354 { "l2_hdr_size", KSTAT_DATA_UINT64 },
355 { "memory_throttle_count", KSTAT_DATA_UINT64 },
356 { "duplicate_buffers", KSTAT_DATA_UINT64 },
357 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
358 { "duplicate_reads", KSTAT_DATA_UINT64 },
359 { "arc_meta_used", KSTAT_DATA_UINT64 },
360 { "arc_meta_limit", KSTAT_DATA_UINT64 },
361 { "arc_meta_max", KSTAT_DATA_UINT64 }
362 };
363
364 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
365
366 #define ARCSTAT_INCR(stat, val) \
367 atomic_add_64(&arc_stats.stat.value.ui64, (val));
368
369 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
370 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
371
372 #define ARCSTAT_MAX(stat, val) { \
373 uint64_t m; \
374 while ((val) > (m = arc_stats.stat.value.ui64) && \
375 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
376 continue; \
377 }
378
379 #define ARCSTAT_MAXSTAT(stat) \
380 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
381
382 /*
383 * We define a macro to allow ARC hits/misses to be easily broken down by
384 * two separate conditions, giving a total of four different subtypes for
385 * each of hits and misses (so eight statistics total).
386 */
387 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
388 if (cond1) { \
389 if (cond2) { \
390 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
391 } else { \
392 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
393 } \
394 } else { \
395 if (cond2) { \
396 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
397 } else { \
398 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
399 } \
400 }
401
402 kstat_t *arc_ksp;
403 static arc_state_t *arc_anon;
404 static arc_state_t *arc_mru;
405 static arc_state_t *arc_mru_ghost;
406 static arc_state_t *arc_mfu;
407 static arc_state_t *arc_mfu_ghost;
408 static arc_state_t *arc_l2c_only;
409
410 /*
411 * There are several ARC variables that are critical to export as kstats --
412 * but we don't want to have to grovel around in the kstat whenever we wish to
413 * manipulate them. For these variables, we therefore define them to be in
414 * terms of the statistic variable. This assures that we are not introducing
415 * the possibility of inconsistency by having shadow copies of the variables,
416 * while still allowing the code to be readable.
417 */
418 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
419 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
420 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
421 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
422 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
423 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
424 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
425 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
426
427 static int arc_no_grow; /* Don't try to grow cache size */
428 static uint64_t arc_tempreserve;
429 static uint64_t arc_loaned_bytes;
430
431 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
432
433 typedef struct arc_callback arc_callback_t;
434
435 struct arc_callback {
436 void *acb_private;
437 arc_done_func_t *acb_done;
438 arc_buf_t *acb_buf;
439 zio_t *acb_zio_dummy;
440 arc_callback_t *acb_next;
441 };
442
443 typedef struct arc_write_callback arc_write_callback_t;
444
445 struct arc_write_callback {
446 void *awcb_private;
447 arc_done_func_t *awcb_ready;
448 arc_done_func_t *awcb_done;
449 arc_buf_t *awcb_buf;
450 };
451
452 struct arc_buf_hdr {
453 /* protected by hash lock */
454 dva_t b_dva;
455 uint64_t b_birth;
456 uint64_t b_cksum0;
457
458 kmutex_t b_freeze_lock;
459 zio_cksum_t *b_freeze_cksum;
460 void *b_thawed;
461
462 arc_buf_hdr_t *b_hash_next;
463 arc_buf_t *b_buf;
464 uint32_t b_flags;
465 uint32_t b_datacnt;
466
467 arc_callback_t *b_acb;
468 kcondvar_t b_cv;
469
470 /* immutable */
471 arc_buf_contents_t b_type;
472 uint64_t b_size;
473 uint64_t b_spa;
474
475 /* protected by arc state mutex */
476 arc_state_t *b_state;
477 list_node_t b_arc_node;
478
479 /* updated atomically */
480 clock_t b_arc_access;
481
482 /* self protecting */
483 refcount_t b_refcnt;
484
485 l2arc_buf_hdr_t *b_l2hdr;
486 list_node_t b_l2node;
487 };
488
489 static arc_buf_t *arc_eviction_list;
490 static kmutex_t arc_eviction_mtx;
491 static arc_buf_hdr_t arc_eviction_hdr;
492 static void arc_get_data_buf(arc_buf_t *buf);
493 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
494 static int arc_evict_needed(arc_buf_contents_t type);
495 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
496 static void arc_buf_watch(arc_buf_t *buf);
497
498 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
499
500 #define GHOST_STATE(state) \
501 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
502 (state) == arc_l2c_only)
503
504 /*
505 * Private ARC flags. These flags are private ARC only flags that will show up
506 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
507 * be passed in as arc_flags in things like arc_read. However, these flags
508 * should never be passed and should only be set by ARC code. When adding new
509 * public flags, make sure not to smash the private ones.
510 */
511
512 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
513 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
514 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
515 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
516 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
517 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
518 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
519 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
520 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
521 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
522
523 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
524 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
525 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
526 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
527 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
528 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
529 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
530 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
531 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
532 (hdr)->b_l2hdr != NULL)
533 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
534 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
535 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
536
537 /*
538 * Other sizes
539 */
540
541 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
542 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
543
544 /*
545 * Hash table routines
546 */
547
548 struct ht_table {
549 arc_buf_hdr_t *hdr;
550 kmutex_t lock;
551 };
552
553 typedef struct buf_hash_table {
554 uint64_t ht_mask;
555 struct ht_table *ht_table;
556 } buf_hash_table_t;
557
558 static buf_hash_table_t buf_hash_table;
559
560 #define BUF_HASH_INDEX(spa, dva, birth) \
561 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
562 #define BUF_HASH_LOCK(idx) (&buf_hash_table.ht_table[idx].lock)
563 #define HDR_LOCK(hdr) \
564 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
565
566 uint64_t zfs_crc64_table[256];
567
568 /*
569 * Level 2 ARC
570 */
571
572 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
573 #define L2ARC_HEADROOM 2 /* num of writes */
574 #define L2ARC_FEED_SECS 1 /* caching interval secs */
575 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
576
577 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
578 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
579
580 /*
581 * L2ARC Performance Tunables
582 */
583 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
584 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
585 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
586 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
587 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
588 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
589 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
590 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
591
592 /*
593 * L2ARC Internals
594 */
595 typedef struct l2arc_dev {
596 vdev_t *l2ad_vdev; /* vdev */
597 spa_t *l2ad_spa; /* spa */
598 uint64_t l2ad_hand; /* next write location */
599 uint64_t l2ad_write; /* desired write size, bytes */
600 uint64_t l2ad_boost; /* warmup write boost, bytes */
601 uint64_t l2ad_start; /* first addr on device */
602 uint64_t l2ad_end; /* last addr on device */
603 uint64_t l2ad_evict; /* last addr eviction reached */
604 boolean_t l2ad_first; /* first sweep through */
605 boolean_t l2ad_writing; /* currently writing */
606 list_t *l2ad_buflist; /* buffer list */
607 list_node_t l2ad_node; /* device list node */
608 } l2arc_dev_t;
609
610 static list_t L2ARC_dev_list; /* device list */
611 static list_t *l2arc_dev_list; /* device list pointer */
612 static kmutex_t l2arc_dev_mtx; /* device list mutex */
613 static l2arc_dev_t *l2arc_dev_last; /* last device used */
614 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
615 static list_t L2ARC_free_on_write; /* free after write buf list */
616 static list_t *l2arc_free_on_write; /* free after write list ptr */
617 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
618 static uint64_t l2arc_ndev; /* number of devices */
619
620 typedef struct l2arc_read_callback {
621 arc_buf_t *l2rcb_buf; /* read buffer */
622 spa_t *l2rcb_spa; /* spa */
623 blkptr_t l2rcb_bp; /* original blkptr */
624 zbookmark_t l2rcb_zb; /* original bookmark */
625 int l2rcb_flags; /* original flags */
626 } l2arc_read_callback_t;
627
628 typedef struct l2arc_write_callback {
629 l2arc_dev_t *l2wcb_dev; /* device info */
630 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
631 } l2arc_write_callback_t;
632
633 struct l2arc_buf_hdr {
634 /* protected by arc_buf_hdr mutex */
635 l2arc_dev_t *b_dev; /* L2ARC device */
636 uint64_t b_daddr; /* disk address, offset byte */
637 };
638
639 typedef struct l2arc_data_free {
640 /* protected by l2arc_free_on_write_mtx */
641 void *l2df_data;
642 size_t l2df_size;
643 void (*l2df_func)(void *, size_t);
644 list_node_t l2df_list_node;
645 } l2arc_data_free_t;
646
647 static kmutex_t l2arc_feed_thr_lock;
648 static kcondvar_t l2arc_feed_thr_cv;
649 static uint8_t l2arc_thread_exit;
650
651 static void l2arc_read_done(zio_t *zio);
652 static void l2arc_hdr_stat_add(void);
653 static void l2arc_hdr_stat_remove(void);
654
655 static uint64_t
656 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
657 {
658 uint8_t *vdva = (uint8_t *)dva;
659 uint64_t crc = -1ULL;
660 int i;
661
662 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
663
664 for (i = 0; i < sizeof (dva_t); i++)
665 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
666
667 crc ^= (spa>>8) ^ birth;
668
669 return (crc);
670 }
671
672 #define BUF_EMPTY(buf) \
673 ((buf)->b_dva.dva_word[0] == 0 && \
674 (buf)->b_dva.dva_word[1] == 0 && \
675 (buf)->b_birth == 0)
676
677 #define BUF_EQUAL(spa, dva, birth, buf) \
678 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
679 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
680 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
681
682 static void
683 buf_discard_identity(arc_buf_hdr_t *hdr)
684 {
685 hdr->b_dva.dva_word[0] = 0;
686 hdr->b_dva.dva_word[1] = 0;
687 hdr->b_birth = 0;
688 hdr->b_cksum0 = 0;
689 }
690
691 static arc_buf_hdr_t *
692 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
693 {
694 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
695 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
696 arc_buf_hdr_t *buf;
697
698 mutex_enter(hash_lock);
699 for (buf = buf_hash_table.ht_table[idx].hdr; buf != NULL;
700 buf = buf->b_hash_next) {
701 if (BUF_EQUAL(spa, dva, birth, buf)) {
702 *lockp = hash_lock;
703 return (buf);
704 }
705 }
706 mutex_exit(hash_lock);
707 *lockp = NULL;
708 return (NULL);
709 }
710
711 /*
712 * Insert an entry into the hash table. If there is already an element
713 * equal to elem in the hash table, then the already existing element
714 * will be returned and the new element will not be inserted.
715 * Otherwise returns NULL.
716 */
717 static arc_buf_hdr_t *
718 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
719 {
720 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
721 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
722 arc_buf_hdr_t *fbuf;
723 uint32_t i;
724
725 ASSERT(!HDR_IN_HASH_TABLE(buf));
726 *lockp = hash_lock;
727 mutex_enter(hash_lock);
728 for (fbuf = buf_hash_table.ht_table[idx].hdr, i = 0; fbuf != NULL;
729 fbuf = fbuf->b_hash_next, i++) {
730 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
731 return (fbuf);
732 }
733
734 buf->b_hash_next = buf_hash_table.ht_table[idx].hdr;
735 buf_hash_table.ht_table[idx].hdr = buf;
736 buf->b_flags |= ARC_IN_HASH_TABLE;
737
738 /* collect some hash table performance data */
739 if (i > 0) {
740 ARCSTAT_BUMP(arcstat_hash_collisions);
741 if (i == 1)
742 ARCSTAT_BUMP(arcstat_hash_chains);
743
744 ARCSTAT_MAX(arcstat_hash_chain_max, i);
745 }
746
747 ARCSTAT_BUMP(arcstat_hash_elements);
748 ARCSTAT_MAXSTAT(arcstat_hash_elements);
749
750 return (NULL);
751 }
752
753 static void
754 buf_hash_remove(arc_buf_hdr_t *buf)
755 {
756 arc_buf_hdr_t *fbuf, **bufp;
757 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
758
759 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
760 ASSERT(HDR_IN_HASH_TABLE(buf));
761
762 bufp = &buf_hash_table.ht_table[idx].hdr;
763 while ((fbuf = *bufp) != buf) {
764 ASSERT(fbuf != NULL);
765 bufp = &fbuf->b_hash_next;
766 }
767 *bufp = buf->b_hash_next;
768 buf->b_hash_next = NULL;
769 buf->b_flags &= ~ARC_IN_HASH_TABLE;
770
771 /* collect some hash table performance data */
772 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
773
774 if (buf_hash_table.ht_table[idx].hdr &&
775 buf_hash_table.ht_table[idx].hdr->b_hash_next == NULL)
776 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
777 }
778
779 /*
780 * Global data structures and functions for the buf kmem cache.
781 */
782 static kmem_cache_t *hdr_cache;
783 static kmem_cache_t *buf_cache;
784
785 static void
786 buf_fini(void)
787 {
788 int i;
789
790 for (i = 0; i < buf_hash_table.ht_mask + 1; i++)
791 mutex_destroy(&buf_hash_table.ht_table[i].lock);
792 kmem_free(buf_hash_table.ht_table,
793 (buf_hash_table.ht_mask + 1) * sizeof (struct ht_table));
794 kmem_cache_destroy(hdr_cache);
795 kmem_cache_destroy(buf_cache);
796 }
797
798 /*
799 * Constructor callback - called when the cache is empty
800 * and a new buf is requested.
801 */
802 /* ARGSUSED */
803 static int
804 hdr_cons(void *vbuf, void *unused, int kmflag)
805 {
806 arc_buf_hdr_t *buf = vbuf;
807
808 bzero(buf, sizeof (arc_buf_hdr_t));
809 refcount_create(&buf->b_refcnt);
810 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
811 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
812 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
813
814 return (0);
815 }
816
817 /* ARGSUSED */
818 static int
819 buf_cons(void *vbuf, void *unused, int kmflag)
820 {
821 arc_buf_t *buf = vbuf;
822
823 bzero(buf, sizeof (arc_buf_t));
824 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
825 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
826
827 return (0);
828 }
829
830 /*
831 * Destructor callback - called when a cached buf is
832 * no longer required.
833 */
834 /* ARGSUSED */
835 static void
836 hdr_dest(void *vbuf, void *unused)
837 {
838 arc_buf_hdr_t *buf = vbuf;
839
840 ASSERT(BUF_EMPTY(buf));
841 refcount_destroy(&buf->b_refcnt);
842 cv_destroy(&buf->b_cv);
843 mutex_destroy(&buf->b_freeze_lock);
844 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
845 }
846
847 /* ARGSUSED */
848 static void
849 buf_dest(void *vbuf, void *unused)
850 {
851 arc_buf_t *buf = vbuf;
852
853 mutex_destroy(&buf->b_evict_lock);
854 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
855 }
856
857 /*
858 * Reclaim callback -- invoked when memory is low.
859 */
860 /* ARGSUSED */
861 static void
862 hdr_recl(void *unused)
863 {
864 dprintf("hdr_recl called\n");
865 /*
866 * umem calls the reclaim func when we destroy the buf cache,
867 * which is after we do arc_fini().
