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 (c) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright 2013 Nexenta Systems, Inc. All rights reserved.
27 */
28
29 /*
30 * DVA-based Adjustable Replacement Cache
31 *
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
36 *
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
47 *
48 * There are times when it is not possible to evict the requested
49 * space. In these circumstances we are unable to adjust the cache
50 * size. To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
53 *
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss. Our model has a variable sized cache. It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
59 * tight.
60 *
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size. So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict. In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
69 *
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
72 */
73
74 /*
75 * The locking model:
76 *
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists. The arc_read() interface
80 * uses method 1, while the internal arc algorithms for
81 * adjusting the cache use method 2. We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
83 * arc list locks.
84 *
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
88 *
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table. It returns
91 * NULL for the mutex if the buffer was not in the table.
92 *
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
95 *
96 * Each arc state also has a mutex which is used to protect the
97 * buffer list associated with the state. When attempting to
98 * obtain a hash table lock while holding an arc list lock you
99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 * the active state mutex must be held before the ghost state mutex.
101 *
102 * Arc buffers may have an associated eviction callback function.
103 * This function will be invoked prior to removing the buffer (e.g.
104 * in arc_do_user_evicts()). Note however that the data associated
105 * with the buffer may be evicted prior to the callback. The callback
106 * must be made with *no locks held* (to prevent deadlock). Additionally,
107 * the users of callbacks must ensure that their private data is
108 * protected from simultaneous callbacks from arc_buf_evict()
109 * and arc_do_user_evicts().
110 *
111 * Note that the majority of the performance stats are manipulated
112 * with atomic operations.
113 *
114 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
115 *
116 * - L2ARC buflist creation
117 * - L2ARC buflist eviction
118 * - L2ARC write completion, which walks L2ARC buflists
119 * - ARC header destruction, as it removes from L2ARC buflists
120 * - ARC header release, as it removes from L2ARC buflists
121 */
122
123 #include <sys/spa.h>
124 #include <sys/zio.h>
125 #include <sys/zio_compress.h>
126 #include <sys/zfs_context.h>
127 #include <sys/arc.h>
128 #include <sys/refcount.h>
129 #include <sys/vdev.h>
130 #include <sys/vdev_impl.h>
131 #include <sys/dsl_pool.h>
132 #ifdef _KERNEL
133 #include <sys/vmsystm.h>
134 #include <vm/anon.h>
135 #include <sys/fs/swapnode.h>
136 #include <sys/dnlc.h>
137 #endif
138 #include <sys/callb.h>
139 #include <sys/kstat.h>
140 #include <zfs_fletcher.h>
141
142 #ifndef _KERNEL
143 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
144 boolean_t arc_watch = B_FALSE;
145 int arc_procfd;
146 #endif
147
148 static kmutex_t arc_reclaim_thr_lock;
149 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
150 static uint8_t arc_thread_exit;
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 /*
161 * The number of iterations through arc_evict_*() before we
162 * drop & reacquire the lock.
163 */
164 int arc_evict_iterations = 100;
165
166 /* number of seconds before growing cache again */
167 static int arc_grow_retry = 60;
168
169 /* shift of arc_c for calculating both min and max arc_p */
170 static int arc_p_min_shift = 4;
171
172 /* log2(fraction of arc to reclaim) */
173 static int arc_shrink_shift = 5;
174
175 /*
176 * minimum lifespan of a prefetch block in clock ticks
177 * (initialized in arc_init())
178 */
179 static int arc_min_prefetch_lifespan;
180
181 /*
182 * If this percent of memory is free, don't throttle.
183 */
184 int arc_lotsfree_percent = 10;
185
186 static int arc_dead;
187
188 /*
189 * The arc has filled available memory and has now warmed up.
190 */
191 static boolean_t arc_warm;
192
193 /*
194 * These tunables are for performance analysis.
195 */
196 uint64_t zfs_arc_max;
197 uint64_t zfs_arc_min;
198 uint64_t zfs_arc_meta_limit = 0;
199 int zfs_arc_grow_retry = 0;
200 int zfs_arc_shrink_shift = 0;
201 int zfs_arc_p_min_shift = 0;
202 int zfs_disable_dup_eviction = 0;
203
204 /*
205 * Note that buffers can be in one of 6 states:
206 * ARC_anon - anonymous (discussed below)
207 * ARC_mru - recently used, currently cached
208 * ARC_mru_ghost - recentely used, no longer in cache
209 * ARC_mfu - frequently used, currently cached
210 * ARC_mfu_ghost - frequently used, no longer in cache
211 * ARC_l2c_only - exists in L2ARC but not other states
212 * When there are no active references to the buffer, they are
213 * are linked onto a list in one of these arc states. These are
214 * the only buffers that can be evicted or deleted. Within each
215 * state there are multiple lists, one for meta-data and one for
216 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
217 * etc.) is tracked separately so that it can be managed more
218 * explicitly: favored over data, limited explicitly.
219 *
220 * Anonymous buffers are buffers that are not associated with
221 * a DVA. These are buffers that hold dirty block copies
222 * before they are written to stable storage. By definition,
223 * they are "ref'd" and are considered part of arc_mru
224 * that cannot be freed. Generally, they will aquire a DVA
225 * as they are written and migrate onto the arc_mru list.
226 *
227 * The ARC_l2c_only state is for buffers that are in the second
228 * level ARC but no longer in any of the ARC_m* lists. The second
229 * level ARC itself may also contain buffers that are in any of
230 * the ARC_m* states - meaning that a buffer can exist in two
231 * places. The reason for the ARC_l2c_only state is to keep the
232 * buffer header in the hash table, so that reads that hit the
233 * second level ARC benefit from these fast lookups.
234 */
235
236 typedef struct arc_state {
237 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
238 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
239 uint64_t arcs_size; /* total amount of data in this state */
240 kmutex_t arcs_mtx;
241 } arc_state_t;
242
243 /* The 6 states: */
244 static arc_state_t ARC_anon;
245 static arc_state_t ARC_mru;
246 static arc_state_t ARC_mru_ghost;
247 static arc_state_t ARC_mfu;
248 static arc_state_t ARC_mfu_ghost;
249 static arc_state_t ARC_l2c_only;
250
251 typedef struct arc_stats {
252 kstat_named_t arcstat_hits;
253 kstat_named_t arcstat_misses;
254 kstat_named_t arcstat_demand_data_hits;
255 kstat_named_t arcstat_demand_data_misses;
256 kstat_named_t arcstat_demand_metadata_hits;
257 kstat_named_t arcstat_demand_metadata_misses;
258 kstat_named_t arcstat_prefetch_data_hits;
259 kstat_named_t arcstat_prefetch_data_misses;
260 kstat_named_t arcstat_prefetch_metadata_hits;
261 kstat_named_t arcstat_prefetch_metadata_misses;
262 kstat_named_t arcstat_mru_hits;
263 kstat_named_t arcstat_mru_ghost_hits;
264 kstat_named_t arcstat_mfu_hits;
265 kstat_named_t arcstat_mfu_ghost_hits;
266 kstat_named_t arcstat_deleted;
267 kstat_named_t arcstat_recycle_miss;
268 /*
269 * Number of buffers that could not be evicted because the hash lock
270 * was held by another thread. The lock may not necessarily be held
271 * by something using the same buffer, since hash locks are shared
272 * by multiple buffers.
273 */
274 kstat_named_t arcstat_mutex_miss;
275 /*
276 * Number of buffers skipped because they have I/O in progress, are
277 * indrect prefetch buffers that have not lived long enough, or are
278 * not from the spa we're trying to evict from.
279 */
280 kstat_named_t arcstat_evict_skip;
281 kstat_named_t arcstat_evict_l2_cached;
282 kstat_named_t arcstat_evict_l2_eligible;
283 kstat_named_t arcstat_evict_l2_ineligible;
284 kstat_named_t arcstat_hash_elements;
285 kstat_named_t arcstat_hash_elements_max;
286 kstat_named_t arcstat_hash_collisions;
287 kstat_named_t arcstat_hash_chains;
288 kstat_named_t arcstat_hash_chain_max;
289 kstat_named_t arcstat_p;
290 kstat_named_t arcstat_c;
291 kstat_named_t arcstat_c_min;
292 kstat_named_t arcstat_c_max;
293 kstat_named_t arcstat_size;
294 kstat_named_t arcstat_hdr_size;
295 kstat_named_t arcstat_data_size;
296 kstat_named_t arcstat_other_size;
297 kstat_named_t arcstat_l2_hits;
298 kstat_named_t arcstat_l2_misses;
299 kstat_named_t arcstat_l2_feeds;
300 kstat_named_t arcstat_l2_rw_clash;
301 kstat_named_t arcstat_l2_read_bytes;
302 kstat_named_t arcstat_l2_write_bytes;
303 kstat_named_t arcstat_l2_writes_sent;
304 kstat_named_t arcstat_l2_writes_done;
305 kstat_named_t arcstat_l2_writes_error;
306 kstat_named_t arcstat_l2_writes_hdr_miss;
307 kstat_named_t arcstat_l2_evict_lock_retry;
308 kstat_named_t arcstat_l2_evict_reading;
309 kstat_named_t arcstat_l2_free_on_write;
310 kstat_named_t arcstat_l2_abort_lowmem;
311 kstat_named_t arcstat_l2_cksum_bad;
312 kstat_named_t arcstat_l2_io_error;
313 kstat_named_t arcstat_l2_size;
314 kstat_named_t arcstat_l2_asize;
315 kstat_named_t arcstat_l2_hdr_size;
316 kstat_named_t arcstat_l2_compress_successes;
317 kstat_named_t arcstat_l2_compress_zeros;
318 kstat_named_t arcstat_l2_compress_failures;
319 kstat_named_t arcstat_memory_throttle_count;
320 kstat_named_t arcstat_duplicate_buffers;
321 kstat_named_t arcstat_duplicate_buffers_size;
322 kstat_named_t arcstat_duplicate_reads;
323 kstat_named_t arcstat_meta_used;
324 kstat_named_t arcstat_meta_limit;
325 kstat_named_t arcstat_meta_max;
326 } arc_stats_t;
327
328 static arc_stats_t arc_stats = {
329 { "hits", KSTAT_DATA_UINT64 },
330 { "misses", KSTAT_DATA_UINT64 },
331 { "demand_data_hits", KSTAT_DATA_UINT64 },
332 { "demand_data_misses", KSTAT_DATA_UINT64 },
333 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
334 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
335 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
336 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
337 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
338 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
339 { "mru_hits", KSTAT_DATA_UINT64 },
340 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
341 { "mfu_hits", KSTAT_DATA_UINT64 },
342 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
343 { "deleted", KSTAT_DATA_UINT64 },
344 { "recycle_miss", KSTAT_DATA_UINT64 },
345 { "mutex_miss", KSTAT_DATA_UINT64 },
346 { "evict_skip", KSTAT_DATA_UINT64 },
347 { "evict_l2_cached", KSTAT_DATA_UINT64 },
348 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
349 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
350 { "hash_elements", KSTAT_DATA_UINT64 },
351 { "hash_elements_max", KSTAT_DATA_UINT64 },
352 { "hash_collisions", KSTAT_DATA_UINT64 },
353 { "hash_chains", KSTAT_DATA_UINT64 },
354 { "hash_chain_max", KSTAT_DATA_UINT64 },
355 { "p", KSTAT_DATA_UINT64 },
356 { "c", KSTAT_DATA_UINT64 },
357 { "c_min", KSTAT_DATA_UINT64 },
358 { "c_max", KSTAT_DATA_UINT64 },
359 { "size", KSTAT_DATA_UINT64 },
360 { "hdr_size", KSTAT_DATA_UINT64 },
361 { "data_size", KSTAT_DATA_UINT64 },
362 { "other_size", KSTAT_DATA_UINT64 },
363 { "l2_hits", KSTAT_DATA_UINT64 },
364 { "l2_misses", KSTAT_DATA_UINT64 },
365 { "l2_feeds", KSTAT_DATA_UINT64 },
366 { "l2_rw_clash", KSTAT_DATA_UINT64 },
367 { "l2_read_bytes", KSTAT_DATA_UINT64 },
368 { "l2_write_bytes", KSTAT_DATA_UINT64 },
369 { "l2_writes_sent", KSTAT_DATA_UINT64 },
370 { "l2_writes_done", KSTAT_DATA_UINT64 },
371 { "l2_writes_error", KSTAT_DATA_UINT64 },
372 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
373 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
374 { "l2_evict_reading", KSTAT_DATA_UINT64 },
375 { "l2_free_on_write", KSTAT_DATA_UINT64 },
376 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
377 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
378 { "l2_io_error", KSTAT_DATA_UINT64 },
379 { "l2_size", KSTAT_DATA_UINT64 },
380 { "l2_asize", KSTAT_DATA_UINT64 },
381 { "l2_hdr_size", KSTAT_DATA_UINT64 },
382 { "l2_compress_successes", KSTAT_DATA_UINT64 },
383 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
384 { "l2_compress_failures", KSTAT_DATA_UINT64 },
385 { "memory_throttle_count", KSTAT_DATA_UINT64 },
386 { "duplicate_buffers", KSTAT_DATA_UINT64 },
387 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
388 { "duplicate_reads", KSTAT_DATA_UINT64 },
389 { "arc_meta_used", KSTAT_DATA_UINT64 },
390 { "arc_meta_limit", KSTAT_DATA_UINT64 },
391 { "arc_meta_max", KSTAT_DATA_UINT64 }
392 };
393
394 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
395
396 #define ARCSTAT_INCR(stat, val) \
397 atomic_add_64(&arc_stats.stat.value.ui64, (val))
398
399 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
400 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
401
402 #define ARCSTAT_MAX(stat, val) { \
403 uint64_t m; \
404 while ((val) > (m = arc_stats.stat.value.ui64) && \
405 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
406 continue; \
407 }
408
409 #define ARCSTAT_MAXSTAT(stat) \
410 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
411
412 /*
413 * We define a macro to allow ARC hits/misses to be easily broken down by
414 * two separate conditions, giving a total of four different subtypes for
415 * each of hits and misses (so eight statistics total).
416 */
417 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
418 if (cond1) { \
419 if (cond2) { \
420 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
421 } else { \
422 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
423 } \
424 } else { \
425 if (cond2) { \
426 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
427 } else { \
428 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
429 } \
430 }
431
432 kstat_t *arc_ksp;
433 static arc_state_t *arc_anon;
434 static arc_state_t *arc_mru;
435 static arc_state_t *arc_mru_ghost;
436 static arc_state_t *arc_mfu;
437 static arc_state_t *arc_mfu_ghost;
438 static arc_state_t *arc_l2c_only;
439
440 /*
441 * There are several ARC variables that are critical to export as kstats --
442 * but we don't want to have to grovel around in the kstat whenever we wish to
443 * manipulate them. For these variables, we therefore define them to be in
444 * terms of the statistic variable. This assures that we are not introducing
445 * the possibility of inconsistency by having shadow copies of the variables,
446 * while still allowing the code to be readable.
447 */
448 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
449 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
450 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
451 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
452 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
453 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
454 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
455 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
456
457 #define L2ARC_IS_VALID_COMPRESS(_c_) \
458 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
459
460 static int arc_no_grow; /* Don't try to grow cache size */
461 static uint64_t arc_tempreserve;
462 static uint64_t arc_loaned_bytes;
463
464 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
465
466 typedef struct arc_callback arc_callback_t;
467
468 struct arc_callback {
469 void *acb_private;
470 arc_done_func_t *acb_done;
471 arc_buf_t *acb_buf;
472 zio_t *acb_zio_dummy;
473 arc_callback_t *acb_next;
474 };
475
476 typedef struct arc_write_callback arc_write_callback_t;
477
478 struct arc_write_callback {
479 void *awcb_private;
480 arc_done_func_t *awcb_ready;
481 arc_done_func_t *awcb_physdone;
482 arc_done_func_t *awcb_done;
483 arc_buf_t *awcb_buf;
484 };
485
486 struct arc_buf_hdr {
487 /* protected by hash lock */
488 dva_t b_dva;
489 uint64_t b_birth;
490 uint64_t b_cksum0;
491
492 kmutex_t b_freeze_lock;
493 zio_cksum_t *b_freeze_cksum;
494 void *b_thawed;
495
496 arc_buf_hdr_t *b_hash_next;
497 arc_buf_t *b_buf;
498 uint32_t b_flags;
499 uint32_t b_datacnt;
500
501 arc_callback_t *b_acb;
502 kcondvar_t b_cv;
503
504 /* immutable */
505 arc_buf_contents_t b_type;
506 uint64_t b_size;
507 uint64_t b_spa;
508
509 /* protected by arc state mutex */
510 arc_state_t *b_state;
511 list_node_t b_arc_node;
512
513 /* updated atomically */
514 clock_t b_arc_access;
515
516 /* self protecting */
517 refcount_t b_refcnt;
518
519 l2arc_buf_hdr_t *b_l2hdr;
520 list_node_t b_l2node;
521 };
522
523 static arc_buf_t *arc_eviction_list;
524 static kmutex_t arc_eviction_mtx;
525 static arc_buf_hdr_t arc_eviction_hdr;
526 static void arc_get_data_buf(arc_buf_t *buf);
527 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
528 static int arc_evict_needed(arc_buf_contents_t type);
529 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
530 static void arc_buf_watch(arc_buf_t *buf);
531
532 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
533
534 #define GHOST_STATE(state) \
535 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
536 (state) == arc_l2c_only)
537
538 /*
539 * Private ARC flags. These flags are private ARC only flags that will show up
540 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
541 * be passed in as arc_flags in things like arc_read. However, these flags
542 * should never be passed and should only be set by ARC code. When adding new
543 * public flags, make sure not to smash the private ones.
544 */
545
546 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
547 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
548 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
549 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
550 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
551 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
552 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
553 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
554 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
555 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
556
557 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
558 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
559 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
560 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
561 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
562 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
563 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
564 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
565 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
566 (hdr)->b_l2hdr != NULL)
567 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
568 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
569 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
570
571 /*
572 * Other sizes
573 */
574
575 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
576 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
577
578 /*
579 * Hash table routines
580 */
581
582 #define HT_LOCK_PAD 64
583
584 struct ht_lock {
585 kmutex_t ht_lock;
586 #ifdef _KERNEL
587 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
588 #endif
589 };
590
591 #define BUF_LOCKS 256
592 typedef struct buf_hash_table {
593 uint64_t ht_mask;
594 arc_buf_hdr_t **ht_table;
595 struct ht_lock ht_locks[BUF_LOCKS];
596 } buf_hash_table_t;
597
598 static buf_hash_table_t buf_hash_table;
599
600 #define BUF_HASH_INDEX(spa, dva, birth) \
601 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
602 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
603 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
604 #define HDR_LOCK(hdr) \
605 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
606
607 uint64_t zfs_crc64_table[256];
608
609 /*
610 * Level 2 ARC
611 */
612
613 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
614 #define L2ARC_HEADROOM 2 /* num of writes */
615 /*
616 * If we discover during ARC scan any buffers to be compressed, we boost
617 * our headroom for the next scanning cycle by this percentage multiple.
