1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 */
27
28 /*
29 * DVA-based Adjustable Replacement Cache
30 *
31 * While much of the theory of operation used here is
32 * based on the self-tuning, low overhead replacement cache
33 * presented by Megiddo and Modha at FAST 2003, there are some
34 * significant differences:
35 *
36 * 1. The Megiddo and Modha model assumes any page is evictable.
37 * Pages in its cache cannot be "locked" into memory. This makes
38 * the eviction algorithm simple: evict the last page in the list.
39 * This also make the performance characteristics easy to reason
40 * about. Our cache is not so simple. At any given moment, some
41 * subset of the blocks in the cache are un-evictable because we
42 * have handed out a reference to them. Blocks are only evictable
43 * when there are no external references active. This makes
44 * eviction far more problematic: we choose to evict the evictable
45 * blocks that are the "lowest" in the list.
46 *
47 * There are times when it is not possible to evict the requested
48 * space. In these circumstances we are unable to adjust the cache
49 * size. To prevent the cache growing unbounded at these times we
50 * implement a "cache throttle" that slows the flow of new data
51 * into the cache until we can make space available.
52 *
53 * 2. The Megiddo and Modha model assumes a fixed cache size.
54 * Pages are evicted when the cache is full and there is a cache
55 * miss. Our model has a variable sized cache. It grows with
56 * high use, but also tries to react to memory pressure from the
57 * operating system: decreasing its size when system memory is
58 * tight.
59 *
60 * 3. The Megiddo and Modha model assumes a fixed page size. All
61 * elements of the cache are therefore exactly the same size. So
62 * when adjusting the cache size following a cache miss, its simply
63 * a matter of choosing a single page to evict. In our model, we
64 * have variable sized cache blocks (rangeing from 512 bytes to
65 * 128K bytes). We therefore choose a set of blocks to evict to make
66 * space for a cache miss that approximates as closely as possible
67 * the space used by the new block.
68 *
69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
70 * by N. Megiddo & D. Modha, FAST 2003
71 */
72
73 /*
74 * The locking model:
75 *
76 * A new reference to a cache buffer can be obtained in two
77 * ways: 1) via a hash table lookup using the DVA as a key,
78 * or 2) via one of the ARC lists. The arc_read() interface
79 * uses method 1, while the internal arc algorithms for
80 * adjusting the cache use method 2. We therefore provide two
81 * types of locks: 1) the hash table lock array, and 2) the
82 * arc list locks.
83 *
84 * Buffers do not have their own mutexes, rather they rely on the
85 * hash table mutexes for the bulk of their protection (i.e. most
86 * fields in the arc_buf_hdr_t are protected by these mutexes).
87 *
88 * buf_hash_find() returns the appropriate mutex (held) when it
89 * locates the requested buffer in the hash table. It returns
90 * NULL for the mutex if the buffer was not in the table.
91 *
92 * buf_hash_remove() expects the appropriate hash mutex to be
93 * already held before it is invoked.
94 *
95 * Each arc state also has a mutex which is used to protect the
96 * buffer list associated with the state. When attempting to
97 * obtain a hash table lock while holding an arc list lock you
98 * must use: mutex_tryenter() to avoid deadlock. Also note that
99 * the active state mutex must be held before the ghost state mutex.
100 *
101 * Arc buffers may have an associated eviction callback function.
102 * This function will be invoked prior to removing the buffer (e.g.
103 * in arc_do_user_evicts()). Note however that the data associated
104 * with the buffer may be evicted prior to the callback. The callback
105 * must be made with *no locks held* (to prevent deadlock). Additionally,
106 * the users of callbacks must ensure that their private data is
107 * protected from simultaneous callbacks from arc_buf_evict()
108 * and arc_do_user_evicts().
109 *
110 * Note that the majority of the performance stats are manipulated
111 * with atomic operations.
112 *
113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
114 *
115 * - L2ARC buflist creation
116 * - L2ARC buflist eviction
117 * - L2ARC write completion, which walks L2ARC buflists
118 * - ARC header destruction, as it removes from L2ARC buflists
119 * - ARC header release, as it removes from L2ARC buflists
120 *
121 * Please note that if you first grab the l2arc_buflist_mtx, you can't do a
122 * mutex_enter on a buffer's hash_lock anymore due to lock inversion. To grab
123 * the hash_lock you must use mutex_tryenter and possibly deal with the buffer
124 * not being available (due to e.g. some other thread holding it while trying
125 * to unconditionally grab the l2arc_buflist_mtx which you are holding). The
126 * inverse situation (first grab hash_lock, then l2arc_buflist_mtx) is safe.
127 */
128
129 #include <sys/spa.h>
130 #include <sys/zio.h>
131 #include <sys/zio_compress.h>
132 #include <sys/zfs_context.h>
133 #include <sys/arc.h>
134 #include <sys/refcount.h>
135 #include <sys/vdev.h>
136 #include <sys/vdev_impl.h>
137 #ifdef _KERNEL
138 #include <sys/vmsystm.h>
139 #include <vm/anon.h>
140 #include <sys/fs/swapnode.h>
141 #include <sys/dnlc.h>
142 #endif
143 #include <sys/callb.h>
144 #include <sys/kstat.h>
145 #include <zfs_fletcher.h>
146
147 #ifndef _KERNEL
148 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
149 boolean_t arc_watch = B_FALSE;
150 int arc_procfd;
151 #endif
152
153 static kmutex_t arc_reclaim_thr_lock;
154 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
155 static uint8_t arc_thread_exit;
156
157 extern int zfs_write_limit_shift;
158 extern uint64_t zfs_write_limit_max;
159 extern kmutex_t zfs_write_limit_lock;
160
161 #define ARC_REDUCE_DNLC_PERCENT 3
162 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
163
164 typedef enum arc_reclaim_strategy {
165 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
166 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
167 } arc_reclaim_strategy_t;
168
169 /* number of seconds before growing cache again */
170 static int arc_grow_retry = 60;
171
172 /* shift of arc_c for calculating both min and max arc_p */
173 static int arc_p_min_shift = 4;
174
175 /* log2(fraction of arc to reclaim) */
176 static int arc_shrink_shift = 5;
177
178 /*
179 * minimum lifespan of a prefetch block in clock ticks
180 * (initialized in arc_init())
181 */
182 static int arc_min_prefetch_lifespan;
183
184 static int arc_dead;
185
186 /*
187 * The arc has filled available memory and has now warmed up.
188 */
189 static boolean_t arc_warm;
190
191 /*
192 * These tunables are for performance analysis.
193 */
194 uint64_t zfs_arc_max;
195 uint64_t zfs_arc_min;
196 uint64_t zfs_arc_meta_limit = 0;
197 int zfs_arc_grow_retry = 0;
198 int zfs_arc_shrink_shift = 0;
199 int zfs_arc_p_min_shift = 0;
200 int zfs_disable_dup_eviction = 0;
201
202 /*
203 * Note that buffers can be in one of 6 states:
204 * ARC_anon - anonymous (discussed below)
205 * ARC_mru - recently used, currently cached
206 * ARC_mru_ghost - recentely used, no longer in cache
207 * ARC_mfu - frequently used, currently cached
208 * ARC_mfu_ghost - frequently used, no longer in cache
209 * ARC_l2c_only - exists in L2ARC but not other states
210 * When there are no active references to the buffer, they are
211 * are linked onto a list in one of these arc states. These are
212 * the only buffers that can be evicted or deleted. Within each
213 * state there are multiple lists, one for meta-data and one for
214 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
215 * etc.) is tracked separately so that it can be managed more
216 * explicitly: favored over data, limited explicitly.
217 *
218 * Anonymous buffers are buffers that are not associated with
219 * a DVA. These are buffers that hold dirty block copies
220 * before they are written to stable storage. By definition,
221 * they are "ref'd" and are considered part of arc_mru
222 * that cannot be freed. Generally, they will aquire a DVA
223 * as they are written and migrate onto the arc_mru list.
224 *
225 * The ARC_l2c_only state is for buffers that are in the second
226 * level ARC but no longer in any of the ARC_m* lists. The second
227 * level ARC itself may also contain buffers that are in any of
228 * the ARC_m* states - meaning that a buffer can exist in two
229 * places. The reason for the ARC_l2c_only state is to keep the
230 * buffer header in the hash table, so that reads that hit the
231 * second level ARC benefit from these fast lookups.
232 */
233
234 typedef struct arc_state {
235 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
236 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
237 uint64_t arcs_size; /* total amount of data in this state */
238 kmutex_t arcs_mtx;
239 } arc_state_t;
240
241 /* The 6 states: */
242 static arc_state_t ARC_anon;
243 static arc_state_t ARC_mru;
244 static arc_state_t ARC_mru_ghost;
245 static arc_state_t ARC_mfu;
246 static arc_state_t ARC_mfu_ghost;
247 static arc_state_t ARC_l2c_only;
248
249 typedef struct arc_stats {
250 kstat_named_t arcstat_hits;
251 kstat_named_t arcstat_misses;
252 kstat_named_t arcstat_demand_data_hits;
253 kstat_named_t arcstat_demand_data_misses;
254 kstat_named_t arcstat_demand_metadata_hits;
255 kstat_named_t arcstat_demand_metadata_misses;
256 kstat_named_t arcstat_prefetch_data_hits;
257 kstat_named_t arcstat_prefetch_data_misses;
258 kstat_named_t arcstat_prefetch_metadata_hits;
259 kstat_named_t arcstat_prefetch_metadata_misses;
260 kstat_named_t arcstat_mru_hits;
261 kstat_named_t arcstat_mru_ghost_hits;
262 kstat_named_t arcstat_mfu_hits;
263 kstat_named_t arcstat_mfu_ghost_hits;
264 kstat_named_t arcstat_deleted;
265 kstat_named_t arcstat_recycle_miss;
266 /*
267 * Number of buffers that could not be evicted because the hash lock
268 * was held by another thread. The lock may not necessarily be held
269 * by something using the same buffer, since hash locks are shared
270 * by multiple buffers.
271 */
272 kstat_named_t arcstat_mutex_miss;
273 /*
274 * Number of buffers skipped because they have I/O in progress, are
275 * indrect prefetch buffers that have not lived long enough, or are
276 * not from the spa we're trying to evict from.
277 */
278 kstat_named_t arcstat_evict_skip;
279 kstat_named_t arcstat_evict_l2_cached;
280 kstat_named_t arcstat_evict_l2_eligible;
281 kstat_named_t arcstat_evict_l2_ineligible;
282 kstat_named_t arcstat_hash_elements;
283 kstat_named_t arcstat_hash_elements_max;
284 kstat_named_t arcstat_hash_collisions;
285 kstat_named_t arcstat_hash_chains;
286 kstat_named_t arcstat_hash_chain_max;
287 kstat_named_t arcstat_p;
288 kstat_named_t arcstat_c;
289 kstat_named_t arcstat_c_min;
290 kstat_named_t arcstat_c_max;
291 kstat_named_t arcstat_size;
292 kstat_named_t arcstat_hdr_size;
293 kstat_named_t arcstat_data_size;
294 kstat_named_t arcstat_other_size;
295 kstat_named_t arcstat_l2_hits;
296 kstat_named_t arcstat_l2_misses;
297 kstat_named_t arcstat_l2_feeds;
298 kstat_named_t arcstat_l2_rw_clash;
299 kstat_named_t arcstat_l2_read_bytes;
300 kstat_named_t arcstat_l2_write_bytes;
301 kstat_named_t arcstat_l2_writes_sent;
302 kstat_named_t arcstat_l2_writes_done;
303 kstat_named_t arcstat_l2_writes_error;
304 kstat_named_t arcstat_l2_writes_hdr_miss;
305 kstat_named_t arcstat_l2_evict_lock_retry;
306 kstat_named_t arcstat_l2_evict_reading;
307 kstat_named_t arcstat_l2_free_on_write;
308 kstat_named_t arcstat_l2_abort_lowmem;
309 kstat_named_t arcstat_l2_cksum_bad;
310 kstat_named_t arcstat_l2_io_error;
311 kstat_named_t arcstat_l2_size;
312 kstat_named_t arcstat_l2_asize;
313 kstat_named_t arcstat_l2_hdr_size;
314 kstat_named_t arcstat_l2_compress_successes;
315 kstat_named_t arcstat_l2_compress_zeros;
316 kstat_named_t arcstat_l2_compress_failures;
317 kstat_named_t arcstat_memory_throttle_count;
318 kstat_named_t arcstat_duplicate_buffers;
319 kstat_named_t arcstat_duplicate_buffers_size;
320 kstat_named_t arcstat_duplicate_reads;
321 kstat_named_t arcstat_meta_used;
322 kstat_named_t arcstat_meta_limit;
323 kstat_named_t arcstat_meta_max;
324 } arc_stats_t;
325
326 static arc_stats_t arc_stats = {
327 { "hits", KSTAT_DATA_UINT64 },
328 { "misses", KSTAT_DATA_UINT64 },
329 { "demand_data_hits", KSTAT_DATA_UINT64 },
330 { "demand_data_misses", KSTAT_DATA_UINT64 },
331 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
332 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
333 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
334 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
335 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
336 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
337 { "mru_hits", KSTAT_DATA_UINT64 },
338 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
339 { "mfu_hits", KSTAT_DATA_UINT64 },
340 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
341 { "deleted", KSTAT_DATA_UINT64 },
342 { "recycle_miss", KSTAT_DATA_UINT64 },
343 { "mutex_miss", KSTAT_DATA_UINT64 },
344 { "evict_skip", KSTAT_DATA_UINT64 },
345 { "evict_l2_cached", KSTAT_DATA_UINT64 },
346 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
347 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
348 { "hash_elements", KSTAT_DATA_UINT64 },
349 { "hash_elements_max", KSTAT_DATA_UINT64 },
350 { "hash_collisions", KSTAT_DATA_UINT64 },
351 { "hash_chains", KSTAT_DATA_UINT64 },
352 { "hash_chain_max", KSTAT_DATA_UINT64 },
353 { "p", KSTAT_DATA_UINT64 },
354 { "c", KSTAT_DATA_UINT64 },
355 { "c_min", KSTAT_DATA_UINT64 },
356 { "c_max", KSTAT_DATA_UINT64 },
357 { "size", KSTAT_DATA_UINT64 },
358 { "hdr_size", KSTAT_DATA_UINT64 },
359 { "data_size", KSTAT_DATA_UINT64 },
360 { "other_size", KSTAT_DATA_UINT64 },
361 { "l2_hits", KSTAT_DATA_UINT64 },
362 { "l2_misses", KSTAT_DATA_UINT64 },
363 { "l2_feeds", KSTAT_DATA_UINT64 },
364 { "l2_rw_clash", KSTAT_DATA_UINT64 },
365 { "l2_read_bytes", KSTAT_DATA_UINT64 },
366 { "l2_write_bytes", KSTAT_DATA_UINT64 },
367 { "l2_writes_sent", KSTAT_DATA_UINT64 },
368 { "l2_writes_done", KSTAT_DATA_UINT64 },
369 { "l2_writes_error", KSTAT_DATA_UINT64 },
370 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
371 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
372 { "l2_evict_reading", KSTAT_DATA_UINT64 },
373 { "l2_free_on_write", KSTAT_DATA_UINT64 },
374 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
375 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
376 { "l2_io_error", KSTAT_DATA_UINT64 },
377 { "l2_size", KSTAT_DATA_UINT64 },
378 { "l2_asize", KSTAT_DATA_UINT64 },
379 { "l2_hdr_size", KSTAT_DATA_UINT64 },
380 { "l2_compress_successes", KSTAT_DATA_UINT64 },
381 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
382 { "l2_compress_failures", KSTAT_DATA_UINT64 },
383 { "memory_throttle_count", KSTAT_DATA_UINT64 },
384 { "duplicate_buffers", KSTAT_DATA_UINT64 },
385 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
386 { "duplicate_reads", KSTAT_DATA_UINT64 },
387 { "arc_meta_used", KSTAT_DATA_UINT64 },
388 { "arc_meta_limit", KSTAT_DATA_UINT64 },
389 { "arc_meta_max", KSTAT_DATA_UINT64 }
390 };
391
392 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
393
394 #define ARCSTAT_INCR(stat, val) \
395 atomic_add_64(&arc_stats.stat.value.ui64, (val))
396
397 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
398 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
399
400 #define ARCSTAT_MAX(stat, val) { \
401 uint64_t m; \
402 while ((val) > (m = arc_stats.stat.value.ui64) && \
403 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
404 continue; \
405 }
406
407 #define ARCSTAT_MAXSTAT(stat) \
408 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
409
410 /*
411 * We define a macro to allow ARC hits/misses to be easily broken down by
412 * two separate conditions, giving a total of four different subtypes for
413 * each of hits and misses (so eight statistics total).
414 */
415 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
416 if (cond1) { \
417 if (cond2) { \
418 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
419 } else { \
420 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
421 } \
422 } else { \
423 if (cond2) { \
424 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
425 } else { \
426 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
427 } \
428 }
429
430 kstat_t *arc_ksp;
431 static arc_state_t *arc_anon;
432 static arc_state_t *arc_mru;
433 static arc_state_t *arc_mru_ghost;
434 static arc_state_t *arc_mfu;
435 static arc_state_t *arc_mfu_ghost;
436 static arc_state_t *arc_l2c_only;
437
438 /*
439 * There are several ARC variables that are critical to export as kstats --
440 * but we don't want to have to grovel around in the kstat whenever we wish to
441 * manipulate them. For these variables, we therefore define them to be in
442 * terms of the statistic variable. This assures that we are not introducing
443 * the possibility of inconsistency by having shadow copies of the variables,
444 * while still allowing the code to be readable.
445 */
446 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
447 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
448 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
449 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
450 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
451 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
452 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
453 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
454
455 #define L2ARC_IS_VALID_COMPRESS(_c_) \
456 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
457
458 static int arc_no_grow; /* Don't try to grow cache size */
459 static uint64_t arc_tempreserve;
460 static uint64_t arc_loaned_bytes;
461
462 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
463
464 typedef struct arc_callback arc_callback_t;
465
466 struct arc_callback {
467 void *acb_private;
468 arc_done_func_t *acb_done;
469 arc_buf_t *acb_buf;
470 zio_t *acb_zio_dummy;
471 arc_callback_t *acb_next;
472 };
473
474 typedef struct arc_write_callback arc_write_callback_t;
475
476 struct arc_write_callback {
477 void *awcb_private;
478 arc_done_func_t *awcb_ready;
479 arc_done_func_t *awcb_done;
480 arc_buf_t *awcb_buf;
481 };
482
483 struct arc_buf_hdr {
484 /* protected by hash lock */
485 dva_t b_dva;
486 uint64_t b_birth;
487 uint64_t b_cksum0;
488
489 kmutex_t b_freeze_lock;
490 zio_cksum_t *b_freeze_cksum;
491 void *b_thawed;
492
493 arc_buf_hdr_t *b_hash_next;
494 arc_buf_t *b_buf;
495 uint32_t b_flags;
496 uint32_t b_datacnt;
497
498 arc_callback_t *b_acb;
499 kcondvar_t b_cv;
500
501 /* immutable */
502 arc_buf_contents_t b_type;
503 uint64_t b_size;
504 uint64_t b_spa;
505
506 /* protected by arc state mutex */
507 arc_state_t *b_state;
508 list_node_t b_arc_node;
509
510 /* updated atomically */
511 clock_t b_arc_access;
512
513 /* self protecting */
514 refcount_t b_refcnt;
515
516 l2arc_buf_hdr_t *b_l2hdr;
517 list_node_t b_l2node;
518 };
519
520 static arc_buf_t *arc_eviction_list;
521 static kmutex_t arc_eviction_mtx;
522 static arc_buf_hdr_t arc_eviction_hdr;
523 static void arc_get_data_buf(arc_buf_t *buf);
524 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
525 static int arc_evict_needed(arc_buf_contents_t type);
526 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
527 static void arc_buf_watch(arc_buf_t *buf);
528
529 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
530
531 #define GHOST_STATE(state) \
532 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
533 (state) == arc_l2c_only)
534
535 /*
536 * Private ARC flags. These flags are private ARC only flags that will show up
537 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
538 * be passed in as arc_flags in things like arc_read. However, these flags
539 * should never be passed and should only be set by ARC code. When adding new
540 * public flags, make sure not to smash the private ones.