868 */
869 if (!arc_dead)
870 cv_signal(&arc_reclaim_thr_cv);
871 }
872
873 static void
874 buf_init(void)
875 {
876 uint64_t *ct;
877 uint64_t hsize = 1ULL << 12;
878 int i, j;
879
880 /*
881 * The hash table is big enough to fill all of physical memory
882 * with an average 64K block size. The table will take up
883 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
884 */
885 while (hsize * 65536 < physmem * PAGESIZE)
886 hsize <<= 1;
887 retry:
888 buf_hash_table.ht_mask = hsize - 1;
889 buf_hash_table.ht_table =
890 kmem_zalloc(hsize * sizeof (struct ht_table), KM_NOSLEEP);
891 if (buf_hash_table.ht_table == NULL) {
892 ASSERT(hsize > (1ULL << 8));
893 hsize >>= 1;
894 goto retry;
895 }
896
897 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
898 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
899 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
900 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
901
902 for (i = 0; i < 256; i++)
903 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
904 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
905
906 for (i = 0; i < hsize; i++) {
907 mutex_init(&buf_hash_table.ht_table[i].lock,
908 NULL, MUTEX_DEFAULT, NULL);
909 }
910 }
911
912 #define ARC_MINTIME (hz>>4) /* 62 ms */
913
914 static void
915 arc_cksum_verify(arc_buf_t *buf)
916 {
917 zio_cksum_t zc;
918
919 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
920 return;
921
922 mutex_enter(&buf->b_hdr->b_freeze_lock);
923 if (buf->b_hdr->b_freeze_cksum == NULL ||
924 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
925 mutex_exit(&buf->b_hdr->b_freeze_lock);
926 return;
927 }
928 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
929 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
930 panic("buffer modified while frozen!");
931 mutex_exit(&buf->b_hdr->b_freeze_lock);
932 }
933
934 static int
935 arc_cksum_equal(arc_buf_t *buf)
936 {
937 zio_cksum_t zc;
938 int equal;
939
940 mutex_enter(&buf->b_hdr->b_freeze_lock);
941 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
942 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
943 mutex_exit(&buf->b_hdr->b_freeze_lock);
944
945 return (equal);
946 }
947
948 static void
949 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
950 {
951 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
952 return;
953
954 mutex_enter(&buf->b_hdr->b_freeze_lock);
955 if (buf->b_hdr->b_freeze_cksum != NULL) {
956 mutex_exit(&buf->b_hdr->b_freeze_lock);
957 return;
958 }
959 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
960 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
961 buf->b_hdr->b_freeze_cksum);
962 mutex_exit(&buf->b_hdr->b_freeze_lock);
963 arc_buf_watch(buf);
964 }
965
966 #ifndef _KERNEL
967 typedef struct procctl {
968 long cmd;
969 prwatch_t prwatch;
970 } procctl_t;
971 #endif
972
973 /* ARGSUSED */
974 static void
975 arc_buf_unwatch(arc_buf_t *buf)
976 {
977 #ifndef _KERNEL
978 if (arc_watch) {
979 int result;
980 procctl_t ctl;
981 ctl.cmd = PCWATCH;
982 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
983 ctl.prwatch.pr_size = 0;
984 ctl.prwatch.pr_wflags = 0;
985 result = write(arc_procfd, &ctl, sizeof (ctl));
986 ASSERT3U(result, ==, sizeof (ctl));
987 }
988 #endif
989 }
990
991 /* ARGSUSED */
992 static void
993 arc_buf_watch(arc_buf_t *buf)
994 {
995 #ifndef _KERNEL
996 if (arc_watch) {
997 int result;
998 procctl_t ctl;
999 ctl.cmd = PCWATCH;
1000 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1001 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1002 ctl.prwatch.pr_wflags = WA_WRITE;
1003 result = write(arc_procfd, &ctl, sizeof (ctl));
1004 ASSERT3U(result, ==, sizeof (ctl));
1005 }
1006 #endif
1007 }
1008
1009 void
1010 arc_buf_thaw(arc_buf_t *buf)
1011 {
1012 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1013 if (buf->b_hdr->b_state != arc_anon)
1014 panic("modifying non-anon buffer!");
1015 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1016 panic("modifying buffer while i/o in progress!");
1017 arc_cksum_verify(buf);
1018 }
1019
1020 mutex_enter(&buf->b_hdr->b_freeze_lock);
1021 if (buf->b_hdr->b_freeze_cksum != NULL) {
1022 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1023 buf->b_hdr->b_freeze_cksum = NULL;
1024 }
1025
1026 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1027 if (buf->b_hdr->b_thawed)
1028 kmem_free(buf->b_hdr->b_thawed, 1);
1029 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1030 }
1031
1032 mutex_exit(&buf->b_hdr->b_freeze_lock);
1033
1034 arc_buf_unwatch(buf);
1035 }
1036
1037 void
1038 arc_buf_freeze(arc_buf_t *buf)
1039 {
1040 kmutex_t *hash_lock;
1041
1042 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1043 return;
1044
1045 hash_lock = HDR_LOCK(buf->b_hdr);
1046 mutex_enter(hash_lock);
1047
1048 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1049 buf->b_hdr->b_state == arc_anon);
1050 arc_cksum_compute(buf, B_FALSE);
1051 mutex_exit(hash_lock);
1052
1053 }
1054
1055 static void
1056 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1057 {
1058 ASSERT(MUTEX_HELD(hash_lock));
1059
1060 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1061 (ab->b_state != arc_anon)) {
1062 uint64_t delta = ab->b_size * ab->b_datacnt;
1063 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1064 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1065
1066 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1067 mutex_enter(&ab->b_state->arcs_mtx);
1068 ASSERT(list_link_active(&ab->b_arc_node));
1069 list_remove(list, ab);
1070 if (GHOST_STATE(ab->b_state)) {
1071 ASSERT0(ab->b_datacnt);
1072 ASSERT3P(ab->b_buf, ==, NULL);
1073 delta = ab->b_size;
1074 }
1075 ASSERT(delta > 0);
1076 ASSERT3U(*size, >=, delta);
1077 atomic_add_64(size, -delta);
1078 mutex_exit(&ab->b_state->arcs_mtx);
1079 /* remove the prefetch flag if we get a reference */
1080 if (ab->b_flags & ARC_PREFETCH)
1081 ab->b_flags &= ~ARC_PREFETCH;
1082 }
1083 }
1084
1085 static int
1086 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1087 {
1088 int cnt;
1089 arc_state_t *state = ab->b_state;
1090
1091 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1092 ASSERT(!GHOST_STATE(state));
1093
1094 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1095 (state != arc_anon)) {
1096 uint64_t *size = &state->arcs_lsize[ab->b_type];
1097
1098 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1099 mutex_enter(&state->arcs_mtx);
1100 ASSERT(!list_link_active(&ab->b_arc_node));
1101 list_insert_head(&state->arcs_list[ab->b_type], ab);
1102 ASSERT(ab->b_datacnt > 0);
1103 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1104 mutex_exit(&state->arcs_mtx);
1105 }
1106 return (cnt);
1107 }
1108
1109 /*
1110 * Move the supplied buffer to the indicated state. The mutex
1111 * for the buffer must be held by the caller.
1112 */
1113 static void
1114 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1115 {
1116 arc_state_t *old_state = ab->b_state;
1117 int64_t refcnt = refcount_count(&ab->b_refcnt);
1118 uint64_t from_delta, to_delta;
1119
1120 ASSERT(MUTEX_HELD(hash_lock));
1121 ASSERT(new_state != old_state);
1122 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1123 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1124 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1125
1126 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1127
1128 /*
1129 * If this buffer is evictable, transfer it from the
1130 * old state list to the new state list.
1131 */
1132 if (refcnt == 0) {
1133 if (old_state != arc_anon) {
1134 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1135 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1136
1137 if (use_mutex)
1138 mutex_enter(&old_state->arcs_mtx);
1139
1140 ASSERT(list_link_active(&ab->b_arc_node));
1141 list_remove(&old_state->arcs_list[ab->b_type], ab);
1142
1143 /*
1144 * If prefetching out of the ghost cache,
1145 * we will have a non-zero datacnt.
1146 */
1147 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1148 /* ghost elements have a ghost size */
1149 ASSERT(ab->b_buf == NULL);
1150 from_delta = ab->b_size;
1151 }
1152 ASSERT3U(*size, >=, from_delta);
1153 atomic_add_64(size, -from_delta);
1154
1155 if (use_mutex)
1156 mutex_exit(&old_state->arcs_mtx);
1157 }
1158 if (new_state != arc_anon) {
1159 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1160 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1161
1162 if (use_mutex)
1163 mutex_enter(&new_state->arcs_mtx);
1164
1165 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1166
1167 /* ghost elements have a ghost size */
1168 if (GHOST_STATE(new_state)) {
1169 ASSERT(ab->b_datacnt == 0);
1170 ASSERT(ab->b_buf == NULL);
1171 to_delta = ab->b_size;
1172 }
1173 atomic_add_64(size, to_delta);
1174
1175 if (use_mutex)
1176 mutex_exit(&new_state->arcs_mtx);
1177 }
1178 }
1179
1180 ASSERT(!BUF_EMPTY(ab));
1181 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1182 buf_hash_remove(ab);
1183
1184 /* adjust state sizes */
1185 if (to_delta)
1186 atomic_add_64(&new_state->arcs_size, to_delta);
1187 if (from_delta) {
1188 ASSERT3U(old_state->arcs_size, >=, from_delta);
1189 atomic_add_64(&old_state->arcs_size, -from_delta);
1190 }
1191 ab->b_state = new_state;
1192
1193 /* adjust l2arc hdr stats */
1194 if (new_state == arc_l2c_only)
1195 l2arc_hdr_stat_add();
1196 else if (old_state == arc_l2c_only)
1197 l2arc_hdr_stat_remove();
1198 }
1199
1200 void
1201 arc_space_consume(uint64_t space, arc_space_type_t type)
1202 {
1203 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1204
1205 switch (type) {
1206 case ARC_SPACE_DATA:
1207 ARCSTAT_INCR(arcstat_data_size, space);
1208 break;
1209 case ARC_SPACE_OTHER:
1210 ARCSTAT_INCR(arcstat_other_size, space);
1211 break;
1212 case ARC_SPACE_HDRS:
1213 ARCSTAT_INCR(arcstat_hdr_size, space);
1214 break;
1215 case ARC_SPACE_L2HDRS:
1216 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1217 break;
1218 }
1219
1220 ARCSTAT_INCR(arcstat_meta_used, space);
1221 atomic_add_64(&arc_size, space);
1222 }
1223
1224 void
1225 arc_space_return(uint64_t space, arc_space_type_t type)
1226 {
1227 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1228
1229 switch (type) {
1230 case ARC_SPACE_DATA:
1231 ARCSTAT_INCR(arcstat_data_size, -space);
1232 break;
1233 case ARC_SPACE_OTHER:
1234 ARCSTAT_INCR(arcstat_other_size, -space);
1235 break;
1236 case ARC_SPACE_HDRS:
1237 ARCSTAT_INCR(arcstat_hdr_size, -space);
1238 break;
1239 case ARC_SPACE_L2HDRS:
1240 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1241 break;
1242 }
1243
1244 ASSERT(arc_meta_used >= space);
1245 if (arc_meta_max < arc_meta_used)
1246 arc_meta_max = arc_meta_used;
1247 ARCSTAT_INCR(arcstat_meta_used, -space);
1248 ASSERT(arc_size >= space);
1249 atomic_add_64(&arc_size, -space);
1250 }
1251
1252 void *
1253 arc_data_buf_alloc(uint64_t size)
1254 {
1255 if (arc_evict_needed(ARC_BUFC_DATA))
1256 cv_signal(&arc_reclaim_thr_cv);
1257 atomic_add_64(&arc_size, size);
1258 return (zio_data_buf_alloc(size));
1259 }
1260
1261 void
1262 arc_data_buf_free(void *buf, uint64_t size)
1263 {
1264 zio_data_buf_free(buf, size);
1265 ASSERT(arc_size >= size);
1266 atomic_add_64(&arc_size, -size);
1267 }
1268
1269 arc_buf_t *
1270 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1271 {
1272 arc_buf_hdr_t *hdr;
1273 arc_buf_t *buf;
1274
1275 ASSERT3U(size, >, 0);
1276 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1277 ASSERT(BUF_EMPTY(hdr));
1278 hdr->b_size = size;
1279 hdr->b_type = type;
1280 hdr->b_spa = spa_load_guid(spa);
1281 hdr->b_state = arc_anon;
1282 hdr->b_arc_access = 0;
1283 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1284 buf->b_hdr = hdr;
1285 buf->b_data = NULL;
1286 buf->b_efunc = NULL;
1287 buf->b_private = NULL;
1288 buf->b_next = NULL;
1289 hdr->b_buf = buf;
1290 arc_get_data_buf(buf);
1291 hdr->b_datacnt = 1;
1292 hdr->b_flags = 0;
1293 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1294 (void) refcount_add(&hdr->b_refcnt, tag);
1295
1296 return (buf);
1297 }
1298
1299 static char *arc_onloan_tag = "onloan";
1300
1301 /*
1302 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1303 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1304 * buffers must be returned to the arc before they can be used by the DMU or
1305 * freed.
1306 */
1307 arc_buf_t *
1308 arc_loan_buf(spa_t *spa, int size)
1309 {
1310 arc_buf_t *buf;
1311
1312 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1313
1314 atomic_add_64(&arc_loaned_bytes, size);
1315 return (buf);
1316 }
1317
1318 /*
1319 * Return a loaned arc buffer to the arc.
1320 */
1321 void
1322 arc_return_buf(arc_buf_t *buf, void *tag)
1323 {
1324 arc_buf_hdr_t *hdr = buf->b_hdr;
1325
1326 ASSERT(buf->b_data != NULL);
1327 (void) refcount_add(&hdr->b_refcnt, tag);
1328 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1329
1330 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1331 }
1332
1333 /* Detach an arc_buf from a dbuf (tag) */
1334 void
1335 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1336 {
1337 arc_buf_hdr_t *hdr;
1338
1339 ASSERT(buf->b_data != NULL);
1340 hdr = buf->b_hdr;
1341 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1342 (void) refcount_remove(&hdr->b_refcnt, tag);
1343 buf->b_efunc = NULL;
1344 buf->b_private = NULL;
1345
1346 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1347 }
1348
1349 static arc_buf_t *
1350 arc_buf_clone(arc_buf_t *from)
1351 {
1352 arc_buf_t *buf;
1353 arc_buf_hdr_t *hdr = from->b_hdr;
1354 uint64_t size = hdr->b_size;
1355
1356 ASSERT(hdr->b_state != arc_anon);
1357
1358 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1359 buf->b_hdr = hdr;
1360 buf->b_data = NULL;
1361 buf->b_efunc = NULL;
1362 buf->b_private = NULL;
1363 buf->b_next = hdr->b_buf;
1364 hdr->b_buf = buf;
1365 arc_get_data_buf(buf);
1366 bcopy(from->b_data, buf->b_data, size);
1367
1368 /*
1369 * This buffer already exists in the arc so create a duplicate
1370 * copy for the caller. If the buffer is associated with user data
1371 * then track the size and number of duplicates. These stats will be
1372 * updated as duplicate buffers are created and destroyed.
1373 */
1374 if (hdr->b_type == ARC_BUFC_DATA) {
1375 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1376 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1377 }
1378 hdr->b_datacnt += 1;
1379 return (buf);
1380 }
1381
1382 void
1383 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1384 {
1385 arc_buf_hdr_t *hdr;
1386 kmutex_t *hash_lock;
1387
1388 /*
1389 * Check to see if this buffer is evicted. Callers
1390 * must verify b_data != NULL to know if the add_ref
1391 * was successful.
1392 */
1393 mutex_enter(&buf->b_evict_lock);
1394 if (buf->b_data == NULL) {
1395 mutex_exit(&buf->b_evict_lock);
1396 return;
1397 }
1398 hash_lock = HDR_LOCK(buf->b_hdr);
1399 mutex_enter(hash_lock);
1400 hdr = buf->b_hdr;
1401 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1402 mutex_exit(&buf->b_evict_lock);
1403
1404 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1405 add_reference(hdr, hash_lock, tag);
1406 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1407 arc_access(hdr, hash_lock);
1408 mutex_exit(hash_lock);
1409 ARCSTAT_BUMP(arcstat_hits);
1410 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1411 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1412 data, metadata, hits);
1413 }
1414
1415 /*
1416 * Free the arc data buffer. If it is an l2arc write in progress,
1417 * the buffer is placed on l2arc_free_on_write to be freed later.
1418 */
1419 static void
1420 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1421 {
1422 arc_buf_hdr_t *hdr = buf->b_hdr;
1423
1424 if (HDR_L2_WRITING(hdr)) {
1425 l2arc_data_free_t *df;
1426 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1427 df->l2df_data = buf->b_data;
1428 df->l2df_size = hdr->b_size;
1429 df->l2df_func = free_func;
1430 mutex_enter(&l2arc_free_on_write_mtx);
1431 list_insert_head(l2arc_free_on_write, df);
1432 mutex_exit(&l2arc_free_on_write_mtx);
1433 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1434 } else {
1435 free_func(buf->b_data, hdr->b_size);
1436 }
1437 }
1438
1439 static void
1440 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1441 {
1442 arc_buf_t **bufp;
1443
1444 /* free up data associated with the buf */
1445 if (buf->b_data) {
1446 arc_state_t *state = buf->b_hdr->b_state;
1447 uint64_t size = buf->b_hdr->b_size;
1448 arc_buf_contents_t type = buf->b_hdr->b_type;
1449
1450 arc_cksum_verify(buf);
1451 arc_buf_unwatch(buf);
1452
1453 if (!recycle) {
1454 if (type == ARC_BUFC_METADATA) {
1455 arc_buf_data_free(buf, zio_buf_free);
1456 arc_space_return(size, ARC_SPACE_DATA);
1457 } else {
1458 ASSERT(type == ARC_BUFC_DATA);
1459 arc_buf_data_free(buf, zio_data_buf_free);
1460 ARCSTAT_INCR(arcstat_data_size, -size);
1461 atomic_add_64(&arc_size, -size);
1462 }
1463 }
1464 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1465 uint64_t *cnt = &state->arcs_lsize[type];
1466
1467 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1468 ASSERT(state != arc_anon);
1469
1470 ASSERT3U(*cnt, >=, size);
1471 atomic_add_64(cnt, -size);
1472 }
1473 ASSERT3U(state->arcs_size, >=, size);
1474 atomic_add_64(&state->arcs_size, -size);
1475 buf->b_data = NULL;
1476
1477 /*
1478 * If we're destroying a duplicate buffer make sure
1479 * that the appropriate statistics are updated.
1480 */
1481 if (buf->b_hdr->b_datacnt > 1 &&
1482 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1483 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1484 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1485 }
1486 ASSERT(buf->b_hdr->b_datacnt > 0);
1487 buf->b_hdr->b_datacnt -= 1;
1488 }
1489
1490 /* only remove the buf if requested */
1491 if (!all)
1492 return;
1493
1494 /* remove the buf from the hdr list */
1495 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1496 continue;
1497 *bufp = buf->b_next;
1498 buf->b_next = NULL;
1499
1500 ASSERT(buf->b_efunc == NULL);
1501
1502 /* clean up the buf */
1503 buf->b_hdr = NULL;
1504 kmem_cache_free(buf_cache, buf);
1505 }
1506
1507 static void
1508 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1509 {
1510 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1511 ASSERT3P(hdr->b_state, ==, arc_anon);
1512 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1513 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1514
1515 if (l2hdr != NULL) {
1516 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1517 /*
1518 * To prevent arc_free() and l2arc_evict() from
1519 * attempting to free the same buffer at the same time,
1520 * a FREE_IN_PROGRESS flag is given to arc_free() to
1521 * give it priority. l2arc_evict() can't destroy this
1522 * header while we are waiting on l2arc_buflist_mtx.