618 */
619 #define L2ARC_HEADROOM_BOOST 200
620 #define L2ARC_FEED_SECS 1 /* caching interval secs */
621 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
622
623 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
624 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
625
626 /* L2ARC Performance Tunables */
627 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
628 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
629 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
630 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
631 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
632 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
633 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
634 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
635 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
636
637 /*
638 * L2ARC Internals
639 */
640 typedef struct l2arc_dev {
641 vdev_t *l2ad_vdev; /* vdev */
642 spa_t *l2ad_spa; /* spa */
643 uint64_t l2ad_hand; /* next write location */
644 uint64_t l2ad_start; /* first addr on device */
645 uint64_t l2ad_end; /* last addr on device */
646 uint64_t l2ad_evict; /* last addr eviction reached */
647 boolean_t l2ad_first; /* first sweep through */
648 boolean_t l2ad_writing; /* currently writing */
649 list_t *l2ad_buflist; /* buffer list */
650 list_node_t l2ad_node; /* device list node */
651 } l2arc_dev_t;
652
653 static list_t L2ARC_dev_list; /* device list */
654 static list_t *l2arc_dev_list; /* device list pointer */
655 static kmutex_t l2arc_dev_mtx; /* device list mutex */
656 static l2arc_dev_t *l2arc_dev_last; /* last device used */
657 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
658 static list_t L2ARC_free_on_write; /* free after write buf list */
659 static list_t *l2arc_free_on_write; /* free after write list ptr */
660 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
661 static uint64_t l2arc_ndev; /* number of devices */
662
663 typedef struct l2arc_read_callback {
664 arc_buf_t *l2rcb_buf; /* read buffer */
665 spa_t *l2rcb_spa; /* spa */
666 blkptr_t l2rcb_bp; /* original blkptr */
667 zbookmark_t l2rcb_zb; /* original bookmark */
668 int l2rcb_flags; /* original flags */
669 enum zio_compress l2rcb_compress; /* applied compress */
670 } l2arc_read_callback_t;
671
672 typedef struct l2arc_write_callback {
673 l2arc_dev_t *l2wcb_dev; /* device info */
674 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
675 } l2arc_write_callback_t;
676
677 struct l2arc_buf_hdr {
678 /* protected by arc_buf_hdr mutex */
679 l2arc_dev_t *b_dev; /* L2ARC device */
680 uint64_t b_daddr; /* disk address, offset byte */
681 /* compression applied to buffer data */
682 enum zio_compress b_compress;
683 /* real alloc'd buffer size depending on b_compress applied */
684 int b_asize;
685 /* temporary buffer holder for in-flight compressed data */
686 void *b_tmp_cdata;
687 };
688
689 typedef struct l2arc_data_free {
690 /* protected by l2arc_free_on_write_mtx */
691 void *l2df_data;
692 size_t l2df_size;
693 void (*l2df_func)(void *, size_t);
694 list_node_t l2df_list_node;
695 } l2arc_data_free_t;
696
697 static kmutex_t l2arc_feed_thr_lock;
698 static kcondvar_t l2arc_feed_thr_cv;
699 static uint8_t l2arc_thread_exit;
700
701 static void l2arc_read_done(zio_t *zio);
702 static void l2arc_hdr_stat_add(void);
703 static void l2arc_hdr_stat_remove(void);
704
705 static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr);
706 static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr,
707 enum zio_compress c);
708 static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab);
709
710 static uint64_t
711 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
712 {
713 uint8_t *vdva = (uint8_t *)dva;
714 uint64_t crc = -1ULL;
715 int i;
716
717 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
718
719 for (i = 0; i < sizeof (dva_t); i++)
720 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
721
722 crc ^= (spa>>8) ^ birth;
723
724 return (crc);
725 }
726
727 #define BUF_EMPTY(buf) \
728 ((buf)->b_dva.dva_word[0] == 0 && \
729 (buf)->b_dva.dva_word[1] == 0 && \
730 (buf)->b_birth == 0)
731
732 #define BUF_EQUAL(spa, dva, birth, buf) \
733 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
734 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
735 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
736
737 static void
738 buf_discard_identity(arc_buf_hdr_t *hdr)
739 {
740 hdr->b_dva.dva_word[0] = 0;
741 hdr->b_dva.dva_word[1] = 0;
742 hdr->b_birth = 0;
743 hdr->b_cksum0 = 0;
744 }
745
746 static arc_buf_hdr_t *
747 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
748 {
749 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
750 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
751 arc_buf_hdr_t *buf;
752
753 mutex_enter(hash_lock);
754 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
755 buf = buf->b_hash_next) {
756 if (BUF_EQUAL(spa, dva, birth, buf)) {
757 *lockp = hash_lock;
758 return (buf);
759 }
760 }
761 mutex_exit(hash_lock);
762 *lockp = NULL;
763 return (NULL);
764 }
765
766 /*
767 * Insert an entry into the hash table. If there is already an element
768 * equal to elem in the hash table, then the already existing element
769 * will be returned and the new element will not be inserted.
770 * Otherwise returns NULL.
771 */
772 static arc_buf_hdr_t *
773 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
774 {
775 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
776 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
777 arc_buf_hdr_t *fbuf;
778 uint32_t i;
779
780 ASSERT(!HDR_IN_HASH_TABLE(buf));
781 *lockp = hash_lock;
782 mutex_enter(hash_lock);
783 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
784 fbuf = fbuf->b_hash_next, i++) {
785 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
786 return (fbuf);
787 }
788
789 buf->b_hash_next = buf_hash_table.ht_table[idx];
790 buf_hash_table.ht_table[idx] = buf;
791 buf->b_flags |= ARC_IN_HASH_TABLE;
792
793 /* collect some hash table performance data */
794 if (i > 0) {
795 ARCSTAT_BUMP(arcstat_hash_collisions);
796 if (i == 1)
797 ARCSTAT_BUMP(arcstat_hash_chains);
798
799 ARCSTAT_MAX(arcstat_hash_chain_max, i);
800 }
801
802 ARCSTAT_BUMP(arcstat_hash_elements);
803 ARCSTAT_MAXSTAT(arcstat_hash_elements);
804
805 return (NULL);
806 }
807
808 static void
809 buf_hash_remove(arc_buf_hdr_t *buf)
810 {
811 arc_buf_hdr_t *fbuf, **bufp;
812 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
813
814 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
815 ASSERT(HDR_IN_HASH_TABLE(buf));
816
817 bufp = &buf_hash_table.ht_table[idx];
818 while ((fbuf = *bufp) != buf) {
819 ASSERT(fbuf != NULL);
820 bufp = &fbuf->b_hash_next;
821 }
822 *bufp = buf->b_hash_next;
823 buf->b_hash_next = NULL;
824 buf->b_flags &= ~ARC_IN_HASH_TABLE;
825
826 /* collect some hash table performance data */
827 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
828
829 if (buf_hash_table.ht_table[idx] &&
830 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
831 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
832 }
833
834 /*
835 * Global data structures and functions for the buf kmem cache.
836 */
837 static kmem_cache_t *hdr_cache;
838 static kmem_cache_t *buf_cache;
839
840 static void
841 buf_fini(void)
842 {
843 int i;
844
845 kmem_free(buf_hash_table.ht_table,
846 (buf_hash_table.ht_mask + 1) * sizeof (void *));
847 for (i = 0; i < BUF_LOCKS; i++)
848 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
849 kmem_cache_destroy(hdr_cache);
850 kmem_cache_destroy(buf_cache);
851 }
852
853 /*
854 * Constructor callback - called when the cache is empty
855 * and a new buf is requested.
856 */
857 /* ARGSUSED */
858 static int
859 hdr_cons(void *vbuf, void *unused, int kmflag)
860 {
861 arc_buf_hdr_t *buf = vbuf;
862
863 bzero(buf, sizeof (arc_buf_hdr_t));
864 refcount_create(&buf->b_refcnt);
865 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
866 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
867 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
868
869 return (0);
870 }
871
872 /* ARGSUSED */
873 static int
874 buf_cons(void *vbuf, void *unused, int kmflag)
875 {
876 arc_buf_t *buf = vbuf;
877
878 bzero(buf, sizeof (arc_buf_t));
879 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
880 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
881
882 return (0);
883 }
884
885 /*
886 * Destructor callback - called when a cached buf is
887 * no longer required.
888 */
889 /* ARGSUSED */
890 static void
891 hdr_dest(void *vbuf, void *unused)
892 {
893 arc_buf_hdr_t *buf = vbuf;
894
895 ASSERT(BUF_EMPTY(buf));
896 refcount_destroy(&buf->b_refcnt);
897 cv_destroy(&buf->b_cv);
898 mutex_destroy(&buf->b_freeze_lock);
899 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
900 }
901
902 /* ARGSUSED */
903 static void
904 buf_dest(void *vbuf, void *unused)
905 {
906 arc_buf_t *buf = vbuf;
907
908 mutex_destroy(&buf->b_evict_lock);
909 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
910 }
911
912 /*
913 * Reclaim callback -- invoked when memory is low.
914 */
915 /* ARGSUSED */
916 static void
917 hdr_recl(void *unused)
918 {
919 dprintf("hdr_recl called\n");
920 /*
921 * umem calls the reclaim func when we destroy the buf cache,
922 * which is after we do arc_fini().
923 */
924 if (!arc_dead)
925 cv_signal(&arc_reclaim_thr_cv);
926 }
927
928 static void
929 buf_init(void)
930 {
931 uint64_t *ct;
932 uint64_t hsize = 1ULL << 12;
933 int i, j;
934
935 /*
936 * The hash table is big enough to fill all of physical memory
937 * with an average 64K block size. The table will take up
938 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
939 */
940 while (hsize * 65536 < physmem * PAGESIZE)
941 hsize <<= 1;
942 retry:
943 buf_hash_table.ht_mask = hsize - 1;
944 buf_hash_table.ht_table =
945 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
946 if (buf_hash_table.ht_table == NULL) {
947 ASSERT(hsize > (1ULL << 8));
948 hsize >>= 1;
949 goto retry;
950 }
951
952 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
953 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
954 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
955 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
956
957 for (i = 0; i < 256; i++)
958 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
959 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
960
961 for (i = 0; i < BUF_LOCKS; i++) {
962 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
963 NULL, MUTEX_DEFAULT, NULL);
964 }
965 }
966
967 #define ARC_MINTIME (hz>>4) /* 62 ms */
968
969 static void
970 arc_cksum_verify(arc_buf_t *buf)
971 {
972 zio_cksum_t zc;
973
974 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
975 return;
976
977 mutex_enter(&buf->b_hdr->b_freeze_lock);
978 if (buf->b_hdr->b_freeze_cksum == NULL ||
979 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
980 mutex_exit(&buf->b_hdr->b_freeze_lock);
981 return;
982 }
983 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
984 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
985 panic("buffer modified while frozen!");
986 mutex_exit(&buf->b_hdr->b_freeze_lock);
987 }
988
989 static int
990 arc_cksum_equal(arc_buf_t *buf)
991 {
992 zio_cksum_t zc;
993 int equal;
994
995 mutex_enter(&buf->b_hdr->b_freeze_lock);
996 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
997 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
998 mutex_exit(&buf->b_hdr->b_freeze_lock);
999
1000 return (equal);
1001 }
1002
1003 static void
1004 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1005 {
1006 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1007 return;
1008
1009 mutex_enter(&buf->b_hdr->b_freeze_lock);
1010 if (buf->b_hdr->b_freeze_cksum != NULL) {
1011 mutex_exit(&buf->b_hdr->b_freeze_lock);
1012 return;
1013 }
1014 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1015 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1016 buf->b_hdr->b_freeze_cksum);
1017 mutex_exit(&buf->b_hdr->b_freeze_lock);
1018 arc_buf_watch(buf);
1019 }
1020
1021 #ifndef _KERNEL
1022 typedef struct procctl {
1023 long cmd;
1024 prwatch_t prwatch;
1025 } procctl_t;
1026 #endif
1027
1028 /* ARGSUSED */
1029 static void
1030 arc_buf_unwatch(arc_buf_t *buf)
1031 {
1032 #ifndef _KERNEL
1033 if (arc_watch) {
1034 int result;
1035 procctl_t ctl;
1036 ctl.cmd = PCWATCH;
1037 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1038 ctl.prwatch.pr_size = 0;
1039 ctl.prwatch.pr_wflags = 0;
1040 result = write(arc_procfd, &ctl, sizeof (ctl));
1041 ASSERT3U(result, ==, sizeof (ctl));
1042 }
1043 #endif
1044 }
1045
1046 /* ARGSUSED */
1047 static void
1048 arc_buf_watch(arc_buf_t *buf)
1049 {
1050 #ifndef _KERNEL
1051 if (arc_watch) {
1052 int result;
1053 procctl_t ctl;
1054 ctl.cmd = PCWATCH;
1055 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1056 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1057 ctl.prwatch.pr_wflags = WA_WRITE;
1058 result = write(arc_procfd, &ctl, sizeof (ctl));
1059 ASSERT3U(result, ==, sizeof (ctl));
1060 }
1061 #endif
1062 }
1063
1064 void
1065 arc_buf_thaw(arc_buf_t *buf)
1066 {
1067 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1068 if (buf->b_hdr->b_state != arc_anon)
1069 panic("modifying non-anon buffer!");
1070 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1071 panic("modifying buffer while i/o in progress!");
1072 arc_cksum_verify(buf);
1073 }
1074
1075 mutex_enter(&buf->b_hdr->b_freeze_lock);
1076 if (buf->b_hdr->b_freeze_cksum != NULL) {
1077 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1078 buf->b_hdr->b_freeze_cksum = NULL;
1079 }
1080
1081 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1082 if (buf->b_hdr->b_thawed)
1083 kmem_free(buf->b_hdr->b_thawed, 1);
1084 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1085 }
1086
1087 mutex_exit(&buf->b_hdr->b_freeze_lock);
1088
1089 arc_buf_unwatch(buf);
1090 }
1091
1092 void
1093 arc_buf_freeze(arc_buf_t *buf)
1094 {
1095 kmutex_t *hash_lock;
1096
1097 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1098 return;
1099
1100 hash_lock = HDR_LOCK(buf->b_hdr);
1101 mutex_enter(hash_lock);
1102
1103 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1104 buf->b_hdr->b_state == arc_anon);
1105 arc_cksum_compute(buf, B_FALSE);
1106 mutex_exit(hash_lock);
1107
1108 }
1109
1110 static void
1111 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1112 {
1113 ASSERT(MUTEX_HELD(hash_lock));
1114
1115 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1116 (ab->b_state != arc_anon)) {
1117 uint64_t delta = ab->b_size * ab->b_datacnt;
1118 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1119 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1120
1121 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1122 mutex_enter(&ab->b_state->arcs_mtx);
1123 ASSERT(list_link_active(&ab->b_arc_node));
1124 list_remove(list, ab);
1125 if (GHOST_STATE(ab->b_state)) {
1126 ASSERT0(ab->b_datacnt);
1127 ASSERT3P(ab->b_buf, ==, NULL);
1128 delta = ab->b_size;
1129 }
1130 ASSERT(delta > 0);
1131 ASSERT3U(*size, >=, delta);
1132 atomic_add_64(size, -delta);
1133 mutex_exit(&ab->b_state->arcs_mtx);
1134 /* remove the prefetch flag if we get a reference */
1135 if (ab->b_flags & ARC_PREFETCH)
1136 ab->b_flags &= ~ARC_PREFETCH;
1137 }
1138 }
1139
1140 static int
1141 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1142 {
1143 int cnt;
1144 arc_state_t *state = ab->b_state;
1145
1146 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1147 ASSERT(!GHOST_STATE(state));
1148
1149 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1150 (state != arc_anon)) {
1151 uint64_t *size = &state->arcs_lsize[ab->b_type];
1152
1153 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1154 mutex_enter(&state->arcs_mtx);
1155 ASSERT(!list_link_active(&ab->b_arc_node));
1156 list_insert_head(&state->arcs_list[ab->b_type], ab);
1157 ASSERT(ab->b_datacnt > 0);
1158 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1159 mutex_exit(&state->arcs_mtx);
1160 }
1161 return (cnt);
1162 }
1163
1164 /*
1165 * Move the supplied buffer to the indicated state. The mutex
1166 * for the buffer must be held by the caller.
1167 */
1168 static void
1169 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1170 {
1171 arc_state_t *old_state = ab->b_state;
1172 int64_t refcnt = refcount_count(&ab->b_refcnt);
1173 uint64_t from_delta, to_delta;
1174
1175 ASSERT(MUTEX_HELD(hash_lock));
1176 ASSERT3P(new_state, !=, old_state);
1177 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1178 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1179 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1180
1181 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1182
1183 /*
1184 * If this buffer is evictable, transfer it from the
1185 * old state list to the new state list.
1186 */
1187 if (refcnt == 0) {
1188 if (old_state != arc_anon) {
1189 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1190 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1191
1192 if (use_mutex)
1193 mutex_enter(&old_state->arcs_mtx);
1194
1195 ASSERT(list_link_active(&ab->b_arc_node));
1196 list_remove(&old_state->arcs_list[ab->b_type], ab);
1197
1198 /*
1199 * If prefetching out of the ghost cache,
1200 * we will have a non-zero datacnt.
1201 */
1202 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1203 /* ghost elements have a ghost size */
1204 ASSERT(ab->b_buf == NULL);
1205 from_delta = ab->b_size;
1206 }
1207 ASSERT3U(*size, >=, from_delta);
1208 atomic_add_64(size, -from_delta);
1209
1210 if (use_mutex)
1211 mutex_exit(&old_state->arcs_mtx);
1212 }
1213 if (new_state != arc_anon) {
1214 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1215 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1216
1217 if (use_mutex)
1218 mutex_enter(&new_state->arcs_mtx);
1219
1220 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1221
1222 /* ghost elements have a ghost size */
1223 if (GHOST_STATE(new_state)) {
1224 ASSERT(ab->b_datacnt == 0);
1225 ASSERT(ab->b_buf == NULL);
1226 to_delta = ab->b_size;
1227 }
1228 atomic_add_64(size, to_delta);
1229
1230 if (use_mutex)
1231 mutex_exit(&new_state->arcs_mtx);
1232 }
1233 }
1234
1235 ASSERT(!BUF_EMPTY(ab));
1236 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1237 buf_hash_remove(ab);
1238
1239 /* adjust state sizes */
1240 if (to_delta)
1241 atomic_add_64(&new_state->arcs_size, to_delta);
1242 if (from_delta) {
1243 ASSERT3U(old_state->arcs_size, >=, from_delta);
1244 atomic_add_64(&old_state->arcs_size, -from_delta);
1245 }
1246 ab->b_state = new_state;
1247
1248 /* adjust l2arc hdr stats */
1249 if (new_state == arc_l2c_only)
1250 l2arc_hdr_stat_add();
1251 else if (old_state == arc_l2c_only)
1252 l2arc_hdr_stat_remove();
1253 }
1254
1255 void
1256 arc_space_consume(uint64_t space, arc_space_type_t type)
1257 {
1258 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1259
1260 switch (type) {
1261 case ARC_SPACE_DATA:
1262 ARCSTAT_INCR(arcstat_data_size, space);
1263 break;
1264 case ARC_SPACE_OTHER:
1265 ARCSTAT_INCR(arcstat_other_size, space);
1266 break;
1267 case ARC_SPACE_HDRS:
1268 ARCSTAT_INCR(arcstat_hdr_size, space);
1269 break;
1270 case ARC_SPACE_L2HDRS:
1271 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1272 break;
1273 }
1274
1275 ARCSTAT_INCR(arcstat_meta_used, space);
1276 atomic_add_64(&arc_size, space);
1277 }
1278
1279 void
1280 arc_space_return(uint64_t space, arc_space_type_t type)
1281 {
1282 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1283
1284 switch (type) {
1285 case ARC_SPACE_DATA:
1286 ARCSTAT_INCR(arcstat_data_size, -space);
1287 break;
1288 case ARC_SPACE_OTHER:
1289 ARCSTAT_INCR(arcstat_other_size, -space);
1290 break;
1291 case ARC_SPACE_HDRS:
1292 ARCSTAT_INCR(arcstat_hdr_size, -space);
1293 break;
1294 case ARC_SPACE_L2HDRS:
1295 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1296 break;
1297 }
1298
1299 ASSERT(arc_meta_used >= space);
1300 if (arc_meta_max < arc_meta_used)
1301 arc_meta_max = arc_meta_used;
1302 ARCSTAT_INCR(arcstat_meta_used, -space);
1303 ASSERT(arc_size >= space);
1304 atomic_add_64(&arc_size, -space);
1305 }
1306
1307 void *
1308 arc_data_buf_alloc(uint64_t size)
1309 {
1310 if (arc_evict_needed(ARC_BUFC_DATA))
1311 cv_signal(&arc_reclaim_thr_cv);
1312 atomic_add_64(&arc_size, size);
1313 return (zio_data_buf_alloc(size));
1314 }
1315
1316 void
1317 arc_data_buf_free(void *buf, uint64_t size)
1318 {
1319 zio_data_buf_free(buf, size);
1320 ASSERT(arc_size >= size);
1321 atomic_add_64(&arc_size, -size);
1322 }
1323
1324 arc_buf_t *
1325 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1326 {
1327 arc_buf_hdr_t *hdr;
1328 arc_buf_t *buf;
1329
1330 ASSERT3U(size, >, 0);
1331 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1332 ASSERT(BUF_EMPTY(hdr));
1333 hdr->b_size = size;
1334 hdr->b_type = type;
1335 hdr->b_spa = spa_load_guid(spa);
1336 hdr->b_state = arc_anon;
1337 hdr->b_arc_access = 0;
1338 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1339 buf->b_hdr = hdr;
1340 buf->b_data = NULL;
1341 buf->b_efunc = NULL;
1342 buf->b_private = NULL;
1343 buf->b_next = NULL;
1344 hdr->b_buf = buf;
1345 arc_get_data_buf(buf);
1346 hdr->b_datacnt = 1;
1347 hdr->b_flags = 0;
1348 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1349 (void) refcount_add(&hdr->b_refcnt, tag);
1350
1351 return (buf);
1352 }
1353
1354 static char *arc_onloan_tag = "onloan";
1355
1356 /*
1357 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1358 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1359 * buffers must be returned to the arc before they can be used by the DMU or
1360 * freed.