541 */
542
543 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
544 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
545 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
546 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
547 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
548 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
549 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
550 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
551 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
552 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
553
554 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
555 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
556 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
557 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
558 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
559 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
560 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
561 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
562 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
563 (hdr)->b_l2hdr != NULL)
564 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
565 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
566 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
567
568 /*
569 * Other sizes
570 */
571
572 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
573 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
574
575 /*
576 * Hash table routines
577 */
578
579 #define HT_LOCK_PAD 64
580
581 struct ht_lock {
582 kmutex_t ht_lock;
583 #ifdef _KERNEL
584 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
585 #endif
586 };
587
588 #define BUF_LOCKS 256
589 typedef struct buf_hash_table {
590 uint64_t ht_mask;
591 arc_buf_hdr_t **ht_table;
592 struct ht_lock ht_locks[BUF_LOCKS];
593 } buf_hash_table_t;
594
595 static buf_hash_table_t buf_hash_table;
596
597 #define BUF_HASH_INDEX(spa, dva, birth) \
598 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
599 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
600 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
601 #define HDR_LOCK(hdr) \
602 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
603
604 uint64_t zfs_crc64_table[256];
605
606 /*
607 * Level 2 ARC
608 */
609
610 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
611 #define L2ARC_HEADROOM 2 /* num of writes */
612 /*
613 * If we discover during ARC scan any buffers to be compressed, we boost
614 * our headroom for the next scanning cycle by this percentage multiple.
615 */
616 #define L2ARC_HEADROOM_BOOST 200
617 #define L2ARC_FEED_SECS 1 /* caching interval secs */
618 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
619
620 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
621 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
622
623 /* L2ARC Performance Tunables */
624 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
625 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
626 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
627 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
628 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
629 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
630 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
631 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
632 boolean_t l2arc_norw = B_FALSE; /* no reads during writes */
633
634 /*
635 * L2ARC Internals
636 */
637 typedef struct l2arc_dev {
638 vdev_t *l2ad_vdev; /* vdev */
639 spa_t *l2ad_spa; /* spa */
640 uint64_t l2ad_hand; /* next write location */
641 uint64_t l2ad_start; /* first addr on device */
642 uint64_t l2ad_end; /* last addr on device */
643 uint64_t l2ad_evict; /* last addr eviction reached */
644 boolean_t l2ad_first; /* first sweep through */
645 boolean_t l2ad_writing; /* currently writing */
646 list_t *l2ad_buflist; /* buffer list */
647 list_node_t l2ad_node; /* device list node */
648 } l2arc_dev_t;
649
650 static list_t L2ARC_dev_list; /* device list */
651 static list_t *l2arc_dev_list; /* device list pointer */
652 static kmutex_t l2arc_dev_mtx; /* device list mutex */
653 static l2arc_dev_t *l2arc_dev_last; /* last device used */
654 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
655 static list_t L2ARC_free_on_write; /* free after write buf list */
656 static list_t *l2arc_free_on_write; /* free after write list ptr */
657 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
658 static uint64_t l2arc_ndev; /* number of devices */
659
660 typedef struct l2arc_read_callback {
661 arc_buf_t *l2rcb_buf; /* read buffer */
662 spa_t *l2rcb_spa; /* spa */
663 blkptr_t l2rcb_bp; /* original blkptr */
664 zbookmark_t l2rcb_zb; /* original bookmark */
665 int l2rcb_flags; /* original flags */
666 enum zio_compress l2rcb_compress; /* applied compress */
667 } l2arc_read_callback_t;
668
669 typedef struct l2arc_write_callback {
670 l2arc_dev_t *l2wcb_dev; /* device info */
671 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
672 } l2arc_write_callback_t;
673
674 struct l2arc_buf_hdr {
675 /* protected by arc_buf_hdr mutex */
676 l2arc_dev_t *b_dev; /* L2ARC device */
677 uint64_t b_daddr; /* disk address, offset byte */
678 /* compression applied to buffer data */
679 enum zio_compress b_compress;
680 /* real alloc'd buffer size depending on b_compress applied */
681 int b_asize;
682 };
683
684 typedef struct l2arc_data_free {
685 /* protected by l2arc_free_on_write_mtx */
686 void *l2df_data;
687 size_t l2df_size;
688 void (*l2df_func)(void *, size_t);
689 list_node_t l2df_list_node;
690 } l2arc_data_free_t;
691
692 static kmutex_t l2arc_feed_thr_lock;
693 static kcondvar_t l2arc_feed_thr_cv;
694 static uint8_t l2arc_thread_exit;
695
696 static void l2arc_read_done(zio_t *zio);
697 static void l2arc_hdr_stat_add(void);
698 static void l2arc_hdr_stat_remove(void);
699
700 static boolean_t l2arc_compress_buf(void *in_data, uint64_t in_sz,
701 void **out_data, uint64_t *out_sz, enum zio_compress *compress);
702 static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr,
703 enum zio_compress c);
704
705 static uint64_t
706 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
707 {
708 uint8_t *vdva = (uint8_t *)dva;
709 uint64_t crc = -1ULL;
710 int i;
711
712 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
713
714 for (i = 0; i < sizeof (dva_t); i++)
715 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
716
717 crc ^= (spa>>8) ^ birth;
718
719 return (crc);
720 }
721
722 #define BUF_EMPTY(buf) \
723 ((buf)->b_dva.dva_word[0] == 0 && \
724 (buf)->b_dva.dva_word[1] == 0 && \
725 (buf)->b_birth == 0)
726
727 #define BUF_EQUAL(spa, dva, birth, buf) \
728 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
729 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
730 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
731
732 static void
733 buf_discard_identity(arc_buf_hdr_t *hdr)
734 {
735 hdr->b_dva.dva_word[0] = 0;
736 hdr->b_dva.dva_word[1] = 0;
737 hdr->b_birth = 0;
738 hdr->b_cksum0 = 0;
739 }
740
741 static arc_buf_hdr_t *
742 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
743 {
744 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
745 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
746 arc_buf_hdr_t *buf;
747
748 mutex_enter(hash_lock);
749 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
750 buf = buf->b_hash_next) {
751 if (BUF_EQUAL(spa, dva, birth, buf)) {
752 *lockp = hash_lock;
753 return (buf);
754 }
755 }
756 mutex_exit(hash_lock);
757 *lockp = NULL;
758 return (NULL);
759 }
760
761 /*
762 * Insert an entry into the hash table. If there is already an element
763 * equal to elem in the hash table, then the already existing element
764 * will be returned and the new element will not be inserted.
765 * Otherwise returns NULL.
766 */
767 static arc_buf_hdr_t *
768 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
769 {
770 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
771 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
772 arc_buf_hdr_t *fbuf;
773 uint32_t i;
774
775 ASSERT(!HDR_IN_HASH_TABLE(buf));
776 *lockp = hash_lock;
777 mutex_enter(hash_lock);
778 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
779 fbuf = fbuf->b_hash_next, i++) {
780 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
781 return (fbuf);
782 }
783
784 buf->b_hash_next = buf_hash_table.ht_table[idx];
785 buf_hash_table.ht_table[idx] = buf;
786 buf->b_flags |= ARC_IN_HASH_TABLE;
787
788 /* collect some hash table performance data */
789 if (i > 0) {
790 ARCSTAT_BUMP(arcstat_hash_collisions);
791 if (i == 1)
792 ARCSTAT_BUMP(arcstat_hash_chains);
793
794 ARCSTAT_MAX(arcstat_hash_chain_max, i);
795 }
796
797 ARCSTAT_BUMP(arcstat_hash_elements);
798 ARCSTAT_MAXSTAT(arcstat_hash_elements);
799
800 return (NULL);
801 }
802
803 static void
804 buf_hash_remove(arc_buf_hdr_t *buf)
805 {
806 arc_buf_hdr_t *fbuf, **bufp;
807 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
808
809 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
810 ASSERT(HDR_IN_HASH_TABLE(buf));
811
812 bufp = &buf_hash_table.ht_table[idx];
813 while ((fbuf = *bufp) != buf) {
814 ASSERT(fbuf != NULL);
815 bufp = &fbuf->b_hash_next;
816 }
817 *bufp = buf->b_hash_next;
818 buf->b_hash_next = NULL;
819 buf->b_flags &= ~ARC_IN_HASH_TABLE;
820
821 /* collect some hash table performance data */
822 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
823
824 if (buf_hash_table.ht_table[idx] &&
825 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
826 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
827 }
828
829 /*
830 * Global data structures and functions for the buf kmem cache.
831 */
832 static kmem_cache_t *hdr_cache;
833 static kmem_cache_t *buf_cache;
834
835 static void
836 buf_fini(void)
837 {
838 int i;
839
840 kmem_free(buf_hash_table.ht_table,
841 (buf_hash_table.ht_mask + 1) * sizeof (void *));
842 for (i = 0; i < BUF_LOCKS; i++)
843 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
844 kmem_cache_destroy(hdr_cache);
845 kmem_cache_destroy(buf_cache);
846 }
847
848 /*
849 * Constructor callback - called when the cache is empty
850 * and a new buf is requested.
851 */
852 /* ARGSUSED */
853 static int
854 hdr_cons(void *vbuf, void *unused, int kmflag)
855 {
856 arc_buf_hdr_t *buf = vbuf;
857
858 bzero(buf, sizeof (arc_buf_hdr_t));
859 refcount_create(&buf->b_refcnt);
860 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
861 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
862 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
863
864 return (0);
865 }
866
867 /* ARGSUSED */
868 static int
869 buf_cons(void *vbuf, void *unused, int kmflag)
870 {
871 arc_buf_t *buf = vbuf;
872
873 bzero(buf, sizeof (arc_buf_t));
874 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
875 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
876
877 return (0);
878 }
879
880 /*
881 * Destructor callback - called when a cached buf is
882 * no longer required.
883 */
884 /* ARGSUSED */
885 static void
886 hdr_dest(void *vbuf, void *unused)
887 {
888 arc_buf_hdr_t *buf = vbuf;
889
890 ASSERT(BUF_EMPTY(buf));
891 refcount_destroy(&buf->b_refcnt);
892 cv_destroy(&buf->b_cv);
893 mutex_destroy(&buf->b_freeze_lock);
894 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
895 }
896
897 /* ARGSUSED */
898 static void
899 buf_dest(void *vbuf, void *unused)
900 {
901 arc_buf_t *buf = vbuf;
902
903 mutex_destroy(&buf->b_evict_lock);
904 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
905 }
906
907 /*
908 * Reclaim callback -- invoked when memory is low.
909 */
910 /* ARGSUSED */
911 static void
912 hdr_recl(void *unused)
913 {
914 dprintf("hdr_recl called\n");
915 /*
916 * umem calls the reclaim func when we destroy the buf cache,
917 * which is after we do arc_fini().
918 */
919 if (!arc_dead)
920 cv_signal(&arc_reclaim_thr_cv);
921 }
922
923 static void
924 buf_init(void)
925 {
926 uint64_t *ct;
927 uint64_t hsize = 1ULL << 12;
928 int i, j;
929
930 /*
931 * The hash table is big enough to fill all of physical memory
932 * with an average 64K block size. The table will take up
933 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
934 */
935 while (hsize * 65536 < physmem * PAGESIZE)
936 hsize <<= 1;
937 retry:
938 buf_hash_table.ht_mask = hsize - 1;
939 buf_hash_table.ht_table =
940 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
941 if (buf_hash_table.ht_table == NULL) {
942 ASSERT(hsize > (1ULL << 8));
943 hsize >>= 1;
944 goto retry;
945 }
946
947 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
948 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
949 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
950 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
951
952 for (i = 0; i < 256; i++)
953 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
954 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
955
956 for (i = 0; i < BUF_LOCKS; i++) {
957 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
958 NULL, MUTEX_DEFAULT, NULL);
959 }
960 }
961
962 #define ARC_MINTIME (hz>>4) /* 62 ms */
963
964 static void
965 arc_cksum_verify(arc_buf_t *buf)
966 {
967 zio_cksum_t zc;
968
969 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
970 return;
971
972 mutex_enter(&buf->b_hdr->b_freeze_lock);
973 if (buf->b_hdr->b_freeze_cksum == NULL ||
974 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
975 mutex_exit(&buf->b_hdr->b_freeze_lock);
976 return;
977 }
978 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
979 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
980 panic("buffer modified while frozen!");
981 mutex_exit(&buf->b_hdr->b_freeze_lock);
982 }
983
984 static int
985 arc_cksum_equal(arc_buf_t *buf)
986 {
987 zio_cksum_t zc;
988 int equal;
989
990 mutex_enter(&buf->b_hdr->b_freeze_lock);
991 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
992 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
993 mutex_exit(&buf->b_hdr->b_freeze_lock);
994
995 return (equal);
996 }
997
998 static void
999 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1000 {
1001 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1002 return;
1003
1004 mutex_enter(&buf->b_hdr->b_freeze_lock);
1005 if (buf->b_hdr->b_freeze_cksum != NULL) {
1006 mutex_exit(&buf->b_hdr->b_freeze_lock);
1007 return;
1008 }
1009 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1010 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1011 buf->b_hdr->b_freeze_cksum);
1012 mutex_exit(&buf->b_hdr->b_freeze_lock);
1013 arc_buf_watch(buf);
1014 }
1015
1016 #ifndef _KERNEL
1017 typedef struct procctl {
1018 long cmd;
1019 prwatch_t prwatch;
1020 } procctl_t;
1021 #endif
1022
1023 /* ARGSUSED */
1024 static void
1025 arc_buf_unwatch(arc_buf_t *buf)
1026 {
1027 #ifndef _KERNEL
1028 if (arc_watch) {
1029 int result;
1030 procctl_t ctl;
1031 ctl.cmd = PCWATCH;
1032 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1033 ctl.prwatch.pr_size = 0;
1034 ctl.prwatch.pr_wflags = 0;
1035 result = write(arc_procfd, &ctl, sizeof (ctl));
1036 ASSERT3U(result, ==, sizeof (ctl));
1037 }
1038 #endif
1039 }
1040
1041 /* ARGSUSED */
1042 static void
1043 arc_buf_watch(arc_buf_t *buf)
1044 {
1045 #ifndef _KERNEL
1046 if (arc_watch) {
1047 int result;
1048 procctl_t ctl;
1049 ctl.cmd = PCWATCH;
1050 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1051 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1052 ctl.prwatch.pr_wflags = WA_WRITE;
1053 result = write(arc_procfd, &ctl, sizeof (ctl));
1054 ASSERT3U(result, ==, sizeof (ctl));
1055 }
1056 #endif
1057 }
1058
1059 void
1060 arc_buf_thaw(arc_buf_t *buf)
1061 {
1062 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1063 if (buf->b_hdr->b_state != arc_anon)
1064 panic("modifying non-anon buffer!");
1065 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1066 panic("modifying buffer while i/o in progress!");
1067 arc_cksum_verify(buf);
1068 }
1069
1070 mutex_enter(&buf->b_hdr->b_freeze_lock);
1071 if (buf->b_hdr->b_freeze_cksum != NULL) {
1072 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1073 buf->b_hdr->b_freeze_cksum = NULL;
1074 }
1075
1076 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1077 if (buf->b_hdr->b_thawed)
1078 kmem_free(buf->b_hdr->b_thawed, 1);
1079 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1080 }
1081
1082 mutex_exit(&buf->b_hdr->b_freeze_lock);
1083
1084 arc_buf_unwatch(buf);
1085 }
1086
1087 void
1088 arc_buf_freeze(arc_buf_t *buf)
1089 {
1090 kmutex_t *hash_lock;
1091
1092 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1093 return;
1094
1095 hash_lock = HDR_LOCK(buf->b_hdr);
1096 mutex_enter(hash_lock);
1097
1098 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1099 buf->b_hdr->b_state == arc_anon);
1100 arc_cksum_compute(buf, B_FALSE);
1101 mutex_exit(hash_lock);
1102
1103 }
1104
1105 static void
1106 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1107 {
1108 ASSERT(MUTEX_HELD(hash_lock));
1109
1110 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1111 (ab->b_state != arc_anon)) {
1112 uint64_t delta = ab->b_size * ab->b_datacnt;
1113 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1114 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1115
1116 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1117 mutex_enter(&ab->b_state->arcs_mtx);
1118 ASSERT(list_link_active(&ab->b_arc_node));
1119 list_remove(list, ab);
1120 if (GHOST_STATE(ab->b_state)) {
1121 ASSERT0(ab->b_datacnt);
1122 ASSERT3P(ab->b_buf, ==, NULL);
1123 delta = ab->b_size;
1124 }
1125 ASSERT(delta > 0);
1126 ASSERT3U(*size, >=, delta);
1127 atomic_add_64(size, -delta);
1128 mutex_exit(&ab->b_state->arcs_mtx);
1129 /* remove the prefetch flag if we get a reference */
1130 if (ab->b_flags & ARC_PREFETCH)
1131 ab->b_flags &= ~ARC_PREFETCH;
1132 }
1133 }
1134
1135 static int
1136 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1137 {
1138 int cnt;
1139 arc_state_t *state = ab->b_state;
1140
1141 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1142 ASSERT(!GHOST_STATE(state));
1143
1144 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1145 (state != arc_anon)) {
1146 uint64_t *size = &state->arcs_lsize[ab->b_type];
1147
1148 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1149 mutex_enter(&state->arcs_mtx);
1150 ASSERT(!list_link_active(&ab->b_arc_node));
1151 list_insert_head(&state->arcs_list[ab->b_type], ab);
1152 ASSERT(ab->b_datacnt > 0);
1153 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1154 mutex_exit(&state->arcs_mtx);
1155 }
1156 return (cnt);
1157 }
1158
1159 /*
1160 * Move the supplied buffer to the indicated state. The mutex
1161 * for the buffer must be held by the caller.