1523 *
1524 * The hdr may be removed from l2ad_buflist before we
1525 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1526 */
1527 if (!buflist_held) {
1528 mutex_enter(&l2arc_buflist_mtx);
1529 l2hdr = hdr->b_l2hdr;
1530 }
1531
1532 if (l2hdr != NULL) {
1533 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1534 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1535 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1536 if (hdr->b_state == arc_l2c_only)
1537 l2arc_hdr_stat_remove();
1538 hdr->b_l2hdr = NULL;
1539 }
1540
1541 if (!buflist_held)
1542 mutex_exit(&l2arc_buflist_mtx);
1543 }
1544
1545 if (!BUF_EMPTY(hdr)) {
1546 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1547 buf_discard_identity(hdr);
1548 }
1549 while (hdr->b_buf) {
1550 arc_buf_t *buf = hdr->b_buf;
1551
1552 if (buf->b_efunc) {
1553 mutex_enter(&arc_eviction_mtx);
1554 mutex_enter(&buf->b_evict_lock);
1555 ASSERT(buf->b_hdr != NULL);
1556 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1557 hdr->b_buf = buf->b_next;
1558 buf->b_hdr = &arc_eviction_hdr;
1559 buf->b_next = arc_eviction_list;
1560 arc_eviction_list = buf;
1561 mutex_exit(&buf->b_evict_lock);
1562 mutex_exit(&arc_eviction_mtx);
1563 } else {
1564 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1565 }
1566 }
1567 if (hdr->b_freeze_cksum != NULL) {
1568 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1569 hdr->b_freeze_cksum = NULL;
1570 }
1571 if (hdr->b_thawed) {
1572 kmem_free(hdr->b_thawed, 1);
1573 hdr->b_thawed = NULL;
1574 }
1575
1576 ASSERT(!list_link_active(&hdr->b_arc_node));
1577 ASSERT3P(hdr->b_hash_next, ==, NULL);
1578 ASSERT3P(hdr->b_acb, ==, NULL);
1579 kmem_cache_free(hdr_cache, hdr);
1580 }
1581
1582 void
1583 arc_buf_free(arc_buf_t *buf, void *tag)
1584 {
1585 arc_buf_hdr_t *hdr = buf->b_hdr;
1586 int hashed = hdr->b_state != arc_anon;
1587
1588 ASSERT(buf->b_efunc == NULL);
1589 ASSERT(buf->b_data != NULL);
1590
1591 if (hashed) {
1592 kmutex_t *hash_lock = HDR_LOCK(hdr);
1593
1594 mutex_enter(hash_lock);
1595 hdr = buf->b_hdr;
1596 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1597
1598 (void) remove_reference(hdr, hash_lock, tag);
1599 if (hdr->b_datacnt > 1) {
1600 arc_buf_destroy(buf, FALSE, TRUE);
1601 } else {
1602 ASSERT(buf == hdr->b_buf);
1603 ASSERT(buf->b_efunc == NULL);
1604 hdr->b_flags |= ARC_BUF_AVAILABLE;
1605 }
1606 mutex_exit(hash_lock);
1607 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1608 int destroy_hdr;
1609 /*
1610 * We are in the middle of an async write. Don't destroy
1611 * this buffer unless the write completes before we finish
1612 * decrementing the reference count.
1613 */
1614 mutex_enter(&arc_eviction_mtx);
1615 (void) remove_reference(hdr, NULL, tag);
1616 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1617 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1618 mutex_exit(&arc_eviction_mtx);
1619 if (destroy_hdr)
1620 arc_hdr_destroy(hdr);
1621 } else {
1622 if (remove_reference(hdr, NULL, tag) > 0)
1623 arc_buf_destroy(buf, FALSE, TRUE);
1624 else
1625 arc_hdr_destroy(hdr);
1626 }
1627 }
1628
1629 boolean_t
1630 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1631 {
1632 arc_buf_hdr_t *hdr = buf->b_hdr;
1633 kmutex_t *hash_lock = HDR_LOCK(hdr);
1634 boolean_t no_callback = (buf->b_efunc == NULL);
1635
1636 if (hdr->b_state == arc_anon) {
1637 ASSERT(hdr->b_datacnt == 1);
1638 arc_buf_free(buf, tag);
1639 return (no_callback);
1640 }
1641
1642 mutex_enter(hash_lock);
1643 hdr = buf->b_hdr;
1644 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1645 ASSERT(hdr->b_state != arc_anon);
1646 ASSERT(buf->b_data != NULL);
1647
1648 (void) remove_reference(hdr, hash_lock, tag);
1649 if (hdr->b_datacnt > 1) {
1650 if (no_callback)
1651 arc_buf_destroy(buf, FALSE, TRUE);
1652 } else if (no_callback) {
1653 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1654 ASSERT(buf->b_efunc == NULL);
1655 hdr->b_flags |= ARC_BUF_AVAILABLE;
1656 }
1657 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1658 refcount_is_zero(&hdr->b_refcnt));
1659 mutex_exit(hash_lock);
1660 return (no_callback);
1661 }
1662
1663 int
1664 arc_buf_size(arc_buf_t *buf)
1665 {
1666 return (buf->b_hdr->b_size);
1667 }
1668
1669 /*
1670 * Called from the DMU to determine if the current buffer should be
1671 * evicted. In order to ensure proper locking, the eviction must be initiated
1672 * from the DMU. Return true if the buffer is associated with user data and
1673 * duplicate buffers still exist.
1674 */
1675 boolean_t
1676 arc_buf_eviction_needed(arc_buf_t *buf)
1677 {
1678 arc_buf_hdr_t *hdr;
1679 boolean_t evict_needed = B_FALSE;
1680
1681 if (zfs_disable_dup_eviction)
1682 return (B_FALSE);
1683
1684 mutex_enter(&buf->b_evict_lock);
1685 hdr = buf->b_hdr;
1686 if (hdr == NULL) {
1687 /*
1688 * We are in arc_do_user_evicts(); let that function
1689 * perform the eviction.
1690 */
1691 ASSERT(buf->b_data == NULL);
1692 mutex_exit(&buf->b_evict_lock);
1693 return (B_FALSE);
1694 } else if (buf->b_data == NULL) {
1695 /*
1696 * We have already been added to the arc eviction list;
1697 * recommend eviction.
1698 */
1699 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1700 mutex_exit(&buf->b_evict_lock);
1701 return (B_TRUE);
1702 }
1703
1704 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1705 evict_needed = B_TRUE;
1706
1707 mutex_exit(&buf->b_evict_lock);
1708 return (evict_needed);
1709 }
1710
1711 /*
1712 * Evict buffers from list until we've removed the specified number of
1713 * bytes. Move the removed buffers to the appropriate evict state.
1714 * If the recycle flag is set, then attempt to "recycle" a buffer:
1715 * - look for a buffer to evict that is `bytes' long.
1716 * - return the data block from this buffer rather than freeing it.
1717 * This flag is used by callers that are trying to make space for a
1718 * new buffer in a full arc cache.
1719 *
1720 * This function makes a "best effort". It skips over any buffers
1721 * it can't get a hash_lock on, and so may not catch all candidates.
1722 * It may also return without evicting as much space as requested.
1723 */
1724 static void *
1725 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1726 arc_buf_contents_t type)
1727 {
1728 arc_state_t *evicted_state;
1729 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1730 arc_buf_hdr_t *ab, *ab_prev = NULL;
1731 list_t *list = &state->arcs_list[type];
1732 kmutex_t *hash_lock;
1733 boolean_t have_lock;
1734 void *stolen = NULL;
1735
1736 ASSERT(state == arc_mru || state == arc_mfu);
1737
1738 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1739
1740 mutex_enter(&state->arcs_mtx);
1741 mutex_enter(&evicted_state->arcs_mtx);
1742
1743 for (ab = list_tail(list); ab; ab = ab_prev) {
1744 ab_prev = list_prev(list, ab);
1745 /* prefetch buffers have a minimum lifespan */
1746 if (HDR_IO_IN_PROGRESS(ab) ||
1747 (spa && ab->b_spa != spa) ||
1748 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1749 ddi_get_lbolt() - ab->b_arc_access <
1750 arc_min_prefetch_lifespan)) {
1751 skipped++;
1752 continue;
1753 }
1754 /* "lookahead" for better eviction candidate */
1755 if (recycle && ab->b_size != bytes &&
1756 ab_prev && ab_prev->b_size == bytes)
1757 continue;
1758 hash_lock = HDR_LOCK(ab);
1759 have_lock = MUTEX_HELD(hash_lock);
1760 if (have_lock || mutex_tryenter(hash_lock)) {
1761 ASSERT0(refcount_count(&ab->b_refcnt));
1762 ASSERT(ab->b_datacnt > 0);
1763 while (ab->b_buf) {
1764 arc_buf_t *buf = ab->b_buf;
1765 if (!mutex_tryenter(&buf->b_evict_lock)) {
1766 missed += 1;
1767 break;
1768 }
1769 if (buf->b_data) {
1770 bytes_evicted += ab->b_size;
1771 if (recycle && ab->b_type == type &&
1772 ab->b_size == bytes &&
1773 !HDR_L2_WRITING(ab)) {
1774 stolen = buf->b_data;
1775 recycle = FALSE;
1776 }
1777 }
1778 if (buf->b_efunc) {
1779 mutex_enter(&arc_eviction_mtx);
1780 arc_buf_destroy(buf,
1781 buf->b_data == stolen, FALSE);
1782 ab->b_buf = buf->b_next;
1783 buf->b_hdr = &arc_eviction_hdr;
1784 buf->b_next = arc_eviction_list;
1785 arc_eviction_list = buf;
1786 mutex_exit(&arc_eviction_mtx);
1787 mutex_exit(&buf->b_evict_lock);
1788 } else {
1789 mutex_exit(&buf->b_evict_lock);
1790 arc_buf_destroy(buf,
1791 buf->b_data == stolen, TRUE);
1792 }
1793 }
1794
1795 if (ab->b_l2hdr) {
1796 ARCSTAT_INCR(arcstat_evict_l2_cached,
1797 ab->b_size);
1798 } else {
1799 if (l2arc_write_eligible(ab->b_spa, ab)) {
1800 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1801 ab->b_size);
1802 } else {
1803 ARCSTAT_INCR(
1804 arcstat_evict_l2_ineligible,
1805 ab->b_size);
1806 }
1807 }
1808
1809 if (ab->b_datacnt == 0) {
1810 arc_change_state(evicted_state, ab, hash_lock);
1811 ASSERT(HDR_IN_HASH_TABLE(ab));
1812 ab->b_flags |= ARC_IN_HASH_TABLE;
1813 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1814 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1815 }
1816 if (!have_lock)
1817 mutex_exit(hash_lock);
1818 if (bytes >= 0 && bytes_evicted >= bytes)
1819 break;
1820 } else {
1821 missed += 1;
1822 }
1823 }
1824
1825 mutex_exit(&evicted_state->arcs_mtx);
1826 mutex_exit(&state->arcs_mtx);
1827
1828 if (bytes_evicted < bytes)
1829 dprintf("only evicted %lld bytes from %x",
1830 (longlong_t)bytes_evicted, state);
1831
1832 if (skipped)
1833 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1834
1835 if (missed)
1836 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1837
1838 /*
1839 * We have just evicted some data into the ghost state, make
1840 * sure we also adjust the ghost state size if necessary.
1841 */
1842 if (arc_no_grow &&
1843 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1844 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1845 arc_mru_ghost->arcs_size - arc_c;
1846
1847 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1848 int64_t todelete =
1849 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1850 arc_evict_ghost(arc_mru_ghost, NULL, todelete);
1851 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1852 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1853 arc_mru_ghost->arcs_size +
1854 arc_mfu_ghost->arcs_size - arc_c);
1855 arc_evict_ghost(arc_mfu_ghost, NULL, todelete);
1856 }
1857 }
1858
1859 return (stolen);
1860 }
1861
1862 /*
1863 * Remove buffers from list until we've removed the specified number of
1864 * bytes. Destroy the buffers that are removed.
1865 */
1866 static void
1867 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1868 {
1869 arc_buf_hdr_t *ab, *ab_prev;
1870 arc_buf_hdr_t marker = { 0 };
1871 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1872 kmutex_t *hash_lock;
1873 uint64_t bytes_deleted = 0;
1874 uint64_t bufs_skipped = 0;
1875
1876 ASSERT(GHOST_STATE(state));
1877 top:
1878 mutex_enter(&state->arcs_mtx);
1879 for (ab = list_tail(list); ab; ab = ab_prev) {
1880 ab_prev = list_prev(list, ab);
1881 if (spa && ab->b_spa != spa)
1882 continue;
1883
1884 /* ignore markers */
1885 if (ab->b_spa == 0)
1886 continue;
1887
1888 hash_lock = HDR_LOCK(ab);
1889 /* caller may be trying to modify this buffer, skip it */
1890 if (MUTEX_HELD(hash_lock))
1891 continue;
1892 if (mutex_tryenter(hash_lock)) {
1893 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1894 ASSERT(ab->b_buf == NULL);
1895 ARCSTAT_BUMP(arcstat_deleted);
1896 bytes_deleted += ab->b_size;
1897
1898 if (ab->b_l2hdr != NULL) {
1899 /*
1900 * This buffer is cached on the 2nd Level ARC;
1901 * don't destroy the header.
1902 */
1903 arc_change_state(arc_l2c_only, ab, hash_lock);
1904 mutex_exit(hash_lock);
1905 } else {
1906 arc_change_state(arc_anon, ab, hash_lock);
1907 mutex_exit(hash_lock);
1908 arc_hdr_destroy(ab);
1909 }
1910
1911 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1912 if (bytes >= 0 && bytes_deleted >= bytes)
1913 break;
1914 } else if (bytes < 0) {
1915 /*
1916 * Insert a list marker and then wait for the
1917 * hash lock to become available. Once its
1918 * available, restart from where we left off.
1919 */
1920 list_insert_after(list, ab, &marker);
1921 mutex_exit(&state->arcs_mtx);
1922 mutex_enter(hash_lock);
1923 mutex_exit(hash_lock);
1924 mutex_enter(&state->arcs_mtx);
1925 ab_prev = list_prev(list, &marker);
1926 list_remove(list, &marker);
1927 } else
1928 bufs_skipped += 1;
1929 }
1930 mutex_exit(&state->arcs_mtx);
1931
1932 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1933 (bytes < 0 || bytes_deleted < bytes)) {
1934 list = &state->arcs_list[ARC_BUFC_METADATA];
1935 goto top;
1936 }
1937
1938 if (bufs_skipped) {
1939 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1940 ASSERT(bytes >= 0);
1941 }
1942
1943 if (bytes_deleted < bytes)
1944 dprintf("only deleted %lld bytes from %p",
1945 (longlong_t)bytes_deleted, state);
1946 }
1947
1948 static void
1949 arc_adjust(void)
1950 {
1951 int64_t adjustment, delta;
1952
1953 /*
1954 * Adjust MRU size
1955 */
1956
1957 adjustment = MIN((int64_t)(arc_size - arc_c),
1958 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
1959 arc_p));
1960
1961 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1962 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
1963 (void) arc_evict(arc_mru, NULL, delta, FALSE, ARC_BUFC_DATA);
1964 adjustment -= delta;
1965 }
1966
1967 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1968 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
1969 (void) arc_evict(arc_mru, NULL, delta, FALSE,
1970 ARC_BUFC_METADATA);
1971 }
1972
1973 /*
1974 * Adjust MFU size
1975 */
1976
1977 adjustment = arc_size - arc_c;
1978
1979 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1980 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
1981 (void) arc_evict(arc_mfu, NULL, delta, FALSE, ARC_BUFC_DATA);
1982 adjustment -= delta;
1983 }
1984
1985 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1986 int64_t delta = MIN(adjustment,
1987 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
1988 (void) arc_evict(arc_mfu, NULL, delta, FALSE,
1989 ARC_BUFC_METADATA);
1990 }
1991
1992 /*
1993 * Adjust ghost lists
1994 */
1995
1996 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
1997
1998 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
1999 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2000 arc_evict_ghost(arc_mru_ghost, NULL, delta);
2001 }
2002
2003 adjustment =
2004 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2005
2006 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2007 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2008 arc_evict_ghost(arc_mfu_ghost, NULL, delta);
2009 }
2010 }
2011
2012 static void
2013 arc_do_user_evicts(void)
2014 {
2015 mutex_enter(&arc_eviction_mtx);
2016 while (arc_eviction_list != NULL) {
2017 arc_buf_t *buf = arc_eviction_list;
2018 arc_eviction_list = buf->b_next;
2019 mutex_enter(&buf->b_evict_lock);
2020 buf->b_hdr = NULL;
2021 mutex_exit(&buf->b_evict_lock);
2022 mutex_exit(&arc_eviction_mtx);
2023
2024 if (buf->b_efunc != NULL)
2025 VERIFY(buf->b_efunc(buf) == 0);
2026
2027 buf->b_efunc = NULL;
2028 buf->b_private = NULL;
2029 kmem_cache_free(buf_cache, buf);
2030 mutex_enter(&arc_eviction_mtx);
2031 }
2032 mutex_exit(&arc_eviction_mtx);
2033 }
2034
2035 /*
2036 * Flush all *evictable* data from the cache for the given spa.
2037 * NOTE: this will not touch "active" (i.e. referenced) data.