1361 */
1362 arc_buf_t *
1363 arc_loan_buf(spa_t *spa, int size)
1364 {
1365 arc_buf_t *buf;
1366
1367 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1368
1369 atomic_add_64(&arc_loaned_bytes, size);
1370 return (buf);
1371 }
1372
1373 /*
1374 * Return a loaned arc buffer to the arc.
1375 */
1376 void
1377 arc_return_buf(arc_buf_t *buf, void *tag)
1378 {
1379 arc_buf_hdr_t *hdr = buf->b_hdr;
1380
1381 ASSERT(buf->b_data != NULL);
1382 (void) refcount_add(&hdr->b_refcnt, tag);
1383 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1384
1385 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1386 }
1387
1388 /* Detach an arc_buf from a dbuf (tag) */
1389 void
1390 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1391 {
1392 arc_buf_hdr_t *hdr;
1393
1394 ASSERT(buf->b_data != NULL);
1395 hdr = buf->b_hdr;
1396 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1397 (void) refcount_remove(&hdr->b_refcnt, tag);
1398 buf->b_efunc = NULL;
1399 buf->b_private = NULL;
1400
1401 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1402 }
1403
1404 static arc_buf_t *
1405 arc_buf_clone(arc_buf_t *from)
1406 {
1407 arc_buf_t *buf;
1408 arc_buf_hdr_t *hdr = from->b_hdr;
1409 uint64_t size = hdr->b_size;
1410
1411 ASSERT(hdr->b_state != arc_anon);
1412
1413 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1414 buf->b_hdr = hdr;
1415 buf->b_data = NULL;
1416 buf->b_efunc = NULL;
1417 buf->b_private = NULL;
1418 buf->b_next = hdr->b_buf;
1419 hdr->b_buf = buf;
1420 arc_get_data_buf(buf);
1421 bcopy(from->b_data, buf->b_data, size);
1422
1423 /*
1424 * This buffer already exists in the arc so create a duplicate
1425 * copy for the caller. If the buffer is associated with user data
1426 * then track the size and number of duplicates. These stats will be
1427 * updated as duplicate buffers are created and destroyed.
1428 */
1429 if (hdr->b_type == ARC_BUFC_DATA) {
1430 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1431 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1432 }
1433 hdr->b_datacnt += 1;
1434 return (buf);
1435 }
1436
1437 void
1438 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1439 {
1440 arc_buf_hdr_t *hdr;
1441 kmutex_t *hash_lock;
1442
1443 /*
1444 * Check to see if this buffer is evicted. Callers
1445 * must verify b_data != NULL to know if the add_ref
1446 * was successful.
1447 */
1448 mutex_enter(&buf->b_evict_lock);
1449 if (buf->b_data == NULL) {
1450 mutex_exit(&buf->b_evict_lock);
1451 return;
1452 }
1453 hash_lock = HDR_LOCK(buf->b_hdr);
1454 mutex_enter(hash_lock);
1455 hdr = buf->b_hdr;
1456 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1457 mutex_exit(&buf->b_evict_lock);
1458
1459 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1460 add_reference(hdr, hash_lock, tag);
1461 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1462 arc_access(hdr, hash_lock);
1463 mutex_exit(hash_lock);
1464 ARCSTAT_BUMP(arcstat_hits);
1465 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1466 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1467 data, metadata, hits);
1468 }
1469
1470 /*
1471 * Free the arc data buffer. If it is an l2arc write in progress,
1472 * the buffer is placed on l2arc_free_on_write to be freed later.
1473 */
1474 static void
1475 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1476 {
1477 arc_buf_hdr_t *hdr = buf->b_hdr;
1478
1479 if (HDR_L2_WRITING(hdr)) {
1480 l2arc_data_free_t *df;
1481 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1482 df->l2df_data = buf->b_data;
1483 df->l2df_size = hdr->b_size;
1484 df->l2df_func = free_func;
1485 mutex_enter(&l2arc_free_on_write_mtx);
1486 list_insert_head(l2arc_free_on_write, df);
1487 mutex_exit(&l2arc_free_on_write_mtx);
1488 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1489 } else {
1490 free_func(buf->b_data, hdr->b_size);
1491 }
1492 }
1493
1494 static void
1495 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1496 {
1497 arc_buf_t **bufp;
1498
1499 /* free up data associated with the buf */
1500 if (buf->b_data) {
1501 arc_state_t *state = buf->b_hdr->b_state;
1502 uint64_t size = buf->b_hdr->b_size;
1503 arc_buf_contents_t type = buf->b_hdr->b_type;
1504
1505 arc_cksum_verify(buf);
1506 arc_buf_unwatch(buf);
1507
1508 if (!recycle) {
1509 if (type == ARC_BUFC_METADATA) {
1510 arc_buf_data_free(buf, zio_buf_free);
1511 arc_space_return(size, ARC_SPACE_DATA);
1512 } else {
1513 ASSERT(type == ARC_BUFC_DATA);
1514 arc_buf_data_free(buf, zio_data_buf_free);
1515 ARCSTAT_INCR(arcstat_data_size, -size);
1516 atomic_add_64(&arc_size, -size);
1517 }
1518 }
1519 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1520 uint64_t *cnt = &state->arcs_lsize[type];
1521
1522 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1523 ASSERT(state != arc_anon);
1524
1525 ASSERT3U(*cnt, >=, size);
1526 atomic_add_64(cnt, -size);
1527 }
1528 ASSERT3U(state->arcs_size, >=, size);
1529 atomic_add_64(&state->arcs_size, -size);
1530 buf->b_data = NULL;
1531
1532 /*
1533 * If we're destroying a duplicate buffer make sure
1534 * that the appropriate statistics are updated.
1535 */
1536 if (buf->b_hdr->b_datacnt > 1 &&
1537 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1538 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1539 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1540 }
1541 ASSERT(buf->b_hdr->b_datacnt > 0);
1542 buf->b_hdr->b_datacnt -= 1;
1543 }
1544
1545 /* only remove the buf if requested */
1546 if (!all)
1547 return;
1548
1549 /* remove the buf from the hdr list */
1550 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1551 continue;
1552 *bufp = buf->b_next;
1553 buf->b_next = NULL;
1554
1555 ASSERT(buf->b_efunc == NULL);
1556
1557 /* clean up the buf */
1558 buf->b_hdr = NULL;
1559 kmem_cache_free(buf_cache, buf);
1560 }
1561
1562 static void
1563 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1564 {
1565 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1566 ASSERT3P(hdr->b_state, ==, arc_anon);
1567 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1568 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1569
1570 if (l2hdr != NULL) {
1571 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1572 /*
1573 * To prevent arc_free() and l2arc_evict() from
1574 * attempting to free the same buffer at the same time,
1575 * a FREE_IN_PROGRESS flag is given to arc_free() to
1576 * give it priority. l2arc_evict() can't destroy this
1577 * header while we are waiting on l2arc_buflist_mtx.
1578 *
1579 * The hdr may be removed from l2ad_buflist before we
1580 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1581 */
1582 if (!buflist_held) {
1583 mutex_enter(&l2arc_buflist_mtx);
1584 l2hdr = hdr->b_l2hdr;
1585 }
1586
1587 if (l2hdr != NULL) {
1588 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1589 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1590 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
1591 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1592 if (hdr->b_state == arc_l2c_only)
1593 l2arc_hdr_stat_remove();
1594 hdr->b_l2hdr = NULL;
1595 }
1596
1597 if (!buflist_held)
1598 mutex_exit(&l2arc_buflist_mtx);
1599 }
1600
1601 if (!BUF_EMPTY(hdr)) {
1602 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1603 buf_discard_identity(hdr);
1604 }
1605 while (hdr->b_buf) {
1606 arc_buf_t *buf = hdr->b_buf;
1607
1608 if (buf->b_efunc) {
1609 mutex_enter(&arc_eviction_mtx);
1610 mutex_enter(&buf->b_evict_lock);
1611 ASSERT(buf->b_hdr != NULL);
1612 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1613 hdr->b_buf = buf->b_next;
1614 buf->b_hdr = &arc_eviction_hdr;
1615 buf->b_next = arc_eviction_list;
1616 arc_eviction_list = buf;
1617 mutex_exit(&buf->b_evict_lock);
1618 mutex_exit(&arc_eviction_mtx);
1619 } else {
1620 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1621 }
1622 }
1623 if (hdr->b_freeze_cksum != NULL) {
1624 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1625 hdr->b_freeze_cksum = NULL;
1626 }
1627 if (hdr->b_thawed) {
1628 kmem_free(hdr->b_thawed, 1);
1629 hdr->b_thawed = NULL;
1630 }
1631
1632 ASSERT(!list_link_active(&hdr->b_arc_node));
1633 ASSERT3P(hdr->b_hash_next, ==, NULL);
1634 ASSERT3P(hdr->b_acb, ==, NULL);
1635 kmem_cache_free(hdr_cache, hdr);
1636 }
1637
1638 void
1639 arc_buf_free(arc_buf_t *buf, void *tag)
1640 {
1641 arc_buf_hdr_t *hdr = buf->b_hdr;
1642 int hashed = hdr->b_state != arc_anon;
1643
1644 ASSERT(buf->b_efunc == NULL);
1645 ASSERT(buf->b_data != NULL);
1646
1647 if (hashed) {
1648 kmutex_t *hash_lock = HDR_LOCK(hdr);
1649
1650 mutex_enter(hash_lock);
1651 hdr = buf->b_hdr;
1652 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1653
1654 (void) remove_reference(hdr, hash_lock, tag);
1655 if (hdr->b_datacnt > 1) {
1656 arc_buf_destroy(buf, FALSE, TRUE);
1657 } else {
1658 ASSERT(buf == hdr->b_buf);
1659 ASSERT(buf->b_efunc == NULL);
1660 hdr->b_flags |= ARC_BUF_AVAILABLE;
1661 }
1662 mutex_exit(hash_lock);
1663 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1664 int destroy_hdr;
1665 /*
1666 * We are in the middle of an async write. Don't destroy
1667 * this buffer unless the write completes before we finish
1668 * decrementing the reference count.
1669 */
1670 mutex_enter(&arc_eviction_mtx);
1671 (void) remove_reference(hdr, NULL, tag);
1672 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1673 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1674 mutex_exit(&arc_eviction_mtx);
1675 if (destroy_hdr)
1676 arc_hdr_destroy(hdr);
1677 } else {
1678 if (remove_reference(hdr, NULL, tag) > 0)
1679 arc_buf_destroy(buf, FALSE, TRUE);
1680 else
1681 arc_hdr_destroy(hdr);
1682 }
1683 }
1684
1685 boolean_t
1686 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1687 {
1688 arc_buf_hdr_t *hdr = buf->b_hdr;
1689 kmutex_t *hash_lock = HDR_LOCK(hdr);
1690 boolean_t no_callback = (buf->b_efunc == NULL);
1691
1692 if (hdr->b_state == arc_anon) {
1693 ASSERT(hdr->b_datacnt == 1);
1694 arc_buf_free(buf, tag);
1695 return (no_callback);
1696 }
1697
1698 mutex_enter(hash_lock);
1699 hdr = buf->b_hdr;
1700 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1701 ASSERT(hdr->b_state != arc_anon);
1702 ASSERT(buf->b_data != NULL);
1703
1704 (void) remove_reference(hdr, hash_lock, tag);
1705 if (hdr->b_datacnt > 1) {
1706 if (no_callback)
1707 arc_buf_destroy(buf, FALSE, TRUE);
1708 } else if (no_callback) {
1709 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1710 ASSERT(buf->b_efunc == NULL);
1711 hdr->b_flags |= ARC_BUF_AVAILABLE;
1712 }
1713 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1714 refcount_is_zero(&hdr->b_refcnt));
1715 mutex_exit(hash_lock);
1716 return (no_callback);
1717 }
1718
1719 int
1720 arc_buf_size(arc_buf_t *buf)
1721 {
1722 return (buf->b_hdr->b_size);
1723 }
1724
1725 /*
1726 * Called from the DMU to determine if the current buffer should be
1727 * evicted. In order to ensure proper locking, the eviction must be initiated
1728 * from the DMU. Return true if the buffer is associated with user data and
1729 * duplicate buffers still exist.
1730 */
1731 boolean_t
1732 arc_buf_eviction_needed(arc_buf_t *buf)
1733 {
1734 arc_buf_hdr_t *hdr;
1735 boolean_t evict_needed = B_FALSE;
1736
1737 if (zfs_disable_dup_eviction)
1738 return (B_FALSE);
1739
1740 mutex_enter(&buf->b_evict_lock);
1741 hdr = buf->b_hdr;
1742 if (hdr == NULL) {
1743 /*
1744 * We are in arc_do_user_evicts(); let that function
1745 * perform the eviction.
1746 */
1747 ASSERT(buf->b_data == NULL);
1748 mutex_exit(&buf->b_evict_lock);
1749 return (B_FALSE);
1750 } else if (buf->b_data == NULL) {
1751 /*
1752 * We have already been added to the arc eviction list;
1753 * recommend eviction.
1754 */
1755 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1756 mutex_exit(&buf->b_evict_lock);
1757 return (B_TRUE);
1758 }
1759
1760 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1761 evict_needed = B_TRUE;
1762
1763 mutex_exit(&buf->b_evict_lock);
1764 return (evict_needed);
1765 }
1766
1767 /*
1768 * Evict buffers from list until we've removed the specified number of
1769 * bytes. Move the removed buffers to the appropriate evict state.
1770 * If the recycle flag is set, then attempt to "recycle" a buffer:
1771 * - look for a buffer to evict that is `bytes' long.
1772 * - return the data block from this buffer rather than freeing it.
1773 * This flag is used by callers that are trying to make space for a
1774 * new buffer in a full arc cache.
1775 *
1776 * This function makes a "best effort". It skips over any buffers
1777 * it can't get a hash_lock on, and so may not catch all candidates.
1778 * It may also return without evicting as much space as requested.
1779 */
1780 static void *
1781 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1782 arc_buf_contents_t type)
1783 {
1784 arc_state_t *evicted_state;
1785 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1786 arc_buf_hdr_t *ab, *ab_prev = NULL;
1787 list_t *list = &state->arcs_list[type];
1788 kmutex_t *hash_lock;
1789 boolean_t have_lock;
1790 void *stolen = NULL;
1791 arc_buf_hdr_t marker = { 0 };
1792 int count = 0;
1793
1794 ASSERT(state == arc_mru || state == arc_mfu);
1795
1796 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1797
1798 mutex_enter(&state->arcs_mtx);
1799 mutex_enter(&evicted_state->arcs_mtx);
1800
1801 for (ab = list_tail(list); ab; ab = ab_prev) {
1802 ab_prev = list_prev(list, ab);
1803 /* prefetch buffers have a minimum lifespan */
1804 if (HDR_IO_IN_PROGRESS(ab) ||
1805 (spa && ab->b_spa != spa) ||
1806 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1807 ddi_get_lbolt() - ab->b_arc_access <
1808 arc_min_prefetch_lifespan)) {
1809 skipped++;
1810 continue;
1811 }
1812 /* "lookahead" for better eviction candidate */
1813 if (recycle && ab->b_size != bytes &&
1814 ab_prev && ab_prev->b_size == bytes)
1815 continue;
1816
1817 /* ignore markers */
1818 if (ab->b_spa == 0)
1819 continue;
1820
1821 /*
1822 * It may take a long time to evict all the bufs requested.
1823 * To avoid blocking all arc activity, periodically drop
1824 * the arcs_mtx and give other threads a chance to run
1825 * before reacquiring the lock.
1826 *
1827 * If we are looking for a buffer to recycle, we are in
1828 * the hot code path, so don't sleep.
1829 */
1830 if (!recycle && count++ > arc_evict_iterations) {
1831 list_insert_after(list, ab, &marker);
1832 mutex_exit(&evicted_state->arcs_mtx);
1833 mutex_exit(&state->arcs_mtx);
1834 kpreempt(KPREEMPT_SYNC);
1835 mutex_enter(&state->arcs_mtx);
1836 mutex_enter(&evicted_state->arcs_mtx);
1837 ab_prev = list_prev(list, &marker);
1838 list_remove(list, &marker);
1839 count = 0;
1840 continue;
1841 }
1842
1843 hash_lock = HDR_LOCK(ab);
1844 have_lock = MUTEX_HELD(hash_lock);
1845 if (have_lock || mutex_tryenter(hash_lock)) {
1846 ASSERT0(refcount_count(&ab->b_refcnt));
1847 ASSERT(ab->b_datacnt > 0);
1848 while (ab->b_buf) {
1849 arc_buf_t *buf = ab->b_buf;
1850 if (!mutex_tryenter(&buf->b_evict_lock)) {
1851 missed += 1;
1852 break;
1853 }
1854 if (buf->b_data) {
1855 bytes_evicted += ab->b_size;
1856 if (recycle && ab->b_type == type &&
1857 ab->b_size == bytes &&
1858 !HDR_L2_WRITING(ab)) {
1859 stolen = buf->b_data;
1860 recycle = FALSE;
1861 }
1862 }
1863 if (buf->b_efunc) {
1864 mutex_enter(&arc_eviction_mtx);
1865 arc_buf_destroy(buf,
1866 buf->b_data == stolen, FALSE);
1867 ab->b_buf = buf->b_next;
1868 buf->b_hdr = &arc_eviction_hdr;
1869 buf->b_next = arc_eviction_list;
1870 arc_eviction_list = buf;
1871 mutex_exit(&arc_eviction_mtx);
1872 mutex_exit(&buf->b_evict_lock);
1873 } else {
1874 mutex_exit(&buf->b_evict_lock);
1875 arc_buf_destroy(buf,
1876 buf->b_data == stolen, TRUE);
1877 }
1878 }
1879
1880 if (ab->b_l2hdr) {
1881 ARCSTAT_INCR(arcstat_evict_l2_cached,
1882 ab->b_size);
1883 } else {
1884 if (l2arc_write_eligible(ab->b_spa, ab)) {
1885 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1886 ab->b_size);
1887 } else {
1888 ARCSTAT_INCR(
1889 arcstat_evict_l2_ineligible,
1890 ab->b_size);
1891 }
1892 }
1893
1894 if (ab->b_datacnt == 0) {
1895 arc_change_state(evicted_state, ab, hash_lock);
1896 ASSERT(HDR_IN_HASH_TABLE(ab));
1897 ab->b_flags |= ARC_IN_HASH_TABLE;
1898 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1899 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1900 }
1901 if (!have_lock)
1902 mutex_exit(hash_lock);
1903 if (bytes >= 0 && bytes_evicted >= bytes)
1904 break;
1905 } else {
1906 missed += 1;
1907 }
1908 }
1909
1910 mutex_exit(&evicted_state->arcs_mtx);
1911 mutex_exit(&state->arcs_mtx);
1912
1913 if (bytes_evicted < bytes)
1914 dprintf("only evicted %lld bytes from %x",
1915 (longlong_t)bytes_evicted, state);
1916
1917 if (skipped)
1918 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1919
1920 if (missed)
1921 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1922
1923 /*
1924 * Note: we have just evicted some data into the ghost state,
1925 * potentially putting the ghost size over the desired size. Rather
1926 * that evicting from the ghost list in this hot code path, leave
1927 * this chore to the arc_reclaim_thread().
1928 */
1929
1930 return (stolen);
1931 }
1932
1933 /*
1934 * Remove buffers from list until we've removed the specified number of
1935 * bytes. Destroy the buffers that are removed.
1936 */
1937 static void
1938 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1939 {
1940 arc_buf_hdr_t *ab, *ab_prev;
1941 arc_buf_hdr_t marker = { 0 };
1942 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1943 kmutex_t *hash_lock;
1944 uint64_t bytes_deleted = 0;
1945 uint64_t bufs_skipped = 0;
1946 int count = 0;
1947
1948 ASSERT(GHOST_STATE(state));
1949 top:
1950 mutex_enter(&state->arcs_mtx);
1951 for (ab = list_tail(list); ab; ab = ab_prev) {
1952 ab_prev = list_prev(list, ab);
1953 if (ab->b_type > ARC_BUFC_NUMTYPES)
1954 panic("invalid ab=%p", (void *)ab);
1955 if (spa && ab->b_spa != spa)
1956 continue;
1957
1958 /* ignore markers */
1959 if (ab->b_spa == 0)
1960 continue;
1961
1962 hash_lock = HDR_LOCK(ab);
1963 /* caller may be trying to modify this buffer, skip it */
1964 if (MUTEX_HELD(hash_lock))
1965 continue;
1966
1967 /*
1968 * It may take a long time to evict all the bufs requested.
1969 * To avoid blocking all arc activity, periodically drop
1970 * the arcs_mtx and give other threads a chance to run
1971 * before reacquiring the lock.