1162 */
1163 static void
1164 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1165 {
1166 arc_state_t *old_state = ab->b_state;
1167 int64_t refcnt = refcount_count(&ab->b_refcnt);
1168 uint64_t from_delta, to_delta;
1169
1170 ASSERT(MUTEX_HELD(hash_lock));
1171 ASSERT(new_state != old_state);
1172 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1173 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1174 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1175
1176 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1177
1178 /*
1179 * If this buffer is evictable, transfer it from the
1180 * old state list to the new state list.
1181 */
1182 if (refcnt == 0) {
1183 if (old_state != arc_anon) {
1184 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1185 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1186
1187 if (use_mutex)
1188 mutex_enter(&old_state->arcs_mtx);
1189
1190 ASSERT(list_link_active(&ab->b_arc_node));
1191 list_remove(&old_state->arcs_list[ab->b_type], ab);
1192
1193 /*
1194 * If prefetching out of the ghost cache,
1195 * we will have a non-zero datacnt.
1196 */
1197 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1198 /* ghost elements have a ghost size */
1199 ASSERT(ab->b_buf == NULL);
1200 from_delta = ab->b_size;
1201 }
1202 ASSERT3U(*size, >=, from_delta);
1203 atomic_add_64(size, -from_delta);
1204
1205 if (use_mutex)
1206 mutex_exit(&old_state->arcs_mtx);
1207 }
1208 if (new_state != arc_anon) {
1209 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1210 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1211
1212 if (use_mutex)
1213 mutex_enter(&new_state->arcs_mtx);
1214
1215 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1216
1217 /* ghost elements have a ghost size */
1218 if (GHOST_STATE(new_state)) {
1219 ASSERT(ab->b_datacnt == 0);
1220 ASSERT(ab->b_buf == NULL);
1221 to_delta = ab->b_size;
1222 }
1223 atomic_add_64(size, to_delta);
1224
1225 if (use_mutex)
1226 mutex_exit(&new_state->arcs_mtx);
1227 }
1228 }
1229
1230 ASSERT(!BUF_EMPTY(ab));
1231 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1232 buf_hash_remove(ab);
1233
1234 /* adjust state sizes */
1235 if (to_delta)
1236 atomic_add_64(&new_state->arcs_size, to_delta);
1237 if (from_delta) {
1238 ASSERT3U(old_state->arcs_size, >=, from_delta);
1239 atomic_add_64(&old_state->arcs_size, -from_delta);
1240 }
1241 ab->b_state = new_state;
1242
1243 /* adjust l2arc hdr stats */
1244 if (new_state == arc_l2c_only)
1245 l2arc_hdr_stat_add();
1246 else if (old_state == arc_l2c_only)
1247 l2arc_hdr_stat_remove();
1248 }
1249
1250 void
1251 arc_space_consume(uint64_t space, arc_space_type_t type)
1252 {
1253 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1254
1255 switch (type) {
1256 case ARC_SPACE_DATA:
1257 ARCSTAT_INCR(arcstat_data_size, space);
1258 break;
1259 case ARC_SPACE_OTHER:
1260 ARCSTAT_INCR(arcstat_other_size, space);
1261 break;
1262 case ARC_SPACE_HDRS:
1263 ARCSTAT_INCR(arcstat_hdr_size, space);
1264 break;
1265 case ARC_SPACE_L2HDRS:
1266 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1267 break;
1268 }
1269
1270 ARCSTAT_INCR(arcstat_meta_used, space);
1271 atomic_add_64(&arc_size, space);
1272 }
1273
1274 void
1275 arc_space_return(uint64_t space, arc_space_type_t type)
1276 {
1277 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1278
1279 switch (type) {
1280 case ARC_SPACE_DATA:
1281 ARCSTAT_INCR(arcstat_data_size, -space);
1282 break;
1283 case ARC_SPACE_OTHER:
1284 ARCSTAT_INCR(arcstat_other_size, -space);
1285 break;
1286 case ARC_SPACE_HDRS:
1287 ARCSTAT_INCR(arcstat_hdr_size, -space);
1288 break;
1289 case ARC_SPACE_L2HDRS:
1290 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1291 break;
1292 }
1293
1294 ASSERT(arc_meta_used >= space);
1295 if (arc_meta_max < arc_meta_used)
1296 arc_meta_max = arc_meta_used;
1297 ARCSTAT_INCR(arcstat_meta_used, -space);
1298 ASSERT(arc_size >= space);
1299 atomic_add_64(&arc_size, -space);
1300 }
1301
1302 void *
1303 arc_data_buf_alloc(uint64_t size)
1304 {
1305 if (arc_evict_needed(ARC_BUFC_DATA))
1306 cv_signal(&arc_reclaim_thr_cv);
1307 atomic_add_64(&arc_size, size);
1308 return (zio_data_buf_alloc(size));
1309 }
1310
1311 void
1312 arc_data_buf_free(void *buf, uint64_t size)
1313 {
1314 zio_data_buf_free(buf, size);
1315 ASSERT(arc_size >= size);
1316 atomic_add_64(&arc_size, -size);
1317 }
1318
1319 arc_buf_t *
1320 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1321 {
1322 arc_buf_hdr_t *hdr;
1323 arc_buf_t *buf;
1324
1325 ASSERT3U(size, >, 0);
1326 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1327 ASSERT(BUF_EMPTY(hdr));
1328 hdr->b_size = size;
1329 hdr->b_type = type;
1330 hdr->b_spa = spa_load_guid(spa);
1331 hdr->b_state = arc_anon;
1332 hdr->b_arc_access = 0;
1333 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1334 buf->b_hdr = hdr;
1335 buf->b_data = NULL;
1336 buf->b_efunc = NULL;
1337 buf->b_private = NULL;
1338 buf->b_next = NULL;
1339 hdr->b_buf = buf;
1340 arc_get_data_buf(buf);
1341 hdr->b_datacnt = 1;
1342 hdr->b_flags = 0;
1343 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1344 (void) refcount_add(&hdr->b_refcnt, tag);
1345
1346 return (buf);
1347 }
1348
1349 static char *arc_onloan_tag = "onloan";
1350
1351 /*
1352 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1353 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1354 * buffers must be returned to the arc before they can be used by the DMU or
1355 * freed.
1356 */
1357 arc_buf_t *
1358 arc_loan_buf(spa_t *spa, int size)
1359 {
1360 arc_buf_t *buf;
1361
1362 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1363
1364 atomic_add_64(&arc_loaned_bytes, size);
1365 return (buf);
1366 }
1367
1368 /*
1369 * Return a loaned arc buffer to the arc.
1370 */
1371 void
1372 arc_return_buf(arc_buf_t *buf, void *tag)
1373 {
1374 arc_buf_hdr_t *hdr = buf->b_hdr;
1375
1376 ASSERT(buf->b_data != NULL);
1377 (void) refcount_add(&hdr->b_refcnt, tag);
1378 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1379
1380 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1381 }
1382
1383 /* Detach an arc_buf from a dbuf (tag) */
1384 void
1385 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1386 {
1387 arc_buf_hdr_t *hdr;
1388
1389 ASSERT(buf->b_data != NULL);
1390 hdr = buf->b_hdr;
1391 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1392 (void) refcount_remove(&hdr->b_refcnt, tag);
1393 buf->b_efunc = NULL;
1394 buf->b_private = NULL;
1395
1396 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1397 }
1398
1399 static arc_buf_t *
1400 arc_buf_clone(arc_buf_t *from)
1401 {
1402 arc_buf_t *buf;
1403 arc_buf_hdr_t *hdr = from->b_hdr;
1404 uint64_t size = hdr->b_size;
1405
1406 ASSERT(hdr->b_state != arc_anon);
1407
1408 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1409 buf->b_hdr = hdr;
1410 buf->b_data = NULL;
1411 buf->b_efunc = NULL;
1412 buf->b_private = NULL;
1413 buf->b_next = hdr->b_buf;
1414 hdr->b_buf = buf;
1415 arc_get_data_buf(buf);
1416 bcopy(from->b_data, buf->b_data, size);
1417
1418 /*
1419 * This buffer already exists in the arc so create a duplicate
1420 * copy for the caller. If the buffer is associated with user data
1421 * then track the size and number of duplicates. These stats will be
1422 * updated as duplicate buffers are created and destroyed.
1423 */
1424 if (hdr->b_type == ARC_BUFC_DATA) {
1425 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1426 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1427 }
1428 hdr->b_datacnt += 1;
1429 return (buf);
1430 }
1431
1432 void
1433 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1434 {
1435 arc_buf_hdr_t *hdr;
1436 kmutex_t *hash_lock;
1437
1438 /*
1439 * Check to see if this buffer is evicted. Callers
1440 * must verify b_data != NULL to know if the add_ref
1441 * was successful.
1442 */
1443 mutex_enter(&buf->b_evict_lock);
1444 if (buf->b_data == NULL) {
1445 mutex_exit(&buf->b_evict_lock);
1446 return;
1447 }
1448 hash_lock = HDR_LOCK(buf->b_hdr);
1449 mutex_enter(hash_lock);
1450 hdr = buf->b_hdr;
1451 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1452 mutex_exit(&buf->b_evict_lock);
1453
1454 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1455 add_reference(hdr, hash_lock, tag);
1456 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1457 arc_access(hdr, hash_lock);
1458 mutex_exit(hash_lock);
1459 ARCSTAT_BUMP(arcstat_hits);
1460 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1461 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1462 data, metadata, hits);
1463 }
1464
1465 /*
1466 * Free the arc data buffer. If it is an l2arc write in progress,
1467 * the buffer is placed on l2arc_free_on_write to be freed later.
1468 */
1469 static void
1470 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1471 {
1472 arc_buf_hdr_t *hdr = buf->b_hdr;
1473
1474 if (HDR_L2_WRITING(hdr)) {
1475 l2arc_data_free_t *df;
1476 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1477 df->l2df_data = buf->b_data;
1478 df->l2df_size = hdr->b_size;
1479 df->l2df_func = free_func;
1480 mutex_enter(&l2arc_free_on_write_mtx);
1481 list_insert_head(l2arc_free_on_write, df);
1482 mutex_exit(&l2arc_free_on_write_mtx);
1483 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1484 } else {
1485 free_func(buf->b_data, hdr->b_size);
1486 }
1487 }
1488
1489 static void
1490 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1491 {
1492 arc_buf_t **bufp;
1493
1494 /* free up data associated with the buf */
1495 if (buf->b_data) {
1496 arc_state_t *state = buf->b_hdr->b_state;
1497 uint64_t size = buf->b_hdr->b_size;
1498 arc_buf_contents_t type = buf->b_hdr->b_type;
1499
1500 arc_cksum_verify(buf);
1501 arc_buf_unwatch(buf);
1502
1503 if (!recycle) {
1504 if (type == ARC_BUFC_METADATA) {
1505 arc_buf_data_free(buf, zio_buf_free);
1506 arc_space_return(size, ARC_SPACE_DATA);
1507 } else {
1508 ASSERT(type == ARC_BUFC_DATA);
1509 arc_buf_data_free(buf, zio_data_buf_free);
1510 ARCSTAT_INCR(arcstat_data_size, -size);
1511 atomic_add_64(&arc_size, -size);
1512 }
1513 }
1514 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1515 uint64_t *cnt = &state->arcs_lsize[type];
1516
1517 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1518 ASSERT(state != arc_anon);
1519
1520 ASSERT3U(*cnt, >=, size);
1521 atomic_add_64(cnt, -size);
1522 }
1523 ASSERT3U(state->arcs_size, >=, size);
1524 atomic_add_64(&state->arcs_size, -size);
1525 buf->b_data = NULL;
1526
1527 /*
1528 * If we're destroying a duplicate buffer make sure
1529 * that the appropriate statistics are updated.
1530 */
1531 if (buf->b_hdr->b_datacnt > 1 &&
1532 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1533 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1534 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1535 }
1536 ASSERT(buf->b_hdr->b_datacnt > 0);
1537 buf->b_hdr->b_datacnt -= 1;
1538 }
1539
1540 /* only remove the buf if requested */
1541 if (!all)
1542 return;
1543
1544 /* remove the buf from the hdr list */
1545 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1546 continue;
1547 *bufp = buf->b_next;
1548 buf->b_next = NULL;
1549
1550 ASSERT(buf->b_efunc == NULL);
1551
1552 /* clean up the buf */
1553 buf->b_hdr = NULL;
1554 kmem_cache_free(buf_cache, buf);
1555 }
1556
1557 static void
1558 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1559 {
1560 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1561 ASSERT3P(hdr->b_state, ==, arc_anon);
1562 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1563 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1564
1565 if (l2hdr != NULL) {
1566 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1567 /*
1568 * To prevent arc_free() and l2arc_evict() from
1569 * attempting to free the same buffer at the same time,
1570 * a FREE_IN_PROGRESS flag is given to arc_free() to
1571 * give it priority. l2arc_evict() can't destroy this
1572 * header while we are waiting on l2arc_buflist_mtx.
1573 *
1574 * The hdr may be removed from l2ad_buflist before we
1575 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1576 */
1577 if (!buflist_held) {
1578 mutex_enter(&l2arc_buflist_mtx);
1579 l2hdr = hdr->b_l2hdr;
1580 }
1581
1582 if (l2hdr != NULL) {
1583 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1584 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1585 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
1586 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1587 if (hdr->b_state == arc_l2c_only)
1588 l2arc_hdr_stat_remove();
1589 hdr->b_l2hdr = NULL;
1590 }
1591
1592 if (!buflist_held)
1593 mutex_exit(&l2arc_buflist_mtx);
1594 }
1595
1596 if (!BUF_EMPTY(hdr)) {
1597 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1598 buf_discard_identity(hdr);
1599 }
1600 while (hdr->b_buf) {
1601 arc_buf_t *buf = hdr->b_buf;
1602
1603 if (buf->b_efunc) {
1604 mutex_enter(&arc_eviction_mtx);
1605 mutex_enter(&buf->b_evict_lock);
1606 ASSERT(buf->b_hdr != NULL);
1607 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1608 hdr->b_buf = buf->b_next;
1609 buf->b_hdr = &arc_eviction_hdr;
1610 buf->b_next = arc_eviction_list;
1611 arc_eviction_list = buf;
1612 mutex_exit(&buf->b_evict_lock);
1613 mutex_exit(&arc_eviction_mtx);
1614 } else {
1615 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1616 }
1617 }
1618 if (hdr->b_freeze_cksum != NULL) {
1619 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1620 hdr->b_freeze_cksum = NULL;
1621 }
1622 if (hdr->b_thawed) {
1623 kmem_free(hdr->b_thawed, 1);
1624 hdr->b_thawed = NULL;
1625 }
1626
1627 ASSERT(!list_link_active(&hdr->b_arc_node));
1628 ASSERT3P(hdr->b_hash_next, ==, NULL);
1629 ASSERT3P(hdr->b_acb, ==, NULL);
1630 kmem_cache_free(hdr_cache, hdr);
1631 }
1632
1633 void
1634 arc_buf_free(arc_buf_t *buf, void *tag)
1635 {
1636 arc_buf_hdr_t *hdr = buf->b_hdr;
1637 int hashed = hdr->b_state != arc_anon;
1638
1639 ASSERT(buf->b_efunc == NULL);
1640 ASSERT(buf->b_data != NULL);
1641
1642 if (hashed) {
1643 kmutex_t *hash_lock = HDR_LOCK(hdr);
1644
1645 mutex_enter(hash_lock);
1646 hdr = buf->b_hdr;
1647 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1648
1649 (void) remove_reference(hdr, hash_lock, tag);
1650 if (hdr->b_datacnt > 1) {
1651 arc_buf_destroy(buf, FALSE, TRUE);
1652 } else {
1653 ASSERT(buf == hdr->b_buf);
1654 ASSERT(buf->b_efunc == NULL);
1655 hdr->b_flags |= ARC_BUF_AVAILABLE;
1656 }
1657 mutex_exit(hash_lock);
1658 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1659 int destroy_hdr;
1660 /*
1661 * We are in the middle of an async write. Don't destroy
1662 * this buffer unless the write completes before we finish
1663 * decrementing the reference count.
1664 */
1665 mutex_enter(&arc_eviction_mtx);
1666 (void) remove_reference(hdr, NULL, tag);
1667 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1668 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1669 mutex_exit(&arc_eviction_mtx);
1670 if (destroy_hdr)
1671 arc_hdr_destroy(hdr);
1672 } else {
1673 if (remove_reference(hdr, NULL, tag) > 0)
1674 arc_buf_destroy(buf, FALSE, TRUE);
1675 else
1676 arc_hdr_destroy(hdr);
1677 }
1678 }
1679
1680 boolean_t
1681 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1682 {
1683 arc_buf_hdr_t *hdr = buf->b_hdr;
1684 kmutex_t *hash_lock = HDR_LOCK(hdr);
1685 boolean_t no_callback = (buf->b_efunc == NULL);
1686
1687 if (hdr->b_state == arc_anon) {
1688 ASSERT(hdr->b_datacnt == 1);
1689 arc_buf_free(buf, tag);
1690 return (no_callback);
1691 }
1692
1693 mutex_enter(hash_lock);
1694 hdr = buf->b_hdr;
1695 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1696 ASSERT(hdr->b_state != arc_anon);
1697 ASSERT(buf->b_data != NULL);
1698
1699 (void) remove_reference(hdr, hash_lock, tag);
1700 if (hdr->b_datacnt > 1) {
1701 if (no_callback)
1702 arc_buf_destroy(buf, FALSE, TRUE);
1703 } else if (no_callback) {
1704 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1705 ASSERT(buf->b_efunc == NULL);
1706 hdr->b_flags |= ARC_BUF_AVAILABLE;
1707 }
1708 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1709 refcount_is_zero(&hdr->b_refcnt));
1710 mutex_exit(hash_lock);
1711 return (no_callback);
1712 }
1713
1714 int
1715 arc_buf_size(arc_buf_t *buf)
1716 {
1717 return (buf->b_hdr->b_size);
1718 }
1719
1720 /*
1721 * Called from the DMU to determine if the current buffer should be
1722 * evicted. In order to ensure proper locking, the eviction must be initiated
1723 * from the DMU. Return true if the buffer is associated with user data and
1724 * duplicate buffers still exist.
1725 */
1726 boolean_t
1727 arc_buf_eviction_needed(arc_buf_t *buf)
1728 {
1729 arc_buf_hdr_t *hdr;
1730 boolean_t evict_needed = B_FALSE;
1731
1732 if (zfs_disable_dup_eviction)
1733 return (B_FALSE);
1734
1735 mutex_enter(&buf->b_evict_lock);
1736 hdr = buf->b_hdr;
1737 if (hdr == NULL) {
1738 /*
1739 * We are in arc_do_user_evicts(); let that function
1740 * perform the eviction.
1741 */
1742 ASSERT(buf->b_data == NULL);
1743 mutex_exit(&buf->b_evict_lock);
1744 return (B_FALSE);
1745 } else if (buf->b_data == NULL) {
1746 /*
1747 * We have already been added to the arc eviction list;
1748 * recommend eviction.
1749 */
1750 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1751 mutex_exit(&buf->b_evict_lock);
1752 return (B_TRUE);
1753 }
1754
1755 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1756 evict_needed = B_TRUE;
1757
1758 mutex_exit(&buf->b_evict_lock);
1759 return (evict_needed);
1760 }
1761
1762 /*
1763 * Evict buffers from list until we've removed the specified number of
1764 * bytes. Move the removed buffers to the appropriate evict state.