2038 */
2039 void
2040 arc_flush(spa_t *spa)
2041 {
2042 uint64_t guid = 0;
2043
2044 if (spa)
2045 guid = spa_load_guid(spa);
2046
2047 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2048 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2049 if (spa)
2050 break;
2051 }
2052 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2053 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2054 if (spa)
2055 break;
2056 }
2057 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2058 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2059 if (spa)
2060 break;
2061 }
2062 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2063 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2064 if (spa)
2065 break;
2066 }
2067
2068 arc_evict_ghost(arc_mru_ghost, guid, -1);
2069 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2070
2071 mutex_enter(&arc_reclaim_thr_lock);
2072 arc_do_user_evicts();
2073 mutex_exit(&arc_reclaim_thr_lock);
2074 ASSERT(spa || arc_eviction_list == NULL);
2075 }
2076
2077 void
2078 arc_shrink(void)
2079 {
2080 if (arc_c > arc_c_min) {
2081 uint64_t to_free;
2082
2083 #ifdef _KERNEL
2084 to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
2085 #else
2086 to_free = arc_c >> arc_shrink_shift;
2087 #endif
2088 if (arc_c > arc_c_min + to_free)
2089 atomic_add_64(&arc_c, -to_free);
2090 else
2091 arc_c = arc_c_min;
2092
2093 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2094 if (arc_c > arc_size)
2095 arc_c = MAX(arc_size, arc_c_min);
2096 if (arc_p > arc_c)
2097 arc_p = (arc_c >> 1);
2098 ASSERT(arc_c >= arc_c_min);
2099 ASSERT((int64_t)arc_p >= 0);
2100 }
2101
2102 if (arc_size > arc_c)
2103 arc_adjust();
2104 }
2105
2106 /*
2107 * Determine if the system is under memory pressure and is asking
2108 * to reclaim memory. A return value of 1 indicates that the system
2109 * is under memory pressure and that the arc should adjust accordingly.
2110 */
2111 static int
2112 arc_reclaim_needed(void)
2113 {
2114 uint64_t extra;
2115
2116 #ifdef _KERNEL
2117
2118 if (needfree)
2119 return (1);
2120
2121 /*
2122 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2123 */
2124 extra = desfree;
2125
2126 /*
2127 * check that we're out of range of the pageout scanner. It starts to
2128 * schedule paging if freemem is less than lotsfree and needfree.
2129 * lotsfree is the high-water mark for pageout, and needfree is the
2130 * number of needed free pages. We add extra pages here to make sure
2131 * the scanner doesn't start up while we're freeing memory.
2132 */
2133 if (freemem < lotsfree + needfree + extra)
2134 return (1);
2135
2136 /*
2137 * check to make sure that swapfs has enough space so that anon
2138 * reservations can still succeed. anon_resvmem() checks that the
2139 * availrmem is greater than swapfs_minfree, and the number of reserved
2140 * swap pages. We also add a bit of extra here just to prevent
2141 * circumstances from getting really dire.
2142 */
2143 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2144 return (1);
2145
2146 #if defined(__i386)
2147 /*
2148 * If we're on an i386 platform, it's possible that we'll exhaust the
2149 * kernel heap space before we ever run out of available physical
2150 * memory. Most checks of the size of the heap_area compare against
2151 * tune.t_minarmem, which is the minimum available real memory that we
2152 * can have in the system. However, this is generally fixed at 25 pages
2153 * which is so low that it's useless. In this comparison, we seek to
2154 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2155 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2156 * free)
2157 */
2158 if (vmem_size(heap_arena, VMEM_FREE) <
2159 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2))
2160 return (1);
2161 #endif
2162
2163 /*
2164 * If zio data pages are being allocated out of a separate heap segment,
2165 * then enforce that the size of available vmem for this arena remains
2166 * above about 1/16th free.
2167 *
2168 * Note: The 1/16th arena free requirement was put in place
2169 * to aggressively evict memory from the arc in order to avoid
2170 * memory fragmentation issues.
2171 */
2172 if (zio_arena != NULL &&
2173 vmem_size(zio_arena, VMEM_FREE) <
2174 (vmem_size(zio_arena, VMEM_ALLOC) >> 4))
2175 return (1);
2176 #else
2177 if (spa_get_random(100) == 0)
2178 return (1);
2179 #endif
2180 return (0);
2181 }
2182
2183 static void
2184 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2185 {
2186 size_t i;
2187 kmem_cache_t *prev_cache = NULL;
2188 kmem_cache_t *prev_data_cache = NULL;
2189 extern kmem_cache_t *zio_buf_cache[];
2190 extern kmem_cache_t *zio_data_buf_cache[];
2191
2192 #ifdef _KERNEL
2193 if (arc_meta_used >= arc_meta_limit) {
2194 /*
2195 * We are exceeding our meta-data cache limit.
2196 * Purge some DNLC entries to release holds on meta-data.
2197 */
2198 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2199 }
2200 #if defined(__i386)
2201 /*
2202 * Reclaim unused memory from all kmem caches.
2203 */
2204 kmem_reap();
2205 #endif
2206 #endif
2207
2208 /*
2209 * An aggressive reclamation will shrink the cache size as well as
2210 * reap free buffers from the arc kmem caches.
2211 */
2212 if (strat == ARC_RECLAIM_AGGR)
2213 arc_shrink();
2214
2215 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2216 if (zio_buf_cache[i] != prev_cache) {
2217 prev_cache = zio_buf_cache[i];
2218 kmem_cache_reap_now(zio_buf_cache[i]);
2219 }
2220 if (zio_data_buf_cache[i] != prev_data_cache) {
2221 prev_data_cache = zio_data_buf_cache[i];
2222 kmem_cache_reap_now(zio_data_buf_cache[i]);
2223 }
2224 }
2225 kmem_cache_reap_now(buf_cache);
2226 kmem_cache_reap_now(hdr_cache);
2227
2228 /*
2229 * Ask the vmem areana to reclaim unused memory from its
2230 * quantum caches.
2231 */
2232 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR)
2233 vmem_qcache_reap(zio_arena);
2234 }
2235
2236 static void
2237 arc_reclaim_thread(void)
2238 {
2239 clock_t growtime = 0;
2240 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2241 callb_cpr_t cpr;
2242
2243 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2244
2245 mutex_enter(&arc_reclaim_thr_lock);
2246 while (arc_thread_exit == 0) {
2247 if (arc_reclaim_needed()) {
2248
2249 if (arc_no_grow) {
2250 if (last_reclaim == ARC_RECLAIM_CONS) {
2251 last_reclaim = ARC_RECLAIM_AGGR;
2252 } else {
2253 last_reclaim = ARC_RECLAIM_CONS;
2254 }
2255 } else {
2256 arc_no_grow = TRUE;
2257 last_reclaim = ARC_RECLAIM_AGGR;
2258 membar_producer();
2259 }
2260
2261 /* reset the growth delay for every reclaim */
2262 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2263
2264 arc_kmem_reap_now(last_reclaim);
2265 arc_warm = B_TRUE;
2266
2267 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2268 arc_no_grow = FALSE;
2269 }
2270
2271 arc_adjust();
2272
2273 if (arc_eviction_list != NULL)
2274 arc_do_user_evicts();
2275
2276 /* block until needed, or one second, whichever is shorter */
2277 CALLB_CPR_SAFE_BEGIN(&cpr);
2278 (void) cv_timedwait(&arc_reclaim_thr_cv,
2279 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2280 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2281 }
2282
2283 arc_thread_exit = 0;
2284 cv_broadcast(&arc_reclaim_thr_cv);
2285 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2286 thread_exit();
2287 }
2288
2289 /*
2290 * Adapt arc info given the number of bytes we are trying to add and
2291 * the state that we are comming from. This function is only called
2292 * when we are adding new content to the cache.
2293 */
2294 static void
2295 arc_adapt(int bytes, arc_state_t *state)
2296 {
2297 int mult;
2298 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2299
2300 if (state == arc_l2c_only)
2301 return;
2302
2303 ASSERT(bytes > 0);
2304 /*
2305 * Adapt the target size of the MRU list:
2306 * - if we just hit in the MRU ghost list, then increase
2307 * the target size of the MRU list.
2308 * - if we just hit in the MFU ghost list, then increase
2309 * the target size of the MFU list by decreasing the
2310 * target size of the MRU list.
2311 */
2312 if (state == arc_mru_ghost) {
2313 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2314 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2315 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2316
2317 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2318 } else if (state == arc_mfu_ghost) {
2319 uint64_t delta;
2320
2321 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2322 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2323 mult = MIN(mult, 10);
2324
2325 delta = MIN(bytes * mult, arc_p);
2326 arc_p = MAX(arc_p_min, arc_p - delta);
2327 }
2328 ASSERT((int64_t)arc_p >= 0);
2329
2330 if (arc_reclaim_needed()) {
2331 cv_signal(&arc_reclaim_thr_cv);
2332 return;
2333 }
2334
2335 if (arc_no_grow)
2336 return;
2337
2338 if (arc_c >= arc_c_max)
2339 return;
2340
2341 /*
2342 * If we're within (2 * maxblocksize) bytes of the target
2343 * cache size, increment the target cache size
2344 */
2345 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2346 atomic_add_64(&arc_c, (int64_t)bytes);
2347 if (arc_c > arc_c_max)
2348 arc_c = arc_c_max;
2349 else if (state == arc_anon)
2350 atomic_add_64(&arc_p, (int64_t)bytes);
2351 if (arc_p > arc_c)
2352 arc_p = arc_c;
2353 }
2354 ASSERT((int64_t)arc_p >= 0);
2355 }
2356
2357 /*
2358 * Check if the cache has reached its limits and eviction is required
2359 * prior to insert.
2360 */
2361 static int
2362 arc_evict_needed(arc_buf_contents_t type)
2363 {
2364 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2365 return (1);
2366
2367 if (arc_reclaim_needed())
2368 return (1);
2369
2370 return (arc_size > arc_c);
2371 }
2372
2373 /*
2374 * The buffer, supplied as the first argument, needs a data block.
2375 * So, if we are at cache max, determine which cache should be victimized.
2376 * We have the following cases:
2377 *
2378 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2379 * In this situation if we're out of space, but the resident size of the MFU is
2380 * under the limit, victimize the MFU cache to satisfy this insertion request.
2381 *
2382 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2383 * Here, we've used up all of the available space for the MRU, so we need to
2384 * evict from our own cache instead. Evict from the set of resident MRU
2385 * entries.
2386 *
2387 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2388 * c minus p represents the MFU space in the cache, since p is the size of the
2389 * cache that is dedicated to the MRU. In this situation there's still space on
2390 * the MFU side, so the MRU side needs to be victimized.
2391 *
2392 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2393 * MFU's resident set is consuming more space than it has been allotted. In
2394 * this situation, we must victimize our own cache, the MFU, for this insertion.
2395 */
2396 static void
2397 arc_get_data_buf(arc_buf_t *buf)
2398 {
2399 arc_state_t *state = buf->b_hdr->b_state;
2400 uint64_t size = buf->b_hdr->b_size;
2401 arc_buf_contents_t type = buf->b_hdr->b_type;
2402
2403 arc_adapt(size, state);
2404
2405 /*
2406 * We have not yet reached cache maximum size,
2407 * just allocate a new buffer.
2408 */
2409 if (!arc_evict_needed(type)) {
2410 if (type == ARC_BUFC_METADATA) {
2411 buf->b_data = zio_buf_alloc(size);
2412 arc_space_consume(size, ARC_SPACE_DATA);
2413 } else {
2414 ASSERT(type == ARC_BUFC_DATA);
2415 buf->b_data = zio_data_buf_alloc(size);
2416 ARCSTAT_INCR(arcstat_data_size, size);
2417 atomic_add_64(&arc_size, size);
2418 }
2419 goto out;
2420 }
2421
2422 /*
2423 * If we are prefetching from the mfu ghost list, this buffer
2424 * will end up on the mru list; so steal space from there.
2425 */
2426 if (state == arc_mfu_ghost)
2427 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2428 else if (state == arc_mru_ghost)
2429 state = arc_mru;
2430
2431 if (state == arc_mru || state == arc_anon) {
2432 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2433 state = (arc_mfu->arcs_lsize[type] >= size &&
2434 arc_p > mru_used) ? arc_mfu : arc_mru;
2435 } else {
2436 /* MFU cases */
2437 uint64_t mfu_space = arc_c - arc_p;
2438 state = (arc_mru->arcs_lsize[type] >= size &&
2439 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2440 }
2441 if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) {
2442 if (type == ARC_BUFC_METADATA) {
2443 buf->b_data = zio_buf_alloc(size);
2444 arc_space_consume(size, ARC_SPACE_DATA);
2445 } else {
2446 ASSERT(type == ARC_BUFC_DATA);
2447 buf->b_data = zio_data_buf_alloc(size);
2448 ARCSTAT_INCR(arcstat_data_size, size);
2449 atomic_add_64(&arc_size, size);
2450 }
2451 ARCSTAT_BUMP(arcstat_recycle_miss);
2452 }
2453 ASSERT(buf->b_data != NULL);
2454 out:
2455 /*
2456 * Update the state size. Note that ghost states have a
2457 * "ghost size" and so don't need to be updated.
2458 */
2459 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2460 arc_buf_hdr_t *hdr = buf->b_hdr;
2461
2462 atomic_add_64(&hdr->b_state->arcs_size, size);
2463 if (list_link_active(&hdr->b_arc_node)) {
2464 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2465 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2466 }
2467 /*
2468 * If we are growing the cache, and we are adding anonymous
2469 * data, and we have outgrown arc_p, update arc_p
2470 */
2471 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2472 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2473 arc_p = MIN(arc_c, arc_p + size);
2474 }
2475 }
2476
2477 /*
2478 * This routine is called whenever a buffer is accessed.
2479 * NOTE: the hash lock is dropped in this function.
2480 */
2481 static void
2482 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2483 {
2484 clock_t now;
2485
2486 ASSERT(MUTEX_HELD(hash_lock));
2487
2488 if (buf->b_state == arc_anon) {
2489 /*
2490 * This buffer is not in the cache, and does not
2491 * appear in our "ghost" list. Add the new buffer
2492 * to the MRU state.
2493 */
2494
2495 ASSERT(buf->b_arc_access == 0);
2496 buf->b_arc_access = ddi_get_lbolt();
2497 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2498 arc_change_state(arc_mru, buf, hash_lock);
2499
2500 } else if (buf->b_state == arc_mru) {
2501 now = ddi_get_lbolt();
2502
2503 /*
2504 * If this buffer is here because of a prefetch, then either:
2505 * - clear the flag if this is a "referencing" read
2506 * (any subsequent access will bump this into the MFU state).
2507 * or
2508 * - move the buffer to the head of the list if this is
2509 * another prefetch (to make it less likely to be evicted).
2510 */
2511 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2512 if (refcount_count(&buf->b_refcnt) == 0) {
2513 ASSERT(list_link_active(&buf->b_arc_node));
2514 } else {
2515 buf->b_flags &= ~ARC_PREFETCH;
2516 ARCSTAT_BUMP(arcstat_mru_hits);
2517 }
2518 buf->b_arc_access = now;
2519 return;
2520 }
2521
2522 /*
2523 * This buffer has been "accessed" only once so far,
2524 * but it is still in the cache. Move it to the MFU
2525 * state.
2526 */
2527 if (now > buf->b_arc_access + ARC_MINTIME) {
2528 /*
2529 * More than 125ms have passed since we
2530 * instantiated this buffer. Move it to the
2531 * most frequently used state.
2532 */
2533 buf->b_arc_access = now;
2534 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2535 arc_change_state(arc_mfu, buf, hash_lock);
2536 }
2537 ARCSTAT_BUMP(arcstat_mru_hits);
2538 } else if (buf->b_state == arc_mru_ghost) {
2539 arc_state_t *new_state;
2540 /*
2541 * This buffer has been "accessed" recently, but
2542 * was evicted from the cache. Move it to the
2543 * MFU state.
2544 */
2545
2546 if (buf->b_flags & ARC_PREFETCH) {
2547 new_state = arc_mru;
2548 if (refcount_count(&buf->b_refcnt) > 0)
2549 buf->b_flags &= ~ARC_PREFETCH;
2550 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2551 } else {
2552 new_state = arc_mfu;
2553 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2554 }
2555
2556 buf->b_arc_access = ddi_get_lbolt();
2557 arc_change_state(new_state, buf, hash_lock);
2558
2559 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2560 } else if (buf->b_state == arc_mfu) {
2561 /*
2562 * This buffer has been accessed more than once and is
2563 * still in the cache. Keep it in the MFU state.
2564 *
2565 * NOTE: an add_reference() that occurred when we did
2566 * the arc_read() will have kicked this off the list.
2567 * If it was a prefetch, we will explicitly move it to
2568 * the head of the list now.
2569 */
2570 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2571 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2572 ASSERT(list_link_active(&buf->b_arc_node));
2573 }
2574 ARCSTAT_BUMP(arcstat_mfu_hits);
2575 buf->b_arc_access = ddi_get_lbolt();
2576 } else if (buf->b_state == arc_mfu_ghost) {
2577 arc_state_t *new_state = arc_mfu;
2578 /*
2579 * This buffer has been accessed more than once but has
2580 * been evicted from the cache. Move it back to the
2581 * MFU state.
2582 */
2583
2584 if (buf->b_flags & ARC_PREFETCH) {
2585 /*
2586 * This is a prefetch access...
2587 * move this block back to the MRU state.