1972 */
1973 if (count++ > arc_evict_iterations) {
1974 list_insert_after(list, ab, &marker);
1975 mutex_exit(&state->arcs_mtx);
1976 kpreempt(KPREEMPT_SYNC);
1977 mutex_enter(&state->arcs_mtx);
1978 ab_prev = list_prev(list, &marker);
1979 list_remove(list, &marker);
1980 count = 0;
1981 continue;
1982 }
1983 if (mutex_tryenter(hash_lock)) {
1984 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1985 ASSERT(ab->b_buf == NULL);
1986 ARCSTAT_BUMP(arcstat_deleted);
1987 bytes_deleted += ab->b_size;
1988
1989 if (ab->b_l2hdr != NULL) {
1990 /*
1991 * This buffer is cached on the 2nd Level ARC;
1992 * don't destroy the header.
1993 */
1994 arc_change_state(arc_l2c_only, ab, hash_lock);
1995 mutex_exit(hash_lock);
1996 } else {
1997 arc_change_state(arc_anon, ab, hash_lock);
1998 mutex_exit(hash_lock);
1999 arc_hdr_destroy(ab);
2000 }
2001
2002 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
2003 if (bytes >= 0 && bytes_deleted >= bytes)
2004 break;
2005 } else if (bytes < 0) {
2006 /*
2007 * Insert a list marker and then wait for the
2008 * hash lock to become available. Once its
2009 * available, restart from where we left off.
2010 */
2011 list_insert_after(list, ab, &marker);
2012 mutex_exit(&state->arcs_mtx);
2013 mutex_enter(hash_lock);
2014 mutex_exit(hash_lock);
2015 mutex_enter(&state->arcs_mtx);
2016 ab_prev = list_prev(list, &marker);
2017 list_remove(list, &marker);
2018 } else {
2019 bufs_skipped += 1;
2020 }
2021
2022 }
2023 mutex_exit(&state->arcs_mtx);
2024
2025 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
2026 (bytes < 0 || bytes_deleted < bytes)) {
2027 list = &state->arcs_list[ARC_BUFC_METADATA];
2028 goto top;
2029 }
2030
2031 if (bufs_skipped) {
2032 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2033 ASSERT(bytes >= 0);
2034 }
2035
2036 if (bytes_deleted < bytes)
2037 dprintf("only deleted %lld bytes from %p",
2038 (longlong_t)bytes_deleted, state);
2039 }
2040
2041 static void
2042 arc_adjust(void)
2043 {
2044 int64_t adjustment, delta;
2045
2046 /*
2047 * Adjust MRU size
2048 */
2049
2050 adjustment = MIN((int64_t)(arc_size - arc_c),
2051 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2052 arc_p));
2053
2054 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2055 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2056 (void) arc_evict(arc_mru, NULL, delta, FALSE, ARC_BUFC_DATA);
2057 adjustment -= delta;
2058 }
2059
2060 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2061 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2062 (void) arc_evict(arc_mru, NULL, delta, FALSE,
2063 ARC_BUFC_METADATA);
2064 }
2065
2066 /*
2067 * Adjust MFU size
2068 */
2069
2070 adjustment = arc_size - arc_c;
2071
2072 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2073 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2074 (void) arc_evict(arc_mfu, NULL, delta, FALSE, ARC_BUFC_DATA);
2075 adjustment -= delta;
2076 }
2077
2078 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2079 int64_t delta = MIN(adjustment,
2080 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2081 (void) arc_evict(arc_mfu, NULL, delta, FALSE,
2082 ARC_BUFC_METADATA);
2083 }
2084
2085 /*
2086 * Adjust ghost lists
2087 */
2088
2089 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2090
2091 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2092 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2093 arc_evict_ghost(arc_mru_ghost, NULL, delta);
2094 }
2095
2096 adjustment =
2097 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2098
2099 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2100 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2101 arc_evict_ghost(arc_mfu_ghost, NULL, delta);
2102 }
2103 }
2104
2105 static void
2106 arc_do_user_evicts(void)
2107 {
2108 mutex_enter(&arc_eviction_mtx);
2109 while (arc_eviction_list != NULL) {
2110 arc_buf_t *buf = arc_eviction_list;
2111 arc_eviction_list = buf->b_next;
2112 mutex_enter(&buf->b_evict_lock);
2113 buf->b_hdr = NULL;
2114 mutex_exit(&buf->b_evict_lock);
2115 mutex_exit(&arc_eviction_mtx);
2116
2117 if (buf->b_efunc != NULL)
2118 VERIFY(buf->b_efunc(buf) == 0);
2119
2120 buf->b_efunc = NULL;
2121 buf->b_private = NULL;
2122 kmem_cache_free(buf_cache, buf);
2123 mutex_enter(&arc_eviction_mtx);
2124 }
2125 mutex_exit(&arc_eviction_mtx);
2126 }
2127
2128 /*
2129 * Flush all *evictable* data from the cache for the given spa.
2130 * NOTE: this will not touch "active" (i.e. referenced) data.
2131 */
2132 void
2133 arc_flush(spa_t *spa)
2134 {
2135 uint64_t guid = 0;
2136
2137 if (spa)
2138 guid = spa_load_guid(spa);
2139
2140 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2141 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2142 if (spa)
2143 break;
2144 }
2145 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2146 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2147 if (spa)
2148 break;
2149 }
2150 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2151 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2152 if (spa)
2153 break;
2154 }
2155 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2156 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2157 if (spa)
2158 break;
2159 }
2160
2161 arc_evict_ghost(arc_mru_ghost, guid, -1);
2162 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2163
2164 mutex_enter(&arc_reclaim_thr_lock);
2165 arc_do_user_evicts();
2166 mutex_exit(&arc_reclaim_thr_lock);
2167 ASSERT(spa || arc_eviction_list == NULL);
2168 }
2169
2170 void
2171 arc_shrink(void)
2172 {
2173 if (arc_c > arc_c_min) {
2174 uint64_t to_free;
2175
2176 #ifdef _KERNEL
2177 to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
2178 #else
2179 to_free = arc_c >> arc_shrink_shift;
2180 #endif
2181 if (arc_c > arc_c_min + to_free)
2182 atomic_add_64(&arc_c, -to_free);
2183 else
2184 arc_c = arc_c_min;
2185
2186 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2187 if (arc_c > arc_size)
2188 arc_c = MAX(arc_size, arc_c_min);
2189 if (arc_p > arc_c)
2190 arc_p = (arc_c >> 1);
2191 ASSERT(arc_c >= arc_c_min);
2192 ASSERT((int64_t)arc_p >= 0);
2193 }
2194
2195 if (arc_size > arc_c)
2196 arc_adjust();
2197 }
2198
2199 /*
2200 * Determine if the system is under memory pressure and is asking
2201 * to reclaim memory. A return value of 1 indicates that the system
2202 * is under memory pressure and that the arc should adjust accordingly.
2203 */
2204 static int
2205 arc_reclaim_needed(void)
2206 {
2207 uint64_t extra;
2208
2209 #ifdef _KERNEL
2210
2211 if (needfree)
2212 return (1);
2213
2214 /*
2215 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2216 */
2217 extra = desfree;
2218
2219 /*
2220 * check that we're out of range of the pageout scanner. It starts to
2221 * schedule paging if freemem is less than lotsfree and needfree.
2222 * lotsfree is the high-water mark for pageout, and needfree is the
2223 * number of needed free pages. We add extra pages here to make sure
2224 * the scanner doesn't start up while we're freeing memory.
2225 */
2226 if (freemem < lotsfree + needfree + extra)
2227 return (1);
2228
2229 /*
2230 * check to make sure that swapfs has enough space so that anon
2231 * reservations can still succeed. anon_resvmem() checks that the
2232 * availrmem is greater than swapfs_minfree, and the number of reserved
2233 * swap pages. We also add a bit of extra here just to prevent
2234 * circumstances from getting really dire.
2235 */
2236 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2237 return (1);
2238
2239 /*
2240 * Check that we have enough availrmem that memory locking (e.g., via
2241 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
2242 * stores the number of pages that cannot be locked; when availrmem
2243 * drops below pages_pp_maximum, page locking mechanisms such as
2244 * page_pp_lock() will fail.)
2245 */
2246 if (availrmem <= pages_pp_maximum)
2247 return (1);
2248
2249 #if defined(__i386)
2250 /*
2251 * If we're on an i386 platform, it's possible that we'll exhaust the
2252 * kernel heap space before we ever run out of available physical
2253 * memory. Most checks of the size of the heap_area compare against
2254 * tune.t_minarmem, which is the minimum available real memory that we
2255 * can have in the system. However, this is generally fixed at 25 pages
2256 * which is so low that it's useless. In this comparison, we seek to
2257 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2258 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2259 * free)
2260 */
2261 if (vmem_size(heap_arena, VMEM_FREE) <
2262 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2))
2263 return (1);
2264 #endif
2265
2266 /*
2267 * If zio data pages are being allocated out of a separate heap segment,
2268 * then enforce that the size of available vmem for this arena remains
2269 * above about 1/16th free.
2270 *
2271 * Note: The 1/16th arena free requirement was put in place
2272 * to aggressively evict memory from the arc in order to avoid
2273 * memory fragmentation issues.
2274 */
2275 if (zio_arena != NULL &&
2276 vmem_size(zio_arena, VMEM_FREE) <
2277 (vmem_size(zio_arena, VMEM_ALLOC) >> 4))
2278 return (1);
2279 #else
2280 if (spa_get_random(100) == 0)
2281 return (1);
2282 #endif
2283 return (0);
2284 }
2285
2286 static void
2287 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2288 {
2289 size_t i;
2290 kmem_cache_t *prev_cache = NULL;
2291 kmem_cache_t *prev_data_cache = NULL;
2292 extern kmem_cache_t *zio_buf_cache[];
2293 extern kmem_cache_t *zio_data_buf_cache[];
2294
2295 #ifdef _KERNEL
2296 if (arc_meta_used >= arc_meta_limit) {
2297 /*
2298 * We are exceeding our meta-data cache limit.
2299 * Purge some DNLC entries to release holds on meta-data.
2300 */
2301 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2302 }
2303 #if defined(__i386)
2304 /*
2305 * Reclaim unused memory from all kmem caches.
2306 */
2307 kmem_reap();
2308 #endif
2309 #endif
2310
2311 /*
2312 * An aggressive reclamation will shrink the cache size as well as
2313 * reap free buffers from the arc kmem caches.
2314 */
2315 if (strat == ARC_RECLAIM_AGGR)
2316 arc_shrink();
2317
2318 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2319 if (zio_buf_cache[i] != prev_cache) {
2320 prev_cache = zio_buf_cache[i];
2321 kmem_cache_reap_now(zio_buf_cache[i]);
2322 }
2323 if (zio_data_buf_cache[i] != prev_data_cache) {
2324 prev_data_cache = zio_data_buf_cache[i];
2325 kmem_cache_reap_now(zio_data_buf_cache[i]);
2326 }
2327 }
2328 kmem_cache_reap_now(buf_cache);
2329 kmem_cache_reap_now(hdr_cache);
2330
2331 /*
2332 * Ask the vmem areana to reclaim unused memory from its
2333 * quantum caches.
2334 */
2335 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR)
2336 vmem_qcache_reap(zio_arena);
2337 }
2338
2339 static void
2340 arc_reclaim_thread(void)
2341 {
2342 clock_t growtime = 0;
2343 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2344 callb_cpr_t cpr;
2345
2346 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2347
2348 mutex_enter(&arc_reclaim_thr_lock);
2349 while (arc_thread_exit == 0) {
2350 if (arc_reclaim_needed()) {
2351
2352 if (arc_no_grow) {
2353 if (last_reclaim == ARC_RECLAIM_CONS) {
2354 last_reclaim = ARC_RECLAIM_AGGR;
2355 } else {
2356 last_reclaim = ARC_RECLAIM_CONS;
2357 }
2358 } else {
2359 arc_no_grow = TRUE;
2360 last_reclaim = ARC_RECLAIM_AGGR;
2361 membar_producer();
2362 }
2363
2364 /* reset the growth delay for every reclaim */
2365 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2366
2367 arc_kmem_reap_now(last_reclaim);
2368 arc_warm = B_TRUE;
2369
2370 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2371 arc_no_grow = FALSE;
2372 }
2373
2374 arc_adjust();
2375
2376 if (arc_eviction_list != NULL)
2377 arc_do_user_evicts();
2378
2379 /* block until needed, or one second, whichever is shorter */
2380 CALLB_CPR_SAFE_BEGIN(&cpr);
2381 (void) cv_timedwait(&arc_reclaim_thr_cv,
2382 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2383 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2384 }
2385
2386 arc_thread_exit = 0;
2387 cv_broadcast(&arc_reclaim_thr_cv);
2388 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2389 thread_exit();
2390 }
2391
2392 /*
2393 * Adapt arc info given the number of bytes we are trying to add and
2394 * the state that we are comming from. This function is only called
2395 * when we are adding new content to the cache.
2396 */
2397 static void
2398 arc_adapt(int bytes, arc_state_t *state)
2399 {
2400 int mult;
2401 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2402
2403 if (state == arc_l2c_only)
2404 return;
2405
2406 ASSERT(bytes > 0);
2407 /*
2408 * Adapt the target size of the MRU list:
2409 * - if we just hit in the MRU ghost list, then increase
2410 * the target size of the MRU list.
2411 * - if we just hit in the MFU ghost list, then increase
2412 * the target size of the MFU list by decreasing the
2413 * target size of the MRU list.
2414 */
2415 if (state == arc_mru_ghost) {
2416 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2417 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2418 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2419
2420 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2421 } else if (state == arc_mfu_ghost) {
2422 uint64_t delta;
2423
2424 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2425 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2426 mult = MIN(mult, 10);
2427
2428 delta = MIN(bytes * mult, arc_p);
2429 arc_p = MAX(arc_p_min, arc_p - delta);
2430 }
2431 ASSERT((int64_t)arc_p >= 0);
2432
2433 if (arc_reclaim_needed()) {
2434 cv_signal(&arc_reclaim_thr_cv);
2435 return;
2436 }
2437
2438 if (arc_no_grow)
2439 return;
2440
2441 if (arc_c >= arc_c_max)
2442 return;
2443
2444 /*
2445 * If we're within (2 * maxblocksize) bytes of the target
2446 * cache size, increment the target cache size
2447 */
2448 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2449 atomic_add_64(&arc_c, (int64_t)bytes);
2450 if (arc_c > arc_c_max)
2451 arc_c = arc_c_max;
2452 else if (state == arc_anon)
2453 atomic_add_64(&arc_p, (int64_t)bytes);
2454 if (arc_p > arc_c)
2455 arc_p = arc_c;
2456 }
2457 ASSERT((int64_t)arc_p >= 0);
2458 }
2459
2460 /*
2461 * Check if the cache has reached its limits and eviction is required
2462 * prior to insert.
2463 */
2464 static int
2465 arc_evict_needed(arc_buf_contents_t type)
2466 {
2467 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2468 return (1);
2469
2470 if (arc_reclaim_needed())
2471 return (1);
2472
2473 return (arc_size > arc_c);
2474 }
2475
2476 /*
2477 * The buffer, supplied as the first argument, needs a data block.
2478 * So, if we are at cache max, determine which cache should be victimized.
2479 * We have the following cases:
2480 *
2481 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2482 * In this situation if we're out of space, but the resident size of the MFU is
2483 * under the limit, victimize the MFU cache to satisfy this insertion request.
2484 *
2485 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2486 * Here, we've used up all of the available space for the MRU, so we need to
2487 * evict from our own cache instead. Evict from the set of resident MRU
2488 * entries.
2489 *
2490 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2491 * c minus p represents the MFU space in the cache, since p is the size of the
2492 * cache that is dedicated to the MRU. In this situation there's still space on
2493 * the MFU side, so the MRU side needs to be victimized.
2494 *
2495 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2496 * MFU's resident set is consuming more space than it has been allotted. In
2497 * this situation, we must victimize our own cache, the MFU, for this insertion.
2498 */
2499 static void
2500 arc_get_data_buf(arc_buf_t *buf)
2501 {
2502 arc_state_t *state = buf->b_hdr->b_state;
2503 uint64_t size = buf->b_hdr->b_size;
2504 arc_buf_contents_t type = buf->b_hdr->b_type;
2505
2506 arc_adapt(size, state);
2507
2508 /*
2509 * We have not yet reached cache maximum size,
2510 * just allocate a new buffer.
2511 */
2512 if (!arc_evict_needed(type)) {
2513 if (type == ARC_BUFC_METADATA) {
2514 buf->b_data = zio_buf_alloc(size);
2515 arc_space_consume(size, ARC_SPACE_DATA);
2516 } else {
2517 ASSERT(type == ARC_BUFC_DATA);
2518 buf->b_data = zio_data_buf_alloc(size);
2519 ARCSTAT_INCR(arcstat_data_size, size);
2520 atomic_add_64(&arc_size, size);
2521 }
2522 goto out;
2523 }
2524
2525 /*
2526 * If we are prefetching from the mfu ghost list, this buffer
2527 * will end up on the mru list; so steal space from there.
2528 */
2529 if (state == arc_mfu_ghost)
2530 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2531 else if (state == arc_mru_ghost)
2532 state = arc_mru;
2533
2534 if (state == arc_mru || state == arc_anon) {
2535 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2536 state = (arc_mfu->arcs_lsize[type] >= size &&
2537 arc_p > mru_used) ? arc_mfu : arc_mru;
2538 } else {
2539 /* MFU cases */
2540 uint64_t mfu_space = arc_c - arc_p;
2541 state = (arc_mru->arcs_lsize[type] >= size &&
2542 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2543 }
2544 if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) {
2545 if (type == ARC_BUFC_METADATA) {
2546 buf->b_data = zio_buf_alloc(size);
2547 arc_space_consume(size, ARC_SPACE_DATA);
2548 } else {
2549 ASSERT(type == ARC_BUFC_DATA);
2550 buf->b_data = zio_data_buf_alloc(size);
2551 ARCSTAT_INCR(arcstat_data_size, size);
2552 atomic_add_64(&arc_size, size);
2553 }
2554 ARCSTAT_BUMP(arcstat_recycle_miss);
2555 }
2556 ASSERT(buf->b_data != NULL);
2557 out:
2558 /*
2559 * Update the state size. Note that ghost states have a
2560 * "ghost size" and so don't need to be updated.
2561 */
2562 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2563 arc_buf_hdr_t *hdr = buf->b_hdr;
2564
2565 atomic_add_64(&hdr->b_state->arcs_size, size);
2566 if (list_link_active(&hdr->b_arc_node)) {
2567 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2568 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2569 }
2570 /*
2571 * If we are growing the cache, and we are adding anonymous
2572 * data, and we have outgrown arc_p, update arc_p
2573 */
2574 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2575 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2576 arc_p = MIN(arc_c, arc_p + size);
2577 }
2578 }
2579
2580 /*
2581 * This routine is called whenever a buffer is accessed.
2582 * NOTE: the hash lock is dropped in this function.
2583 */
2584 static void
2585 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2586 {
2587 clock_t now;
2588
2589 ASSERT(MUTEX_HELD(hash_lock));
2590
2591 if (buf->b_state == arc_anon) {
2592 /*
2593 * This buffer is not in the cache, and does not
2594 * appear in our "ghost" list. Add the new buffer
2595 * to the MRU state.
2596 */
2597
2598 ASSERT(buf->b_arc_access == 0);
2599 buf->b_arc_access = ddi_get_lbolt();
2600 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2601 arc_change_state(arc_mru, buf, hash_lock);
2602
2603 } else if (buf->b_state == arc_mru) {
2604 now = ddi_get_lbolt();
2605
2606 /*
2607 * If this buffer is here because of a prefetch, then either:
2608 * - clear the flag if this is a "referencing" read
2609 * (any subsequent access will bump this into the MFU state).
2610 * or
2611 * - move the buffer to the head of the list if this is
2612 * another prefetch (to make it less likely to be evicted).
2613 */
2614 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2615 if (refcount_count(&buf->b_refcnt) == 0) {
2616 ASSERT(list_link_active(&buf->b_arc_node));
2617 } else {
2618 buf->b_flags &= ~ARC_PREFETCH;
2619 ARCSTAT_BUMP(arcstat_mru_hits);
2620 }
2621 buf->b_arc_access = now;
2622 return;
2623 }
2624
2625 /*
2626 * This buffer has been "accessed" only once so far,
2627 * but it is still in the cache. Move it to the MFU
2628 * state.
2629 */
2630 if (now > buf->b_arc_access + ARC_MINTIME) {
2631 /*
2632 * More than 125ms have passed since we
2633 * instantiated this buffer. Move it to the
2634 * most frequently used state.
2635 */
2636 buf->b_arc_access = now;
2637 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2638 arc_change_state(arc_mfu, buf, hash_lock);
2639 }
2640 ARCSTAT_BUMP(arcstat_mru_hits);
2641 } else if (buf->b_state == arc_mru_ghost) {
2642 arc_state_t *new_state;
2643 /*
2644 * This buffer has been "accessed" recently, but
2645 * was evicted from the cache. Move it to the
2646 * MFU state.