1765 * If the recycle flag is set, then attempt to "recycle" a buffer:
1766 * - look for a buffer to evict that is `bytes' long.
1767 * - return the data block from this buffer rather than freeing it.
1768 * This flag is used by callers that are trying to make space for a
1769 * new buffer in a full arc cache.
1770 *
1771 * This function makes a "best effort". It skips over any buffers
1772 * it can't get a hash_lock on, and so may not catch all candidates.
1773 * It may also return without evicting as much space as requested.
1774 */
1775 static void *
1776 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1777 arc_buf_contents_t type)
1778 {
1779 arc_state_t *evicted_state;
1780 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1781 arc_buf_hdr_t *ab, *ab_prev = NULL;
1782 list_t *list = &state->arcs_list[type];
1783 kmutex_t *hash_lock;
1784 boolean_t have_lock;
1785 void *stolen = NULL;
1786
1787 ASSERT(state == arc_mru || state == arc_mfu);
1788
1789 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1790
1791 mutex_enter(&state->arcs_mtx);
1792 mutex_enter(&evicted_state->arcs_mtx);
1793
1794 for (ab = list_tail(list); ab; ab = ab_prev) {
1795 ab_prev = list_prev(list, ab);
1796 /* prefetch buffers have a minimum lifespan */
1797 if (HDR_IO_IN_PROGRESS(ab) ||
1798 (spa && ab->b_spa != spa) ||
1799 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1800 ddi_get_lbolt() - ab->b_arc_access <
1801 arc_min_prefetch_lifespan)) {
1802 skipped++;
1803 continue;
1804 }
1805 /* "lookahead" for better eviction candidate */
1806 if (recycle && ab->b_size != bytes &&
1807 ab_prev && ab_prev->b_size == bytes)
1808 continue;
1809 hash_lock = HDR_LOCK(ab);
1810 have_lock = MUTEX_HELD(hash_lock);
1811 if (have_lock || mutex_tryenter(hash_lock)) {
1812 ASSERT0(refcount_count(&ab->b_refcnt));
1813 ASSERT(ab->b_datacnt > 0);
1814 while (ab->b_buf) {
1815 arc_buf_t *buf = ab->b_buf;
1816 if (!mutex_tryenter(&buf->b_evict_lock)) {
1817 missed += 1;
1818 break;
1819 }
1820 if (buf->b_data) {
1821 bytes_evicted += ab->b_size;
1822 if (recycle && ab->b_type == type &&
1823 ab->b_size == bytes &&
1824 !HDR_L2_WRITING(ab)) {
1825 stolen = buf->b_data;
1826 recycle = FALSE;
1827 }
1828 }
1829 if (buf->b_efunc) {
1830 mutex_enter(&arc_eviction_mtx);
1831 arc_buf_destroy(buf,
1832 buf->b_data == stolen, FALSE);
1833 ab->b_buf = buf->b_next;
1834 buf->b_hdr = &arc_eviction_hdr;
1835 buf->b_next = arc_eviction_list;
1836 arc_eviction_list = buf;
1837 mutex_exit(&arc_eviction_mtx);
1838 mutex_exit(&buf->b_evict_lock);
1839 } else {
1840 mutex_exit(&buf->b_evict_lock);
1841 arc_buf_destroy(buf,
1842 buf->b_data == stolen, TRUE);
1843 }
1844 }
1845
1846 if (ab->b_l2hdr) {
1847 ARCSTAT_INCR(arcstat_evict_l2_cached,
1848 ab->b_size);
1849 } else {
1850 if (l2arc_write_eligible(ab->b_spa, ab)) {
1851 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1852 ab->b_size);
1853 } else {
1854 ARCSTAT_INCR(
1855 arcstat_evict_l2_ineligible,
1856 ab->b_size);
1857 }
1858 }
1859
1860 if (ab->b_datacnt == 0) {
1861 arc_change_state(evicted_state, ab, hash_lock);
1862 ASSERT(HDR_IN_HASH_TABLE(ab));
1863 ab->b_flags |= ARC_IN_HASH_TABLE;
1864 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1865 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1866 }
1867 if (!have_lock)
1868 mutex_exit(hash_lock);
1869 if (bytes >= 0 && bytes_evicted >= bytes)
1870 break;
1871 } else {
1872 missed += 1;
1873 }
1874 }
1875
1876 mutex_exit(&evicted_state->arcs_mtx);
1877 mutex_exit(&state->arcs_mtx);
1878
1879 if (bytes_evicted < bytes)
1880 dprintf("only evicted %lld bytes from %x",
1881 (longlong_t)bytes_evicted, state);
1882
1883 if (skipped)
1884 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1885
1886 if (missed)
1887 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1888
1889 /*
1890 * We have just evicted some data into the ghost state, make
1891 * sure we also adjust the ghost state size if necessary.
1892 */
1893 if (arc_no_grow &&
1894 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1895 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1896 arc_mru_ghost->arcs_size - arc_c;
1897
1898 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1899 int64_t todelete =
1900 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1901 arc_evict_ghost(arc_mru_ghost, NULL, todelete);
1902 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1903 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1904 arc_mru_ghost->arcs_size +
1905 arc_mfu_ghost->arcs_size - arc_c);
1906 arc_evict_ghost(arc_mfu_ghost, NULL, todelete);
1907 }
1908 }
1909
1910 return (stolen);
1911 }
1912
1913 /*
1914 * Remove buffers from list until we've removed the specified number of
1915 * bytes. Destroy the buffers that are removed.
1916 */
1917 static void
1918 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1919 {
1920 arc_buf_hdr_t *ab, *ab_prev;
1921 arc_buf_hdr_t marker = { 0 };
1922 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1923 kmutex_t *hash_lock;
1924 uint64_t bytes_deleted = 0;
1925 uint64_t bufs_skipped = 0;
1926
1927 ASSERT(GHOST_STATE(state));
1928 top:
1929 mutex_enter(&state->arcs_mtx);
1930 for (ab = list_tail(list); ab; ab = ab_prev) {
1931 ab_prev = list_prev(list, ab);
1932 if (spa && ab->b_spa != spa)
1933 continue;
1934
1935 /* ignore markers */
1936 if (ab->b_spa == 0)
1937 continue;
1938
1939 hash_lock = HDR_LOCK(ab);
1940 /* caller may be trying to modify this buffer, skip it */
1941 if (MUTEX_HELD(hash_lock))
1942 continue;
1943 if (mutex_tryenter(hash_lock)) {
1944 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1945 ASSERT(ab->b_buf == NULL);
1946 ARCSTAT_BUMP(arcstat_deleted);
1947 bytes_deleted += ab->b_size;
1948
1949 if (ab->b_l2hdr != NULL) {
1950 /*
1951 * This buffer is cached on the 2nd Level ARC;
1952 * don't destroy the header.
1953 */
1954 arc_change_state(arc_l2c_only, ab, hash_lock);
1955 mutex_exit(hash_lock);
1956 } else {
1957 arc_change_state(arc_anon, ab, hash_lock);
1958 mutex_exit(hash_lock);
1959 arc_hdr_destroy(ab);
1960 }
1961
1962 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1963 if (bytes >= 0 && bytes_deleted >= bytes)
1964 break;
1965 } else if (bytes < 0) {
1966 /*
1967 * Insert a list marker and then wait for the
1968 * hash lock to become available. Once its
1969 * available, restart from where we left off.
1970 */
1971 list_insert_after(list, ab, &marker);
1972 mutex_exit(&state->arcs_mtx);
1973 mutex_enter(hash_lock);
1974 mutex_exit(hash_lock);
1975 mutex_enter(&state->arcs_mtx);
1976 ab_prev = list_prev(list, &marker);
1977 list_remove(list, &marker);
1978 } else
1979 bufs_skipped += 1;
1980 }
1981 mutex_exit(&state->arcs_mtx);
1982
1983 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1984 (bytes < 0 || bytes_deleted < bytes)) {
1985 list = &state->arcs_list[ARC_BUFC_METADATA];
1986 goto top;
1987 }
1988
1989 if (bufs_skipped) {
1990 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1991 ASSERT(bytes >= 0);
1992 }
1993
1994 if (bytes_deleted < bytes)
1995 dprintf("only deleted %lld bytes from %p",
1996 (longlong_t)bytes_deleted, state);
1997 }
1998
1999 static void
2000 arc_adjust(void)
2001 {
2002 int64_t adjustment, delta;
2003
2004 /*
2005 * Adjust MRU size
2006 */
2007
2008 adjustment = MIN((int64_t)(arc_size - arc_c),
2009 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2010 arc_p));
2011
2012 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2013 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2014 (void) arc_evict(arc_mru, NULL, delta, FALSE, ARC_BUFC_DATA);
2015 adjustment -= delta;
2016 }
2017
2018 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2019 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2020 (void) arc_evict(arc_mru, NULL, delta, FALSE,
2021 ARC_BUFC_METADATA);
2022 }
2023
2024 /*
2025 * Adjust MFU size
2026 */
2027
2028 adjustment = arc_size - arc_c;
2029
2030 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2031 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2032 (void) arc_evict(arc_mfu, NULL, delta, FALSE, ARC_BUFC_DATA);
2033 adjustment -= delta;
2034 }
2035
2036 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2037 int64_t delta = MIN(adjustment,
2038 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2039 (void) arc_evict(arc_mfu, NULL, delta, FALSE,
2040 ARC_BUFC_METADATA);
2041 }
2042
2043 /*
2044 * Adjust ghost lists
2045 */
2046
2047 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2048
2049 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2050 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2051 arc_evict_ghost(arc_mru_ghost, NULL, delta);
2052 }
2053
2054 adjustment =
2055 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2056
2057 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2058 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2059 arc_evict_ghost(arc_mfu_ghost, NULL, delta);
2060 }
2061 }
2062
2063 static void
2064 arc_do_user_evicts(void)
2065 {
2066 mutex_enter(&arc_eviction_mtx);
2067 while (arc_eviction_list != NULL) {
2068 arc_buf_t *buf = arc_eviction_list;
2069 arc_eviction_list = buf->b_next;
2070 mutex_enter(&buf->b_evict_lock);
2071 buf->b_hdr = NULL;
2072 mutex_exit(&buf->b_evict_lock);
2073 mutex_exit(&arc_eviction_mtx);
2074
2075 if (buf->b_efunc != NULL)
2076 VERIFY(buf->b_efunc(buf) == 0);
2077
2078 buf->b_efunc = NULL;
2079 buf->b_private = NULL;
2080 kmem_cache_free(buf_cache, buf);
2081 mutex_enter(&arc_eviction_mtx);
2082 }
2083 mutex_exit(&arc_eviction_mtx);
2084 }
2085
2086 /*
2087 * Flush all *evictable* data from the cache for the given spa.
2088 * NOTE: this will not touch "active" (i.e. referenced) data.
2089 */
2090 void
2091 arc_flush(spa_t *spa)
2092 {
2093 uint64_t guid = 0;
2094
2095 if (spa)
2096 guid = spa_load_guid(spa);
2097
2098 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2099 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2100 if (spa)
2101 break;
2102 }
2103 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2104 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2105 if (spa)
2106 break;
2107 }
2108 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2109 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2110 if (spa)
2111 break;
2112 }
2113 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2114 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2115 if (spa)
2116 break;
2117 }
2118
2119 arc_evict_ghost(arc_mru_ghost, guid, -1);
2120 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2121
2122 mutex_enter(&arc_reclaim_thr_lock);
2123 arc_do_user_evicts();
2124 mutex_exit(&arc_reclaim_thr_lock);
2125 ASSERT(spa || arc_eviction_list == NULL);
2126 }
2127
2128 void
2129 arc_shrink(void)
2130 {
2131 if (arc_c > arc_c_min) {
2132 uint64_t to_free;
2133
2134 #ifdef _KERNEL
2135 to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
2136 #else
2137 to_free = arc_c >> arc_shrink_shift;
2138 #endif
2139 if (arc_c > arc_c_min + to_free)
2140 atomic_add_64(&arc_c, -to_free);
2141 else
2142 arc_c = arc_c_min;
2143
2144 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2145 if (arc_c > arc_size)
2146 arc_c = MAX(arc_size, arc_c_min);
2147 if (arc_p > arc_c)
2148 arc_p = (arc_c >> 1);
2149 ASSERT(arc_c >= arc_c_min);
2150 ASSERT((int64_t)arc_p >= 0);
2151 }
2152
2153 if (arc_size > arc_c)
2154 arc_adjust();
2155 }
2156
2157 /*
2158 * Determine if the system is under memory pressure and is asking
2159 * to reclaim memory. A return value of 1 indicates that the system
2160 * is under memory pressure and that the arc should adjust accordingly.
2161 */
2162 static int
2163 arc_reclaim_needed(void)
2164 {
2165 uint64_t extra;
2166
2167 #ifdef _KERNEL
2168
2169 if (needfree)
2170 return (1);
2171
2172 /*
2173 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2174 */
2175 extra = desfree;
2176
2177 /*
2178 * check that we're out of range of the pageout scanner. It starts to
2179 * schedule paging if freemem is less than lotsfree and needfree.
2180 * lotsfree is the high-water mark for pageout, and needfree is the
2181 * number of needed free pages. We add extra pages here to make sure
2182 * the scanner doesn't start up while we're freeing memory.
2183 */
2184 if (freemem < lotsfree + needfree + extra)
2185 return (1);
2186
2187 /*
2188 * check to make sure that swapfs has enough space so that anon
2189 * reservations can still succeed. anon_resvmem() checks that the
2190 * availrmem is greater than swapfs_minfree, and the number of reserved
2191 * swap pages. We also add a bit of extra here just to prevent
2192 * circumstances from getting really dire.
2193 */
2194 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2195 return (1);
2196
2197 #if defined(__i386)
2198 /*
2199 * If we're on an i386 platform, it's possible that we'll exhaust the
2200 * kernel heap space before we ever run out of available physical
2201 * memory. Most checks of the size of the heap_area compare against
2202 * tune.t_minarmem, which is the minimum available real memory that we
2203 * can have in the system. However, this is generally fixed at 25 pages
2204 * which is so low that it's useless. In this comparison, we seek to
2205 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2206 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2207 * free)
2208 */
2209 if (vmem_size(heap_arena, VMEM_FREE) <
2210 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2))
2211 return (1);
2212 #endif
2213
2214 /*
2215 * If zio data pages are being allocated out of a separate heap segment,
2216 * then enforce that the size of available vmem for this arena remains
2217 * above about 1/16th free.
2218 *
2219 * Note: The 1/16th arena free requirement was put in place
2220 * to aggressively evict memory from the arc in order to avoid
2221 * memory fragmentation issues.
2222 */
2223 if (zio_arena != NULL &&
2224 vmem_size(zio_arena, VMEM_FREE) <
2225 (vmem_size(zio_arena, VMEM_ALLOC) >> 4))
2226 return (1);
2227 #else
2228 if (spa_get_random(100) == 0)
2229 return (1);
2230 #endif
2231 return (0);
2232 }
2233
2234 static void
2235 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2236 {
2237 size_t i;
2238 kmem_cache_t *prev_cache = NULL;
2239 kmem_cache_t *prev_data_cache = NULL;
2240 extern kmem_cache_t *zio_buf_cache[];
2241 extern kmem_cache_t *zio_data_buf_cache[];
2242
2243 #ifdef _KERNEL
2244 if (arc_meta_used >= arc_meta_limit) {
2245 /*
2246 * We are exceeding our meta-data cache limit.
2247 * Purge some DNLC entries to release holds on meta-data.
2248 */
2249 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2250 }
2251 #if defined(__i386)
2252 /*
2253 * Reclaim unused memory from all kmem caches.
2254 */
2255 kmem_reap();
2256 #endif
2257 #endif
2258
2259 /*
2260 * An aggressive reclamation will shrink the cache size as well as
2261 * reap free buffers from the arc kmem caches.
2262 */
2263 if (strat == ARC_RECLAIM_AGGR)
2264 arc_shrink();
2265
2266 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2267 if (zio_buf_cache[i] != prev_cache) {
2268 prev_cache = zio_buf_cache[i];
2269 kmem_cache_reap_now(zio_buf_cache[i]);
2270 }
2271 if (zio_data_buf_cache[i] != prev_data_cache) {
2272 prev_data_cache = zio_data_buf_cache[i];
2273 kmem_cache_reap_now(zio_data_buf_cache[i]);
2274 }
2275 }
2276 kmem_cache_reap_now(buf_cache);
2277 kmem_cache_reap_now(hdr_cache);
2278
2279 /*
2280 * Ask the vmem areana to reclaim unused memory from its
2281 * quantum caches.
2282 */
2283 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR)
2284 vmem_qcache_reap(zio_arena);
2285 }
2286
2287 static void
2288 arc_reclaim_thread(void)
2289 {
2290 clock_t growtime = 0;
2291 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2292 callb_cpr_t cpr;
2293
2294 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2295
2296 mutex_enter(&arc_reclaim_thr_lock);
2297 while (arc_thread_exit == 0) {
2298 if (arc_reclaim_needed()) {
2299
2300 if (arc_no_grow) {
2301 if (last_reclaim == ARC_RECLAIM_CONS) {
2302 last_reclaim = ARC_RECLAIM_AGGR;
2303 } else {
2304 last_reclaim = ARC_RECLAIM_CONS;
2305 }
2306 } else {
2307 arc_no_grow = TRUE;
2308 last_reclaim = ARC_RECLAIM_AGGR;
2309 membar_producer();
2310 }
2311
2312 /* reset the growth delay for every reclaim */
2313 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2314
2315 arc_kmem_reap_now(last_reclaim);
2316 arc_warm = B_TRUE;
2317
2318 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2319 arc_no_grow = FALSE;
2320 }
2321
2322 arc_adjust();
2323
2324 if (arc_eviction_list != NULL)
2325 arc_do_user_evicts();
2326
2327 /* block until needed, or one second, whichever is shorter */
2328 CALLB_CPR_SAFE_BEGIN(&cpr);
2329 (void) cv_timedwait(&arc_reclaim_thr_cv,
2330 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2331 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2332 }
2333
2334 arc_thread_exit = 0;
2335 cv_broadcast(&arc_reclaim_thr_cv);
2336 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2337 thread_exit();
2338 }
2339
2340 /*
2341 * Adapt arc info given the number of bytes we are trying to add and
2342 * the state that we are comming from. This function is only called
2343 * when we are adding new content to the cache.
2344 */
2345 static void
2346 arc_adapt(int bytes, arc_state_t *state)
2347 {
2348 int mult;
2349 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2350
2351 if (state == arc_l2c_only)
2352 return;
2353
2354 ASSERT(bytes > 0);
2355 /*
2356 * Adapt the target size of the MRU list:
2357 * - if we just hit in the MRU ghost list, then increase
2358 * the target size of the MRU list.
2359 * - if we just hit in the MFU ghost list, then increase
2360 * the target size of the MFU list by decreasing the
2361 * target size of the MRU list.