2588 */
2589 ASSERT0(refcount_count(&buf->b_refcnt));
2590 new_state = arc_mru;
2591 }
2592
2593 buf->b_arc_access = ddi_get_lbolt();
2594 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2595 arc_change_state(new_state, buf, hash_lock);
2596
2597 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2598 } else if (buf->b_state == arc_l2c_only) {
2599 /*
2600 * This buffer is on the 2nd Level ARC.
2601 */
2602
2603 buf->b_arc_access = ddi_get_lbolt();
2604 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2605 arc_change_state(arc_mfu, buf, hash_lock);
2606 } else {
2607 ASSERT(!"invalid arc state");
2608 }
2609 }
2610
2611 /* a generic arc_done_func_t which you can use */
2612 /* ARGSUSED */
2613 void
2614 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2615 {
2616 if (zio == NULL || zio->io_error == 0)
2617 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2618 VERIFY(arc_buf_remove_ref(buf, arg));
2619 }
2620
2621 /* a generic arc_done_func_t */
2622 void
2623 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2624 {
2625 arc_buf_t **bufp = arg;
2626 if (zio && zio->io_error) {
2627 VERIFY(arc_buf_remove_ref(buf, arg));
2628 *bufp = NULL;
2629 } else {
2630 *bufp = buf;
2631 ASSERT(buf->b_data);
2632 }
2633 }
2634
2635 static void
2636 arc_read_done(zio_t *zio)
2637 {
2638 arc_buf_hdr_t *hdr, *found;
2639 arc_buf_t *buf;
2640 arc_buf_t *abuf; /* buffer we're assigning to callback */
2641 kmutex_t *hash_lock;
2642 arc_callback_t *callback_list, *acb;
2643 int freeable = FALSE;
2644
2645 buf = zio->io_private;
2646 hdr = buf->b_hdr;
2647
2648 /*
2649 * The hdr was inserted into hash-table and removed from lists
2650 * prior to starting I/O. We should find this header, since
2651 * it's in the hash table, and it should be legit since it's
2652 * not possible to evict it during the I/O. The only possible
2653 * reason for it not to be found is if we were freed during the
2654 * read.
2655 */
2656 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2657 &hash_lock);
2658
2659 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2660 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2661 (found == hdr && HDR_L2_READING(hdr)));
2662
2663 hdr->b_flags &= ~ARC_L2_EVICTED;
2664 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2665 hdr->b_flags &= ~ARC_L2CACHE;
2666
2667 /* byteswap if necessary */
2668 callback_list = hdr->b_acb;
2669 ASSERT(callback_list != NULL);
2670 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2671 dmu_object_byteswap_t bswap =
2672 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2673 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2674 byteswap_uint64_array :
2675 dmu_ot_byteswap[bswap].ob_func;
2676 func(buf->b_data, hdr->b_size);
2677 }
2678
2679 arc_cksum_compute(buf, B_FALSE);
2680 arc_buf_watch(buf);
2681
2682 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2683 /*
2684 * Only call arc_access on anonymous buffers. This is because
2685 * if we've issued an I/O for an evicted buffer, we've already
2686 * called arc_access (to prevent any simultaneous readers from
2687 * getting confused).
2688 */
2689 arc_access(hdr, hash_lock);
2690 }
2691
2692 /* create copies of the data buffer for the callers */
2693 abuf = buf;
2694 for (acb = callback_list; acb; acb = acb->acb_next) {
2695 if (acb->acb_done) {
2696 if (abuf == NULL) {
2697 ARCSTAT_BUMP(arcstat_duplicate_reads);
2698 abuf = arc_buf_clone(buf);
2699 }
2700 acb->acb_buf = abuf;
2701 abuf = NULL;
2702 }
2703 }
2704 hdr->b_acb = NULL;
2705 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2706 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2707 if (abuf == buf) {
2708 ASSERT(buf->b_efunc == NULL);
2709 ASSERT(hdr->b_datacnt == 1);
2710 hdr->b_flags |= ARC_BUF_AVAILABLE;
2711 }
2712
2713 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2714
2715 if (zio->io_error != 0) {
2716 hdr->b_flags |= ARC_IO_ERROR;
2717 if (hdr->b_state != arc_anon)
2718 arc_change_state(arc_anon, hdr, hash_lock);
2719 if (HDR_IN_HASH_TABLE(hdr))
2720 buf_hash_remove(hdr);
2721 freeable = refcount_is_zero(&hdr->b_refcnt);
2722 }
2723
2724 /*
2725 * Broadcast before we drop the hash_lock to avoid the possibility
2726 * that the hdr (and hence the cv) might be freed before we get to
2727 * the cv_broadcast().
2728 */
2729 cv_broadcast(&hdr->b_cv);
2730
2731 if (hash_lock) {
2732 mutex_exit(hash_lock);
2733 } else {
2734 /*
2735 * This block was freed while we waited for the read to
2736 * complete. It has been removed from the hash table and
2737 * moved to the anonymous state (so that it won't show up
2738 * in the cache).
2739 */
2740 ASSERT3P(hdr->b_state, ==, arc_anon);
2741 freeable = refcount_is_zero(&hdr->b_refcnt);
2742 }
2743
2744 /* execute each callback and free its structure */
2745 while ((acb = callback_list) != NULL) {
2746 if (acb->acb_done)
2747 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2748
2749 if (acb->acb_zio_dummy != NULL) {
2750 acb->acb_zio_dummy->io_error = zio->io_error;
2751 zio_nowait(acb->acb_zio_dummy);
2752 }
2753
2754 callback_list = acb->acb_next;
2755 kmem_free(acb, sizeof (arc_callback_t));
2756 }
2757
2758 if (freeable)
2759 arc_hdr_destroy(hdr);
2760 }
2761
2762 /*
2763 * "Read" the block at the specified DVA (in bp) via the
2764 * cache. If the block is found in the cache, invoke the provided
2765 * callback immediately and return. Note that the `zio' parameter
2766 * in the callback will be NULL in this case, since no IO was
2767 * required. If the block is not in the cache pass the read request
2768 * on to the spa with a substitute callback function, so that the
2769 * requested block will be added to the cache.
2770 *
2771 * If a read request arrives for a block that has a read in-progress,
2772 * either wait for the in-progress read to complete (and return the
2773 * results); or, if this is a read with a "done" func, add a record
2774 * to the read to invoke the "done" func when the read completes,
2775 * and return; or just return.
2776 *
2777 * arc_read_done() will invoke all the requested "done" functions
2778 * for readers of this block.
2779 */
2780 int
2781 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
2782 void *private, int priority, int zio_flags, uint32_t *arc_flags,
2783 const zbookmark_t *zb)
2784 {
2785 arc_buf_hdr_t *hdr;
2786 arc_buf_t *buf = NULL;
2787 kmutex_t *hash_lock;
2788 zio_t *rzio;
2789 uint64_t guid = spa_load_guid(spa);
2790
2791 top:
2792 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2793 &hash_lock);
2794 if (hdr && hdr->b_datacnt > 0) {
2795
2796 *arc_flags |= ARC_CACHED;
2797
2798 if (HDR_IO_IN_PROGRESS(hdr)) {
2799
2800 if (*arc_flags & ARC_WAIT) {
2801 cv_wait(&hdr->b_cv, hash_lock);
2802 mutex_exit(hash_lock);
2803 goto top;
2804 }
2805 ASSERT(*arc_flags & ARC_NOWAIT);
2806
2807 if (done) {
2808 arc_callback_t *acb = NULL;
2809
2810 acb = kmem_zalloc(sizeof (arc_callback_t),
2811 KM_SLEEP);
2812 acb->acb_done = done;
2813 acb->acb_private = private;
2814 if (pio != NULL)
2815 acb->acb_zio_dummy = zio_null(pio,
2816 spa, NULL, NULL, NULL, zio_flags);
2817
2818 ASSERT(acb->acb_done != NULL);
2819 acb->acb_next = hdr->b_acb;
2820 hdr->b_acb = acb;
2821 add_reference(hdr, hash_lock, private);
2822 mutex_exit(hash_lock);
2823 return (0);
2824 }
2825 mutex_exit(hash_lock);
2826 return (0);
2827 }
2828
2829 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2830
2831 if (done) {
2832 add_reference(hdr, hash_lock, private);
2833 /*
2834 * If this block is already in use, create a new
2835 * copy of the data so that we will be guaranteed
2836 * that arc_release() will always succeed.
2837 */
2838 buf = hdr->b_buf;
2839 ASSERT(buf);
2840 ASSERT(buf->b_data);
2841 if (HDR_BUF_AVAILABLE(hdr)) {
2842 ASSERT(buf->b_efunc == NULL);
2843 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2844 } else {
2845 buf = arc_buf_clone(buf);
2846 }
2847
2848 } else if (*arc_flags & ARC_PREFETCH &&
2849 refcount_count(&hdr->b_refcnt) == 0) {
2850 hdr->b_flags |= ARC_PREFETCH;
2851 }
2852 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2853 arc_access(hdr, hash_lock);
2854 if (*arc_flags & ARC_L2CACHE)
2855 hdr->b_flags |= ARC_L2CACHE;
2856 mutex_exit(hash_lock);
2857 ARCSTAT_BUMP(arcstat_hits);
2858 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2859 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2860 data, metadata, hits);
2861
2862 if (done)
2863 done(NULL, buf, private);
2864 } else {
2865 uint64_t size = BP_GET_LSIZE(bp);
2866 arc_callback_t *acb;
2867 vdev_t *vd = NULL;
2868 uint64_t addr = 0;
2869 boolean_t devw = B_FALSE;
2870
2871 if (hdr == NULL) {
2872 /* this block is not in the cache */
2873 arc_buf_hdr_t *exists;
2874 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2875 buf = arc_buf_alloc(spa, size, private, type);
2876 hdr = buf->b_hdr;
2877 hdr->b_dva = *BP_IDENTITY(bp);
2878 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
2879 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2880 exists = buf_hash_insert(hdr, &hash_lock);
2881 if (exists) {
2882 /* somebody beat us to the hash insert */
2883 mutex_exit(hash_lock);
2884 buf_discard_identity(hdr);
2885 (void) arc_buf_remove_ref(buf, private);
2886 goto top; /* restart the IO request */
2887 }
2888 /* if this is a prefetch, we don't have a reference */
2889 if (*arc_flags & ARC_PREFETCH) {
2890 (void) remove_reference(hdr, hash_lock,
2891 private);
2892 hdr->b_flags |= ARC_PREFETCH;
2893 }
2894 if (*arc_flags & ARC_L2CACHE)
2895 hdr->b_flags |= ARC_L2CACHE;
2896 if (BP_GET_LEVEL(bp) > 0)
2897 hdr->b_flags |= ARC_INDIRECT;
2898 } else {
2899 /* this block is in the ghost cache */
2900 ASSERT(GHOST_STATE(hdr->b_state));
2901 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2902 ASSERT0(refcount_count(&hdr->b_refcnt));
2903 ASSERT(hdr->b_buf == NULL);
2904
2905 /* if this is a prefetch, we don't have a reference */
2906 if (*arc_flags & ARC_PREFETCH)
2907 hdr->b_flags |= ARC_PREFETCH;
2908 else
2909 add_reference(hdr, hash_lock, private);
2910 if (*arc_flags & ARC_L2CACHE)
2911 hdr->b_flags |= ARC_L2CACHE;
2912 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2913 buf->b_hdr = hdr;
2914 buf->b_data = NULL;
2915 buf->b_efunc = NULL;
2916 buf->b_private = NULL;
2917 buf->b_next = NULL;
2918 hdr->b_buf = buf;
2919 ASSERT(hdr->b_datacnt == 0);
2920 hdr->b_datacnt = 1;
2921 arc_get_data_buf(buf);
2922 arc_access(hdr, hash_lock);
2923 }
2924
2925 ASSERT(!GHOST_STATE(hdr->b_state));
2926
2927 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
2928 acb->acb_done = done;
2929 acb->acb_private = private;
2930
2931 ASSERT(hdr->b_acb == NULL);
2932 hdr->b_acb = acb;
2933 hdr->b_flags |= ARC_IO_IN_PROGRESS;
2934
2935 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
2936 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
2937 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
2938 addr = hdr->b_l2hdr->b_daddr;
2939 /*
2940 * Lock out device removal.
2941 */
2942 if (vdev_is_dead(vd) ||
2943 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
2944 vd = NULL;
2945 }
2946
2947 mutex_exit(hash_lock);
2948
2949 ASSERT3U(hdr->b_size, ==, size);
2950 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
2951 uint64_t, size, zbookmark_t *, zb);
2952 ARCSTAT_BUMP(arcstat_misses);
2953 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2954 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2955 data, metadata, misses);
2956
2957 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
2958 /*
2959 * Read from the L2ARC if the following are true:
2960 * 1. The L2ARC vdev was previously cached.
2961 * 2. This buffer still has L2ARC metadata.
2962 * 3. This buffer isn't currently writing to the L2ARC.
2963 * 4. The L2ARC entry wasn't evicted, which may
2964 * also have invalidated the vdev.
2965 * 5. This isn't prefetch and l2arc_noprefetch is set.
2966 */
2967 if (hdr->b_l2hdr != NULL &&
2968 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
2969 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
2970 l2arc_read_callback_t *cb;
2971
2972 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
2973 ARCSTAT_BUMP(arcstat_l2_hits);
2974
2975 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
2976 KM_SLEEP);
2977 cb->l2rcb_buf = buf;
2978 cb->l2rcb_spa = spa;
2979 cb->l2rcb_bp = *bp;
2980 cb->l2rcb_zb = *zb;
2981 cb->l2rcb_flags = zio_flags;
2982
2983 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
2984 addr + size < vd->vdev_psize -
2985 VDEV_LABEL_END_SIZE);
2986
2987 /*
2988 * l2arc read. The SCL_L2ARC lock will be
2989 * released by l2arc_read_done().
2990 */
2991 rzio = zio_read_phys(pio, vd, addr, size,
2992 buf->b_data, ZIO_CHECKSUM_OFF,
2993 l2arc_read_done, cb, priority, zio_flags |
2994 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
2995 ZIO_FLAG_DONT_PROPAGATE |
2996 ZIO_FLAG_DONT_RETRY, B_FALSE);
2997 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
2998 zio_t *, rzio);
2999 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
3000
3001 if (*arc_flags & ARC_NOWAIT) {
3002 zio_nowait(rzio);
3003 return (0);
3004 }
3005
3006 ASSERT(*arc_flags & ARC_WAIT);
3007 if (zio_wait(rzio) == 0)
3008 return (0);
3009
3010 /* l2arc read error; goto zio_read() */
3011 } else {
3012 DTRACE_PROBE1(l2arc__miss,
3013 arc_buf_hdr_t *, hdr);
3014 ARCSTAT_BUMP(arcstat_l2_misses);
3015 if (HDR_L2_WRITING(hdr))
3016 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3017 spa_config_exit(spa, SCL_L2ARC, vd);
3018 }
3019 } else {
3020 if (vd != NULL)
3021 spa_config_exit(spa, SCL_L2ARC, vd);
3022 if (l2arc_ndev != 0) {
3023 DTRACE_PROBE1(l2arc__miss,
3024 arc_buf_hdr_t *, hdr);
3025 ARCSTAT_BUMP(arcstat_l2_misses);
3026 }
3027 }
3028
3029 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3030 arc_read_done, buf, priority, zio_flags, zb);
3031
3032 if (*arc_flags & ARC_WAIT)
3033 return (zio_wait(rzio));
3034
3035 ASSERT(*arc_flags & ARC_NOWAIT);
3036 zio_nowait(rzio);
3037 }
3038 return (0);
3039 }
3040
3041 void
3042 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3043 {
3044 ASSERT(buf->b_hdr != NULL);
3045 ASSERT(buf->b_hdr->b_state != arc_anon);
3046 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3047 ASSERT(buf->b_efunc == NULL);
3048 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3049
3050 buf->b_efunc = func;
3051 buf->b_private = private;
3052 }
3053
3054 /*
3055 * This is used by the DMU to let the ARC know that a buffer is
3056 * being evicted, so the ARC should clean up. If this arc buf
3057 * is not yet in the evicted state, it will be put there.
3058 */
3059 int
3060 arc_buf_evict(arc_buf_t *buf)
3061 {
3062 arc_buf_hdr_t *hdr;
3063 kmutex_t *hash_lock;
3064 arc_buf_t **bufp;
3065
3066 mutex_enter(&buf->b_evict_lock);
3067 hdr = buf->b_hdr;
3068 if (hdr == NULL) {
3069 /*
3070 * We are in arc_do_user_evicts().
3071 */
3072 ASSERT(buf->b_data == NULL);
3073 mutex_exit(&buf->b_evict_lock);
3074 return (0);
3075 } else if (buf->b_data == NULL) {
3076 arc_buf_t copy = *buf; /* structure assignment */
3077 /*
3078 * We are on the eviction list; process this buffer now
3079 * but let arc_do_user_evicts() do the reaping.