2647 */
2648
2649 if (buf->b_flags & ARC_PREFETCH) {
2650 new_state = arc_mru;
2651 if (refcount_count(&buf->b_refcnt) > 0)
2652 buf->b_flags &= ~ARC_PREFETCH;
2653 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2654 } else {
2655 new_state = arc_mfu;
2656 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2657 }
2658
2659 buf->b_arc_access = ddi_get_lbolt();
2660 arc_change_state(new_state, buf, hash_lock);
2661
2662 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2663 } else if (buf->b_state == arc_mfu) {
2664 /*
2665 * This buffer has been accessed more than once and is
2666 * still in the cache. Keep it in the MFU state.
2667 *
2668 * NOTE: an add_reference() that occurred when we did
2669 * the arc_read() will have kicked this off the list.
2670 * If it was a prefetch, we will explicitly move it to
2671 * the head of the list now.
2672 */
2673 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2674 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2675 ASSERT(list_link_active(&buf->b_arc_node));
2676 }
2677 ARCSTAT_BUMP(arcstat_mfu_hits);
2678 buf->b_arc_access = ddi_get_lbolt();
2679 } else if (buf->b_state == arc_mfu_ghost) {
2680 arc_state_t *new_state = arc_mfu;
2681 /*
2682 * This buffer has been accessed more than once but has
2683 * been evicted from the cache. Move it back to the
2684 * MFU state.
2685 */
2686
2687 if (buf->b_flags & ARC_PREFETCH) {
2688 /*
2689 * This is a prefetch access...
2690 * move this block back to the MRU state.
2691 */
2692 ASSERT0(refcount_count(&buf->b_refcnt));
2693 new_state = arc_mru;
2694 }
2695
2696 buf->b_arc_access = ddi_get_lbolt();
2697 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2698 arc_change_state(new_state, buf, hash_lock);
2699
2700 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2701 } else if (buf->b_state == arc_l2c_only) {
2702 /*
2703 * This buffer is on the 2nd Level ARC.
2704 */
2705
2706 buf->b_arc_access = ddi_get_lbolt();
2707 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2708 arc_change_state(arc_mfu, buf, hash_lock);
2709 } else {
2710 ASSERT(!"invalid arc state");
2711 }
2712 }
2713
2714 /* a generic arc_done_func_t which you can use */
2715 /* ARGSUSED */
2716 void
2717 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2718 {
2719 if (zio == NULL || zio->io_error == 0)
2720 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2721 VERIFY(arc_buf_remove_ref(buf, arg));
2722 }
2723
2724 /* a generic arc_done_func_t */
2725 void
2726 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2727 {
2728 arc_buf_t **bufp = arg;
2729 if (zio && zio->io_error) {
2730 VERIFY(arc_buf_remove_ref(buf, arg));
2731 *bufp = NULL;
2732 } else {
2733 *bufp = buf;
2734 ASSERT(buf->b_data);
2735 }
2736 }
2737
2738 static void
2739 arc_read_done(zio_t *zio)
2740 {
2741 arc_buf_hdr_t *hdr, *found;
2742 arc_buf_t *buf;
2743 arc_buf_t *abuf; /* buffer we're assigning to callback */
2744 kmutex_t *hash_lock;
2745 arc_callback_t *callback_list, *acb;
2746 int freeable = FALSE;
2747
2748 buf = zio->io_private;
2749 hdr = buf->b_hdr;
2750
2751 /*
2752 * The hdr was inserted into hash-table and removed from lists
2753 * prior to starting I/O. We should find this header, since
2754 * it's in the hash table, and it should be legit since it's
2755 * not possible to evict it during the I/O. The only possible
2756 * reason for it not to be found is if we were freed during the
2757 * read.
2758 */
2759 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2760 &hash_lock);
2761
2762 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2763 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2764 (found == hdr && HDR_L2_READING(hdr)));
2765
2766 hdr->b_flags &= ~ARC_L2_EVICTED;
2767 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2768 hdr->b_flags &= ~ARC_L2CACHE;
2769
2770 /* byteswap if necessary */
2771 callback_list = hdr->b_acb;
2772 ASSERT(callback_list != NULL);
2773 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2774 dmu_object_byteswap_t bswap =
2775 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2776 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2777 byteswap_uint64_array :
2778 dmu_ot_byteswap[bswap].ob_func;
2779 func(buf->b_data, hdr->b_size);
2780 }
2781
2782 arc_cksum_compute(buf, B_FALSE);
2783 arc_buf_watch(buf);
2784
2785 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2786 /*
2787 * Only call arc_access on anonymous buffers. This is because
2788 * if we've issued an I/O for an evicted buffer, we've already
2789 * called arc_access (to prevent any simultaneous readers from
2790 * getting confused).
2791 */
2792 arc_access(hdr, hash_lock);
2793 }
2794
2795 /* create copies of the data buffer for the callers */
2796 abuf = buf;
2797 for (acb = callback_list; acb; acb = acb->acb_next) {
2798 if (acb->acb_done) {
2799 if (abuf == NULL) {
2800 ARCSTAT_BUMP(arcstat_duplicate_reads);
2801 abuf = arc_buf_clone(buf);
2802 }
2803 acb->acb_buf = abuf;
2804 abuf = NULL;
2805 }
2806 }
2807 hdr->b_acb = NULL;
2808 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2809 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2810 if (abuf == buf) {
2811 ASSERT(buf->b_efunc == NULL);
2812 ASSERT(hdr->b_datacnt == 1);
2813 hdr->b_flags |= ARC_BUF_AVAILABLE;
2814 }
2815
2816 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2817
2818 if (zio->io_error != 0) {
2819 hdr->b_flags |= ARC_IO_ERROR;
2820 if (hdr->b_state != arc_anon)
2821 arc_change_state(arc_anon, hdr, hash_lock);
2822 if (HDR_IN_HASH_TABLE(hdr))
2823 buf_hash_remove(hdr);
2824 freeable = refcount_is_zero(&hdr->b_refcnt);
2825 }
2826
2827 /*
2828 * Broadcast before we drop the hash_lock to avoid the possibility
2829 * that the hdr (and hence the cv) might be freed before we get to
2830 * the cv_broadcast().
2831 */
2832 cv_broadcast(&hdr->b_cv);
2833
2834 if (hash_lock) {
2835 mutex_exit(hash_lock);
2836 } else {
2837 /*
2838 * This block was freed while we waited for the read to
2839 * complete. It has been removed from the hash table and
2840 * moved to the anonymous state (so that it won't show up
2841 * in the cache).
2842 */
2843 ASSERT3P(hdr->b_state, ==, arc_anon);
2844 freeable = refcount_is_zero(&hdr->b_refcnt);
2845 }
2846
2847 /* execute each callback and free its structure */
2848 while ((acb = callback_list) != NULL) {
2849 if (acb->acb_done)
2850 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2851
2852 if (acb->acb_zio_dummy != NULL) {
2853 acb->acb_zio_dummy->io_error = zio->io_error;
2854 zio_nowait(acb->acb_zio_dummy);
2855 }
2856
2857 callback_list = acb->acb_next;
2858 kmem_free(acb, sizeof (arc_callback_t));
2859 }
2860
2861 if (freeable)
2862 arc_hdr_destroy(hdr);
2863 }
2864
2865 /*
2866 * "Read" the block at the specified DVA (in bp) via the
2867 * cache. If the block is found in the cache, invoke the provided
2868 * callback immediately and return. Note that the `zio' parameter
2869 * in the callback will be NULL in this case, since no IO was
2870 * required. If the block is not in the cache pass the read request
2871 * on to the spa with a substitute callback function, so that the
2872 * requested block will be added to the cache.
2873 *
2874 * If a read request arrives for a block that has a read in-progress,
2875 * either wait for the in-progress read to complete (and return the
2876 * results); or, if this is a read with a "done" func, add a record
2877 * to the read to invoke the "done" func when the read completes,
2878 * and return; or just return.
2879 *
2880 * arc_read_done() will invoke all the requested "done" functions
2881 * for readers of this block.
2882 */
2883 int
2884 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
2885 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags,
2886 const zbookmark_t *zb)
2887 {
2888 arc_buf_hdr_t *hdr;
2889 arc_buf_t *buf = NULL;
2890 kmutex_t *hash_lock;
2891 zio_t *rzio;
2892 uint64_t guid = spa_load_guid(spa);
2893
2894 top:
2895 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2896 &hash_lock);
2897 if (hdr && hdr->b_datacnt > 0) {
2898
2899 *arc_flags |= ARC_CACHED;
2900
2901 if (HDR_IO_IN_PROGRESS(hdr)) {
2902
2903 if (*arc_flags & ARC_WAIT) {
2904 cv_wait(&hdr->b_cv, hash_lock);
2905 mutex_exit(hash_lock);
2906 goto top;
2907 }
2908 ASSERT(*arc_flags & ARC_NOWAIT);
2909
2910 if (done) {
2911 arc_callback_t *acb = NULL;
2912
2913 acb = kmem_zalloc(sizeof (arc_callback_t),
2914 KM_SLEEP);
2915 acb->acb_done = done;
2916 acb->acb_private = private;
2917 if (pio != NULL)
2918 acb->acb_zio_dummy = zio_null(pio,
2919 spa, NULL, NULL, NULL, zio_flags);
2920
2921 ASSERT(acb->acb_done != NULL);
2922 acb->acb_next = hdr->b_acb;
2923 hdr->b_acb = acb;
2924 add_reference(hdr, hash_lock, private);
2925 mutex_exit(hash_lock);
2926 return (0);
2927 }
2928 mutex_exit(hash_lock);
2929 return (0);
2930 }
2931
2932 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2933
2934 if (done) {
2935 add_reference(hdr, hash_lock, private);
2936 /*
2937 * If this block is already in use, create a new
2938 * copy of the data so that we will be guaranteed
2939 * that arc_release() will always succeed.
2940 */
2941 buf = hdr->b_buf;
2942 ASSERT(buf);
2943 ASSERT(buf->b_data);
2944 if (HDR_BUF_AVAILABLE(hdr)) {
2945 ASSERT(buf->b_efunc == NULL);
2946 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2947 } else {
2948 buf = arc_buf_clone(buf);
2949 }
2950
2951 } else if (*arc_flags & ARC_PREFETCH &&
2952 refcount_count(&hdr->b_refcnt) == 0) {
2953 hdr->b_flags |= ARC_PREFETCH;
2954 }
2955 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2956 arc_access(hdr, hash_lock);
2957 if (*arc_flags & ARC_L2CACHE)
2958 hdr->b_flags |= ARC_L2CACHE;
2959 if (*arc_flags & ARC_L2COMPRESS)
2960 hdr->b_flags |= ARC_L2COMPRESS;
2961 mutex_exit(hash_lock);
2962 ARCSTAT_BUMP(arcstat_hits);
2963 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2964 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2965 data, metadata, hits);
2966
2967 if (done)
2968 done(NULL, buf, private);
2969 } else {
2970 uint64_t size = BP_GET_LSIZE(bp);
2971 arc_callback_t *acb;
2972 vdev_t *vd = NULL;
2973 uint64_t addr = 0;
2974 boolean_t devw = B_FALSE;
2975 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
2976 uint64_t b_asize = 0;
2977
2978 if (hdr == NULL) {
2979 /* this block is not in the cache */
2980 arc_buf_hdr_t *exists;
2981 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2982 buf = arc_buf_alloc(spa, size, private, type);
2983 hdr = buf->b_hdr;
2984 hdr->b_dva = *BP_IDENTITY(bp);
2985 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
2986 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2987 exists = buf_hash_insert(hdr, &hash_lock);
2988 if (exists) {
2989 /* somebody beat us to the hash insert */
2990 mutex_exit(hash_lock);
2991 buf_discard_identity(hdr);
2992 (void) arc_buf_remove_ref(buf, private);
2993 goto top; /* restart the IO request */
2994 }
2995 /* if this is a prefetch, we don't have a reference */
2996 if (*arc_flags & ARC_PREFETCH) {
2997 (void) remove_reference(hdr, hash_lock,
2998 private);
2999 hdr->b_flags |= ARC_PREFETCH;
3000 }
3001 if (*arc_flags & ARC_L2CACHE)
3002 hdr->b_flags |= ARC_L2CACHE;
3003 if (*arc_flags & ARC_L2COMPRESS)
3004 hdr->b_flags |= ARC_L2COMPRESS;
3005 if (BP_GET_LEVEL(bp) > 0)
3006 hdr->b_flags |= ARC_INDIRECT;
3007 } else {
3008 /* this block is in the ghost cache */
3009 ASSERT(GHOST_STATE(hdr->b_state));
3010 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3011 ASSERT0(refcount_count(&hdr->b_refcnt));
3012 ASSERT(hdr->b_buf == NULL);
3013
3014 /* if this is a prefetch, we don't have a reference */
3015 if (*arc_flags & ARC_PREFETCH)
3016 hdr->b_flags |= ARC_PREFETCH;
3017 else
3018 add_reference(hdr, hash_lock, private);
3019 if (*arc_flags & ARC_L2CACHE)
3020 hdr->b_flags |= ARC_L2CACHE;
3021 if (*arc_flags & ARC_L2COMPRESS)
3022 hdr->b_flags |= ARC_L2COMPRESS;
3023 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3024 buf->b_hdr = hdr;
3025 buf->b_data = NULL;
3026 buf->b_efunc = NULL;
3027 buf->b_private = NULL;
3028 buf->b_next = NULL;
3029 hdr->b_buf = buf;
3030 ASSERT(hdr->b_datacnt == 0);
3031 hdr->b_datacnt = 1;
3032 arc_get_data_buf(buf);
3033 arc_access(hdr, hash_lock);
3034 }
3035
3036 ASSERT(!GHOST_STATE(hdr->b_state));
3037
3038 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
3039 acb->acb_done = done;
3040 acb->acb_private = private;
3041
3042 ASSERT(hdr->b_acb == NULL);
3043 hdr->b_acb = acb;
3044 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3045
3046 if (hdr->b_l2hdr != NULL &&
3047 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3048 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3049 addr = hdr->b_l2hdr->b_daddr;
3050 b_compress = hdr->b_l2hdr->b_compress;
3051 b_asize = hdr->b_l2hdr->b_asize;
3052 /*
3053 * Lock out device removal.
3054 */
3055 if (vdev_is_dead(vd) ||
3056 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3057 vd = NULL;
3058 }
3059
3060 mutex_exit(hash_lock);
3061
3062 /*
3063 * At this point, we have a level 1 cache miss. Try again in
3064 * L2ARC if possible.
3065 */
3066 ASSERT3U(hdr->b_size, ==, size);
3067 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3068 uint64_t, size, zbookmark_t *, zb);
3069 ARCSTAT_BUMP(arcstat_misses);
3070 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3071 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3072 data, metadata, misses);
3073
3074 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3075 /*
3076 * Read from the L2ARC if the following are true:
3077 * 1. The L2ARC vdev was previously cached.
3078 * 2. This buffer still has L2ARC metadata.
3079 * 3. This buffer isn't currently writing to the L2ARC.
3080 * 4. The L2ARC entry wasn't evicted, which may
3081 * also have invalidated the vdev.
3082 * 5. This isn't prefetch and l2arc_noprefetch is set.
3083 */
3084 if (hdr->b_l2hdr != NULL &&
3085 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3086 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3087 l2arc_read_callback_t *cb;
3088
3089 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3090 ARCSTAT_BUMP(arcstat_l2_hits);
3091
3092 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3093 KM_SLEEP);
3094 cb->l2rcb_buf = buf;
3095 cb->l2rcb_spa = spa;
3096 cb->l2rcb_bp = *bp;
3097 cb->l2rcb_zb = *zb;
3098 cb->l2rcb_flags = zio_flags;
3099 cb->l2rcb_compress = b_compress;
3100
3101 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
3102 addr + size < vd->vdev_psize -
3103 VDEV_LABEL_END_SIZE);
3104
3105 /*
3106 * l2arc read. The SCL_L2ARC lock will be
3107 * released by l2arc_read_done().
3108 * Issue a null zio if the underlying buffer
3109 * was squashed to zero size by compression.
3110 */
3111 if (b_compress == ZIO_COMPRESS_EMPTY) {
3112 rzio = zio_null(pio, spa, vd,
3113 l2arc_read_done, cb,
3114 zio_flags | ZIO_FLAG_DONT_CACHE |
3115 ZIO_FLAG_CANFAIL |
3116 ZIO_FLAG_DONT_PROPAGATE |
3117 ZIO_FLAG_DONT_RETRY);
3118 } else {
3119 rzio = zio_read_phys(pio, vd, addr,
3120 b_asize, buf->b_data,
3121 ZIO_CHECKSUM_OFF,
3122 l2arc_read_done, cb, priority,
3123 zio_flags | ZIO_FLAG_DONT_CACHE |
3124 ZIO_FLAG_CANFAIL |
3125 ZIO_FLAG_DONT_PROPAGATE |
3126 ZIO_FLAG_DONT_RETRY, B_FALSE);
3127 }
3128 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3129 zio_t *, rzio);
3130 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
3131
3132 if (*arc_flags & ARC_NOWAIT) {
3133 zio_nowait(rzio);
3134 return (0);
3135 }
3136
3137 ASSERT(*arc_flags & ARC_WAIT);
3138 if (zio_wait(rzio) == 0)
3139 return (0);
3140
3141 /* l2arc read error; goto zio_read() */
3142 } else {
3143 DTRACE_PROBE1(l2arc__miss,
3144 arc_buf_hdr_t *, hdr);
3145 ARCSTAT_BUMP(arcstat_l2_misses);
3146 if (HDR_L2_WRITING(hdr))
3147 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3148 spa_config_exit(spa, SCL_L2ARC, vd);
3149 }
3150 } else {
3151 if (vd != NULL)
3152 spa_config_exit(spa, SCL_L2ARC, vd);
3153 if (l2arc_ndev != 0) {
3154 DTRACE_PROBE1(l2arc__miss,
3155 arc_buf_hdr_t *, hdr);
3156 ARCSTAT_BUMP(arcstat_l2_misses);
3157 }
3158 }
3159
3160 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3161 arc_read_done, buf, priority, zio_flags, zb);
3162
3163 if (*arc_flags & ARC_WAIT)
3164 return (zio_wait(rzio));
3165
3166 ASSERT(*arc_flags & ARC_NOWAIT);
3167 zio_nowait(rzio);
3168 }
3169 return (0);
3170 }
3171
3172 void
3173 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3174 {
3175 ASSERT(buf->b_hdr != NULL);
3176 ASSERT(buf->b_hdr->b_state != arc_anon);
3177 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3178 ASSERT(buf->b_efunc == NULL);
3179 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3180
3181 buf->b_efunc = func;
3182 buf->b_private = private;
3183 }
3184
3185 /*
3186 * Notify the arc that a block was freed, and thus will never be used again.
3187 */
3188 void
3189 arc_freed(spa_t *spa, const blkptr_t *bp)
3190 {
3191 arc_buf_hdr_t *hdr;
3192 kmutex_t *hash_lock;
3193 uint64_t guid = spa_load_guid(spa);
3194
3195 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3196 &hash_lock);
3197 if (hdr == NULL)
3198 return;
3199 if (HDR_BUF_AVAILABLE(hdr)) {
3200 arc_buf_t *buf = hdr->b_buf;
3201 add_reference(hdr, hash_lock, FTAG);
3202 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3203 mutex_exit(hash_lock);
3204
3205 arc_release(buf, FTAG);
3206 (void) arc_buf_remove_ref(buf, FTAG);
3207 } else {
3208 mutex_exit(hash_lock);
3209 }
3210
3211 }
3212
3213 /*
3214 * This is used by the DMU to let the ARC know that a buffer is
3215 * being evicted, so the ARC should clean up. If this arc buf
3216 * is not yet in the evicted state, it will be put there.
3217 */
3218 int
3219 arc_buf_evict(arc_buf_t *buf)
3220 {
3221 arc_buf_hdr_t *hdr;
3222 kmutex_t *hash_lock;
3223 arc_buf_t **bufp;
3224
3225 mutex_enter(&buf->b_evict_lock);
3226 hdr = buf->b_hdr;
3227 if (hdr == NULL) {
3228 /*
3229 * We are in arc_do_user_evicts().
3230 */
3231 ASSERT(buf->b_data == NULL);
3232 mutex_exit(&buf->b_evict_lock);
3233 return (0);
3234 } else if (buf->b_data == NULL) {
3235 arc_buf_t copy = *buf; /* structure assignment */
3236 /*
3237 * We are on the eviction list; process this buffer now
3238 * but let arc_do_user_evicts() do the reaping.