2362 */
2363 if (state == arc_mru_ghost) {
2364 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2365 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2366 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2367
2368 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2369 } else if (state == arc_mfu_ghost) {
2370 uint64_t delta;
2371
2372 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2373 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2374 mult = MIN(mult, 10);
2375
2376 delta = MIN(bytes * mult, arc_p);
2377 arc_p = MAX(arc_p_min, arc_p - delta);
2378 }
2379 ASSERT((int64_t)arc_p >= 0);
2380
2381 if (arc_reclaim_needed()) {
2382 cv_signal(&arc_reclaim_thr_cv);
2383 return;
2384 }
2385
2386 if (arc_no_grow)
2387 return;
2388
2389 if (arc_c >= arc_c_max)
2390 return;
2391
2392 /*
2393 * If we're within (2 * maxblocksize) bytes of the target
2394 * cache size, increment the target cache size
2395 */
2396 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2397 atomic_add_64(&arc_c, (int64_t)bytes);
2398 if (arc_c > arc_c_max)
2399 arc_c = arc_c_max;
2400 else if (state == arc_anon)
2401 atomic_add_64(&arc_p, (int64_t)bytes);
2402 if (arc_p > arc_c)
2403 arc_p = arc_c;
2404 }
2405 ASSERT((int64_t)arc_p >= 0);
2406 }
2407
2408 /*
2409 * Check if the cache has reached its limits and eviction is required
2410 * prior to insert.
2411 */
2412 static int
2413 arc_evict_needed(arc_buf_contents_t type)
2414 {
2415 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2416 return (1);
2417
2418 if (arc_reclaim_needed())
2419 return (1);
2420
2421 return (arc_size > arc_c);
2422 }
2423
2424 /*
2425 * The buffer, supplied as the first argument, needs a data block.
2426 * So, if we are at cache max, determine which cache should be victimized.
2427 * We have the following cases:
2428 *
2429 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2430 * In this situation if we're out of space, but the resident size of the MFU is
2431 * under the limit, victimize the MFU cache to satisfy this insertion request.
2432 *
2433 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2434 * Here, we've used up all of the available space for the MRU, so we need to
2435 * evict from our own cache instead. Evict from the set of resident MRU
2436 * entries.
2437 *
2438 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2439 * c minus p represents the MFU space in the cache, since p is the size of the
2440 * cache that is dedicated to the MRU. In this situation there's still space on
2441 * the MFU side, so the MRU side needs to be victimized.
2442 *
2443 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2444 * MFU's resident set is consuming more space than it has been allotted. In
2445 * this situation, we must victimize our own cache, the MFU, for this insertion.
2446 */
2447 static void
2448 arc_get_data_buf(arc_buf_t *buf)
2449 {
2450 arc_state_t *state = buf->b_hdr->b_state;
2451 uint64_t size = buf->b_hdr->b_size;
2452 arc_buf_contents_t type = buf->b_hdr->b_type;
2453
2454 arc_adapt(size, state);
2455
2456 /*
2457 * We have not yet reached cache maximum size,
2458 * just allocate a new buffer.
2459 */
2460 if (!arc_evict_needed(type)) {
2461 if (type == ARC_BUFC_METADATA) {
2462 buf->b_data = zio_buf_alloc(size);
2463 arc_space_consume(size, ARC_SPACE_DATA);
2464 } else {
2465 ASSERT(type == ARC_BUFC_DATA);
2466 buf->b_data = zio_data_buf_alloc(size);
2467 ARCSTAT_INCR(arcstat_data_size, size);
2468 atomic_add_64(&arc_size, size);
2469 }
2470 goto out;
2471 }
2472
2473 /*
2474 * If we are prefetching from the mfu ghost list, this buffer
2475 * will end up on the mru list; so steal space from there.
2476 */
2477 if (state == arc_mfu_ghost)
2478 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2479 else if (state == arc_mru_ghost)
2480 state = arc_mru;
2481
2482 if (state == arc_mru || state == arc_anon) {
2483 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2484 state = (arc_mfu->arcs_lsize[type] >= size &&
2485 arc_p > mru_used) ? arc_mfu : arc_mru;
2486 } else {
2487 /* MFU cases */
2488 uint64_t mfu_space = arc_c - arc_p;
2489 state = (arc_mru->arcs_lsize[type] >= size &&
2490 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2491 }
2492 if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) {
2493 if (type == ARC_BUFC_METADATA) {
2494 buf->b_data = zio_buf_alloc(size);
2495 arc_space_consume(size, ARC_SPACE_DATA);
2496 } else {
2497 ASSERT(type == ARC_BUFC_DATA);
2498 buf->b_data = zio_data_buf_alloc(size);
2499 ARCSTAT_INCR(arcstat_data_size, size);
2500 atomic_add_64(&arc_size, size);
2501 }
2502 ARCSTAT_BUMP(arcstat_recycle_miss);
2503 }
2504 ASSERT(buf->b_data != NULL);
2505 out:
2506 /*
2507 * Update the state size. Note that ghost states have a
2508 * "ghost size" and so don't need to be updated.
2509 */
2510 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2511 arc_buf_hdr_t *hdr = buf->b_hdr;
2512
2513 atomic_add_64(&hdr->b_state->arcs_size, size);
2514 if (list_link_active(&hdr->b_arc_node)) {
2515 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2516 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2517 }
2518 /*
2519 * If we are growing the cache, and we are adding anonymous
2520 * data, and we have outgrown arc_p, update arc_p
2521 */
2522 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2523 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2524 arc_p = MIN(arc_c, arc_p + size);
2525 }
2526 }
2527
2528 /*
2529 * This routine is called whenever a buffer is accessed.
2530 * NOTE: the hash lock is dropped in this function.
2531 */
2532 static void
2533 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2534 {
2535 clock_t now;
2536
2537 ASSERT(MUTEX_HELD(hash_lock));
2538
2539 if (buf->b_state == arc_anon) {
2540 /*
2541 * This buffer is not in the cache, and does not
2542 * appear in our "ghost" list. Add the new buffer
2543 * to the MRU state.
2544 */
2545
2546 ASSERT(buf->b_arc_access == 0);
2547 buf->b_arc_access = ddi_get_lbolt();
2548 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2549 arc_change_state(arc_mru, buf, hash_lock);
2550
2551 } else if (buf->b_state == arc_mru) {
2552 now = ddi_get_lbolt();
2553
2554 /*
2555 * If this buffer is here because of a prefetch, then either:
2556 * - clear the flag if this is a "referencing" read
2557 * (any subsequent access will bump this into the MFU state).
2558 * or
2559 * - move the buffer to the head of the list if this is
2560 * another prefetch (to make it less likely to be evicted).
2561 */
2562 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2563 if (refcount_count(&buf->b_refcnt) == 0) {
2564 ASSERT(list_link_active(&buf->b_arc_node));
2565 } else {
2566 buf->b_flags &= ~ARC_PREFETCH;
2567 ARCSTAT_BUMP(arcstat_mru_hits);
2568 }
2569 buf->b_arc_access = now;
2570 return;
2571 }
2572
2573 /*
2574 * This buffer has been "accessed" only once so far,
2575 * but it is still in the cache. Move it to the MFU
2576 * state.
2577 */
2578 if (now > buf->b_arc_access + ARC_MINTIME) {
2579 /*
2580 * More than 125ms have passed since we
2581 * instantiated this buffer. Move it to the
2582 * most frequently used state.
2583 */
2584 buf->b_arc_access = now;
2585 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2586 arc_change_state(arc_mfu, buf, hash_lock);
2587 }
2588 ARCSTAT_BUMP(arcstat_mru_hits);
2589 } else if (buf->b_state == arc_mru_ghost) {
2590 arc_state_t *new_state;
2591 /*
2592 * This buffer has been "accessed" recently, but
2593 * was evicted from the cache. Move it to the
2594 * MFU state.
2595 */
2596
2597 if (buf->b_flags & ARC_PREFETCH) {
2598 new_state = arc_mru;
2599 if (refcount_count(&buf->b_refcnt) > 0)
2600 buf->b_flags &= ~ARC_PREFETCH;
2601 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2602 } else {
2603 new_state = arc_mfu;
2604 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2605 }
2606
2607 buf->b_arc_access = ddi_get_lbolt();
2608 arc_change_state(new_state, buf, hash_lock);
2609
2610 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2611 } else if (buf->b_state == arc_mfu) {
2612 /*
2613 * This buffer has been accessed more than once and is
2614 * still in the cache. Keep it in the MFU state.
2615 *
2616 * NOTE: an add_reference() that occurred when we did
2617 * the arc_read() will have kicked this off the list.
2618 * If it was a prefetch, we will explicitly move it to
2619 * the head of the list now.
2620 */
2621 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2622 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2623 ASSERT(list_link_active(&buf->b_arc_node));
2624 }
2625 ARCSTAT_BUMP(arcstat_mfu_hits);
2626 buf->b_arc_access = ddi_get_lbolt();
2627 } else if (buf->b_state == arc_mfu_ghost) {
2628 arc_state_t *new_state = arc_mfu;
2629 /*
2630 * This buffer has been accessed more than once but has
2631 * been evicted from the cache. Move it back to the
2632 * MFU state.
2633 */
2634
2635 if (buf->b_flags & ARC_PREFETCH) {
2636 /*
2637 * This is a prefetch access...
2638 * move this block back to the MRU state.
2639 */
2640 ASSERT0(refcount_count(&buf->b_refcnt));
2641 new_state = arc_mru;
2642 }
2643
2644 buf->b_arc_access = ddi_get_lbolt();
2645 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2646 arc_change_state(new_state, buf, hash_lock);
2647
2648 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2649 } else if (buf->b_state == arc_l2c_only) {
2650 /*
2651 * This buffer is on the 2nd Level ARC.
2652 */
2653
2654 buf->b_arc_access = ddi_get_lbolt();
2655 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2656 arc_change_state(arc_mfu, buf, hash_lock);
2657 } else {
2658 ASSERT(!"invalid arc state");
2659 }
2660 }
2661
2662 /* a generic arc_done_func_t which you can use */
2663 /* ARGSUSED */
2664 void
2665 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2666 {
2667 if (zio == NULL || zio->io_error == 0)
2668 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2669 VERIFY(arc_buf_remove_ref(buf, arg));
2670 }
2671
2672 /* a generic arc_done_func_t */
2673 void
2674 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2675 {
2676 arc_buf_t **bufp = arg;
2677 if (zio && zio->io_error) {
2678 VERIFY(arc_buf_remove_ref(buf, arg));
2679 *bufp = NULL;
2680 } else {
2681 *bufp = buf;
2682 ASSERT(buf->b_data);
2683 }
2684 }
2685
2686 static void
2687 arc_read_done(zio_t *zio)
2688 {
2689 arc_buf_hdr_t *hdr, *found;
2690 arc_buf_t *buf;
2691 arc_buf_t *abuf; /* buffer we're assigning to callback */
2692 kmutex_t *hash_lock;
2693 arc_callback_t *callback_list, *acb;
2694 int freeable = FALSE;
2695
2696 buf = zio->io_private;
2697 hdr = buf->b_hdr;
2698
2699 /*
2700 * The hdr was inserted into hash-table and removed from lists
2701 * prior to starting I/O. We should find this header, since
2702 * it's in the hash table, and it should be legit since it's
2703 * not possible to evict it during the I/O. The only possible
2704 * reason for it not to be found is if we were freed during the
2705 * read.
2706 */
2707 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2708 &hash_lock);
2709
2710 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2711 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2712 (found == hdr && HDR_L2_READING(hdr)));
2713
2714 hdr->b_flags &= ~ARC_L2_EVICTED;
2715 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2716 hdr->b_flags &= ~ARC_L2CACHE;
2717
2718 /* byteswap if necessary */
2719 callback_list = hdr->b_acb;
2720 ASSERT(callback_list != NULL);
2721 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2722 dmu_object_byteswap_t bswap =
2723 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2724 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2725 byteswap_uint64_array :
2726 dmu_ot_byteswap[bswap].ob_func;
2727 func(buf->b_data, hdr->b_size);
2728 }
2729
2730 arc_cksum_compute(buf, B_FALSE);
2731 arc_buf_watch(buf);
2732
2733 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2734 /*
2735 * Only call arc_access on anonymous buffers. This is because
2736 * if we've issued an I/O for an evicted buffer, we've already
2737 * called arc_access (to prevent any simultaneous readers from
2738 * getting confused).
2739 */
2740 arc_access(hdr, hash_lock);
2741 }
2742
2743 /* create copies of the data buffer for the callers */
2744 abuf = buf;
2745 for (acb = callback_list; acb; acb = acb->acb_next) {
2746 if (acb->acb_done) {
2747 if (abuf == NULL) {
2748 ARCSTAT_BUMP(arcstat_duplicate_reads);
2749 abuf = arc_buf_clone(buf);
2750 }
2751 acb->acb_buf = abuf;
2752 abuf = NULL;
2753 }
2754 }
2755 hdr->b_acb = NULL;
2756 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2757 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2758 if (abuf == buf) {
2759 ASSERT(buf->b_efunc == NULL);
2760 ASSERT(hdr->b_datacnt == 1);
2761 hdr->b_flags |= ARC_BUF_AVAILABLE;
2762 }
2763
2764 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2765
2766 if (zio->io_error != 0) {
2767 hdr->b_flags |= ARC_IO_ERROR;
2768 if (hdr->b_state != arc_anon)
2769 arc_change_state(arc_anon, hdr, hash_lock);
2770 if (HDR_IN_HASH_TABLE(hdr))
2771 buf_hash_remove(hdr);
2772 freeable = refcount_is_zero(&hdr->b_refcnt);
2773 }
2774
2775 /*
2776 * Broadcast before we drop the hash_lock to avoid the possibility
2777 * that the hdr (and hence the cv) might be freed before we get to
2778 * the cv_broadcast().
2779 */
2780 cv_broadcast(&hdr->b_cv);
2781
2782 if (hash_lock) {
2783 mutex_exit(hash_lock);
2784 } else {
2785 /*
2786 * This block was freed while we waited for the read to
2787 * complete. It has been removed from the hash table and
2788 * moved to the anonymous state (so that it won't show up
2789 * in the cache).
2790 */
2791 ASSERT3P(hdr->b_state, ==, arc_anon);
2792 freeable = refcount_is_zero(&hdr->b_refcnt);
2793 }
2794
2795 /* execute each callback and free its structure */
2796 while ((acb = callback_list) != NULL) {
2797 if (acb->acb_done)
2798 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2799
2800 if (acb->acb_zio_dummy != NULL) {
2801 acb->acb_zio_dummy->io_error = zio->io_error;
2802 zio_nowait(acb->acb_zio_dummy);
2803 }
2804
2805 callback_list = acb->acb_next;
2806 kmem_free(acb, sizeof (arc_callback_t));
2807 }
2808
2809 if (freeable)
2810 arc_hdr_destroy(hdr);
2811 }
2812
2813 /*
2814 * "Read" the block at the specified DVA (in bp) via the
2815 * cache. If the block is found in the cache, invoke the provided
2816 * callback immediately and return. Note that the `zio' parameter
2817 * in the callback will be NULL in this case, since no IO was
2818 * required. If the block is not in the cache pass the read request
2819 * on to the spa with a substitute callback function, so that the
2820 * requested block will be added to the cache.
2821 *
2822 * If a read request arrives for a block that has a read in-progress,
2823 * either wait for the in-progress read to complete (and return the
2824 * results); or, if this is a read with a "done" func, add a record
2825 * to the read to invoke the "done" func when the read completes,
2826 * and return; or just return.
2827 *
2828 * arc_read_done() will invoke all the requested "done" functions
2829 * for readers of this block.
2830 */
2831 int
2832 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
2833 void *private, int priority, int zio_flags, uint32_t *arc_flags,
2834 const zbookmark_t *zb)
2835 {
2836 arc_buf_hdr_t *hdr;
2837 arc_buf_t *buf = NULL;
2838 kmutex_t *hash_lock;
2839 zio_t *rzio;
2840 uint64_t guid = spa_load_guid(spa);
2841
2842 top:
2843 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2844 &hash_lock);
2845 if (hdr && hdr->b_datacnt > 0) {
2846
2847 *arc_flags |= ARC_CACHED;
2848
2849 if (HDR_IO_IN_PROGRESS(hdr)) {
2850
2851 if (*arc_flags & ARC_WAIT) {
2852 cv_wait(&hdr->b_cv, hash_lock);
2853 mutex_exit(hash_lock);
2854 goto top;
2855 }
2856 ASSERT(*arc_flags & ARC_NOWAIT);
2857
2858 if (done) {
2859 arc_callback_t *acb = NULL;
2860
2861 acb = kmem_zalloc(sizeof (arc_callback_t),
2862 KM_SLEEP);
2863 acb->acb_done = done;
2864 acb->acb_private = private;
2865 if (pio != NULL)
2866 acb->acb_zio_dummy = zio_null(pio,
2867 spa, NULL, NULL, NULL, zio_flags);
2868
2869 ASSERT(acb->acb_done != NULL);
2870 acb->acb_next = hdr->b_acb;
2871 hdr->b_acb = acb;
2872 add_reference(hdr, hash_lock, private);
2873 mutex_exit(hash_lock);
2874 return (0);
2875 }
2876 mutex_exit(hash_lock);
2877 return (0);
2878 }
2879
2880 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2881
2882 if (done) {
2883 add_reference(hdr, hash_lock, private);
2884 /*
2885 * If this block is already in use, create a new
2886 * copy of the data so that we will be guaranteed
2887 * that arc_release() will always succeed.