3080 */
3081 buf->b_efunc = NULL;
3082 mutex_exit(&buf->b_evict_lock);
3083 VERIFY(copy.b_efunc(©) == 0);
3084 return (1);
3085 }
3086 hash_lock = HDR_LOCK(hdr);
3087 mutex_enter(hash_lock);
3088 hdr = buf->b_hdr;
3089 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3090
3091 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3092 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3093
3094 /*
3095 * Pull this buffer off of the hdr
3096 */
3097 bufp = &hdr->b_buf;
3098 while (*bufp != buf)
3099 bufp = &(*bufp)->b_next;
3100 *bufp = buf->b_next;
3101
3102 ASSERT(buf->b_data != NULL);
3103 arc_buf_destroy(buf, FALSE, FALSE);
3104
3105 if (hdr->b_datacnt == 0) {
3106 arc_state_t *old_state = hdr->b_state;
3107 arc_state_t *evicted_state;
3108
3109 ASSERT(hdr->b_buf == NULL);
3110 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3111
3112 evicted_state =
3113 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3114
3115 mutex_enter(&old_state->arcs_mtx);
3116 mutex_enter(&evicted_state->arcs_mtx);
3117
3118 arc_change_state(evicted_state, hdr, hash_lock);
3119 ASSERT(HDR_IN_HASH_TABLE(hdr));
3120 hdr->b_flags |= ARC_IN_HASH_TABLE;
3121 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3122
3123 mutex_exit(&evicted_state->arcs_mtx);
3124 mutex_exit(&old_state->arcs_mtx);
3125 }
3126 mutex_exit(hash_lock);
3127 mutex_exit(&buf->b_evict_lock);
3128
3129 VERIFY(buf->b_efunc(buf) == 0);
3130 buf->b_efunc = NULL;
3131 buf->b_private = NULL;
3132 buf->b_hdr = NULL;
3133 buf->b_next = NULL;
3134 kmem_cache_free(buf_cache, buf);
3135 return (1);
3136 }
3137
3138 /*
3139 * Release this buffer from the cache. This must be done
3140 * after a read and prior to modifying the buffer contents.
3141 * If the buffer has more than one reference, we must make
3142 * a new hdr for the buffer.
3143 */
3144 void
3145 arc_release(arc_buf_t *buf, void *tag)
3146 {
3147 arc_buf_hdr_t *hdr;
3148 kmutex_t *hash_lock = NULL;
3149 l2arc_buf_hdr_t *l2hdr;
3150 uint64_t buf_size;
3151
3152 /*
3153 * It would be nice to assert that if it's DMU metadata (level >
3154 * 0 || it's the dnode file), then it must be syncing context.
3155 * But we don't know that information at this level.
3156 */
3157
3158 mutex_enter(&buf->b_evict_lock);
3159 hdr = buf->b_hdr;
3160
3161 /* this buffer is not on any list */
3162 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3163
3164 if (hdr->b_state == arc_anon) {
3165 /* this buffer is already released */
3166 ASSERT(buf->b_efunc == NULL);
3167 } else {
3168 hash_lock = HDR_LOCK(hdr);
3169 mutex_enter(hash_lock);
3170 hdr = buf->b_hdr;
3171 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3172 }
3173
3174 l2hdr = hdr->b_l2hdr;
3175 if (l2hdr) {
3176 mutex_enter(&l2arc_buflist_mtx);
3177 hdr->b_l2hdr = NULL;
3178 }
3179 buf_size = hdr->b_size;
3180
3181 /*
3182 * Do we have more than one buf?
3183 */
3184 if (hdr->b_datacnt > 1) {
3185 arc_buf_hdr_t *nhdr;
3186 arc_buf_t **bufp;
3187 uint64_t blksz = hdr->b_size;
3188 uint64_t spa = hdr->b_spa;
3189 arc_buf_contents_t type = hdr->b_type;
3190 uint32_t flags = hdr->b_flags;
3191
3192 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3193 /*
3194 * Pull the data off of this hdr and attach it to
3195 * a new anonymous hdr.
3196 */
3197 (void) remove_reference(hdr, hash_lock, tag);
3198 bufp = &hdr->b_buf;
3199 while (*bufp != buf)
3200 bufp = &(*bufp)->b_next;
3201 *bufp = buf->b_next;
3202 buf->b_next = NULL;
3203
3204 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3205 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3206 if (refcount_is_zero(&hdr->b_refcnt)) {
3207 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3208 ASSERT3U(*size, >=, hdr->b_size);
3209 atomic_add_64(size, -hdr->b_size);
3210 }
3211
3212 /*
3213 * We're releasing a duplicate user data buffer, update
3214 * our statistics accordingly.
3215 */
3216 if (hdr->b_type == ARC_BUFC_DATA) {
3217 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3218 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3219 -hdr->b_size);
3220 }
3221 hdr->b_datacnt -= 1;
3222 arc_cksum_verify(buf);
3223 arc_buf_unwatch(buf);
3224
3225 mutex_exit(hash_lock);
3226
3227 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3228 nhdr->b_size = blksz;
3229 nhdr->b_spa = spa;
3230 nhdr->b_type = type;
3231 nhdr->b_buf = buf;
3232 nhdr->b_state = arc_anon;
3233 nhdr->b_arc_access = 0;
3234 nhdr->b_flags = flags & ARC_L2_WRITING;
3235 nhdr->b_l2hdr = NULL;
3236 nhdr->b_datacnt = 1;
3237 nhdr->b_freeze_cksum = NULL;
3238 (void) refcount_add(&nhdr->b_refcnt, tag);
3239 buf->b_hdr = nhdr;
3240 mutex_exit(&buf->b_evict_lock);
3241 atomic_add_64(&arc_anon->arcs_size, blksz);
3242 } else {
3243 mutex_exit(&buf->b_evict_lock);
3244 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3245 ASSERT(!list_link_active(&hdr->b_arc_node));
3246 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3247 if (hdr->b_state != arc_anon)
3248 arc_change_state(arc_anon, hdr, hash_lock);
3249 hdr->b_arc_access = 0;
3250 if (hash_lock)
3251 mutex_exit(hash_lock);
3252
3253 buf_discard_identity(hdr);
3254 arc_buf_thaw(buf);
3255 }
3256 buf->b_efunc = NULL;
3257 buf->b_private = NULL;
3258
3259 if (l2hdr) {
3260 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3261 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3262 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3263 mutex_exit(&l2arc_buflist_mtx);
3264 }
3265 }
3266
3267 int
3268 arc_released(arc_buf_t *buf)
3269 {
3270 int released;
3271
3272 mutex_enter(&buf->b_evict_lock);
3273 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3274 mutex_exit(&buf->b_evict_lock);
3275 return (released);
3276 }
3277
3278 int
3279 arc_has_callback(arc_buf_t *buf)
3280 {
3281 int callback;
3282
3283 mutex_enter(&buf->b_evict_lock);
3284 callback = (buf->b_efunc != NULL);
3285 mutex_exit(&buf->b_evict_lock);
3286 return (callback);
3287 }
3288
3289 #ifdef ZFS_DEBUG
3290 int
3291 arc_referenced(arc_buf_t *buf)
3292 {
3293 int referenced;
3294
3295 mutex_enter(&buf->b_evict_lock);
3296 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3297 mutex_exit(&buf->b_evict_lock);
3298 return (referenced);
3299 }
3300 #endif
3301
3302 static void
3303 arc_write_ready(zio_t *zio)
3304 {
3305 arc_write_callback_t *callback = zio->io_private;
3306 arc_buf_t *buf = callback->awcb_buf;
3307 arc_buf_hdr_t *hdr = buf->b_hdr;
3308
3309 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3310 callback->awcb_ready(zio, buf, callback->awcb_private);
3311
3312 /*
3313 * If the IO is already in progress, then this is a re-write
3314 * attempt, so we need to thaw and re-compute the cksum.
3315 * It is the responsibility of the callback to handle the
3316 * accounting for any re-write attempt.
3317 */
3318 if (HDR_IO_IN_PROGRESS(hdr)) {
3319 mutex_enter(&hdr->b_freeze_lock);
3320 if (hdr->b_freeze_cksum != NULL) {
3321 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3322 hdr->b_freeze_cksum = NULL;
3323 }
3324 mutex_exit(&hdr->b_freeze_lock);
3325 }
3326 arc_cksum_compute(buf, B_FALSE);
3327 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3328 }
3329
3330 static void
3331 arc_write_done(zio_t *zio)
3332 {
3333 arc_write_callback_t *callback = zio->io_private;
3334 arc_buf_t *buf = callback->awcb_buf;
3335 arc_buf_hdr_t *hdr = buf->b_hdr;
3336
3337 ASSERT(hdr->b_acb == NULL);
3338
3339 if (zio->io_error == 0) {
3340 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3341 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3342 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3343 } else {
3344 ASSERT(BUF_EMPTY(hdr));
3345 }
3346
3347 /*
3348 * If the block to be written was all-zero, we may have
3349 * compressed it away. In this case no write was performed
3350 * so there will be no dva/birth/checksum. The buffer must
3351 * therefore remain anonymous (and uncached).
3352 */
3353 if (!BUF_EMPTY(hdr)) {
3354 arc_buf_hdr_t *exists;
3355 kmutex_t *hash_lock;
3356
3357 ASSERT(zio->io_error == 0);
3358
3359 arc_cksum_verify(buf);
3360
3361 exists = buf_hash_insert(hdr, &hash_lock);
3362 if (exists) {
3363 /*
3364 * This can only happen if we overwrite for
3365 * sync-to-convergence, because we remove
3366 * buffers from the hash table when we arc_free().
3367 */
3368 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3369 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3370 panic("bad overwrite, hdr=%p exists=%p",
3371 (void *)hdr, (void *)exists);
3372 ASSERT(refcount_is_zero(&exists->b_refcnt));
3373 arc_change_state(arc_anon, exists, hash_lock);
3374 mutex_exit(hash_lock);
3375 arc_hdr_destroy(exists);
3376 exists = buf_hash_insert(hdr, &hash_lock);
3377 ASSERT3P(exists, ==, NULL);
3378 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3379 /* nopwrite */
3380 ASSERT(zio->io_prop.zp_nopwrite);
3381 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3382 panic("bad nopwrite, hdr=%p exists=%p",
3383 (void *)hdr, (void *)exists);
3384 } else {
3385 /* Dedup */
3386 ASSERT(hdr->b_datacnt == 1);
3387 ASSERT(hdr->b_state == arc_anon);
3388 ASSERT(BP_GET_DEDUP(zio->io_bp));
3389 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3390 }
3391 }
3392 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3393 /* if it's not anon, we are doing a scrub */
3394 if (!exists && hdr->b_state == arc_anon)
3395 arc_access(hdr, hash_lock);
3396 mutex_exit(hash_lock);
3397 } else {
3398 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3399 }
3400
3401 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3402 callback->awcb_done(zio, buf, callback->awcb_private);
3403
3404 kmem_free(callback, sizeof (arc_write_callback_t));
3405 }
3406
3407 zio_t *
3408 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3409 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
3410 arc_done_func_t *ready, arc_done_func_t *done, void *private,
3411 int priority, int zio_flags, const zbookmark_t *zb)
3412 {
3413 arc_buf_hdr_t *hdr = buf->b_hdr;
3414 arc_write_callback_t *callback;
3415 zio_t *zio;
3416
3417 ASSERT(ready != NULL);
3418 ASSERT(done != NULL);
3419 ASSERT(!HDR_IO_ERROR(hdr));
3420 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3421 ASSERT(hdr->b_acb == NULL);
3422 if (l2arc)
3423 hdr->b_flags |= ARC_L2CACHE;
3424 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3425 callback->awcb_ready = ready;
3426 callback->awcb_done = done;
3427 callback->awcb_private = private;
3428 callback->awcb_buf = buf;
3429
3430 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3431 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3432
3433 return (zio);
3434 }
3435
3436 static int
3437 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3438 {
3439 #ifdef _KERNEL
3440 uint64_t available_memory = ptob(freemem);
3441 static uint64_t page_load = 0;
3442 static uint64_t last_txg = 0;
3443
3444 #if defined(__i386)
3445 available_memory =
3446 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3447 #endif
3448 if (available_memory >= zfs_write_limit_max)
3449 return (0);
3450
3451 if (txg > last_txg) {
3452 last_txg = txg;
3453 page_load = 0;
3454 }
3455 /*
3456 * If we are in pageout, we know that memory is already tight,
3457 * the arc is already going to be evicting, so we just want to
3458 * continue to let page writes occur as quickly as possible.
3459 */
3460 if (curproc == proc_pageout) {
3461 if (page_load > MAX(ptob(minfree), available_memory) / 4)
3462 return (SET_ERROR(ERESTART));
3463 /* Note: reserve is inflated, so we deflate */
3464 page_load += reserve / 8;
3465 return (0);
3466 } else if (page_load > 0 && arc_reclaim_needed()) {
3467 /* memory is low, delay before restarting */
3468 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3469 return (SET_ERROR(EAGAIN));
3470 }
3471 page_load = 0;
3472
3473 if (arc_size > arc_c_min) {
3474 uint64_t evictable_memory =
3475 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3476 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3477 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3478 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3479 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3480 }
3481
3482 if (inflight_data > available_memory / 4) {
3483 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3484 return (SET_ERROR(ERESTART));
3485 }
3486 #endif
3487 return (0);
3488 }
3489
3490 void
3491 arc_tempreserve_clear(uint64_t reserve)
3492 {
3493 atomic_add_64(&arc_tempreserve, -reserve);
3494 ASSERT((int64_t)arc_tempreserve >= 0);
3495 }
3496
3497 int
3498 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3499 {
3500 int error;
3501 uint64_t anon_size;
3502
3503 #ifdef ZFS_DEBUG
3504 /*
3505 * Once in a while, fail for no reason. Everything should cope.
3506 */
3507 if (spa_get_random(10000) == 0) {
3508 dprintf("forcing random failure\n");
3509 return (SET_ERROR(ERESTART));
3510 }
3511 #endif
3512 if (reserve > arc_c/4 && !arc_no_grow)
3513 arc_c = MIN(arc_c_max, reserve * 4);
3514 if (reserve > arc_c)
3515 return (SET_ERROR(ENOMEM));
3516
3517 /*
3518 * Don't count loaned bufs as in flight dirty data to prevent long
3519 * network delays from blocking transactions that are ready to be
3520 * assigned to a txg.
3521 */
3522 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3523
3524 /*
3525 * Writes will, almost always, require additional memory allocations
3526 * in order to compress/encrypt/etc the data. We therefor need to
3527 * make sure that there is sufficient available memory for this.
3528 */
3529 if (error = arc_memory_throttle(reserve, anon_size, txg))
3530 return (error);
3531
3532 /*
3533 * Throttle writes when the amount of dirty data in the cache
3534 * gets too large. We try to keep the cache less than half full
3535 * of dirty blocks so that our sync times don't grow too large.
3536 * Note: if two requests come in concurrently, we might let them
3537 * both succeed, when one of them should fail. Not a huge deal.
3538 */
3539
3540 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3541 anon_size > arc_c / 4) {
3542 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3543 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3544 arc_tempreserve>>10,
3545 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3546 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3547 reserve>>10, arc_c>>10);
3548 return (SET_ERROR(ERESTART));
3549 }
3550 atomic_add_64(&arc_tempreserve, reserve);
3551 return (0);
3552 }
3553
3554 void
3555 arc_init(void)
3556 {
3557 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3558 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3559
3560 /* Convert seconds to clock ticks */
3561 arc_min_prefetch_lifespan = 1 * hz;
3562
3563 /* Start out with 1/8 of all memory */
3564 arc_c = physmem * PAGESIZE / 8;
3565
3566 #ifdef _KERNEL
3567 /*
3568 * On architectures where the physical memory can be larger
3569 * than the addressable space (intel in 32-bit mode), we may
3570 * need to limit the cache to 1/8 of VM size.