3239 */
3240 buf->b_efunc = NULL;
3241 mutex_exit(&buf->b_evict_lock);
3242 VERIFY(copy.b_efunc(©) == 0);
3243 return (1);
3244 }
3245 hash_lock = HDR_LOCK(hdr);
3246 mutex_enter(hash_lock);
3247 hdr = buf->b_hdr;
3248 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3249
3250 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3251 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3252
3253 /*
3254 * Pull this buffer off of the hdr
3255 */
3256 bufp = &hdr->b_buf;
3257 while (*bufp != buf)
3258 bufp = &(*bufp)->b_next;
3259 *bufp = buf->b_next;
3260
3261 ASSERT(buf->b_data != NULL);
3262 arc_buf_destroy(buf, FALSE, FALSE);
3263
3264 if (hdr->b_datacnt == 0) {
3265 arc_state_t *old_state = hdr->b_state;
3266 arc_state_t *evicted_state;
3267
3268 ASSERT(hdr->b_buf == NULL);
3269 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3270
3271 evicted_state =
3272 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3273
3274 mutex_enter(&old_state->arcs_mtx);
3275 mutex_enter(&evicted_state->arcs_mtx);
3276
3277 arc_change_state(evicted_state, hdr, hash_lock);
3278 ASSERT(HDR_IN_HASH_TABLE(hdr));
3279 hdr->b_flags |= ARC_IN_HASH_TABLE;
3280 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3281
3282 mutex_exit(&evicted_state->arcs_mtx);
3283 mutex_exit(&old_state->arcs_mtx);
3284 }
3285 mutex_exit(hash_lock);
3286 mutex_exit(&buf->b_evict_lock);
3287
3288 VERIFY(buf->b_efunc(buf) == 0);
3289 buf->b_efunc = NULL;
3290 buf->b_private = NULL;
3291 buf->b_hdr = NULL;
3292 buf->b_next = NULL;
3293 kmem_cache_free(buf_cache, buf);
3294 return (1);
3295 }
3296
3297 /*
3298 * Release this buffer from the cache, making it an anonymous buffer. This
3299 * must be done after a read and prior to modifying the buffer contents.
3300 * If the buffer has more than one reference, we must make
3301 * a new hdr for the buffer.
3302 */
3303 void
3304 arc_release(arc_buf_t *buf, void *tag)
3305 {
3306 arc_buf_hdr_t *hdr;
3307 kmutex_t *hash_lock = NULL;
3308 l2arc_buf_hdr_t *l2hdr;
3309 uint64_t buf_size;
3310
3311 /*
3312 * It would be nice to assert that if it's DMU metadata (level >
3313 * 0 || it's the dnode file), then it must be syncing context.
3314 * But we don't know that information at this level.
3315 */
3316
3317 mutex_enter(&buf->b_evict_lock);
3318 hdr = buf->b_hdr;
3319
3320 /* this buffer is not on any list */
3321 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3322
3323 if (hdr->b_state == arc_anon) {
3324 /* this buffer is already released */
3325 ASSERT(buf->b_efunc == NULL);
3326 } else {
3327 hash_lock = HDR_LOCK(hdr);
3328 mutex_enter(hash_lock);
3329 hdr = buf->b_hdr;
3330 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3331 }
3332
3333 l2hdr = hdr->b_l2hdr;
3334 if (l2hdr) {
3335 mutex_enter(&l2arc_buflist_mtx);
3336 hdr->b_l2hdr = NULL;
3337 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3338 }
3339 buf_size = hdr->b_size;
3340
3341 /*
3342 * Do we have more than one buf?
3343 */
3344 if (hdr->b_datacnt > 1) {
3345 arc_buf_hdr_t *nhdr;
3346 arc_buf_t **bufp;
3347 uint64_t blksz = hdr->b_size;
3348 uint64_t spa = hdr->b_spa;
3349 arc_buf_contents_t type = hdr->b_type;
3350 uint32_t flags = hdr->b_flags;
3351
3352 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3353 /*
3354 * Pull the data off of this hdr and attach it to
3355 * a new anonymous hdr.
3356 */
3357 (void) remove_reference(hdr, hash_lock, tag);
3358 bufp = &hdr->b_buf;
3359 while (*bufp != buf)
3360 bufp = &(*bufp)->b_next;
3361 *bufp = buf->b_next;
3362 buf->b_next = NULL;
3363
3364 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3365 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3366 if (refcount_is_zero(&hdr->b_refcnt)) {
3367 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3368 ASSERT3U(*size, >=, hdr->b_size);
3369 atomic_add_64(size, -hdr->b_size);
3370 }
3371
3372 /*
3373 * We're releasing a duplicate user data buffer, update
3374 * our statistics accordingly.
3375 */
3376 if (hdr->b_type == ARC_BUFC_DATA) {
3377 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3378 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3379 -hdr->b_size);
3380 }
3381 hdr->b_datacnt -= 1;
3382 arc_cksum_verify(buf);
3383 arc_buf_unwatch(buf);
3384
3385 mutex_exit(hash_lock);
3386
3387 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3388 nhdr->b_size = blksz;
3389 nhdr->b_spa = spa;
3390 nhdr->b_type = type;
3391 nhdr->b_buf = buf;
3392 nhdr->b_state = arc_anon;
3393 nhdr->b_arc_access = 0;
3394 nhdr->b_flags = flags & ARC_L2_WRITING;
3395 nhdr->b_l2hdr = NULL;
3396 nhdr->b_datacnt = 1;
3397 nhdr->b_freeze_cksum = NULL;
3398 (void) refcount_add(&nhdr->b_refcnt, tag);
3399 buf->b_hdr = nhdr;
3400 mutex_exit(&buf->b_evict_lock);
3401 atomic_add_64(&arc_anon->arcs_size, blksz);
3402 } else {
3403 mutex_exit(&buf->b_evict_lock);
3404 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3405 ASSERT(!list_link_active(&hdr->b_arc_node));
3406 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3407 if (hdr->b_state != arc_anon)
3408 arc_change_state(arc_anon, hdr, hash_lock);
3409 hdr->b_arc_access = 0;
3410 if (hash_lock)
3411 mutex_exit(hash_lock);
3412
3413 buf_discard_identity(hdr);
3414 arc_buf_thaw(buf);
3415 }
3416 buf->b_efunc = NULL;
3417 buf->b_private = NULL;
3418
3419 if (l2hdr) {
3420 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
3421 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3422 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3423 mutex_exit(&l2arc_buflist_mtx);
3424 }
3425 }
3426
3427 int
3428 arc_released(arc_buf_t *buf)
3429 {
3430 int released;
3431
3432 mutex_enter(&buf->b_evict_lock);
3433 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3434 mutex_exit(&buf->b_evict_lock);
3435 return (released);
3436 }
3437
3438 int
3439 arc_has_callback(arc_buf_t *buf)
3440 {
3441 int callback;
3442
3443 mutex_enter(&buf->b_evict_lock);
3444 callback = (buf->b_efunc != NULL);
3445 mutex_exit(&buf->b_evict_lock);
3446 return (callback);
3447 }
3448
3449 #ifdef ZFS_DEBUG
3450 int
3451 arc_referenced(arc_buf_t *buf)
3452 {
3453 int referenced;
3454
3455 mutex_enter(&buf->b_evict_lock);
3456 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3457 mutex_exit(&buf->b_evict_lock);
3458 return (referenced);
3459 }
3460 #endif
3461
3462 static void
3463 arc_write_ready(zio_t *zio)
3464 {
3465 arc_write_callback_t *callback = zio->io_private;
3466 arc_buf_t *buf = callback->awcb_buf;
3467 arc_buf_hdr_t *hdr = buf->b_hdr;
3468
3469 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3470 callback->awcb_ready(zio, buf, callback->awcb_private);
3471
3472 /*
3473 * If the IO is already in progress, then this is a re-write
3474 * attempt, so we need to thaw and re-compute the cksum.
3475 * It is the responsibility of the callback to handle the
3476 * accounting for any re-write attempt.
3477 */
3478 if (HDR_IO_IN_PROGRESS(hdr)) {
3479 mutex_enter(&hdr->b_freeze_lock);
3480 if (hdr->b_freeze_cksum != NULL) {
3481 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3482 hdr->b_freeze_cksum = NULL;
3483 }
3484 mutex_exit(&hdr->b_freeze_lock);
3485 }
3486 arc_cksum_compute(buf, B_FALSE);
3487 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3488 }
3489
3490 /*
3491 * The SPA calls this callback for each physical write that happens on behalf
3492 * of a logical write. See the comment in dbuf_write_physdone() for details.
3493 */
3494 static void
3495 arc_write_physdone(zio_t *zio)
3496 {
3497 arc_write_callback_t *cb = zio->io_private;
3498 if (cb->awcb_physdone != NULL)
3499 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
3500 }
3501
3502 static void
3503 arc_write_done(zio_t *zio)
3504 {
3505 arc_write_callback_t *callback = zio->io_private;
3506 arc_buf_t *buf = callback->awcb_buf;
3507 arc_buf_hdr_t *hdr = buf->b_hdr;
3508
3509 ASSERT(hdr->b_acb == NULL);
3510
3511 if (zio->io_error == 0) {
3512 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3513 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3514 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3515 } else {
3516 ASSERT(BUF_EMPTY(hdr));
3517 }
3518
3519 /*
3520 * If the block to be written was all-zero, we may have
3521 * compressed it away. In this case no write was performed
3522 * so there will be no dva/birth/checksum. The buffer must
3523 * therefore remain anonymous (and uncached).
3524 */
3525 if (!BUF_EMPTY(hdr)) {
3526 arc_buf_hdr_t *exists;
3527 kmutex_t *hash_lock;
3528
3529 ASSERT(zio->io_error == 0);
3530
3531 arc_cksum_verify(buf);
3532
3533 exists = buf_hash_insert(hdr, &hash_lock);
3534 if (exists) {
3535 /*
3536 * This can only happen if we overwrite for
3537 * sync-to-convergence, because we remove
3538 * buffers from the hash table when we arc_free().
3539 */
3540 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3541 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3542 panic("bad overwrite, hdr=%p exists=%p",
3543 (void *)hdr, (void *)exists);
3544 ASSERT(refcount_is_zero(&exists->b_refcnt));
3545 arc_change_state(arc_anon, exists, hash_lock);
3546 mutex_exit(hash_lock);
3547 arc_hdr_destroy(exists);
3548 exists = buf_hash_insert(hdr, &hash_lock);
3549 ASSERT3P(exists, ==, NULL);
3550 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3551 /* nopwrite */
3552 ASSERT(zio->io_prop.zp_nopwrite);
3553 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3554 panic("bad nopwrite, hdr=%p exists=%p",
3555 (void *)hdr, (void *)exists);
3556 } else {
3557 /* Dedup */
3558 ASSERT(hdr->b_datacnt == 1);
3559 ASSERT(hdr->b_state == arc_anon);
3560 ASSERT(BP_GET_DEDUP(zio->io_bp));
3561 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3562 }
3563 }
3564 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3565 /* if it's not anon, we are doing a scrub */
3566 if (!exists && hdr->b_state == arc_anon)
3567 arc_access(hdr, hash_lock);
3568 mutex_exit(hash_lock);
3569 } else {
3570 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3571 }
3572
3573 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3574 callback->awcb_done(zio, buf, callback->awcb_private);
3575
3576 kmem_free(callback, sizeof (arc_write_callback_t));
3577 }
3578
3579 zio_t *
3580 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3581 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
3582 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
3583 arc_done_func_t *done, void *private, zio_priority_t priority,
3584 int zio_flags, const zbookmark_t *zb)
3585 {
3586 arc_buf_hdr_t *hdr = buf->b_hdr;
3587 arc_write_callback_t *callback;
3588 zio_t *zio;
3589
3590 ASSERT(ready != NULL);
3591 ASSERT(done != NULL);
3592 ASSERT(!HDR_IO_ERROR(hdr));
3593 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3594 ASSERT(hdr->b_acb == NULL);
3595 if (l2arc)
3596 hdr->b_flags |= ARC_L2CACHE;
3597 if (l2arc_compress)
3598 hdr->b_flags |= ARC_L2COMPRESS;
3599 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3600 callback->awcb_ready = ready;
3601 callback->awcb_physdone = physdone;
3602 callback->awcb_done = done;
3603 callback->awcb_private = private;
3604 callback->awcb_buf = buf;
3605
3606 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3607 arc_write_ready, arc_write_physdone, arc_write_done, callback,
3608 priority, zio_flags, zb);
3609
3610 return (zio);
3611 }
3612
3613 static int
3614 arc_memory_throttle(uint64_t reserve, uint64_t txg)
3615 {
3616 #ifdef _KERNEL
3617 uint64_t available_memory = ptob(freemem);
3618 static uint64_t page_load = 0;
3619 static uint64_t last_txg = 0;
3620
3621 #if defined(__i386)
3622 available_memory =
3623 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3624 #endif
3625
3626 if (freemem > physmem * arc_lotsfree_percent / 100)
3627 return (0);
3628
3629 if (txg > last_txg) {
3630 last_txg = txg;
3631 page_load = 0;
3632 }
3633 /*
3634 * If we are in pageout, we know that memory is already tight,
3635 * the arc is already going to be evicting, so we just want to
3636 * continue to let page writes occur as quickly as possible.
3637 */
3638 if (curproc == proc_pageout) {
3639 if (page_load > MAX(ptob(minfree), available_memory) / 4)
3640 return (SET_ERROR(ERESTART));
3641 /* Note: reserve is inflated, so we deflate */
3642 page_load += reserve / 8;
3643 return (0);
3644 } else if (page_load > 0 && arc_reclaim_needed()) {
3645 /* memory is low, delay before restarting */
3646 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3647 return (SET_ERROR(EAGAIN));
3648 }
3649 page_load = 0;
3650 #endif
3651 return (0);
3652 }
3653
3654 void
3655 arc_tempreserve_clear(uint64_t reserve)
3656 {
3657 atomic_add_64(&arc_tempreserve, -reserve);
3658 ASSERT((int64_t)arc_tempreserve >= 0);
3659 }
3660
3661 int
3662 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3663 {
3664 int error;
3665 uint64_t anon_size;
3666
3667 if (reserve > arc_c/4 && !arc_no_grow)
3668 arc_c = MIN(arc_c_max, reserve * 4);
3669 if (reserve > arc_c)
3670 return (SET_ERROR(ENOMEM));
3671
3672 /*
3673 * Don't count loaned bufs as in flight dirty data to prevent long
3674 * network delays from blocking transactions that are ready to be
3675 * assigned to a txg.
3676 */
3677 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3678
3679 /*
3680 * Writes will, almost always, require additional memory allocations
3681 * in order to compress/encrypt/etc the data. We therefore need to
3682 * make sure that there is sufficient available memory for this.
3683 */
3684 error = arc_memory_throttle(reserve, txg);
3685 if (error != 0)
3686 return (error);
3687
3688 /*
3689 * Throttle writes when the amount of dirty data in the cache
3690 * gets too large. We try to keep the cache less than half full
3691 * of dirty blocks so that our sync times don't grow too large.
3692 * Note: if two requests come in concurrently, we might let them
3693 * both succeed, when one of them should fail. Not a huge deal.
3694 */
3695
3696 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3697 anon_size > arc_c / 4) {
3698 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3699 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3700 arc_tempreserve>>10,
3701 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3702 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3703 reserve>>10, arc_c>>10);
3704 return (SET_ERROR(ERESTART));
3705 }
3706 atomic_add_64(&arc_tempreserve, reserve);
3707 return (0);
3708 }
3709
3710 void
3711 arc_init(void)
3712 {
3713 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3714 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3715
3716 /* Convert seconds to clock ticks */
3717 arc_min_prefetch_lifespan = 1 * hz;
3718
3719 /* Start out with 1/8 of all memory */
3720 arc_c = physmem * PAGESIZE / 8;
3721
3722 #ifdef _KERNEL
3723 /*
3724 * On architectures where the physical memory can be larger
3725 * than the addressable space (intel in 32-bit mode), we may
3726 * need to limit the cache to 1/8 of VM size.
3727 */
3728 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3729 #endif
3730
3731 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3732 arc_c_min = MAX(arc_c / 4, 64<<20);
3733 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
3734 if (arc_c * 8 >= 1<<30)
3735 arc_c_max = (arc_c * 8) - (1<<30);
3736 else
3737 arc_c_max = arc_c_min;
3738 arc_c_max = MAX(arc_c * 6, arc_c_max);
3739
3740 /*
3741 * Allow the tunables to override our calculations if they are
3742 * reasonable (ie. over 64MB)
3743 */
3744 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3745 arc_c_max = zfs_arc_max;
3746 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3747 arc_c_min = zfs_arc_min;
3748
3749 arc_c = arc_c_max;
3750 arc_p = (arc_c >> 1);
3751
3752 /* limit meta-data to 1/4 of the arc capacity */
3753 arc_meta_limit = arc_c_max / 4;
3754
3755 /* Allow the tunable to override if it is reasonable */
3756 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3757 arc_meta_limit = zfs_arc_meta_limit;
3758
3759 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3760 arc_c_min = arc_meta_limit / 2;
3761
3762 if (zfs_arc_grow_retry > 0)
3763 arc_grow_retry = zfs_arc_grow_retry;
3764
3765 if (zfs_arc_shrink_shift > 0)
3766 arc_shrink_shift = zfs_arc_shrink_shift;
3767
3768 if (zfs_arc_p_min_shift > 0)
3769 arc_p_min_shift = zfs_arc_p_min_shift;
3770
3771 /* if kmem_flags are set, lets try to use less memory */
3772 if (kmem_debugging())
3773 arc_c = arc_c / 2;
3774 if (arc_c < arc_c_min)
3775 arc_c = arc_c_min;
3776
3777 arc_anon = &ARC_anon;
3778 arc_mru = &ARC_mru;
3779 arc_mru_ghost = &ARC_mru_ghost;
3780 arc_mfu = &ARC_mfu;
3781 arc_mfu_ghost = &ARC_mfu_ghost;
3782 arc_l2c_only = &ARC_l2c_only;
3783 arc_size = 0;
3784
3785 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3786 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3787 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3788 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3789 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3790 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3791
3792 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3793 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3794 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3795 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3796 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3797 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3798 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3799 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3800 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3801 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3802 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3803 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3804 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3805 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3806 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3807 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3808 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3809 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3810 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3811 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3812
3813 buf_init();
3814
3815 arc_thread_exit = 0;
3816 arc_eviction_list = NULL;
3817 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3818 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3819
3820 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3821 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3822
3823 if (arc_ksp != NULL) {
3824 arc_ksp->ks_data = &arc_stats;
3825 kstat_install(arc_ksp);
3826 }
3827
3828 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3829 TS_RUN, minclsyspri);
3830
3831 arc_dead = FALSE;
3832 arc_warm = B_FALSE;
3833
3834 /*
3835 * Calculate maximum amount of dirty data per pool.
3836 *
3837 * If it has been set by /etc/system, take that.
3838 * Otherwise, use a percentage of physical memory defined by
3839 * zfs_dirty_data_max_percent (default 10%) with a cap at
3840 * zfs_dirty_data_max_max (default 4GB).
3841 */
3842 if (zfs_dirty_data_max == 0) {
3843 zfs_dirty_data_max = physmem * PAGESIZE *
3844 zfs_dirty_data_max_percent / 100;
3845 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
3846 zfs_dirty_data_max_max);
3847 }
3848 }
3849
3850 void
3851 arc_fini(void)
3852 {
3853 mutex_enter(&arc_reclaim_thr_lock);
3854 arc_thread_exit = 1;
3855 while (arc_thread_exit != 0)
3856 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3857 mutex_exit(&arc_reclaim_thr_lock);
3858
3859 arc_flush(NULL);
3860
3861 arc_dead = TRUE;
3862
3863 if (arc_ksp != NULL) {
3864 kstat_delete(arc_ksp);
3865 arc_ksp = NULL;
3866 }
3867
3868 mutex_destroy(&arc_eviction_mtx);
3869 mutex_destroy(&arc_reclaim_thr_lock);
3870 cv_destroy(&arc_reclaim_thr_cv);
3871
3872 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3873 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3874 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3875 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3876 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3877 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3878 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3879 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3880
3881 mutex_destroy(&arc_anon->arcs_mtx);
3882 mutex_destroy(&arc_mru->arcs_mtx);
3883 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3884 mutex_destroy(&arc_mfu->arcs_mtx);
3885 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3886 mutex_destroy(&arc_l2c_only->arcs_mtx);
3887
3888 buf_fini();
3889
3890 ASSERT(arc_loaned_bytes == 0);
3891 }
3892
3893 /*
3894 * Level 2 ARC
3895 *
3896 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3897 * It uses dedicated storage devices to hold cached data, which are populated
3898 * using large infrequent writes. The main role of this cache is to boost
3899 * the performance of random read workloads. The intended L2ARC devices
3900 * include short-stroked disks, solid state disks, and other media with
3901 * substantially faster read latency than disk.