2888 */
2889 buf = hdr->b_buf;
2890 ASSERT(buf);
2891 ASSERT(buf->b_data);
2892 if (HDR_BUF_AVAILABLE(hdr)) {
2893 ASSERT(buf->b_efunc == NULL);
2894 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2895 } else {
2896 buf = arc_buf_clone(buf);
2897 }
2898
2899 } else if (*arc_flags & ARC_PREFETCH &&
2900 refcount_count(&hdr->b_refcnt) == 0) {
2901 hdr->b_flags |= ARC_PREFETCH;
2902 }
2903 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2904 arc_access(hdr, hash_lock);
2905 if (*arc_flags & ARC_L2CACHE)
2906 hdr->b_flags |= ARC_L2CACHE;
2907 if (*arc_flags & ARC_L2COMPRESS)
2908 hdr->b_flags |= ARC_L2COMPRESS;
2909 mutex_exit(hash_lock);
2910 ARCSTAT_BUMP(arcstat_hits);
2911 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2912 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2913 data, metadata, hits);
2914
2915 if (done)
2916 done(NULL, buf, private);
2917 } else {
2918 uint64_t size = BP_GET_LSIZE(bp);
2919 arc_callback_t *acb;
2920 vdev_t *vd = NULL;
2921 uint64_t addr = 0;
2922 boolean_t devw = B_FALSE;
2923
2924 if (hdr == NULL) {
2925 /* this block is not in the cache */
2926 arc_buf_hdr_t *exists;
2927 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2928 buf = arc_buf_alloc(spa, size, private, type);
2929 hdr = buf->b_hdr;
2930 hdr->b_dva = *BP_IDENTITY(bp);
2931 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
2932 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2933 exists = buf_hash_insert(hdr, &hash_lock);
2934 if (exists) {
2935 /* somebody beat us to the hash insert */
2936 mutex_exit(hash_lock);
2937 buf_discard_identity(hdr);
2938 (void) arc_buf_remove_ref(buf, private);
2939 goto top; /* restart the IO request */
2940 }
2941 /* if this is a prefetch, we don't have a reference */
2942 if (*arc_flags & ARC_PREFETCH) {
2943 (void) remove_reference(hdr, hash_lock,
2944 private);
2945 hdr->b_flags |= ARC_PREFETCH;
2946 }
2947 if (*arc_flags & ARC_L2CACHE)
2948 hdr->b_flags |= ARC_L2CACHE;
2949 if (*arc_flags & ARC_L2COMPRESS)
2950 hdr->b_flags |= ARC_L2COMPRESS;
2951 if (BP_GET_LEVEL(bp) > 0)
2952 hdr->b_flags |= ARC_INDIRECT;
2953 } else {
2954 /* this block is in the ghost cache */
2955 ASSERT(GHOST_STATE(hdr->b_state));
2956 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2957 ASSERT0(refcount_count(&hdr->b_refcnt));
2958 ASSERT(hdr->b_buf == NULL);
2959
2960 /* if this is a prefetch, we don't have a reference */
2961 if (*arc_flags & ARC_PREFETCH)
2962 hdr->b_flags |= ARC_PREFETCH;
2963 else
2964 add_reference(hdr, hash_lock, private);
2965 if (*arc_flags & ARC_L2CACHE)
2966 hdr->b_flags |= ARC_L2CACHE;
2967 if (*arc_flags & ARC_L2COMPRESS)
2968 hdr->b_flags |= ARC_L2COMPRESS;
2969 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2970 buf->b_hdr = hdr;
2971 buf->b_data = NULL;
2972 buf->b_efunc = NULL;
2973 buf->b_private = NULL;
2974 buf->b_next = NULL;
2975 hdr->b_buf = buf;
2976 ASSERT(hdr->b_datacnt == 0);
2977 hdr->b_datacnt = 1;
2978 arc_get_data_buf(buf);
2979 arc_access(hdr, hash_lock);
2980 }
2981
2982 ASSERT(!GHOST_STATE(hdr->b_state));
2983
2984 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
2985 acb->acb_done = done;
2986 acb->acb_private = private;
2987
2988 ASSERT(hdr->b_acb == NULL);
2989 hdr->b_acb = acb;
2990 hdr->b_flags |= ARC_IO_IN_PROGRESS;
2991
2992 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
2993 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
2994 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
2995 addr = hdr->b_l2hdr->b_daddr;
2996 /*
2997 * Lock out device removal.
2998 */
2999 if (vdev_is_dead(vd) ||
3000 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3001 vd = NULL;
3002 }
3003
3004 mutex_exit(hash_lock);
3005
3006 /*
3007 * At this point, we have a level 1 cache miss. Try again in
3008 * L2ARC if possible.
3009 */
3010 ASSERT3U(hdr->b_size, ==, size);
3011 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3012 uint64_t, size, zbookmark_t *, zb);
3013 ARCSTAT_BUMP(arcstat_misses);
3014 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3015 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3016 data, metadata, misses);
3017
3018 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3019 /*
3020 * Read from the L2ARC if the following are true:
3021 * 1. The L2ARC vdev was previously cached.
3022 * 2. This buffer still has L2ARC metadata.
3023 * 3. This buffer isn't currently writing to the L2ARC.
3024 * 4. The L2ARC entry wasn't evicted, which may
3025 * also have invalidated the vdev.
3026 * 5. This isn't prefetch and l2arc_noprefetch is set.
3027 */
3028 if (hdr->b_l2hdr != NULL &&
3029 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3030 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3031 l2arc_read_callback_t *cb;
3032
3033 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3034 ARCSTAT_BUMP(arcstat_l2_hits);
3035
3036 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3037 KM_SLEEP);
3038 cb->l2rcb_buf = buf;
3039 cb->l2rcb_spa = spa;
3040 cb->l2rcb_bp = *bp;
3041 cb->l2rcb_zb = *zb;
3042 cb->l2rcb_flags = zio_flags;
3043 cb->l2rcb_compress = hdr->b_l2hdr->b_compress;
3044
3045 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
3046 addr + size < vd->vdev_psize -
3047 VDEV_LABEL_END_SIZE);
3048
3049 /*
3050 * l2arc read. The SCL_L2ARC lock will be
3051 * released by l2arc_read_done().
3052 * Issue a null zio if the underlying buffer
3053 * was squashed to zero size by compression.
3054 */
3055 if (hdr->b_l2hdr->b_compress ==
3056 ZIO_COMPRESS_EMPTY) {
3057 rzio = zio_null(pio, spa, vd,
3058 l2arc_read_done, cb,
3059 zio_flags | ZIO_FLAG_DONT_CACHE |
3060 ZIO_FLAG_CANFAIL |
3061 ZIO_FLAG_DONT_PROPAGATE |
3062 ZIO_FLAG_DONT_RETRY);
3063 } else {
3064 rzio = zio_read_phys(pio, vd, addr,
3065 hdr->b_l2hdr->b_asize,
3066 buf->b_data, ZIO_CHECKSUM_OFF,
3067 l2arc_read_done, cb, priority,
3068 zio_flags | ZIO_FLAG_DONT_CACHE |
3069 ZIO_FLAG_CANFAIL |
3070 ZIO_FLAG_DONT_PROPAGATE |
3071 ZIO_FLAG_DONT_RETRY, B_FALSE);
3072 }
3073 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3074 zio_t *, rzio);
3075 ARCSTAT_INCR(arcstat_l2_read_bytes,
3076 hdr->b_l2hdr->b_asize);
3077
3078 if (*arc_flags & ARC_NOWAIT) {
3079 zio_nowait(rzio);
3080 return (0);
3081 }
3082
3083 ASSERT(*arc_flags & ARC_WAIT);
3084 if (zio_wait(rzio) == 0)
3085 return (0);
3086
3087 /* l2arc read error; goto zio_read() */
3088 } else {
3089 DTRACE_PROBE1(l2arc__miss,
3090 arc_buf_hdr_t *, hdr);
3091 ARCSTAT_BUMP(arcstat_l2_misses);
3092 if (HDR_L2_WRITING(hdr))
3093 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3094 spa_config_exit(spa, SCL_L2ARC, vd);
3095 }
3096 } else {
3097 if (vd != NULL)
3098 spa_config_exit(spa, SCL_L2ARC, vd);
3099 if (l2arc_ndev != 0) {
3100 DTRACE_PROBE1(l2arc__miss,
3101 arc_buf_hdr_t *, hdr);
3102 ARCSTAT_BUMP(arcstat_l2_misses);
3103 }
3104 }
3105
3106 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3107 arc_read_done, buf, priority, zio_flags, zb);
3108
3109 if (*arc_flags & ARC_WAIT)
3110 return (zio_wait(rzio));
3111
3112 ASSERT(*arc_flags & ARC_NOWAIT);
3113 zio_nowait(rzio);
3114 }
3115 return (0);
3116 }
3117
3118 void
3119 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3120 {
3121 ASSERT(buf->b_hdr != NULL);
3122 ASSERT(buf->b_hdr->b_state != arc_anon);
3123 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3124 ASSERT(buf->b_efunc == NULL);
3125 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3126
3127 buf->b_efunc = func;
3128 buf->b_private = private;
3129 }
3130
3131 /*
3132 * Notify the arc that a block was freed, and thus will never be used again.
3133 */
3134 void
3135 arc_freed(spa_t *spa, const blkptr_t *bp)
3136 {
3137 arc_buf_hdr_t *hdr;
3138 kmutex_t *hash_lock;
3139 uint64_t guid = spa_load_guid(spa);
3140
3141 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3142 &hash_lock);
3143 if (hdr == NULL)
3144 return;
3145 if (HDR_BUF_AVAILABLE(hdr)) {
3146 arc_buf_t *buf = hdr->b_buf;
3147 add_reference(hdr, hash_lock, FTAG);
3148 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3149 mutex_exit(hash_lock);
3150
3151 arc_release(buf, FTAG);
3152 (void) arc_buf_remove_ref(buf, FTAG);
3153 } else {
3154 mutex_exit(hash_lock);
3155 }
3156
3157 }
3158
3159 /*
3160 * This is used by the DMU to let the ARC know that a buffer is
3161 * being evicted, so the ARC should clean up. If this arc buf
3162 * is not yet in the evicted state, it will be put there.
3163 */
3164 int
3165 arc_buf_evict(arc_buf_t *buf)
3166 {
3167 arc_buf_hdr_t *hdr;
3168 kmutex_t *hash_lock;
3169 arc_buf_t **bufp;
3170
3171 mutex_enter(&buf->b_evict_lock);
3172 hdr = buf->b_hdr;
3173 if (hdr == NULL) {
3174 /*
3175 * We are in arc_do_user_evicts().
3176 */
3177 ASSERT(buf->b_data == NULL);
3178 mutex_exit(&buf->b_evict_lock);
3179 return (0);
3180 } else if (buf->b_data == NULL) {
3181 arc_buf_t copy = *buf; /* structure assignment */
3182 /*
3183 * We are on the eviction list; process this buffer now
3184 * but let arc_do_user_evicts() do the reaping.
3185 */
3186 buf->b_efunc = NULL;
3187 mutex_exit(&buf->b_evict_lock);
3188 VERIFY(copy.b_efunc(©) == 0);
3189 return (1);
3190 }
3191 hash_lock = HDR_LOCK(hdr);
3192 mutex_enter(hash_lock);
3193 hdr = buf->b_hdr;
3194 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3195
3196 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3197 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3198
3199 /*
3200 * Pull this buffer off of the hdr
3201 */
3202 bufp = &hdr->b_buf;
3203 while (*bufp != buf)
3204 bufp = &(*bufp)->b_next;
3205 *bufp = buf->b_next;
3206
3207 ASSERT(buf->b_data != NULL);
3208 arc_buf_destroy(buf, FALSE, FALSE);
3209
3210 if (hdr->b_datacnt == 0) {
3211 arc_state_t *old_state = hdr->b_state;
3212 arc_state_t *evicted_state;
3213
3214 ASSERT(hdr->b_buf == NULL);
3215 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3216
3217 evicted_state =
3218 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3219
3220 mutex_enter(&old_state->arcs_mtx);
3221 mutex_enter(&evicted_state->arcs_mtx);
3222
3223 arc_change_state(evicted_state, hdr, hash_lock);
3224 ASSERT(HDR_IN_HASH_TABLE(hdr));
3225 hdr->b_flags |= ARC_IN_HASH_TABLE;
3226 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3227
3228 mutex_exit(&evicted_state->arcs_mtx);
3229 mutex_exit(&old_state->arcs_mtx);
3230 }
3231 mutex_exit(hash_lock);
3232 mutex_exit(&buf->b_evict_lock);
3233
3234 VERIFY(buf->b_efunc(buf) == 0);
3235 buf->b_efunc = NULL;
3236 buf->b_private = NULL;
3237 buf->b_hdr = NULL;
3238 buf->b_next = NULL;
3239 kmem_cache_free(buf_cache, buf);
3240 return (1);
3241 }
3242
3243 /*
3244 * Release this buffer from the cache, making it an anonymous buffer. This
3245 * must be done after a read and prior to modifying the buffer contents.
3246 * If the buffer has more than one reference, we must make
3247 * a new hdr for the buffer.
3248 */
3249 void
3250 arc_release(arc_buf_t *buf, void *tag)
3251 {
3252 arc_buf_hdr_t *hdr;
3253 kmutex_t *hash_lock = NULL;
3254 l2arc_buf_hdr_t *l2hdr;
3255 uint64_t buf_size;
3256
3257 /*
3258 * It would be nice to assert that if it's DMU metadata (level >
3259 * 0 || it's the dnode file), then it must be syncing context.
3260 * But we don't know that information at this level.
3261 */
3262
3263 mutex_enter(&buf->b_evict_lock);
3264 hdr = buf->b_hdr;
3265
3266 /* this buffer is not on any list */
3267 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3268
3269 if (hdr->b_state == arc_anon) {
3270 /* this buffer is already released */
3271 ASSERT(buf->b_efunc == NULL);
3272 } else {
3273 hash_lock = HDR_LOCK(hdr);
3274 mutex_enter(hash_lock);
3275 hdr = buf->b_hdr;
3276 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3277 }
3278
3279 l2hdr = hdr->b_l2hdr;
3280 if (l2hdr) {
3281 mutex_enter(&l2arc_buflist_mtx);
3282 hdr->b_l2hdr = NULL;
3283 }
3284 buf_size = hdr->b_size;
3285
3286 /*
3287 * Do we have more than one buf?
3288 */
3289 if (hdr->b_datacnt > 1) {
3290 arc_buf_hdr_t *nhdr;
3291 arc_buf_t **bufp;
3292 uint64_t blksz = hdr->b_size;
3293 uint64_t spa = hdr->b_spa;
3294 arc_buf_contents_t type = hdr->b_type;
3295 uint32_t flags = hdr->b_flags;
3296
3297 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3298 /*
3299 * Pull the data off of this hdr and attach it to
3300 * a new anonymous hdr.
3301 */
3302 (void) remove_reference(hdr, hash_lock, tag);
3303 bufp = &hdr->b_buf;
3304 while (*bufp != buf)
3305 bufp = &(*bufp)->b_next;
3306 *bufp = buf->b_next;
3307 buf->b_next = NULL;
3308
3309 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3310 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3311 if (refcount_is_zero(&hdr->b_refcnt)) {
3312 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3313 ASSERT3U(*size, >=, hdr->b_size);
3314 atomic_add_64(size, -hdr->b_size);
3315 }
3316
3317 /*
3318 * We're releasing a duplicate user data buffer, update
3319 * our statistics accordingly.
3320 */
3321 if (hdr->b_type == ARC_BUFC_DATA) {
3322 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3323 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3324 -hdr->b_size);
3325 }
3326 hdr->b_datacnt -= 1;
3327 arc_cksum_verify(buf);
3328 arc_buf_unwatch(buf);
3329
3330 mutex_exit(hash_lock);
3331
3332 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3333 nhdr->b_size = blksz;
3334 nhdr->b_spa = spa;
3335 nhdr->b_type = type;
3336 nhdr->b_buf = buf;
3337 nhdr->b_state = arc_anon;
3338 nhdr->b_arc_access = 0;
3339 nhdr->b_flags = flags & ARC_L2_WRITING;
3340 nhdr->b_l2hdr = NULL;
3341 nhdr->b_datacnt = 1;
3342 nhdr->b_freeze_cksum = NULL;
3343 (void) refcount_add(&nhdr->b_refcnt, tag);
3344 buf->b_hdr = nhdr;
3345 mutex_exit(&buf->b_evict_lock);
3346 atomic_add_64(&arc_anon->arcs_size, blksz);
3347 } else {
3348 mutex_exit(&buf->b_evict_lock);
3349 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3350 ASSERT(!list_link_active(&hdr->b_arc_node));
3351 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3352 if (hdr->b_state != arc_anon)
3353 arc_change_state(arc_anon, hdr, hash_lock);
3354 hdr->b_arc_access = 0;
3355 if (hash_lock)
3356 mutex_exit(hash_lock);
3357
3358 buf_discard_identity(hdr);
3359 arc_buf_thaw(buf);
3360 }
3361 buf->b_efunc = NULL;
3362 buf->b_private = NULL;
3363
3364 if (l2hdr) {
3365 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
3366 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3367 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3368 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3369 mutex_exit(&l2arc_buflist_mtx);
3370 }
3371 }
3372
3373 int
3374 arc_released(arc_buf_t *buf)
3375 {
3376 int released;
3377
3378 mutex_enter(&buf->b_evict_lock);
3379 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3380 mutex_exit(&buf->b_evict_lock);
3381 return (released);
3382 }
3383
3384 int
3385 arc_has_callback(arc_buf_t *buf)
3386 {
3387 int callback;
3388
3389 mutex_enter(&buf->b_evict_lock);
3390 callback = (buf->b_efunc != NULL);
3391 mutex_exit(&buf->b_evict_lock);
3392 return (callback);
3393 }
3394
3395 #ifdef ZFS_DEBUG
3396 int
3397 arc_referenced(arc_buf_t *buf)
3398 {
3399 int referenced;
3400
3401 mutex_enter(&buf->b_evict_lock);
3402 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3403 mutex_exit(&buf->b_evict_lock);
3404 return (referenced);
3405 }
3406 #endif
3407
3408 static void
3409 arc_write_ready(zio_t *zio)
3410 {
3411 arc_write_callback_t *callback = zio->io_private;
3412 arc_buf_t *buf = callback->awcb_buf;
3413 arc_buf_hdr_t *hdr = buf->b_hdr;
3414
3415 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3416 callback->awcb_ready(zio, buf, callback->awcb_private);
3417
3418 /*
3419 * If the IO is already in progress, then this is a re-write
3420 * attempt, so we need to thaw and re-compute the cksum.
3421 * It is the responsibility of the callback to handle the
3422 * accounting for any re-write attempt.
3423 */
3424 if (HDR_IO_IN_PROGRESS(hdr)) {
3425 mutex_enter(&hdr->b_freeze_lock);
3426 if (hdr->b_freeze_cksum != NULL) {
3427 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3428 hdr->b_freeze_cksum = NULL;
3429 }
3430 mutex_exit(&hdr->b_freeze_lock);
3431 }
3432 arc_cksum_compute(buf, B_FALSE);
3433 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3434 }
3435
3436 static void
3437 arc_write_done(zio_t *zio)
3438 {
3439 arc_write_callback_t *callback = zio->io_private;
3440 arc_buf_t *buf = callback->awcb_buf;
3441 arc_buf_hdr_t *hdr = buf->b_hdr;
3442
3443 ASSERT(hdr->b_acb == NULL);
3444
3445 if (zio->io_error == 0) {
3446 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3447 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3448 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3449 } else {
3450 ASSERT(BUF_EMPTY(hdr));
3451 }
3452
3453 /*
3454 * If the block to be written was all-zero, we may have
3455 * compressed it away. In this case no write was performed
3456 * so there will be no dva/birth/checksum. The buffer must
3457 * therefore remain anonymous (and uncached).
3458 */
3459 if (!BUF_EMPTY(hdr)) {
3460 arc_buf_hdr_t *exists;
3461 kmutex_t *hash_lock;
3462
3463 ASSERT(zio->io_error == 0);
3464
3465 arc_cksum_verify(buf);
3466
3467 exists = buf_hash_insert(hdr, &hash_lock);
3468 if (exists) {
3469 /*
3470 * This can only happen if we overwrite for
3471 * sync-to-convergence, because we remove
3472 * buffers from the hash table when we arc_free().