3571 */
3572 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3573 #endif
3574
3575 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3576 arc_c_min = MAX(arc_c / 4, 64<<20);
3577 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
3578 if (arc_c * 8 >= 1<<30)
3579 arc_c_max = (arc_c * 8) - (1<<30);
3580 else
3581 arc_c_max = arc_c_min;
3582 arc_c_max = MAX(arc_c * 6, arc_c_max);
3583
3584 /*
3585 * Allow the tunables to override our calculations if they are
3586 * reasonable (ie. over 64MB)
3587 */
3588 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3589 arc_c_max = zfs_arc_max;
3590 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3591 arc_c_min = zfs_arc_min;
3592
3593 arc_c = arc_c_max;
3594 arc_p = (arc_c >> 1);
3595
3596 /* limit meta-data to 1/4 of the arc capacity */
3597 arc_meta_limit = arc_c_max / 4;
3598
3599 /* Allow the tunable to override if it is reasonable */
3600 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3601 arc_meta_limit = zfs_arc_meta_limit;
3602
3603 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3604 arc_c_min = arc_meta_limit / 2;
3605
3606 if (zfs_arc_grow_retry > 0)
3607 arc_grow_retry = zfs_arc_grow_retry;
3608
3609 if (zfs_arc_shrink_shift > 0)
3610 arc_shrink_shift = zfs_arc_shrink_shift;
3611
3612 if (zfs_arc_p_min_shift > 0)
3613 arc_p_min_shift = zfs_arc_p_min_shift;
3614
3615 /* if kmem_flags are set, lets try to use less memory */
3616 if (kmem_debugging())
3617 arc_c = arc_c / 2;
3618 if (arc_c < arc_c_min)
3619 arc_c = arc_c_min;
3620
3621 arc_anon = &ARC_anon;
3622 arc_mru = &ARC_mru;
3623 arc_mru_ghost = &ARC_mru_ghost;
3624 arc_mfu = &ARC_mfu;
3625 arc_mfu_ghost = &ARC_mfu_ghost;
3626 arc_l2c_only = &ARC_l2c_only;
3627 arc_size = 0;
3628
3629 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3630 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3631 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3632 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3633 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3634 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3635
3636 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3637 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3638 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3639 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3640 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3641 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3642 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3643 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3644 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3645 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3646 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3647 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3648 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3649 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3650 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3651 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3652 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3653 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3654 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3655 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3656
3657 buf_init();
3658
3659 arc_thread_exit = 0;
3660 arc_eviction_list = NULL;
3661 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3662 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3663
3664 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3665 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3666
3667 if (arc_ksp != NULL) {
3668 arc_ksp->ks_data = &arc_stats;
3669 kstat_install(arc_ksp);
3670 }
3671
3672 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3673 TS_RUN, minclsyspri);
3674
3675 arc_dead = FALSE;
3676 arc_warm = B_FALSE;
3677
3678 if (zfs_write_limit_max == 0)
3679 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3680 else
3681 zfs_write_limit_shift = 0;
3682 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3683 }
3684
3685 void
3686 arc_fini(void)
3687 {
3688 mutex_enter(&arc_reclaim_thr_lock);
3689 arc_thread_exit = 1;
3690 while (arc_thread_exit != 0)
3691 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3692 mutex_exit(&arc_reclaim_thr_lock);
3693
3694 arc_flush(NULL);
3695
3696 arc_dead = TRUE;
3697
3698 if (arc_ksp != NULL) {
3699 kstat_delete(arc_ksp);
3700 arc_ksp = NULL;
3701 }
3702
3703 mutex_destroy(&arc_eviction_mtx);
3704 mutex_destroy(&arc_reclaim_thr_lock);
3705 cv_destroy(&arc_reclaim_thr_cv);
3706
3707 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3708 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3709 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3710 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3711 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3712 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3713 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3714 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3715
3716 mutex_destroy(&arc_anon->arcs_mtx);
3717 mutex_destroy(&arc_mru->arcs_mtx);
3718 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3719 mutex_destroy(&arc_mfu->arcs_mtx);
3720 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3721 mutex_destroy(&arc_l2c_only->arcs_mtx);
3722
3723 mutex_destroy(&zfs_write_limit_lock);
3724
3725 buf_fini();
3726
3727 ASSERT(arc_loaned_bytes == 0);
3728 }
3729
3730 /*
3731 * Level 2 ARC
3732 *
3733 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3734 * It uses dedicated storage devices to hold cached data, which are populated
3735 * using large infrequent writes. The main role of this cache is to boost
3736 * the performance of random read workloads. The intended L2ARC devices
3737 * include short-stroked disks, solid state disks, and other media with
3738 * substantially faster read latency than disk.
3739 *
3740 * +-----------------------+
3741 * | ARC |
3742 * +-----------------------+
3743 * | ^ ^
3744 * | | |
3745 * l2arc_feed_thread() arc_read()
3746 * | | |
3747 * | l2arc read |
3748 * V | |
3749 * +---------------+ |
3750 * | L2ARC | |
3751 * +---------------+ |
3752 * | ^ |
3753 * l2arc_write() | |
3754 * | | |
3755 * V | |
3756 * +-------+ +-------+
3757 * | vdev | | vdev |
3758 * | cache | | cache |
3759 * +-------+ +-------+
3760 * +=========+ .-----.
3761 * : L2ARC : |-_____-|
3762 * : devices : | Disks |
3763 * +=========+ `-_____-'
3764 *
3765 * Read requests are satisfied from the following sources, in order:
3766 *
3767 * 1) ARC
3768 * 2) vdev cache of L2ARC devices
3769 * 3) L2ARC devices
3770 * 4) vdev cache of disks
3771 * 5) disks
3772 *
3773 * Some L2ARC device types exhibit extremely slow write performance.
3774 * To accommodate for this there are some significant differences between
3775 * the L2ARC and traditional cache design:
3776 *
3777 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3778 * the ARC behave as usual, freeing buffers and placing headers on ghost
3779 * lists. The ARC does not send buffers to the L2ARC during eviction as
3780 * this would add inflated write latencies for all ARC memory pressure.
3781 *
3782 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3783 * It does this by periodically scanning buffers from the eviction-end of
3784 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3785 * not already there. It scans until a headroom of buffers is satisfied,
3786 * which itself is a buffer for ARC eviction. The thread that does this is
3787 * l2arc_feed_thread(), illustrated below; example sizes are included to
3788 * provide a better sense of ratio than this diagram:
3789 *
3790 * head --> tail
3791 * +---------------------+----------+
3792 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
3793 * +---------------------+----------+ | o L2ARC eligible
3794 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
3795 * +---------------------+----------+ |
3796 * 15.9 Gbytes ^ 32 Mbytes |
3797 * headroom |
3798 * l2arc_feed_thread()
3799 * |
3800 * l2arc write hand <--[oooo]--'
3801 * | 8 Mbyte
3802 * | write max
3803 * V
3804 * +==============================+
3805 * L2ARC dev |####|#|###|###| |####| ... |
3806 * +==============================+
3807 * 32 Gbytes
3808 *
3809 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
3810 * evicted, then the L2ARC has cached a buffer much sooner than it probably
3811 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
3812 * safe to say that this is an uncommon case, since buffers at the end of
3813 * the ARC lists have moved there due to inactivity.
3814 *
3815 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
3816 * then the L2ARC simply misses copying some buffers. This serves as a
3817 * pressure valve to prevent heavy read workloads from both stalling the ARC
3818 * with waits and clogging the L2ARC with writes. This also helps prevent
3819 * the potential for the L2ARC to churn if it attempts to cache content too
3820 * quickly, such as during backups of the entire pool.
3821 *
3822 * 5. After system boot and before the ARC has filled main memory, there are
3823 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
3824 * lists can remain mostly static. Instead of searching from tail of these
3825 * lists as pictured, the l2arc_feed_thread() will search from the list heads
3826 * for eligible buffers, greatly increasing its chance of finding them.
3827 *
3828 * The L2ARC device write speed is also boosted during this time so that
3829 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
3830 * there are no L2ARC reads, and no fear of degrading read performance
3831 * through increased writes.
3832 *
3833 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
3834 * the vdev queue can aggregate them into larger and fewer writes. Each
3835 * device is written to in a rotor fashion, sweeping writes through
3836 * available space then repeating.
3837 *
3838 * 7. The L2ARC does not store dirty content. It never needs to flush
3839 * write buffers back to disk based storage.
3840 *
3841 * 8. If an ARC buffer is written (and dirtied) which also exists in the
3842 * L2ARC, the now stale L2ARC buffer is immediately dropped.
3843 *
3844 * The performance of the L2ARC can be tweaked by a number of tunables, which
3845 * may be necessary for different workloads:
3846 *
3847 * l2arc_write_max max write bytes per interval
3848 * l2arc_write_boost extra write bytes during device warmup
3849 * l2arc_noprefetch skip caching prefetched buffers
3850 * l2arc_headroom number of max device writes to precache
3851 * l2arc_feed_secs seconds between L2ARC writing
3852 *
3853 * Tunables may be removed or added as future performance improvements are
3854 * integrated, and also may become zpool properties.
3855 *
3856 * There are three key functions that control how the L2ARC warms up:
3857 *
3858 * l2arc_write_eligible() check if a buffer is eligible to cache
3859 * l2arc_write_size() calculate how much to write
3860 * l2arc_write_interval() calculate sleep delay between writes
3861 *
3862 * These three functions determine what to write, how much, and how quickly
3863 * to send writes.
3864 */
3865
3866 static boolean_t
3867 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
3868 {
3869 /*
3870 * A buffer is *not* eligible for the L2ARC if it:
3871 * 1. belongs to a different spa.
3872 * 2. is already cached on the L2ARC.
3873 * 3. has an I/O in progress (it may be an incomplete read).
3874 * 4. is flagged not eligible (zfs property).
3875 */
3876 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
3877 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
3878 return (B_FALSE);
3879
3880 return (B_TRUE);
3881 }
3882
3883 static uint64_t
3884 l2arc_write_size(l2arc_dev_t *dev)
3885 {
3886 uint64_t size;
3887
3888 size = dev->l2ad_write;
3889
3890 if (arc_warm == B_FALSE)
3891 size += dev->l2ad_boost;
3892
3893 return (size);
3894
3895 }
3896
3897 static clock_t
3898 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
3899 {
3900 clock_t interval, next, now;
3901
3902 /*
3903 * If the ARC lists are busy, increase our write rate; if the
3904 * lists are stale, idle back. This is achieved by checking
3905 * how much we previously wrote - if it was more than half of
3906 * what we wanted, schedule the next write much sooner.
3907 */
3908 if (l2arc_feed_again && wrote > (wanted / 2))
3909 interval = (hz * l2arc_feed_min_ms) / 1000;
3910 else
3911 interval = hz * l2arc_feed_secs;
3912
3913 now = ddi_get_lbolt();
3914 next = MAX(now, MIN(now + interval, began + interval));
3915
3916 return (next);
3917 }
3918
3919 static void
3920 l2arc_hdr_stat_add(void)
3921 {
3922 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
3923 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
3924 }
3925
3926 static void
3927 l2arc_hdr_stat_remove(void)
3928 {
3929 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
3930 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
3931 }
3932
3933 /*
3934 * Cycle through L2ARC devices. This is how L2ARC load balances.
3935 * If a device is returned, this also returns holding the spa config lock.
3936 */
3937 static l2arc_dev_t *
3938 l2arc_dev_get_next(void)
3939 {
3940 l2arc_dev_t *first, *next = NULL;
3941
3942 /*
3943 * Lock out the removal of spas (spa_namespace_lock), then removal
3944 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
3945 * both locks will be dropped and a spa config lock held instead.
3946 */
3947 mutex_enter(&spa_namespace_lock);
3948 mutex_enter(&l2arc_dev_mtx);
3949
3950 /* if there are no vdevs, there is nothing to do */
3951 if (l2arc_ndev == 0)
3952 goto out;
3953
3954 first = NULL;
3955 next = l2arc_dev_last;
3956 do {
3957 /* loop around the list looking for a non-faulted vdev */
3958 if (next == NULL) {
3959 next = list_head(l2arc_dev_list);
3960 } else {
3961 next = list_next(l2arc_dev_list, next);
3962 if (next == NULL)
3963 next = list_head(l2arc_dev_list);
3964 }
3965
3966 /* if we have come back to the start, bail out */
3967 if (first == NULL)
3968 first = next;
3969 else if (next == first)
3970 break;
3971
3972 } while (vdev_is_dead(next->l2ad_vdev));
3973
3974 /* if we were unable to find any usable vdevs, return NULL */
3975 if (vdev_is_dead(next->l2ad_vdev))
3976 next = NULL;
3977
3978 l2arc_dev_last = next;
3979
3980 out:
3981 mutex_exit(&l2arc_dev_mtx);
3982
3983 /*
3984 * Grab the config lock to prevent the 'next' device from being
3985 * removed while we are writing to it.
3986 */
3987 if (next != NULL)
3988 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
3989 mutex_exit(&spa_namespace_lock);
3990
3991 return (next);
3992 }
3993
3994 /*
3995 * Free buffers that were tagged for destruction.
3996 */
3997 static void
3998 l2arc_do_free_on_write()
3999 {
4000 list_t *buflist;
4001 l2arc_data_free_t *df, *df_prev;
4002
4003 mutex_enter(&l2arc_free_on_write_mtx);
4004 buflist = l2arc_free_on_write;
4005
4006 for (df = list_tail(buflist); df; df = df_prev) {
4007 df_prev = list_prev(buflist, df);
4008 ASSERT(df->l2df_data != NULL);
4009 ASSERT(df->l2df_func != NULL);
4010 df->l2df_func(df->l2df_data, df->l2df_size);
4011 list_remove(buflist, df);
4012 kmem_free(df, sizeof (l2arc_data_free_t));
4013 }
4014
4015 mutex_exit(&l2arc_free_on_write_mtx);
4016 }
4017
4018 /*
4019 * A write to a cache device has completed. Update all headers to allow
4020 * reads from these buffers to begin.
4021 */
4022 static void
4023 l2arc_write_done(zio_t *zio)
4024 {
4025 l2arc_write_callback_t *cb;
4026 l2arc_dev_t *dev;
4027 list_t *buflist;
4028 arc_buf_hdr_t *head, *ab, *ab_prev;
4029 l2arc_buf_hdr_t *abl2;
4030 kmutex_t *hash_lock;
4031
4032 cb = zio->io_private;
4033 ASSERT(cb != NULL);
4034 dev = cb->l2wcb_dev;
4035 ASSERT(dev != NULL);
4036 head = cb->l2wcb_head;
4037 ASSERT(head != NULL);
4038 buflist = dev->l2ad_buflist;
4039 ASSERT(buflist != NULL);
4040 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4041 l2arc_write_callback_t *, cb);
4042
4043 if (zio->io_error != 0)
4044 ARCSTAT_BUMP(arcstat_l2_writes_error);
4045
4046 mutex_enter(&l2arc_buflist_mtx);
4047
4048 /*
4049 * All writes completed, or an error was hit.
4050 */
4051 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4052 ab_prev = list_prev(buflist, ab);
4053
4054 hash_lock = HDR_LOCK(ab);
4055 if (!mutex_tryenter(hash_lock)) {
4056 /*
4057 * This buffer misses out. It may be in a stage
4058 * of eviction. Its ARC_L2_WRITING flag will be
4059 * left set, denying reads to this buffer.
4060 */
4061 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4062 continue;
4063 }
4064
4065 if (zio->io_error != 0) {
4066 /*
4067 * Error - drop L2ARC entry.
4068 */
4069 list_remove(buflist, ab);
4070 abl2 = ab->b_l2hdr;
4071 ab->b_l2hdr = NULL;
4072 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4073 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4074 }
4075
4076 /*
4077 * Allow ARC to begin reads to this L2ARC entry.
4078 */
4079 ab->b_flags &= ~ARC_L2_WRITING;
4080
4081 mutex_exit(hash_lock);
4082 }
4083
4084 atomic_inc_64(&l2arc_writes_done);
4085 list_remove(buflist, head);
4086 kmem_cache_free(hdr_cache, head);
4087 mutex_exit(&l2arc_buflist_mtx);
4088
4089 l2arc_do_free_on_write();
4090
4091 kmem_free(cb, sizeof (l2arc_write_callback_t));
4092 }
4093
4094 /*
4095 * A read to a cache device completed. Validate buffer contents before
4096 * handing over to the regular ARC routines.
4097 */
4098 static void
4099 l2arc_read_done(zio_t *zio)
4100 {
4101 l2arc_read_callback_t *cb;
4102 arc_buf_hdr_t *hdr;
4103 arc_buf_t *buf;
4104 kmutex_t *hash_lock;
4105 int equal;
4106
4107 ASSERT(zio->io_vd != NULL);
4108 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4109
4110 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4111
4112 cb = zio->io_private;
4113 ASSERT(cb != NULL);
4114 buf = cb->l2rcb_buf;
4115 ASSERT(buf != NULL);
4116
4117 hash_lock = HDR_LOCK(buf->b_hdr);
4118 mutex_enter(hash_lock);
4119 hdr = buf->b_hdr;
4120 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4121
4122 /*
4123 * Check this survived the L2ARC journey.
4124 */
4125 equal = arc_cksum_equal(buf);
4126 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4127 mutex_exit(hash_lock);
4128 zio->io_private = buf;
4129 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4130 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4131 arc_read_done(zio);
4132 } else {
4133 mutex_exit(hash_lock);
4134 /*
4135 * Buffer didn't survive caching. Increment stats and
4136 * reissue to the original storage device.
4137 */
4138 if (zio->io_error != 0) {
4139 ARCSTAT_BUMP(arcstat_l2_io_error);
4140 } else {
4141 zio->io_error = SET_ERROR(EIO);
4142 }
4143 if (!equal)
4144 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4145
4146 /*
4147 * If there's no waiter, issue an async i/o to the primary
4148 * storage now. If there *is* a waiter, the caller must
4149 * issue the i/o in a context where it's OK to block.
4150 */
4151 if (zio->io_waiter == NULL) {
4152 zio_t *pio = zio_unique_parent(zio);
4153
4154 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4155
4156 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4157 buf->b_data, zio->io_size, arc_read_done, buf,
4158 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4159 }
4160 }
4161
4162 kmem_free(cb, sizeof (l2arc_read_callback_t));
4163 }
4164
4165 /*
4166 * This is the list priority from which the L2ARC will search for pages to
4167 * cache. This is used within loops (0..3) to cycle through lists in the
4168 * desired order. This order can have a significant effect on cache
4169 * performance.
4170 *
4171 * Currently the metadata lists are hit first, MFU then MRU, followed by
4172 * the data lists. This function returns a locked list, and also returns
4173 * the lock pointer.
4174 */
4175 static list_t *
4176 l2arc_list_locked(int list_num, kmutex_t **lock)
4177 {
4178 list_t *list = NULL;
4179
4180 ASSERT(list_num >= 0 && list_num <= 3);
4181
4182 switch (list_num) {
4183 case 0:
4184 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4185 *lock = &arc_mfu->arcs_mtx;
4186 break;
4187 case 1:
4188 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4189 *lock = &arc_mru->arcs_mtx;
4190 break;
4191 case 2:
4192 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4193 *lock = &arc_mfu->arcs_mtx;
4194 break;
4195 case 3:
4196 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4197 *lock = &arc_mru->arcs_mtx;
4198 break;
4199 }
4200
4201 ASSERT(!(MUTEX_HELD(*lock)));
4202 mutex_enter(*lock);
4203 return (list);
4204 }
4205
4206 /*
4207 * Evict buffers from the device write hand to the distance specified in
4208 * bytes. This distance may span populated buffers, it may span nothing.