3902 *
3903 * +-----------------------+
3904 * | ARC |
3905 * +-----------------------+
3906 * | ^ ^
3907 * | | |
3908 * l2arc_feed_thread() arc_read()
3909 * | | |
3910 * | l2arc read |
3911 * V | |
3912 * +---------------+ |
3913 * | L2ARC | |
3914 * +---------------+ |
3915 * | ^ |
3916 * l2arc_write() | |
3917 * | | |
3918 * V | |
3919 * +-------+ +-------+
3920 * | vdev | | vdev |
3921 * | cache | | cache |
3922 * +-------+ +-------+
3923 * +=========+ .-----.
3924 * : L2ARC : |-_____-|
3925 * : devices : | Disks |
3926 * +=========+ `-_____-'
3927 *
3928 * Read requests are satisfied from the following sources, in order:
3929 *
3930 * 1) ARC
3931 * 2) vdev cache of L2ARC devices
3932 * 3) L2ARC devices
3933 * 4) vdev cache of disks
3934 * 5) disks
3935 *
3936 * Some L2ARC device types exhibit extremely slow write performance.
3937 * To accommodate for this there are some significant differences between
3938 * the L2ARC and traditional cache design:
3939 *
3940 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3941 * the ARC behave as usual, freeing buffers and placing headers on ghost
3942 * lists. The ARC does not send buffers to the L2ARC during eviction as
3943 * this would add inflated write latencies for all ARC memory pressure.
3944 *
3945 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3946 * It does this by periodically scanning buffers from the eviction-end of
3947 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3948 * not already there. It scans until a headroom of buffers is satisfied,
3949 * which itself is a buffer for ARC eviction. If a compressible buffer is
3950 * found during scanning and selected for writing to an L2ARC device, we
3951 * temporarily boost scanning headroom during the next scan cycle to make
3952 * sure we adapt to compression effects (which might significantly reduce
3953 * the data volume we write to L2ARC). The thread that does this is
3954 * l2arc_feed_thread(), illustrated below; example sizes are included to
3955 * provide a better sense of ratio than this diagram:
3956 *
3957 * head --> tail
3958 * +---------------------+----------+
3959 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
3960 * +---------------------+----------+ | o L2ARC eligible
3961 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
3962 * +---------------------+----------+ |
3963 * 15.9 Gbytes ^ 32 Mbytes |
3964 * headroom |
3965 * l2arc_feed_thread()
3966 * |
3967 * l2arc write hand <--[oooo]--'
3968 * | 8 Mbyte
3969 * | write max
3970 * V
3971 * +==============================+
3972 * L2ARC dev |####|#|###|###| |####| ... |
3973 * +==============================+
3974 * 32 Gbytes
3975 *
3976 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
3977 * evicted, then the L2ARC has cached a buffer much sooner than it probably
3978 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
3979 * safe to say that this is an uncommon case, since buffers at the end of
3980 * the ARC lists have moved there due to inactivity.
3981 *
3982 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
3983 * then the L2ARC simply misses copying some buffers. This serves as a
3984 * pressure valve to prevent heavy read workloads from both stalling the ARC
3985 * with waits and clogging the L2ARC with writes. This also helps prevent
3986 * the potential for the L2ARC to churn if it attempts to cache content too
3987 * quickly, such as during backups of the entire pool.
3988 *
3989 * 5. After system boot and before the ARC has filled main memory, there are
3990 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
3991 * lists can remain mostly static. Instead of searching from tail of these
3992 * lists as pictured, the l2arc_feed_thread() will search from the list heads
3993 * for eligible buffers, greatly increasing its chance of finding them.
3994 *
3995 * The L2ARC device write speed is also boosted during this time so that
3996 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
3997 * there are no L2ARC reads, and no fear of degrading read performance
3998 * through increased writes.
3999 *
4000 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4001 * the vdev queue can aggregate them into larger and fewer writes. Each
4002 * device is written to in a rotor fashion, sweeping writes through
4003 * available space then repeating.
4004 *
4005 * 7. The L2ARC does not store dirty content. It never needs to flush
4006 * write buffers back to disk based storage.
4007 *
4008 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4009 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4010 *
4011 * The performance of the L2ARC can be tweaked by a number of tunables, which
4012 * may be necessary for different workloads:
4013 *
4014 * l2arc_write_max max write bytes per interval
4015 * l2arc_write_boost extra write bytes during device warmup
4016 * l2arc_noprefetch skip caching prefetched buffers
4017 * l2arc_headroom number of max device writes to precache
4018 * l2arc_headroom_boost when we find compressed buffers during ARC
4019 * scanning, we multiply headroom by this
4020 * percentage factor for the next scan cycle,
4021 * since more compressed buffers are likely to
4022 * be present
4023 * l2arc_feed_secs seconds between L2ARC writing
4024 *
4025 * Tunables may be removed or added as future performance improvements are
4026 * integrated, and also may become zpool properties.
4027 *
4028 * There are three key functions that control how the L2ARC warms up:
4029 *
4030 * l2arc_write_eligible() check if a buffer is eligible to cache
4031 * l2arc_write_size() calculate how much to write
4032 * l2arc_write_interval() calculate sleep delay between writes
4033 *
4034 * These three functions determine what to write, how much, and how quickly
4035 * to send writes.
4036 */
4037
4038 static boolean_t
4039 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4040 {
4041 /*
4042 * A buffer is *not* eligible for the L2ARC if it:
4043 * 1. belongs to a different spa.
4044 * 2. is already cached on the L2ARC.
4045 * 3. has an I/O in progress (it may be an incomplete read).
4046 * 4. is flagged not eligible (zfs property).
4047 */
4048 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
4049 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
4050 return (B_FALSE);
4051
4052 return (B_TRUE);
4053 }
4054
4055 static uint64_t
4056 l2arc_write_size(void)
4057 {
4058 uint64_t size;
4059
4060 /*
4061 * Make sure our globals have meaningful values in case the user
4062 * altered them.
4063 */
4064 size = l2arc_write_max;
4065 if (size == 0) {
4066 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
4067 "be greater than zero, resetting it to the default (%d)",
4068 L2ARC_WRITE_SIZE);
4069 size = l2arc_write_max = L2ARC_WRITE_SIZE;
4070 }
4071
4072 if (arc_warm == B_FALSE)
4073 size += l2arc_write_boost;
4074
4075 return (size);
4076
4077 }
4078
4079 static clock_t
4080 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4081 {
4082 clock_t interval, next, now;
4083
4084 /*
4085 * If the ARC lists are busy, increase our write rate; if the
4086 * lists are stale, idle back. This is achieved by checking
4087 * how much we previously wrote - if it was more than half of
4088 * what we wanted, schedule the next write much sooner.
4089 */
4090 if (l2arc_feed_again && wrote > (wanted / 2))
4091 interval = (hz * l2arc_feed_min_ms) / 1000;
4092 else
4093 interval = hz * l2arc_feed_secs;
4094
4095 now = ddi_get_lbolt();
4096 next = MAX(now, MIN(now + interval, began + interval));
4097
4098 return (next);
4099 }
4100
4101 static void
4102 l2arc_hdr_stat_add(void)
4103 {
4104 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4105 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4106 }
4107
4108 static void
4109 l2arc_hdr_stat_remove(void)
4110 {
4111 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4112 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4113 }
4114
4115 /*
4116 * Cycle through L2ARC devices. This is how L2ARC load balances.
4117 * If a device is returned, this also returns holding the spa config lock.
4118 */
4119 static l2arc_dev_t *
4120 l2arc_dev_get_next(void)
4121 {
4122 l2arc_dev_t *first, *next = NULL;
4123
4124 /*
4125 * Lock out the removal of spas (spa_namespace_lock), then removal
4126 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4127 * both locks will be dropped and a spa config lock held instead.
4128 */
4129 mutex_enter(&spa_namespace_lock);
4130 mutex_enter(&l2arc_dev_mtx);
4131
4132 /* if there are no vdevs, there is nothing to do */
4133 if (l2arc_ndev == 0)
4134 goto out;
4135
4136 first = NULL;
4137 next = l2arc_dev_last;
4138 do {
4139 /* loop around the list looking for a non-faulted vdev */
4140 if (next == NULL) {
4141 next = list_head(l2arc_dev_list);
4142 } else {
4143 next = list_next(l2arc_dev_list, next);
4144 if (next == NULL)
4145 next = list_head(l2arc_dev_list);
4146 }
4147
4148 /* if we have come back to the start, bail out */
4149 if (first == NULL)
4150 first = next;
4151 else if (next == first)
4152 break;
4153
4154 } while (vdev_is_dead(next->l2ad_vdev));
4155
4156 /* if we were unable to find any usable vdevs, return NULL */
4157 if (vdev_is_dead(next->l2ad_vdev))
4158 next = NULL;
4159
4160 l2arc_dev_last = next;
4161
4162 out:
4163 mutex_exit(&l2arc_dev_mtx);
4164
4165 /*
4166 * Grab the config lock to prevent the 'next' device from being
4167 * removed while we are writing to it.
4168 */
4169 if (next != NULL)
4170 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4171 mutex_exit(&spa_namespace_lock);
4172
4173 return (next);
4174 }
4175
4176 /*
4177 * Free buffers that were tagged for destruction.
4178 */
4179 static void
4180 l2arc_do_free_on_write()
4181 {
4182 list_t *buflist;
4183 l2arc_data_free_t *df, *df_prev;
4184
4185 mutex_enter(&l2arc_free_on_write_mtx);
4186 buflist = l2arc_free_on_write;
4187
4188 for (df = list_tail(buflist); df; df = df_prev) {
4189 df_prev = list_prev(buflist, df);
4190 ASSERT(df->l2df_data != NULL);
4191 ASSERT(df->l2df_func != NULL);
4192 df->l2df_func(df->l2df_data, df->l2df_size);
4193 list_remove(buflist, df);
4194 kmem_free(df, sizeof (l2arc_data_free_t));
4195 }
4196
4197 mutex_exit(&l2arc_free_on_write_mtx);
4198 }
4199
4200 /*
4201 * A write to a cache device has completed. Update all headers to allow
4202 * reads from these buffers to begin.
4203 */
4204 static void
4205 l2arc_write_done(zio_t *zio)
4206 {
4207 l2arc_write_callback_t *cb;
4208 l2arc_dev_t *dev;
4209 list_t *buflist;
4210 arc_buf_hdr_t *head, *ab, *ab_prev;
4211 l2arc_buf_hdr_t *abl2;
4212 kmutex_t *hash_lock;
4213
4214 cb = zio->io_private;
4215 ASSERT(cb != NULL);
4216 dev = cb->l2wcb_dev;
4217 ASSERT(dev != NULL);
4218 head = cb->l2wcb_head;
4219 ASSERT(head != NULL);
4220 buflist = dev->l2ad_buflist;
4221 ASSERT(buflist != NULL);
4222 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4223 l2arc_write_callback_t *, cb);
4224
4225 if (zio->io_error != 0)
4226 ARCSTAT_BUMP(arcstat_l2_writes_error);
4227
4228 mutex_enter(&l2arc_buflist_mtx);
4229
4230 /*
4231 * All writes completed, or an error was hit.
4232 */
4233 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4234 ab_prev = list_prev(buflist, ab);
4235
4236 hash_lock = HDR_LOCK(ab);
4237 if (!mutex_tryenter(hash_lock)) {
4238 /*
4239 * This buffer misses out. It may be in a stage
4240 * of eviction. Its ARC_L2_WRITING flag will be
4241 * left set, denying reads to this buffer.
4242 */
4243 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4244 continue;
4245 }
4246
4247 abl2 = ab->b_l2hdr;
4248
4249 /*
4250 * Release the temporary compressed buffer as soon as possible.
4251 */
4252 if (abl2->b_compress != ZIO_COMPRESS_OFF)
4253 l2arc_release_cdata_buf(ab);
4254
4255 if (zio->io_error != 0) {
4256 /*
4257 * Error - drop L2ARC entry.
4258 */
4259 list_remove(buflist, ab);
4260 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4261 ab->b_l2hdr = NULL;
4262 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4263 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4264 }
4265
4266 /*
4267 * Allow ARC to begin reads to this L2ARC entry.
4268 */
4269 ab->b_flags &= ~ARC_L2_WRITING;
4270
4271 mutex_exit(hash_lock);
4272 }
4273
4274 atomic_inc_64(&l2arc_writes_done);
4275 list_remove(buflist, head);
4276 kmem_cache_free(hdr_cache, head);
4277 mutex_exit(&l2arc_buflist_mtx);
4278
4279 l2arc_do_free_on_write();
4280
4281 kmem_free(cb, sizeof (l2arc_write_callback_t));
4282 }
4283
4284 /*
4285 * A read to a cache device completed. Validate buffer contents before
4286 * handing over to the regular ARC routines.
4287 */
4288 static void
4289 l2arc_read_done(zio_t *zio)
4290 {
4291 l2arc_read_callback_t *cb;
4292 arc_buf_hdr_t *hdr;
4293 arc_buf_t *buf;
4294 kmutex_t *hash_lock;
4295 int equal;
4296
4297 ASSERT(zio->io_vd != NULL);
4298 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4299
4300 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4301
4302 cb = zio->io_private;
4303 ASSERT(cb != NULL);
4304 buf = cb->l2rcb_buf;
4305 ASSERT(buf != NULL);
4306
4307 hash_lock = HDR_LOCK(buf->b_hdr);
4308 mutex_enter(hash_lock);
4309 hdr = buf->b_hdr;
4310 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4311
4312 /*
4313 * If the buffer was compressed, decompress it first.
4314 */
4315 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
4316 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
4317 ASSERT(zio->io_data != NULL);
4318
4319 /*
4320 * Check this survived the L2ARC journey.
4321 */
4322 equal = arc_cksum_equal(buf);
4323 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4324 mutex_exit(hash_lock);
4325 zio->io_private = buf;
4326 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4327 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4328 arc_read_done(zio);
4329 } else {
4330 mutex_exit(hash_lock);
4331 /*
4332 * Buffer didn't survive caching. Increment stats and
4333 * reissue to the original storage device.
4334 */
4335 if (zio->io_error != 0) {
4336 ARCSTAT_BUMP(arcstat_l2_io_error);
4337 } else {
4338 zio->io_error = SET_ERROR(EIO);
4339 }
4340 if (!equal)
4341 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4342
4343 /*
4344 * If there's no waiter, issue an async i/o to the primary
4345 * storage now. If there *is* a waiter, the caller must
4346 * issue the i/o in a context where it's OK to block.
4347 */
4348 if (zio->io_waiter == NULL) {
4349 zio_t *pio = zio_unique_parent(zio);
4350
4351 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4352
4353 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4354 buf->b_data, zio->io_size, arc_read_done, buf,
4355 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4356 }
4357 }
4358
4359 kmem_free(cb, sizeof (l2arc_read_callback_t));
4360 }
4361
4362 /*
4363 * This is the list priority from which the L2ARC will search for pages to
4364 * cache. This is used within loops (0..3) to cycle through lists in the
4365 * desired order. This order can have a significant effect on cache
4366 * performance.
4367 *
4368 * Currently the metadata lists are hit first, MFU then MRU, followed by
4369 * the data lists. This function returns a locked list, and also returns
4370 * the lock pointer.
4371 */
4372 static list_t *
4373 l2arc_list_locked(int list_num, kmutex_t **lock)
4374 {
4375 list_t *list = NULL;
4376
4377 ASSERT(list_num >= 0 && list_num <= 3);
4378
4379 switch (list_num) {
4380 case 0:
4381 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4382 *lock = &arc_mfu->arcs_mtx;
4383 break;
4384 case 1:
4385 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4386 *lock = &arc_mru->arcs_mtx;
4387 break;
4388 case 2:
4389 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4390 *lock = &arc_mfu->arcs_mtx;
4391 break;
4392 case 3:
4393 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4394 *lock = &arc_mru->arcs_mtx;
4395 break;
4396 }
4397
4398 ASSERT(!(MUTEX_HELD(*lock)));
4399 mutex_enter(*lock);
4400 return (list);
4401 }
4402
4403 /*
4404 * Evict buffers from the device write hand to the distance specified in
4405 * bytes. This distance may span populated buffers, it may span nothing.
4406 * This is clearing a region on the L2ARC device ready for writing.
4407 * If the 'all' boolean is set, every buffer is evicted.
4408 */
4409 static void
4410 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4411 {
4412 list_t *buflist;
4413 l2arc_buf_hdr_t *abl2;
4414 arc_buf_hdr_t *ab, *ab_prev;
4415 kmutex_t *hash_lock;
4416 uint64_t taddr;
4417
4418 buflist = dev->l2ad_buflist;
4419
4420 if (buflist == NULL)
4421 return;
4422
4423 if (!all && dev->l2ad_first) {
4424 /*
4425 * This is the first sweep through the device. There is
4426 * nothing to evict.
4427 */
4428 return;
4429 }
4430
4431 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4432 /*
4433 * When nearing the end of the device, evict to the end
4434 * before the device write hand jumps to the start.
4435 */
4436 taddr = dev->l2ad_end;
4437 } else {
4438 taddr = dev->l2ad_hand + distance;
4439 }
4440 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4441 uint64_t, taddr, boolean_t, all);
4442
4443 top:
4444 mutex_enter(&l2arc_buflist_mtx);
4445 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4446 ab_prev = list_prev(buflist, ab);
4447
4448 hash_lock = HDR_LOCK(ab);
4449 if (!mutex_tryenter(hash_lock)) {
4450 /*
4451 * Missed the hash lock. Retry.
4452 */
4453 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4454 mutex_exit(&l2arc_buflist_mtx);
4455 mutex_enter(hash_lock);
4456 mutex_exit(hash_lock);
4457 goto top;
4458 }
4459
4460 if (HDR_L2_WRITE_HEAD(ab)) {
4461 /*
4462 * We hit a write head node. Leave it for
4463 * l2arc_write_done().
4464 */
4465 list_remove(buflist, ab);
4466 mutex_exit(hash_lock);
4467 continue;
4468 }
4469
4470 if (!all && ab->b_l2hdr != NULL &&
4471 (ab->b_l2hdr->b_daddr > taddr ||
4472 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4473 /*
4474 * We've evicted to the target address,
4475 * or the end of the device.
4476 */
4477 mutex_exit(hash_lock);
4478 break;
4479 }
4480
4481 if (HDR_FREE_IN_PROGRESS(ab)) {
4482 /*
4483 * Already on the path to destruction.
4484 */
4485 mutex_exit(hash_lock);
4486 continue;
4487 }
4488
4489 if (ab->b_state == arc_l2c_only) {
4490 ASSERT(!HDR_L2_READING(ab));
4491 /*
4492 * This doesn't exist in the ARC. Destroy.
4493 * arc_hdr_destroy() will call list_remove()
4494 * and decrement arcstat_l2_size.
4495 */
4496 arc_change_state(arc_anon, ab, hash_lock);
4497 arc_hdr_destroy(ab);
4498 } else {
4499 /*
4500 * Invalidate issued or about to be issued
4501 * reads, since we may be about to write
4502 * over this location.
4503 */
4504 if (HDR_L2_READING(ab)) {
4505 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4506 ab->b_flags |= ARC_L2_EVICTED;
4507 }
4508
4509 /*
4510 * Tell ARC this no longer exists in L2ARC.
4511 */
4512 if (ab->b_l2hdr != NULL) {
4513 abl2 = ab->b_l2hdr;
4514 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4515 ab->b_l2hdr = NULL;
4516 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4517 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4518 }
4519 list_remove(buflist, ab);
4520
4521 /*
4522 * This may have been leftover after a
4523 * failed write.
4524 */
4525 ab->b_flags &= ~ARC_L2_WRITING;
4526 }
4527 mutex_exit(hash_lock);
4528 }
4529 mutex_exit(&l2arc_buflist_mtx);
4530
4531 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4532 dev->l2ad_evict = taddr;
4533 }
4534
4535 /*
4536 * Find and write ARC buffers to the L2ARC device.
4537 *
4538 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4539 * for reading until they have completed writing.
4540 * The headroom_boost is an in-out parameter used to maintain headroom boost
4541 * state between calls to this function.
4542 *
4543 * Returns the number of bytes actually written (which may be smaller than
4544 * the delta by which the device hand has changed due to alignment).