3473 */
3474 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3475 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3476 panic("bad overwrite, hdr=%p exists=%p",
3477 (void *)hdr, (void *)exists);
3478 ASSERT(refcount_is_zero(&exists->b_refcnt));
3479 arc_change_state(arc_anon, exists, hash_lock);
3480 mutex_exit(hash_lock);
3481 arc_hdr_destroy(exists);
3482 exists = buf_hash_insert(hdr, &hash_lock);
3483 ASSERT3P(exists, ==, NULL);
3484 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3485 /* nopwrite */
3486 ASSERT(zio->io_prop.zp_nopwrite);
3487 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3488 panic("bad nopwrite, hdr=%p exists=%p",
3489 (void *)hdr, (void *)exists);
3490 } else {
3491 /* Dedup */
3492 ASSERT(hdr->b_datacnt == 1);
3493 ASSERT(hdr->b_state == arc_anon);
3494 ASSERT(BP_GET_DEDUP(zio->io_bp));
3495 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3496 }
3497 }
3498 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3499 /* if it's not anon, we are doing a scrub */
3500 if (!exists && hdr->b_state == arc_anon)
3501 arc_access(hdr, hash_lock);
3502 mutex_exit(hash_lock);
3503 } else {
3504 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3505 }
3506
3507 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3508 callback->awcb_done(zio, buf, callback->awcb_private);
3509
3510 kmem_free(callback, sizeof (arc_write_callback_t));
3511 }
3512
3513 zio_t *
3514 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3515 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
3516 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *done,
3517 void *private, int priority, int zio_flags, const zbookmark_t *zb)
3518 {
3519 arc_buf_hdr_t *hdr = buf->b_hdr;
3520 arc_write_callback_t *callback;
3521 zio_t *zio;
3522
3523 ASSERT(ready != NULL);
3524 ASSERT(done != NULL);
3525 ASSERT(!HDR_IO_ERROR(hdr));
3526 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3527 ASSERT(hdr->b_acb == NULL);
3528 if (l2arc)
3529 hdr->b_flags |= ARC_L2CACHE;
3530 if (l2arc_compress)
3531 hdr->b_flags |= ARC_L2COMPRESS;
3532 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3533 callback->awcb_ready = ready;
3534 callback->awcb_done = done;
3535 callback->awcb_private = private;
3536 callback->awcb_buf = buf;
3537
3538 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3539 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3540
3541 return (zio);
3542 }
3543
3544 static int
3545 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3546 {
3547 #ifdef _KERNEL
3548 uint64_t available_memory = ptob(freemem);
3549 static uint64_t page_load = 0;
3550 static uint64_t last_txg = 0;
3551
3552 #if defined(__i386)
3553 available_memory =
3554 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3555 #endif
3556 if (available_memory >= zfs_write_limit_max)
3557 return (0);
3558
3559 if (txg > last_txg) {
3560 last_txg = txg;
3561 page_load = 0;
3562 }
3563 /*
3564 * If we are in pageout, we know that memory is already tight,
3565 * the arc is already going to be evicting, so we just want to
3566 * continue to let page writes occur as quickly as possible.
3567 */
3568 if (curproc == proc_pageout) {
3569 if (page_load > MAX(ptob(minfree), available_memory) / 4)
3570 return (SET_ERROR(ERESTART));
3571 /* Note: reserve is inflated, so we deflate */
3572 page_load += reserve / 8;
3573 return (0);
3574 } else if (page_load > 0 && arc_reclaim_needed()) {
3575 /* memory is low, delay before restarting */
3576 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3577 return (SET_ERROR(EAGAIN));
3578 }
3579 page_load = 0;
3580
3581 if (arc_size > arc_c_min) {
3582 uint64_t evictable_memory =
3583 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3584 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3585 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3586 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3587 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3588 }
3589
3590 if (inflight_data > available_memory / 4) {
3591 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3592 return (SET_ERROR(ERESTART));
3593 }
3594 #endif
3595 return (0);
3596 }
3597
3598 void
3599 arc_tempreserve_clear(uint64_t reserve)
3600 {
3601 atomic_add_64(&arc_tempreserve, -reserve);
3602 ASSERT((int64_t)arc_tempreserve >= 0);
3603 }
3604
3605 int
3606 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3607 {
3608 int error;
3609 uint64_t anon_size;
3610
3611 #ifdef ZFS_DEBUG
3612 /*
3613 * Once in a while, fail for no reason. Everything should cope.
3614 */
3615 if (spa_get_random(10000) == 0) {
3616 dprintf("forcing random failure\n");
3617 return (SET_ERROR(ERESTART));
3618 }
3619 #endif
3620 if (reserve > arc_c/4 && !arc_no_grow)
3621 arc_c = MIN(arc_c_max, reserve * 4);
3622 if (reserve > arc_c)
3623 return (SET_ERROR(ENOMEM));
3624
3625 /*
3626 * Don't count loaned bufs as in flight dirty data to prevent long
3627 * network delays from blocking transactions that are ready to be
3628 * assigned to a txg.
3629 */
3630 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3631
3632 /*
3633 * Writes will, almost always, require additional memory allocations
3634 * in order to compress/encrypt/etc the data. We therefore need to
3635 * make sure that there is sufficient available memory for this.
3636 */
3637 if (error = arc_memory_throttle(reserve, anon_size, txg))
3638 return (error);
3639
3640 /*
3641 * Throttle writes when the amount of dirty data in the cache
3642 * gets too large. We try to keep the cache less than half full
3643 * of dirty blocks so that our sync times don't grow too large.
3644 * Note: if two requests come in concurrently, we might let them
3645 * both succeed, when one of them should fail. Not a huge deal.
3646 */
3647
3648 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3649 anon_size > arc_c / 4) {
3650 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3651 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3652 arc_tempreserve>>10,
3653 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3654 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3655 reserve>>10, arc_c>>10);
3656 return (SET_ERROR(ERESTART));
3657 }
3658 atomic_add_64(&arc_tempreserve, reserve);
3659 return (0);
3660 }
3661
3662 void
3663 arc_init(void)
3664 {
3665 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3666 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3667
3668 /* Convert seconds to clock ticks */
3669 arc_min_prefetch_lifespan = 1 * hz;
3670
3671 /* Start out with 1/8 of all memory */
3672 arc_c = physmem * PAGESIZE / 8;
3673
3674 #ifdef _KERNEL
3675 /*
3676 * On architectures where the physical memory can be larger
3677 * than the addressable space (intel in 32-bit mode), we may
3678 * need to limit the cache to 1/8 of VM size.
3679 */
3680 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3681 #endif
3682
3683 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3684 arc_c_min = MAX(arc_c / 4, 64<<20);
3685 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
3686 if (arc_c * 8 >= 1<<30)
3687 arc_c_max = (arc_c * 8) - (1<<30);
3688 else
3689 arc_c_max = arc_c_min;
3690 arc_c_max = MAX(arc_c * 6, arc_c_max);
3691
3692 /*
3693 * Allow the tunables to override our calculations if they are
3694 * reasonable (ie. over 64MB)
3695 */
3696 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3697 arc_c_max = zfs_arc_max;
3698 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3699 arc_c_min = zfs_arc_min;
3700
3701 arc_c = arc_c_max;
3702 arc_p = (arc_c >> 1);
3703
3704 /* limit meta-data to 1/4 of the arc capacity */
3705 arc_meta_limit = arc_c_max / 4;
3706
3707 /* Allow the tunable to override if it is reasonable */
3708 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3709 arc_meta_limit = zfs_arc_meta_limit;
3710
3711 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3712 arc_c_min = arc_meta_limit / 2;
3713
3714 if (zfs_arc_grow_retry > 0)
3715 arc_grow_retry = zfs_arc_grow_retry;
3716
3717 if (zfs_arc_shrink_shift > 0)
3718 arc_shrink_shift = zfs_arc_shrink_shift;
3719
3720 if (zfs_arc_p_min_shift > 0)
3721 arc_p_min_shift = zfs_arc_p_min_shift;
3722
3723 /* if kmem_flags are set, lets try to use less memory */
3724 if (kmem_debugging())
3725 arc_c = arc_c / 2;
3726 if (arc_c < arc_c_min)
3727 arc_c = arc_c_min;
3728
3729 arc_anon = &ARC_anon;
3730 arc_mru = &ARC_mru;
3731 arc_mru_ghost = &ARC_mru_ghost;
3732 arc_mfu = &ARC_mfu;
3733 arc_mfu_ghost = &ARC_mfu_ghost;
3734 arc_l2c_only = &ARC_l2c_only;
3735 arc_size = 0;
3736
3737 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3738 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3739 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3740 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3741 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3742 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3743
3744 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3745 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3746 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3747 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3748 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3749 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3750 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3751 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3752 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3753 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3754 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3755 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3756 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3757 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3758 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3759 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3760 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3761 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3762 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3763 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3764
3765 buf_init();
3766
3767 arc_thread_exit = 0;
3768 arc_eviction_list = NULL;
3769 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3770 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3771
3772 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3773 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3774
3775 if (arc_ksp != NULL) {
3776 arc_ksp->ks_data = &arc_stats;
3777 kstat_install(arc_ksp);
3778 }
3779
3780 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3781 TS_RUN, minclsyspri);
3782
3783 arc_dead = FALSE;
3784 arc_warm = B_FALSE;
3785
3786 if (zfs_write_limit_max == 0)
3787 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3788 else
3789 zfs_write_limit_shift = 0;
3790 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3791 }
3792
3793 void
3794 arc_fini(void)
3795 {
3796 mutex_enter(&arc_reclaim_thr_lock);
3797 arc_thread_exit = 1;
3798 while (arc_thread_exit != 0)
3799 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3800 mutex_exit(&arc_reclaim_thr_lock);
3801
3802 arc_flush(NULL);
3803
3804 arc_dead = TRUE;
3805
3806 if (arc_ksp != NULL) {
3807 kstat_delete(arc_ksp);
3808 arc_ksp = NULL;
3809 }
3810
3811 mutex_destroy(&arc_eviction_mtx);
3812 mutex_destroy(&arc_reclaim_thr_lock);
3813 cv_destroy(&arc_reclaim_thr_cv);
3814
3815 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3816 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3817 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3818 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3819 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3820 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3821 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3822 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3823
3824 mutex_destroy(&arc_anon->arcs_mtx);
3825 mutex_destroy(&arc_mru->arcs_mtx);
3826 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3827 mutex_destroy(&arc_mfu->arcs_mtx);
3828 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3829 mutex_destroy(&arc_l2c_only->arcs_mtx);
3830
3831 mutex_destroy(&zfs_write_limit_lock);
3832
3833 buf_fini();
3834
3835 ASSERT(arc_loaned_bytes == 0);
3836 }
3837
3838 /*
3839 * Level 2 ARC
3840 *
3841 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3842 * It uses dedicated storage devices to hold cached data, which are populated
3843 * using large infrequent writes. The main role of this cache is to boost
3844 * the performance of random read workloads. The intended L2ARC devices
3845 * include short-stroked disks, solid state disks, and other media with
3846 * substantially faster read latency than disk.
3847 *
3848 * +-----------------------+
3849 * | ARC |
3850 * +-----------------------+
3851 * | ^ ^
3852 * | | |
3853 * l2arc_feed_thread() arc_read()
3854 * | | |
3855 * | l2arc read |
3856 * V | |
3857 * +---------------+ |
3858 * | L2ARC | |
3859 * +---------------+ |
3860 * | ^ |
3861 * l2arc_write() | |
3862 * | | |
3863 * V | |
3864 * +-------+ +-------+
3865 * | vdev | | vdev |
3866 * | cache | | cache |
3867 * +-------+ +-------+
3868 * +=========+ .-----.
3869 * : L2ARC : |-_____-|
3870 * : devices : | Disks |
3871 * +=========+ `-_____-'
3872 *
3873 * Read requests are satisfied from the following sources, in order:
3874 *
3875 * 1) ARC
3876 * 2) vdev cache of L2ARC devices
3877 * 3) L2ARC devices
3878 * 4) vdev cache of disks
3879 * 5) disks
3880 *
3881 * Some L2ARC device types exhibit extremely slow write performance.
3882 * To accommodate for this there are some significant differences between
3883 * the L2ARC and traditional cache design:
3884 *
3885 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3886 * the ARC behave as usual, freeing buffers and placing headers on ghost
3887 * lists. The ARC does not send buffers to the L2ARC during eviction as
3888 * this would add inflated write latencies for all ARC memory pressure.
3889 *
3890 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3891 * It does this by periodically scanning buffers from the eviction-end of
3892 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3893 * not already there. It scans until a headroom of buffers is satisfied,
3894 * which itself is a buffer for ARC eviction. If a compressible buffer is
3895 * found during scanning and selected for writing to an L2ARC device, we
3896 * temporarily boost scanning headroom during the next scan cycle to make
3897 * sure we adapt to compression effects (which might significantly reduce
3898 * the data volume we write to L2ARC). The thread that does this is
3899 * l2arc_feed_thread(), illustrated below; example sizes are included to
3900 * provide a better sense of ratio than this diagram:
3901 *
3902 * head --> tail
3903 * +---------------------+----------+
3904 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
3905 * +---------------------+----------+ | o L2ARC eligible
3906 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
3907 * +---------------------+----------+ |
3908 * 15.9 Gbytes ^ 32 Mbytes |
3909 * headroom |
3910 * l2arc_feed_thread()
3911 * |
3912 * l2arc write hand <--[oooo]--'
3913 * | 8 Mbyte
3914 * | write max
3915 * V
3916 * +==============================+
3917 * L2ARC dev |####|#|###|###| |####| ... |
3918 * +==============================+
3919 * 32 Gbytes
3920 *
3921 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
3922 * evicted, then the L2ARC has cached a buffer much sooner than it probably
3923 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
3924 * safe to say that this is an uncommon case, since buffers at the end of
3925 * the ARC lists have moved there due to inactivity.
3926 *
3927 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
3928 * then the L2ARC simply misses copying some buffers. This serves as a
3929 * pressure valve to prevent heavy read workloads from both stalling the ARC
3930 * with waits and clogging the L2ARC with writes. This also helps prevent
3931 * the potential for the L2ARC to churn if it attempts to cache content too
3932 * quickly, such as during backups of the entire pool.
3933 *
3934 * 5. After system boot and before the ARC has filled main memory, there are
3935 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
3936 * lists can remain mostly static. Instead of searching from tail of these
3937 * lists as pictured, the l2arc_feed_thread() will search from the list heads
3938 * for eligible buffers, greatly increasing its chance of finding them.
3939 *
3940 * The L2ARC device write speed is also boosted during this time so that
3941 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
3942 * there are no L2ARC reads, and no fear of degrading read performance
3943 * through increased writes.
3944 *
3945 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
3946 * the vdev queue can aggregate them into larger and fewer writes. Each
3947 * device is written to in a rotor fashion, sweeping writes through
3948 * available space then repeating.
3949 *
3950 * 7. The L2ARC does not store dirty content. It never needs to flush
3951 * write buffers back to disk based storage.
3952 *
3953 * 8. If an ARC buffer is written (and dirtied) which also exists in the
3954 * L2ARC, the now stale L2ARC buffer is immediately dropped.
3955 *
3956 * The performance of the L2ARC can be tweaked by a number of tunables, which
3957 * may be necessary for different workloads:
3958 *
3959 * l2arc_write_max max write bytes per interval
3960 * l2arc_write_boost extra write bytes during device warmup
3961 * l2arc_noprefetch skip caching prefetched buffers
3962 * l2arc_headroom number of max device writes to precache
3963 * l2arc_headroom_boost when we find compressed buffers during ARC
3964 * scanning, we multiply headroom by this
3965 * percentage factor for the next scan cycle,
3966 * since more compressed buffers are likely to
3967 * be present
3968 * l2arc_feed_secs seconds between L2ARC writing
3969 *
3970 * Tunables may be removed or added as future performance improvements are
3971 * integrated, and also may become zpool properties.
3972 *
3973 * There are three key functions that control how the L2ARC warms up:
3974 *
3975 * l2arc_write_eligible() check if a buffer is eligible to cache
3976 * l2arc_write_size() calculate how much to write
3977 * l2arc_write_interval() calculate sleep delay between writes
3978 *
3979 * These three functions determine what to write, how much, and how quickly
3980 * to send writes.
3981 */
3982
3983 static boolean_t
3984 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
3985 {
3986 /*
3987 * A buffer is *not* eligible for the L2ARC if it:
3988 * 1. belongs to a different spa.
3989 * 2. is already cached on the L2ARC.
3990 * 3. has an I/O in progress (it may be an incomplete read).
3991 * 4. is flagged not eligible (zfs property).
3992 */
3993 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
3994 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
3995 return (B_FALSE);
3996
3997 return (B_TRUE);
3998 }
3999
4000 static uint64_t
4001 l2arc_write_size(void)
4002 {
4003 uint64_t size;
4004
4005 /*
4006 * Make sure our globals have meaningful values in case the user
4007 * altered them.
4008 */
4009 size = l2arc_write_max;
4010 if (size == 0) {
4011 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
4012 "be greater than zero, resetting it to the default (%d)",
4013 L2ARC_WRITE_SIZE);
4014 size = l2arc_write_max = L2ARC_WRITE_SIZE;
4015 }
4016
4017 if (arc_warm == B_FALSE)
4018 size += l2arc_write_boost;
4019
4020 return (size);
4021
4022 }
4023
4024 static clock_t
4025 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4026 {
4027 clock_t interval, next, now;
4028
4029 /*
4030 * If the ARC lists are busy, increase our write rate; if the
4031 * lists are stale, idle back. This is achieved by checking
4032 * how much we previously wrote - if it was more than half of
4033 * what we wanted, schedule the next write much sooner.
4034 */
4035 if (l2arc_feed_again && wrote > (wanted / 2))
4036 interval = (hz * l2arc_feed_min_ms) / 1000;
4037 else
4038 interval = hz * l2arc_feed_secs;
4039
4040 now = ddi_get_lbolt();
4041 next = MAX(now, MIN(now + interval, began + interval));
4042
4043 return (next);
4044 }
4045
4046 static void
4047 l2arc_hdr_stat_add(void)
4048 {
4049 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4050 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4051 }
4052
4053 static void
4054 l2arc_hdr_stat_remove(void)
4055 {
4056 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4057 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4058 }
4059
4060 /*
4061 * Cycle through L2ARC devices. This is how L2ARC load balances.
4062 * If a device is returned, this also returns holding the spa config lock.
4063 */
4064 static l2arc_dev_t *
4065 l2arc_dev_get_next(void)
4066 {
4067 l2arc_dev_t *first, *next = NULL;
4068
4069 /*
4070 * Lock out the removal of spas (spa_namespace_lock), then removal
4071 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4072 * both locks will be dropped and a spa config lock held instead.
4073 */
4074 mutex_enter(&spa_namespace_lock);
4075 mutex_enter(&l2arc_dev_mtx);
4076
4077 /* if there are no vdevs, there is nothing to do */
4078 if (l2arc_ndev == 0)
4079 goto out;
4080
4081 first = NULL;
4082 next = l2arc_dev_last;
4083 do {
4084 /* loop around the list looking for a non-faulted vdev */
4085 if (next == NULL) {
4086 next = list_head(l2arc_dev_list);
4087 } else {
4088 next = list_next(l2arc_dev_list, next);
4089 if (next == NULL)
4090 next = list_head(l2arc_dev_list);
4091 }
4092
4093 /* if we have come back to the start, bail out */
4094 if (first == NULL)
4095 first = next;
4096 else if (next == first)
4097 break;
4098
4099 } while (vdev_is_dead(next->l2ad_vdev));
4100
4101 /* if we were unable to find any usable vdevs, return NULL */
4102 if (vdev_is_dead(next->l2ad_vdev))
4103 next = NULL;
4104
4105 l2arc_dev_last = next;
4106
4107 out:
4108 mutex_exit(&l2arc_dev_mtx);
4109
4110 /*
4111 * Grab the config lock to prevent the 'next' device from being
4112 * removed while we are writing to it.