4209 * This is clearing a region on the L2ARC device ready for writing.
4210 * If the 'all' boolean is set, every buffer is evicted.
4211 */
4212 static void
4213 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4214 {
4215 list_t *buflist;
4216 l2arc_buf_hdr_t *abl2;
4217 arc_buf_hdr_t *ab, *ab_prev;
4218 kmutex_t *hash_lock;
4219 uint64_t taddr;
4220
4221 buflist = dev->l2ad_buflist;
4222
4223 if (buflist == NULL)
4224 return;
4225
4226 if (!all && dev->l2ad_first) {
4227 /*
4228 * This is the first sweep through the device. There is
4229 * nothing to evict.
4230 */
4231 return;
4232 }
4233
4234 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4235 /*
4236 * When nearing the end of the device, evict to the end
4237 * before the device write hand jumps to the start.
4238 */
4239 taddr = dev->l2ad_end;
4240 } else {
4241 taddr = dev->l2ad_hand + distance;
4242 }
4243 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4244 uint64_t, taddr, boolean_t, all);
4245
4246 top:
4247 mutex_enter(&l2arc_buflist_mtx);
4248 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4249 ab_prev = list_prev(buflist, ab);
4250
4251 hash_lock = HDR_LOCK(ab);
4252 if (!mutex_tryenter(hash_lock)) {
4253 /*
4254 * Missed the hash lock. Retry.
4255 */
4256 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4257 mutex_exit(&l2arc_buflist_mtx);
4258 mutex_enter(hash_lock);
4259 mutex_exit(hash_lock);
4260 goto top;
4261 }
4262
4263 if (HDR_L2_WRITE_HEAD(ab)) {
4264 /*
4265 * We hit a write head node. Leave it for
4266 * l2arc_write_done().
4267 */
4268 list_remove(buflist, ab);
4269 mutex_exit(hash_lock);
4270 continue;
4271 }
4272
4273 if (!all && ab->b_l2hdr != NULL &&
4274 (ab->b_l2hdr->b_daddr > taddr ||
4275 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4276 /*
4277 * We've evicted to the target address,
4278 * or the end of the device.
4279 */
4280 mutex_exit(hash_lock);
4281 break;
4282 }
4283
4284 if (HDR_FREE_IN_PROGRESS(ab)) {
4285 /*
4286 * Already on the path to destruction.
4287 */
4288 mutex_exit(hash_lock);
4289 continue;
4290 }
4291
4292 if (ab->b_state == arc_l2c_only) {
4293 ASSERT(!HDR_L2_READING(ab));
4294 /*
4295 * This doesn't exist in the ARC. Destroy.
4296 * arc_hdr_destroy() will call list_remove()
4297 * and decrement arcstat_l2_size.
4298 */
4299 arc_change_state(arc_anon, ab, hash_lock);
4300 arc_hdr_destroy(ab);
4301 } else {
4302 /*
4303 * Invalidate issued or about to be issued
4304 * reads, since we may be about to write
4305 * over this location.
4306 */
4307 if (HDR_L2_READING(ab)) {
4308 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4309 ab->b_flags |= ARC_L2_EVICTED;
4310 }
4311
4312 /*
4313 * Tell ARC this no longer exists in L2ARC.
4314 */
4315 if (ab->b_l2hdr != NULL) {
4316 abl2 = ab->b_l2hdr;
4317 ab->b_l2hdr = NULL;
4318 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4319 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4320 }
4321 list_remove(buflist, ab);
4322
4323 /*
4324 * This may have been leftover after a
4325 * failed write.
4326 */
4327 ab->b_flags &= ~ARC_L2_WRITING;
4328 }
4329 mutex_exit(hash_lock);
4330 }
4331 mutex_exit(&l2arc_buflist_mtx);
4332
4333 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4334 dev->l2ad_evict = taddr;
4335 }
4336
4337 /*
4338 * Find and write ARC buffers to the L2ARC device.
4339 *
4340 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4341 * for reading until they have completed writing.
4342 */
4343 static uint64_t
4344 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4345 {
4346 arc_buf_hdr_t *ab, *ab_prev, *head;
4347 l2arc_buf_hdr_t *hdrl2;
4348 list_t *list;
4349 uint64_t passed_sz, write_sz, buf_sz, headroom;
4350 void *buf_data;
4351 kmutex_t *hash_lock, *list_lock;
4352 boolean_t have_lock, full;
4353 l2arc_write_callback_t *cb;
4354 zio_t *pio, *wzio;
4355 uint64_t guid = spa_load_guid(spa);
4356
4357 ASSERT(dev->l2ad_vdev != NULL);
4358
4359 pio = NULL;
4360 write_sz = 0;
4361 full = B_FALSE;
4362 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4363 head->b_flags |= ARC_L2_WRITE_HEAD;
4364
4365 /*
4366 * Copy buffers for L2ARC writing.
4367 */
4368 mutex_enter(&l2arc_buflist_mtx);
4369 for (int try = 0; try <= 3; try++) {
4370 list = l2arc_list_locked(try, &list_lock);
4371 passed_sz = 0;
4372
4373 /*
4374 * L2ARC fast warmup.
4375 *
4376 * Until the ARC is warm and starts to evict, read from the
4377 * head of the ARC lists rather than the tail.
4378 */
4379 headroom = target_sz * l2arc_headroom;
4380 if (arc_warm == B_FALSE)
4381 ab = list_head(list);
4382 else
4383 ab = list_tail(list);
4384
4385 for (; ab; ab = ab_prev) {
4386 if (arc_warm == B_FALSE)
4387 ab_prev = list_next(list, ab);
4388 else
4389 ab_prev = list_prev(list, ab);
4390
4391 hash_lock = HDR_LOCK(ab);
4392 have_lock = MUTEX_HELD(hash_lock);
4393 if (!have_lock && !mutex_tryenter(hash_lock)) {
4394 /*
4395 * Skip this buffer rather than waiting.
4396 */
4397 continue;
4398 }
4399
4400 passed_sz += ab->b_size;
4401 if (passed_sz > headroom) {
4402 /*
4403 * Searched too far.
4404 */
4405 mutex_exit(hash_lock);
4406 break;
4407 }
4408
4409 if (!l2arc_write_eligible(guid, ab)) {
4410 mutex_exit(hash_lock);
4411 continue;
4412 }
4413
4414 if ((write_sz + ab->b_size) > target_sz) {
4415 full = B_TRUE;
4416 mutex_exit(hash_lock);
4417 break;
4418 }
4419
4420 if (pio == NULL) {
4421 /*
4422 * Insert a dummy header on the buflist so
4423 * l2arc_write_done() can find where the
4424 * write buffers begin without searching.
4425 */
4426 list_insert_head(dev->l2ad_buflist, head);
4427
4428 cb = kmem_alloc(
4429 sizeof (l2arc_write_callback_t), KM_SLEEP);
4430 cb->l2wcb_dev = dev;
4431 cb->l2wcb_head = head;
4432 pio = zio_root(spa, l2arc_write_done, cb,
4433 ZIO_FLAG_CANFAIL);
4434 }
4435
4436 /*
4437 * Create and add a new L2ARC header.
4438 */
4439 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4440 hdrl2->b_dev = dev;
4441 hdrl2->b_daddr = dev->l2ad_hand;
4442
4443 ab->b_flags |= ARC_L2_WRITING;
4444 ab->b_l2hdr = hdrl2;
4445 list_insert_head(dev->l2ad_buflist, ab);
4446 buf_data = ab->b_buf->b_data;
4447 buf_sz = ab->b_size;
4448
4449 /*
4450 * Compute and store the buffer cksum before
4451 * writing. On debug the cksum is verified first.
4452 */
4453 arc_cksum_verify(ab->b_buf);
4454 arc_cksum_compute(ab->b_buf, B_TRUE);
4455
4456 mutex_exit(hash_lock);
4457
4458 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4459 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4460 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4461 ZIO_FLAG_CANFAIL, B_FALSE);
4462
4463 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4464 zio_t *, wzio);
4465 (void) zio_nowait(wzio);
4466
4467 /*
4468 * Keep the clock hand suitably device-aligned.
4469 */
4470 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4471
4472 write_sz += buf_sz;
4473 dev->l2ad_hand += buf_sz;
4474 }
4475
4476 mutex_exit(list_lock);
4477
4478 if (full == B_TRUE)
4479 break;
4480 }
4481 mutex_exit(&l2arc_buflist_mtx);
4482
4483 if (pio == NULL) {
4484 ASSERT0(write_sz);
4485 kmem_cache_free(hdr_cache, head);
4486 return (0);
4487 }
4488
4489 ASSERT3U(write_sz, <=, target_sz);
4490 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4491 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4492 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4493 vdev_space_update(dev->l2ad_vdev, write_sz, 0, 0);
4494
4495 /*
4496 * Bump device hand to the device start if it is approaching the end.
4497 * l2arc_evict() will already have evicted ahead for this case.
4498 */
4499 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4500 vdev_space_update(dev->l2ad_vdev,
4501 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4502 dev->l2ad_hand = dev->l2ad_start;
4503 dev->l2ad_evict = dev->l2ad_start;
4504 dev->l2ad_first = B_FALSE;
4505 }
4506
4507 dev->l2ad_writing = B_TRUE;
4508 (void) zio_wait(pio);
4509 dev->l2ad_writing = B_FALSE;
4510
4511 return (write_sz);
4512 }
4513
4514 /*
4515 * This thread feeds the L2ARC at regular intervals. This is the beating
4516 * heart of the L2ARC.
4517 */
4518 static void
4519 l2arc_feed_thread(void)
4520 {
4521 callb_cpr_t cpr;
4522 l2arc_dev_t *dev;
4523 spa_t *spa;
4524 uint64_t size, wrote;
4525 clock_t begin, next = ddi_get_lbolt();
4526
4527 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4528
4529 mutex_enter(&l2arc_feed_thr_lock);
4530
4531 while (l2arc_thread_exit == 0) {
4532 CALLB_CPR_SAFE_BEGIN(&cpr);
4533 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4534 next);
4535 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4536 next = ddi_get_lbolt() + hz;
4537
4538 /*
4539 * Quick check for L2ARC devices.
4540 */
4541 mutex_enter(&l2arc_dev_mtx);
4542 if (l2arc_ndev == 0) {
4543 mutex_exit(&l2arc_dev_mtx);
4544 continue;
4545 }
4546 mutex_exit(&l2arc_dev_mtx);
4547 begin = ddi_get_lbolt();
4548
4549 /*
4550 * This selects the next l2arc device to write to, and in
4551 * doing so the next spa to feed from: dev->l2ad_spa. This
4552 * will return NULL if there are now no l2arc devices or if
4553 * they are all faulted.
4554 *
4555 * If a device is returned, its spa's config lock is also
4556 * held to prevent device removal. l2arc_dev_get_next()
4557 * will grab and release l2arc_dev_mtx.
4558 */
4559 if ((dev = l2arc_dev_get_next()) == NULL)
4560 continue;
4561
4562 spa = dev->l2ad_spa;
4563 ASSERT(spa != NULL);
4564
4565 /*
4566 * If the pool is read-only then force the feed thread to
4567 * sleep a little longer.
4568 */
4569 if (!spa_writeable(spa)) {
4570 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4571 spa_config_exit(spa, SCL_L2ARC, dev);
4572 continue;
4573 }
4574
4575 /*
4576 * Avoid contributing to memory pressure.
4577 */
4578 if (arc_reclaim_needed()) {
4579 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4580 spa_config_exit(spa, SCL_L2ARC, dev);
4581 continue;
4582 }
4583
4584 ARCSTAT_BUMP(arcstat_l2_feeds);
4585
4586 size = l2arc_write_size(dev);
4587
4588 /*
4589 * Evict L2ARC buffers that will be overwritten.
4590 */
4591 l2arc_evict(dev, size, B_FALSE);
4592
4593 /*
4594 * Write ARC buffers.
4595 */
4596 wrote = l2arc_write_buffers(spa, dev, size);
4597
4598 /*
4599 * Calculate interval between writes.
4600 */
4601 next = l2arc_write_interval(begin, size, wrote);
4602 spa_config_exit(spa, SCL_L2ARC, dev);
4603 }
4604
4605 l2arc_thread_exit = 0;
4606 cv_broadcast(&l2arc_feed_thr_cv);
4607 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4608 thread_exit();
4609 }
4610
4611 boolean_t
4612 l2arc_vdev_present(vdev_t *vd)
4613 {
4614 l2arc_dev_t *dev;
4615
4616 mutex_enter(&l2arc_dev_mtx);
4617 for (dev = list_head(l2arc_dev_list); dev != NULL;
4618 dev = list_next(l2arc_dev_list, dev)) {
4619 if (dev->l2ad_vdev == vd)
4620 break;
4621 }
4622 mutex_exit(&l2arc_dev_mtx);
4623
4624 return (dev != NULL);
4625 }
4626
4627 /*
4628 * Add a vdev for use by the L2ARC. By this point the spa has already
4629 * validated the vdev and opened it.
4630 */
4631 void
4632 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
4633 {
4634 l2arc_dev_t *adddev;
4635
4636 ASSERT(!l2arc_vdev_present(vd));
4637
4638 /*
4639 * Create a new l2arc device entry.
4640 */
4641 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
4642 adddev->l2ad_spa = spa;
4643 adddev->l2ad_vdev = vd;
4644 adddev->l2ad_write = l2arc_write_max;
4645 adddev->l2ad_boost = l2arc_write_boost;
4646 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
4647 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
4648 adddev->l2ad_hand = adddev->l2ad_start;
4649 adddev->l2ad_evict = adddev->l2ad_start;
4650 adddev->l2ad_first = B_TRUE;
4651 adddev->l2ad_writing = B_FALSE;
4652 ASSERT3U(adddev->l2ad_write, >, 0);
4653
4654 /*
4655 * This is a list of all ARC buffers that are still valid on the
4656 * device.
4657 */
4658 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
4659 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
4660 offsetof(arc_buf_hdr_t, b_l2node));
4661
4662 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
4663
4664 /*
4665 * Add device to global list
4666 */
4667 mutex_enter(&l2arc_dev_mtx);
4668 list_insert_head(l2arc_dev_list, adddev);
4669 atomic_inc_64(&l2arc_ndev);
4670 mutex_exit(&l2arc_dev_mtx);
4671 }
4672
4673 /*
4674 * Remove a vdev from the L2ARC.
4675 */
4676 void
4677 l2arc_remove_vdev(vdev_t *vd)
4678 {
4679 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
4680
4681 /*
4682 * Find the device by vdev
4683 */
4684 mutex_enter(&l2arc_dev_mtx);
4685 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
4686 nextdev = list_next(l2arc_dev_list, dev);
4687 if (vd == dev->l2ad_vdev) {
4688 remdev = dev;
4689 break;
4690 }
4691 }
4692 ASSERT(remdev != NULL);
4693
4694 /*
4695 * Remove device from global list
4696 */
4697 list_remove(l2arc_dev_list, remdev);
4698 l2arc_dev_last = NULL; /* may have been invalidated */
4699 atomic_dec_64(&l2arc_ndev);
4700 mutex_exit(&l2arc_dev_mtx);
4701
4702 /*
4703 * Clear all buflists and ARC references. L2ARC device flush.
4704 */
4705 l2arc_evict(remdev, 0, B_TRUE);
4706 list_destroy(remdev->l2ad_buflist);
4707 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
4708 kmem_free(remdev, sizeof (l2arc_dev_t));
4709 }
4710
4711 void
4712 l2arc_init(void)
4713 {
4714 l2arc_thread_exit = 0;
4715 l2arc_ndev = 0;
4716 l2arc_writes_sent = 0;
4717 l2arc_writes_done = 0;
4718
4719 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4720 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
4721 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
4722 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
4723 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
4724
4725 l2arc_dev_list = &L2ARC_dev_list;
4726 l2arc_free_on_write = &L2ARC_free_on_write;
4727 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
4728 offsetof(l2arc_dev_t, l2ad_node));
4729 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
4730 offsetof(l2arc_data_free_t, l2df_list_node));
4731 }
4732
4733 void
4734 l2arc_fini(void)
4735 {
4736 /*
4737 * This is called from dmu_fini(), which is called from spa_fini();
4738 * Because of this, we can assume that all l2arc devices have
4739 * already been removed when the pools themselves were removed.
4740 */
4741
4742 l2arc_do_free_on_write();
4743
4744 mutex_destroy(&l2arc_feed_thr_lock);
4745 cv_destroy(&l2arc_feed_thr_cv);
4746 mutex_destroy(&l2arc_dev_mtx);
4747 mutex_destroy(&l2arc_buflist_mtx);
4748 mutex_destroy(&l2arc_free_on_write_mtx);
4749
4750 list_destroy(l2arc_dev_list);
4751 list_destroy(l2arc_free_on_write);
4752 }
4753
4754 void
4755 l2arc_start(void)
4756 {
4757 if (!(spa_mode_global & FWRITE))
4758 return;
4759
4760 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
4761 TS_RUN, minclsyspri);
4762 }
4763
4764 void
4765 l2arc_stop(void)
4766 {
4767 if (!(spa_mode_global & FWRITE))
4768 return;
4769
4770 mutex_enter(&l2arc_feed_thr_lock);
4771 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
4772 l2arc_thread_exit = 1;
4773 while (l2arc_thread_exit != 0)
4774 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
4775 mutex_exit(&l2arc_feed_thr_lock);
4776 }