4545 */
4546 static uint64_t
4547 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
4548 boolean_t *headroom_boost)
4549 {
4550 arc_buf_hdr_t *ab, *ab_prev, *head;
4551 list_t *list;
4552 uint64_t write_asize, write_psize, write_sz, headroom,
4553 buf_compress_minsz;
4554 void *buf_data;
4555 kmutex_t *list_lock;
4556 boolean_t full;
4557 l2arc_write_callback_t *cb;
4558 zio_t *pio, *wzio;
4559 uint64_t guid = spa_load_guid(spa);
4560 const boolean_t do_headroom_boost = *headroom_boost;
4561
4562 ASSERT(dev->l2ad_vdev != NULL);
4563
4564 /* Lower the flag now, we might want to raise it again later. */
4565 *headroom_boost = B_FALSE;
4566
4567 pio = NULL;
4568 write_sz = write_asize = write_psize = 0;
4569 full = B_FALSE;
4570 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4571 head->b_flags |= ARC_L2_WRITE_HEAD;
4572
4573 /*
4574 * We will want to try to compress buffers that are at least 2x the
4575 * device sector size.
4576 */
4577 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
4578
4579 /*
4580 * Copy buffers for L2ARC writing.
4581 */
4582 mutex_enter(&l2arc_buflist_mtx);
4583 for (int try = 0; try <= 3; try++) {
4584 uint64_t passed_sz = 0;
4585
4586 list = l2arc_list_locked(try, &list_lock);
4587
4588 /*
4589 * L2ARC fast warmup.
4590 *
4591 * Until the ARC is warm and starts to evict, read from the
4592 * head of the ARC lists rather than the tail.
4593 */
4594 if (arc_warm == B_FALSE)
4595 ab = list_head(list);
4596 else
4597 ab = list_tail(list);
4598
4599 headroom = target_sz * l2arc_headroom;
4600 if (do_headroom_boost)
4601 headroom = (headroom * l2arc_headroom_boost) / 100;
4602
4603 for (; ab; ab = ab_prev) {
4604 l2arc_buf_hdr_t *l2hdr;
4605 kmutex_t *hash_lock;
4606 uint64_t buf_sz;
4607
4608 if (arc_warm == B_FALSE)
4609 ab_prev = list_next(list, ab);
4610 else
4611 ab_prev = list_prev(list, ab);
4612
4613 hash_lock = HDR_LOCK(ab);
4614 if (!mutex_tryenter(hash_lock)) {
4615 /*
4616 * Skip this buffer rather than waiting.
4617 */
4618 continue;
4619 }
4620
4621 passed_sz += ab->b_size;
4622 if (passed_sz > headroom) {
4623 /*
4624 * Searched too far.
4625 */
4626 mutex_exit(hash_lock);
4627 break;
4628 }
4629
4630 if (!l2arc_write_eligible(guid, ab)) {
4631 mutex_exit(hash_lock);
4632 continue;
4633 }
4634
4635 if ((write_sz + ab->b_size) > target_sz) {
4636 full = B_TRUE;
4637 mutex_exit(hash_lock);
4638 break;
4639 }
4640
4641 if (pio == NULL) {
4642 /*
4643 * Insert a dummy header on the buflist so
4644 * l2arc_write_done() can find where the
4645 * write buffers begin without searching.
4646 */
4647 list_insert_head(dev->l2ad_buflist, head);
4648
4649 cb = kmem_alloc(
4650 sizeof (l2arc_write_callback_t), KM_SLEEP);
4651 cb->l2wcb_dev = dev;
4652 cb->l2wcb_head = head;
4653 pio = zio_root(spa, l2arc_write_done, cb,
4654 ZIO_FLAG_CANFAIL);
4655 }
4656
4657 /*
4658 * Create and add a new L2ARC header.
4659 */
4660 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4661 l2hdr->b_dev = dev;
4662 ab->b_flags |= ARC_L2_WRITING;
4663
4664 /*
4665 * Temporarily stash the data buffer in b_tmp_cdata.
4666 * The subsequent write step will pick it up from
4667 * there. This is because can't access ab->b_buf
4668 * without holding the hash_lock, which we in turn
4669 * can't access without holding the ARC list locks
4670 * (which we want to avoid during compression/writing).
4671 */
4672 l2hdr->b_compress = ZIO_COMPRESS_OFF;
4673 l2hdr->b_asize = ab->b_size;
4674 l2hdr->b_tmp_cdata = ab->b_buf->b_data;
4675
4676 buf_sz = ab->b_size;
4677 ab->b_l2hdr = l2hdr;
4678
4679 list_insert_head(dev->l2ad_buflist, ab);
4680
4681 /*
4682 * Compute and store the buffer cksum before
4683 * writing. On debug the cksum is verified first.
4684 */
4685 arc_cksum_verify(ab->b_buf);
4686 arc_cksum_compute(ab->b_buf, B_TRUE);
4687
4688 mutex_exit(hash_lock);
4689
4690 write_sz += buf_sz;
4691 }
4692
4693 mutex_exit(list_lock);
4694
4695 if (full == B_TRUE)
4696 break;
4697 }
4698
4699 /* No buffers selected for writing? */
4700 if (pio == NULL) {
4701 ASSERT0(write_sz);
4702 mutex_exit(&l2arc_buflist_mtx);
4703 kmem_cache_free(hdr_cache, head);
4704 return (0);
4705 }
4706
4707 /*
4708 * Now start writing the buffers. We're starting at the write head
4709 * and work backwards, retracing the course of the buffer selector
4710 * loop above.
4711 */
4712 for (ab = list_prev(dev->l2ad_buflist, head); ab;
4713 ab = list_prev(dev->l2ad_buflist, ab)) {
4714 l2arc_buf_hdr_t *l2hdr;
4715 uint64_t buf_sz;
4716
4717 /*
4718 * We shouldn't need to lock the buffer here, since we flagged
4719 * it as ARC_L2_WRITING in the previous step, but we must take
4720 * care to only access its L2 cache parameters. In particular,
4721 * ab->b_buf may be invalid by now due to ARC eviction.
4722 */
4723 l2hdr = ab->b_l2hdr;
4724 l2hdr->b_daddr = dev->l2ad_hand;
4725
4726 if ((ab->b_flags & ARC_L2COMPRESS) &&
4727 l2hdr->b_asize >= buf_compress_minsz) {
4728 if (l2arc_compress_buf(l2hdr)) {
4729 /*
4730 * If compression succeeded, enable headroom
4731 * boost on the next scan cycle.
4732 */
4733 *headroom_boost = B_TRUE;
4734 }
4735 }
4736
4737 /*
4738 * Pick up the buffer data we had previously stashed away
4739 * (and now potentially also compressed).
4740 */
4741 buf_data = l2hdr->b_tmp_cdata;
4742 buf_sz = l2hdr->b_asize;
4743
4744 /* Compression may have squashed the buffer to zero length. */
4745 if (buf_sz != 0) {
4746 uint64_t buf_p_sz;
4747
4748 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4749 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4750 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4751 ZIO_FLAG_CANFAIL, B_FALSE);
4752
4753 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4754 zio_t *, wzio);
4755 (void) zio_nowait(wzio);
4756
4757 write_asize += buf_sz;
4758 /*
4759 * Keep the clock hand suitably device-aligned.
4760 */
4761 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4762 write_psize += buf_p_sz;
4763 dev->l2ad_hand += buf_p_sz;
4764 }
4765 }
4766
4767 mutex_exit(&l2arc_buflist_mtx);
4768
4769 ASSERT3U(write_asize, <=, target_sz);
4770 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4771 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
4772 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4773 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
4774 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
4775
4776 /*
4777 * Bump device hand to the device start if it is approaching the end.
4778 * l2arc_evict() will already have evicted ahead for this case.
4779 */
4780 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4781 vdev_space_update(dev->l2ad_vdev,
4782 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4783 dev->l2ad_hand = dev->l2ad_start;
4784 dev->l2ad_evict = dev->l2ad_start;
4785 dev->l2ad_first = B_FALSE;
4786 }
4787
4788 dev->l2ad_writing = B_TRUE;
4789 (void) zio_wait(pio);
4790 dev->l2ad_writing = B_FALSE;
4791
4792 return (write_asize);
4793 }
4794
4795 /*
4796 * Compresses an L2ARC buffer.
4797 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its
4798 * size in l2hdr->b_asize. This routine tries to compress the data and
4799 * depending on the compression result there are three possible outcomes:
4800 * *) The buffer was incompressible. The original l2hdr contents were left
4801 * untouched and are ready for writing to an L2 device.
4802 * *) The buffer was all-zeros, so there is no need to write it to an L2
4803 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
4804 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
4805 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
4806 * data buffer which holds the compressed data to be written, and b_asize
4807 * tells us how much data there is. b_compress is set to the appropriate
4808 * compression algorithm. Once writing is done, invoke
4809 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
4810 *
4811 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
4812 * buffer was incompressible).
4813 */
4814 static boolean_t
4815 l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr)
4816 {
4817 void *cdata;
4818 size_t csize, len;
4819
4820 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF);
4821 ASSERT(l2hdr->b_tmp_cdata != NULL);
4822
4823 len = l2hdr->b_asize;
4824 cdata = zio_data_buf_alloc(len);
4825 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata,
4826 cdata, l2hdr->b_asize);
4827
4828 if (csize == 0) {
4829 /* zero block, indicate that there's nothing to write */
4830 zio_data_buf_free(cdata, len);
4831 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
4832 l2hdr->b_asize = 0;
4833 l2hdr->b_tmp_cdata = NULL;
4834 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
4835 return (B_TRUE);
4836 } else if (csize > 0 && csize < len) {
4837 /*
4838 * Compression succeeded, we'll keep the cdata around for
4839 * writing and release it afterwards.
4840 */
4841 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
4842 l2hdr->b_asize = csize;
4843 l2hdr->b_tmp_cdata = cdata;
4844 ARCSTAT_BUMP(arcstat_l2_compress_successes);
4845 return (B_TRUE);
4846 } else {
4847 /*
4848 * Compression failed, release the compressed buffer.
4849 * l2hdr will be left unmodified.
4850 */
4851 zio_data_buf_free(cdata, len);
4852 ARCSTAT_BUMP(arcstat_l2_compress_failures);
4853 return (B_FALSE);
4854 }
4855 }
4856
4857 /*
4858 * Decompresses a zio read back from an l2arc device. On success, the
4859 * underlying zio's io_data buffer is overwritten by the uncompressed
4860 * version. On decompression error (corrupt compressed stream), the
4861 * zio->io_error value is set to signal an I/O error.
4862 *
4863 * Please note that the compressed data stream is not checksummed, so
4864 * if the underlying device is experiencing data corruption, we may feed
4865 * corrupt data to the decompressor, so the decompressor needs to be
4866 * able to handle this situation (LZ4 does).
4867 */
4868 static void
4869 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
4870 {
4871 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
4872
4873 if (zio->io_error != 0) {
4874 /*
4875 * An io error has occured, just restore the original io
4876 * size in preparation for a main pool read.
4877 */
4878 zio->io_orig_size = zio->io_size = hdr->b_size;
4879 return;
4880 }
4881
4882 if (c == ZIO_COMPRESS_EMPTY) {
4883 /*
4884 * An empty buffer results in a null zio, which means we
4885 * need to fill its io_data after we're done restoring the
4886 * buffer's contents.
4887 */
4888 ASSERT(hdr->b_buf != NULL);
4889 bzero(hdr->b_buf->b_data, hdr->b_size);
4890 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data;
4891 } else {
4892 ASSERT(zio->io_data != NULL);
4893 /*
4894 * We copy the compressed data from the start of the arc buffer
4895 * (the zio_read will have pulled in only what we need, the
4896 * rest is garbage which we will overwrite at decompression)
4897 * and then decompress back to the ARC data buffer. This way we
4898 * can minimize copying by simply decompressing back over the
4899 * original compressed data (rather than decompressing to an
4900 * aux buffer and then copying back the uncompressed buffer,
4901 * which is likely to be much larger).
4902 */
4903 uint64_t csize;
4904 void *cdata;
4905
4906 csize = zio->io_size;
4907 cdata = zio_data_buf_alloc(csize);
4908 bcopy(zio->io_data, cdata, csize);
4909 if (zio_decompress_data(c, cdata, zio->io_data, csize,
4910 hdr->b_size) != 0)
4911 zio->io_error = EIO;
4912 zio_data_buf_free(cdata, csize);
4913 }
4914
4915 /* Restore the expected uncompressed IO size. */
4916 zio->io_orig_size = zio->io_size = hdr->b_size;
4917 }
4918
4919 /*
4920 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
4921 * This buffer serves as a temporary holder of compressed data while
4922 * the buffer entry is being written to an l2arc device. Once that is
4923 * done, we can dispose of it.
4924 */
4925 static void
4926 l2arc_release_cdata_buf(arc_buf_hdr_t *ab)
4927 {
4928 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr;
4929
4930 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) {
4931 /*
4932 * If the data was compressed, then we've allocated a
4933 * temporary buffer for it, so now we need to release it.
4934 */
4935 ASSERT(l2hdr->b_tmp_cdata != NULL);
4936 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size);
4937 }
4938 l2hdr->b_tmp_cdata = NULL;
4939 }
4940
4941 /*
4942 * This thread feeds the L2ARC at regular intervals. This is the beating
4943 * heart of the L2ARC.
4944 */
4945 static void
4946 l2arc_feed_thread(void)
4947 {
4948 callb_cpr_t cpr;
4949 l2arc_dev_t *dev;
4950 spa_t *spa;
4951 uint64_t size, wrote;
4952 clock_t begin, next = ddi_get_lbolt();
4953 boolean_t headroom_boost = B_FALSE;
4954
4955 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4956
4957 mutex_enter(&l2arc_feed_thr_lock);
4958
4959 while (l2arc_thread_exit == 0) {
4960 CALLB_CPR_SAFE_BEGIN(&cpr);
4961 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4962 next);
4963 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4964 next = ddi_get_lbolt() + hz;
4965
4966 /*
4967 * Quick check for L2ARC devices.
4968 */
4969 mutex_enter(&l2arc_dev_mtx);
4970 if (l2arc_ndev == 0) {
4971 mutex_exit(&l2arc_dev_mtx);
4972 continue;
4973 }
4974 mutex_exit(&l2arc_dev_mtx);
4975 begin = ddi_get_lbolt();
4976
4977 /*
4978 * This selects the next l2arc device to write to, and in
4979 * doing so the next spa to feed from: dev->l2ad_spa. This
4980 * will return NULL if there are now no l2arc devices or if
4981 * they are all faulted.
4982 *
4983 * If a device is returned, its spa's config lock is also
4984 * held to prevent device removal. l2arc_dev_get_next()
4985 * will grab and release l2arc_dev_mtx.
4986 */
4987 if ((dev = l2arc_dev_get_next()) == NULL)
4988 continue;
4989
4990 spa = dev->l2ad_spa;
4991 ASSERT(spa != NULL);
4992
4993 /*
4994 * If the pool is read-only then force the feed thread to
4995 * sleep a little longer.
4996 */
4997 if (!spa_writeable(spa)) {
4998 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4999 spa_config_exit(spa, SCL_L2ARC, dev);
5000 continue;
5001 }
5002
5003 /*
5004 * Avoid contributing to memory pressure.
5005 */
5006 if (arc_reclaim_needed()) {
5007 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
5008 spa_config_exit(spa, SCL_L2ARC, dev);
5009 continue;
5010 }
5011
5012 ARCSTAT_BUMP(arcstat_l2_feeds);
5013
5014 size = l2arc_write_size();
5015
5016 /*
5017 * Evict L2ARC buffers that will be overwritten.
5018 */
5019 l2arc_evict(dev, size, B_FALSE);
5020
5021 /*
5022 * Write ARC buffers.
5023 */
5024 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
5025
5026 /*
5027 * Calculate interval between writes.
5028 */
5029 next = l2arc_write_interval(begin, size, wrote);
5030 spa_config_exit(spa, SCL_L2ARC, dev);
5031 }
5032
5033 l2arc_thread_exit = 0;
5034 cv_broadcast(&l2arc_feed_thr_cv);
5035 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
5036 thread_exit();
5037 }
5038
5039 boolean_t
5040 l2arc_vdev_present(vdev_t *vd)
5041 {
5042 l2arc_dev_t *dev;
5043
5044 mutex_enter(&l2arc_dev_mtx);
5045 for (dev = list_head(l2arc_dev_list); dev != NULL;
5046 dev = list_next(l2arc_dev_list, dev)) {
5047 if (dev->l2ad_vdev == vd)
5048 break;
5049 }
5050 mutex_exit(&l2arc_dev_mtx);
5051
5052 return (dev != NULL);
5053 }
5054
5055 /*
5056 * Add a vdev for use by the L2ARC. By this point the spa has already
5057 * validated the vdev and opened it.
5058 */
5059 void
5060 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
5061 {
5062 l2arc_dev_t *adddev;
5063
5064 ASSERT(!l2arc_vdev_present(vd));
5065
5066 /*
5067 * Create a new l2arc device entry.
5068 */
5069 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
5070 adddev->l2ad_spa = spa;
5071 adddev->l2ad_vdev = vd;
5072 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
5073 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
5074 adddev->l2ad_hand = adddev->l2ad_start;
5075 adddev->l2ad_evict = adddev->l2ad_start;
5076 adddev->l2ad_first = B_TRUE;
5077 adddev->l2ad_writing = B_FALSE;
5078
5079 /*
5080 * This is a list of all ARC buffers that are still valid on the
5081 * device.
5082 */
5083 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
5084 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
5085 offsetof(arc_buf_hdr_t, b_l2node));
5086
5087 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
5088
5089 /*
5090 * Add device to global list
5091 */
5092 mutex_enter(&l2arc_dev_mtx);
5093 list_insert_head(l2arc_dev_list, adddev);
5094 atomic_inc_64(&l2arc_ndev);
5095 mutex_exit(&l2arc_dev_mtx);
5096 }
5097
5098 /*
5099 * Remove a vdev from the L2ARC.
5100 */
5101 void
5102 l2arc_remove_vdev(vdev_t *vd)
5103 {
5104 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
5105
5106 /*
5107 * Find the device by vdev
5108 */
5109 mutex_enter(&l2arc_dev_mtx);
5110 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
5111 nextdev = list_next(l2arc_dev_list, dev);
5112 if (vd == dev->l2ad_vdev) {
5113 remdev = dev;
5114 break;
5115 }
5116 }
5117 ASSERT(remdev != NULL);
5118
5119 /*
5120 * Remove device from global list
5121 */
5122 list_remove(l2arc_dev_list, remdev);
5123 l2arc_dev_last = NULL; /* may have been invalidated */
5124 atomic_dec_64(&l2arc_ndev);
5125 mutex_exit(&l2arc_dev_mtx);
5126
5127 /*
5128 * Clear all buflists and ARC references. L2ARC device flush.
5129 */
5130 l2arc_evict(remdev, 0, B_TRUE);
5131 list_destroy(remdev->l2ad_buflist);
5132 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5133 kmem_free(remdev, sizeof (l2arc_dev_t));
5134 }
5135
5136 void
5137 l2arc_init(void)
5138 {
5139 l2arc_thread_exit = 0;
5140 l2arc_ndev = 0;
5141 l2arc_writes_sent = 0;
5142 l2arc_writes_done = 0;
5143
5144 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5145 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5146 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5147 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5148 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5149
5150 l2arc_dev_list = &L2ARC_dev_list;
5151 l2arc_free_on_write = &L2ARC_free_on_write;
5152 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5153 offsetof(l2arc_dev_t, l2ad_node));
5154 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5155 offsetof(l2arc_data_free_t, l2df_list_node));
5156 }
5157
5158 void
5159 l2arc_fini(void)
5160 {
5161 /*
5162 * This is called from dmu_fini(), which is called from spa_fini();
5163 * Because of this, we can assume that all l2arc devices have
5164 * already been removed when the pools themselves were removed.
5165 */
5166
5167 l2arc_do_free_on_write();
5168
5169 mutex_destroy(&l2arc_feed_thr_lock);
5170 cv_destroy(&l2arc_feed_thr_cv);
5171 mutex_destroy(&l2arc_dev_mtx);
5172 mutex_destroy(&l2arc_buflist_mtx);
5173 mutex_destroy(&l2arc_free_on_write_mtx);
5174
5175 list_destroy(l2arc_dev_list);
5176 list_destroy(l2arc_free_on_write);
5177 }
5178
5179 void
5180 l2arc_start(void)
5181 {
5182 if (!(spa_mode_global & FWRITE))
5183 return;
5184
5185 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5186 TS_RUN, minclsyspri);
5187 }
5188
5189 void
5190 l2arc_stop(void)
5191 {
5192 if (!(spa_mode_global & FWRITE))
5193 return;
5194
5195 mutex_enter(&l2arc_feed_thr_lock);
5196 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5197 l2arc_thread_exit = 1;
5198 while (l2arc_thread_exit != 0)
5199 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5200 mutex_exit(&l2arc_feed_thr_lock);
5201 }