4113 */
4114 if (next != NULL)
4115 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4116 mutex_exit(&spa_namespace_lock);
4117
4118 return (next);
4119 }
4120
4121 /*
4122 * Free buffers that were tagged for destruction.
4123 */
4124 static void
4125 l2arc_do_free_on_write(void)
4126 {
4127 list_t *buflist;
4128 l2arc_data_free_t *df, *df_prev;
4129
4130 mutex_enter(&l2arc_free_on_write_mtx);
4131 buflist = l2arc_free_on_write;
4132
4133 for (df = list_tail(buflist); df; df = df_prev) {
4134 df_prev = list_prev(buflist, df);
4135 ASSERT(df->l2df_data != NULL);
4136 ASSERT(df->l2df_func != NULL);
4137 df->l2df_func(df->l2df_data, df->l2df_size);
4138 list_remove(buflist, df);
4139 kmem_free(df, sizeof (l2arc_data_free_t));
4140 }
4141
4142 mutex_exit(&l2arc_free_on_write_mtx);
4143 }
4144
4145 /*
4146 * A write to a cache device has completed. Update all headers to allow
4147 * reads from these buffers to begin.
4148 */
4149 static void
4150 l2arc_write_done(zio_t *zio)
4151 {
4152 l2arc_write_callback_t *cb;
4153 l2arc_dev_t *dev;
4154 list_t *buflist;
4155 arc_buf_hdr_t *head, *ab;
4156
4157 struct defer_done_entry {
4158 arc_buf_hdr_t *dde_buf;
4159 list_node_t dde_node;
4160 } *dde, *dde_next;
4161 list_t defer_done_list;
4162
4163 cb = zio->io_private;
4164 ASSERT(cb != NULL);
4165 dev = cb->l2wcb_dev;
4166 ASSERT(dev != NULL);
4167 head = cb->l2wcb_head;
4168 ASSERT(head != NULL);
4169 buflist = dev->l2ad_buflist;
4170 ASSERT(buflist != NULL);
4171 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4172 l2arc_write_callback_t *, cb);
4173
4174 if (zio->io_error != 0)
4175 ARCSTAT_BUMP(arcstat_l2_writes_error);
4176
4177 mutex_enter(&l2arc_buflist_mtx);
4178
4179 /*
4180 * All writes completed, or an error was hit.
4181 */
4182 list_create(&defer_done_list, sizeof (*dde),
4183 offsetof(struct defer_done_entry, dde_node));
4184 for (ab = list_prev(buflist, head); ab; ab = list_prev(buflist, ab)) {
4185 /*
4186 * Can't pause here to grab hash_lock while also holding
4187 * l2arc_buflist_mtx, so place the buffers on a temporary
4188 * thread-local list for later processing.
4189 */
4190 dde = kmem_alloc(sizeof (*dde), KM_SLEEP);
4191 dde->dde_buf = ab;
4192 list_insert_tail(&defer_done_list, dde);
4193 }
4194
4195 atomic_inc_64(&l2arc_writes_done);
4196 list_remove(buflist, head);
4197 kmem_cache_free(hdr_cache, head);
4198 mutex_exit(&l2arc_buflist_mtx);
4199
4200 /*
4201 * Now process the buffers. We're not holding l2arc_buflist_mtx
4202 * anymore, so we can do a regular mutex_enter on the hash_lock.
4203 */
4204 for (dde = list_head(&defer_done_list); dde != NULL; dde = dde_next) {
4205 kmutex_t *hash_lock;
4206
4207 dde_next = list_next(&defer_done_list, dde);
4208 ab = dde->dde_buf;
4209 hash_lock = HDR_LOCK(ab);
4210
4211 mutex_enter(hash_lock);
4212
4213 if (zio->io_error != 0) {
4214 /*
4215 * Error - drop L2ARC entry.
4216 */
4217 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr;
4218 mutex_enter(&l2arc_buflist_mtx);
4219 list_remove(buflist, ab);
4220 mutex_exit(&l2arc_buflist_mtx);
4221 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
4222 ab->b_l2hdr = NULL;
4223 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
4224 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4225 }
4226
4227 /*
4228 * Allow ARC to begin reads to this L2ARC entry.
4229 */
4230 ab->b_flags &= ~ARC_L2_WRITING;
4231
4232 mutex_exit(hash_lock);
4233
4234 list_remove(&defer_done_list, dde);
4235 }
4236 list_destroy(&defer_done_list);
4237
4238 l2arc_do_free_on_write();
4239
4240 kmem_free(cb, sizeof (l2arc_write_callback_t));
4241 }
4242
4243 /*
4244 * A read to a cache device completed. Validate buffer contents before
4245 * handing over to the regular ARC routines.
4246 */
4247 static void
4248 l2arc_read_done(zio_t *zio)
4249 {
4250 l2arc_read_callback_t *cb;
4251 arc_buf_hdr_t *hdr;
4252 arc_buf_t *buf;
4253 kmutex_t *hash_lock;
4254 int equal;
4255
4256 ASSERT(zio->io_vd != NULL);
4257 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4258
4259 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4260
4261 cb = zio->io_private;
4262 ASSERT(cb != NULL);
4263 buf = cb->l2rcb_buf;
4264 ASSERT(buf != NULL);
4265
4266 hash_lock = HDR_LOCK(buf->b_hdr);
4267 mutex_enter(hash_lock);
4268 hdr = buf->b_hdr;
4269 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4270
4271 /*
4272 * If the buffer was compressed, decompress it first.
4273 */
4274 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
4275 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
4276 ASSERT(zio->io_data != NULL);
4277
4278 /*
4279 * Check this survived the L2ARC journey.
4280 */
4281 equal = arc_cksum_equal(buf);
4282 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4283 mutex_exit(hash_lock);
4284 zio->io_private = buf;
4285 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4286 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4287 arc_read_done(zio);
4288 } else {
4289 mutex_exit(hash_lock);
4290 /*
4291 * Buffer didn't survive caching. Increment stats and
4292 * reissue to the original storage device.
4293 */
4294 if (zio->io_error != 0) {
4295 ARCSTAT_BUMP(arcstat_l2_io_error);
4296 } else {
4297 zio->io_error = SET_ERROR(EIO);
4298 }
4299 if (!equal)
4300 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4301
4302 /*
4303 * If there's no waiter, issue an async i/o to the primary
4304 * storage now. If there *is* a waiter, the caller must
4305 * issue the i/o in a context where it's OK to block.
4306 */
4307 if (zio->io_waiter == NULL) {
4308 zio_t *pio = zio_unique_parent(zio);
4309
4310 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4311
4312 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4313 buf->b_data, zio->io_size, arc_read_done, buf,
4314 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4315 }
4316 }
4317
4318 kmem_free(cb, sizeof (l2arc_read_callback_t));
4319 }
4320
4321 /*
4322 * This is the list priority from which the L2ARC will search for pages to
4323 * cache. This is used within loops (0..3) to cycle through lists in the
4324 * desired order. This order can have a significant effect on cache
4325 * performance.
4326 *
4327 * Currently the metadata lists are hit first, MFU then MRU, followed by
4328 * the data lists. This function returns a locked list, and also returns
4329 * the lock pointer.
4330 */
4331 static list_t *
4332 l2arc_list_locked(int list_num, kmutex_t **lock)
4333 {
4334 list_t *list = NULL;
4335
4336 ASSERT(list_num >= 0 && list_num <= 3);
4337
4338 switch (list_num) {
4339 case 0:
4340 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4341 *lock = &arc_mfu->arcs_mtx;
4342 break;
4343 case 1:
4344 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4345 *lock = &arc_mru->arcs_mtx;
4346 break;
4347 case 2:
4348 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4349 *lock = &arc_mfu->arcs_mtx;
4350 break;
4351 case 3:
4352 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4353 *lock = &arc_mru->arcs_mtx;
4354 break;
4355 }
4356
4357 ASSERT(!(MUTEX_HELD(*lock)));
4358 mutex_enter(*lock);
4359 return (list);
4360 }
4361
4362 /*
4363 * Evict buffers from the device write hand to the distance specified in
4364 * bytes. This distance may span populated buffers, it may span nothing.
4365 * This is clearing a region on the L2ARC device ready for writing.
4366 * If the 'all' boolean is set, every buffer is evicted.
4367 */
4368 static void
4369 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4370 {
4371 list_t *buflist;
4372 l2arc_buf_hdr_t *l2hdr;
4373 arc_buf_hdr_t *ab, *ab_prev;
4374 kmutex_t *hash_lock;
4375 uint64_t taddr;
4376
4377 buflist = dev->l2ad_buflist;
4378
4379 if (buflist == NULL)
4380 return;
4381
4382 if (!all && dev->l2ad_first) {
4383 /*
4384 * This is the first sweep through the device. There is
4385 * nothing to evict.
4386 */
4387 return;
4388 }
4389
4390 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4391 /*
4392 * When nearing the end of the device, evict to the end
4393 * before the device write hand jumps to the start.
4394 */
4395 taddr = dev->l2ad_end;
4396 } else {
4397 taddr = dev->l2ad_hand + distance;
4398 }
4399 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4400 uint64_t, taddr, boolean_t, all);
4401
4402 top:
4403 mutex_enter(&l2arc_buflist_mtx);
4404 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4405 ab_prev = list_prev(buflist, ab);
4406
4407 hash_lock = HDR_LOCK(ab);
4408 if (!mutex_tryenter(hash_lock)) {
4409 /*
4410 * Missed the hash lock. Retry.
4411 */
4412 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4413 mutex_exit(&l2arc_buflist_mtx);
4414 mutex_enter(hash_lock);
4415 mutex_exit(hash_lock);
4416 goto top;
4417 }
4418
4419 if (HDR_L2_WRITE_HEAD(ab)) {
4420 /*
4421 * We hit a write head node. Leave it for
4422 * l2arc_write_done().
4423 */
4424 list_remove(buflist, ab);
4425 mutex_exit(hash_lock);
4426 continue;
4427 }
4428
4429 if (!all && ab->b_l2hdr != NULL &&
4430 (ab->b_l2hdr->b_daddr > taddr ||
4431 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4432 /*
4433 * We've evicted to the target address,
4434 * or the end of the device.
4435 */
4436 mutex_exit(hash_lock);
4437 break;
4438 }
4439
4440 if (HDR_FREE_IN_PROGRESS(ab)) {
4441 /*
4442 * Already on the path to destruction.
4443 */
4444 mutex_exit(hash_lock);
4445 continue;
4446 }
4447
4448 if (ab->b_state == arc_l2c_only) {
4449 ASSERT(!HDR_L2_READING(ab));
4450 /*
4451 * This doesn't exist in the ARC. Destroy.
4452 * arc_hdr_destroy() will call list_remove()
4453 * and decrement arcstat_l2_size.
4454 */
4455 arc_change_state(arc_anon, ab, hash_lock);
4456 arc_hdr_destroy(ab);
4457 } else {
4458 /*
4459 * Invalidate issued or about to be issued
4460 * reads, since we may be about to write
4461 * over this location.
4462 */
4463 if (HDR_L2_READING(ab)) {
4464 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4465 ab->b_flags |= ARC_L2_EVICTED;
4466 }
4467
4468 /*
4469 * Tell ARC this no longer exists in L2ARC.
4470 */
4471 if (ab->b_l2hdr != NULL) {
4472 l2hdr = ab->b_l2hdr;
4473 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
4474 ab->b_l2hdr = NULL;
4475 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
4476 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4477 }
4478 list_remove(buflist, ab);
4479
4480 /*
4481 * This may have been leftover after a
4482 * failed write.
4483 */
4484 ab->b_flags &= ~ARC_L2_WRITING;
4485 }
4486 mutex_exit(hash_lock);
4487 }
4488 mutex_exit(&l2arc_buflist_mtx);
4489
4490 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4491 dev->l2ad_evict = taddr;
4492 }
4493
4494 /*
4495 * Find and write ARC buffers to the L2ARC device.
4496 *
4497 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4498 * for reading until they have completed writing.
4499 * The headroom_boost is an in-out parameter used to maintain headroom boost
4500 * state between calls to this function.
4501 *
4502 * Returns the number of bytes actually written (which may be smaller than
4503 * the delta by which the device hand has changed due to alignment).
4504 */
4505 static uint64_t
4506 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
4507 boolean_t *headroom_boost)
4508 {
4509 arc_buf_hdr_t *ab, *ab_prev, *head;
4510 list_t *list;
4511 uint64_t write_asize, write_psize, write_sz, headroom,
4512 buf_compress_minsz;
4513 void *buf_data;
4514 kmutex_t *list_lock;
4515 boolean_t full;
4516 l2arc_write_callback_t *cb;
4517 zio_t *pio, *wzio;
4518 uint64_t guid = spa_load_guid(spa);
4519 const boolean_t do_headroom_boost = *headroom_boost;
4520 struct defer_write_entry {
4521 arc_buf_hdr_t *dwe_buf;
4522 void *dwe_orig_data;
4523 uint64_t dwe_orig_size;
4524 list_node_t *dwe_node;
4525 } *dwe, *dwe_next;
4526 list_t defer_write_list;
4527
4528 ASSERT(dev->l2ad_vdev != NULL);
4529
4530 /* Lower the flag now, we might want to raise it again later. */
4531 *headroom_boost = B_FALSE;
4532
4533 pio = NULL;
4534 write_sz = write_asize = write_psize = 0;
4535 full = B_FALSE;
4536 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4537 head->b_flags |= ARC_L2_WRITE_HEAD;
4538
4539 /*
4540 * We will want to try to compress buffers that are at least 2x the
4541 * device sector size.
4542 */
4543 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
4544
4545 /*
4546 * Copy buffers for L2ARC writing.
4547 */
4548 list_create(&defer_write_list, sizeof (*dwe),
4549 offsetof(struct defer_write_entry, dwe_node));
4550 mutex_enter(&l2arc_buflist_mtx);
4551 for (int try = 0; try <= 3; try++) {
4552 uint64_t passed_sz = 0;
4553
4554 list = l2arc_list_locked(try, &list_lock);
4555
4556 /*
4557 * L2ARC fast warmup.
4558 *
4559 * Until the ARC is warm and starts to evict, read from the
4560 * head of the ARC lists rather than the tail.
4561 */
4562 if (arc_warm == B_FALSE)
4563 ab = list_head(list);
4564 else
4565 ab = list_tail(list);
4566
4567 headroom = target_sz * l2arc_headroom;
4568 if (do_headroom_boost)
4569 headroom = (headroom * l2arc_headroom_boost) / 100;
4570
4571 for (; ab; ab = ab_prev) {
4572 l2arc_buf_hdr_t *l2hdr;
4573 kmutex_t *hash_lock;
4574
4575 if (arc_warm == B_FALSE)
4576 ab_prev = list_next(list, ab);
4577 else
4578 ab_prev = list_prev(list, ab);
4579
4580 hash_lock = HDR_LOCK(ab);
4581 if (!mutex_tryenter(hash_lock)) {
4582 /*
4583 * Skip this buffer rather than waiting.
4584 */
4585 continue;
4586 }
4587
4588 passed_sz += ab->b_size;
4589 if (passed_sz > headroom) {
4590 /*
4591 * Searched too far.
4592 */
4593 mutex_exit(hash_lock);
4594 break;
4595 }
4596
4597 if (!l2arc_write_eligible(guid, ab)) {
4598 mutex_exit(hash_lock);
4599 continue;
4600 }
4601
4602 if ((write_sz + ab->b_size) > target_sz) {
4603 full = B_TRUE;
4604 mutex_exit(hash_lock);
4605 break;
4606 }
4607
4608 if (pio == NULL) {
4609 /*
4610 * Insert a dummy header on the buflist so
4611 * l2arc_write_done() can find where the
4612 * write buffers begin without searching.
4613 */
4614 list_insert_head(dev->l2ad_buflist, head);
4615
4616 cb = kmem_alloc(
4617 sizeof (l2arc_write_callback_t), KM_SLEEP);
4618 cb->l2wcb_dev = dev;
4619 cb->l2wcb_head = head;
4620 pio = zio_root(spa, l2arc_write_done, cb,
4621 ZIO_FLAG_CANFAIL);
4622 }
4623
4624 /*
4625 * Create and add a new L2ARC header.
4626 */
4627 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4628 l2hdr->b_dev = dev;
4629 ab->b_flags |= ARC_L2_WRITING;
4630 l2hdr->b_compress = ZIO_COMPRESS_OFF;
4631 l2hdr->b_asize = ab->b_size;
4632
4633 /*
4634 * Temporarily stash the buffer in defer_write_entries.
4635 * The subsequent write step will pick it up from
4636 * there. This is because we can't access ab->b_buf
4637 * without holding the hash_lock, which we in turn
4638 * can't access without holding the ARC list locks
4639 * while walking the ARC lists (we want to avoid
4640 * holding these locks during compression/writing).
4641 */
4642 dwe = kmem_alloc(sizeof (*dwe), KM_SLEEP);
4643 dwe->dwe_buf = ab;
4644 dwe->dwe_orig_data = ab->b_buf->b_data;
4645 dwe->dwe_orig_size = ab->b_size;
4646
4647 ab->b_l2hdr = l2hdr;
4648
4649 list_insert_head(dev->l2ad_buflist, ab);
4650 list_insert_tail(&defer_write_list, dwe);
4651
4652 /*
4653 * Compute and store the buffer cksum before
4654 * writing. On debug the cksum is verified first.
4655 */
4656 arc_cksum_verify(ab->b_buf);
4657 arc_cksum_compute(ab->b_buf, B_TRUE);
4658
4659 mutex_exit(hash_lock);
4660
4661 write_sz += dwe->dwe_orig_size;
4662 }
4663
4664 mutex_exit(list_lock);
4665
4666 if (full == B_TRUE)
4667 break;
4668 }
4669
4670 /* No buffers selected for writing? */
4671 if (pio == NULL) {
4672 ASSERT0(write_sz);
4673 mutex_exit(&l2arc_buflist_mtx);
4674 kmem_cache_free(hdr_cache, head);
4675 list_destroy(&defer_write_list);
4676 return (0);
4677 }
4678
4679 mutex_exit(&l2arc_buflist_mtx);
4680
4681 /*
4682 * Now start writing the buffers. We're starting at the write head
4683 * and work backwards, retracing the course of the buffer selector
4684 * loop above.
4685 */
4686 for (dwe = list_head(&defer_write_list); dwe != NULL; dwe = dwe_next) {
4687 l2arc_buf_hdr_t *l2hdr;
4688 uint64_t buf_sz;
4689
4690 dwe_next = list_next(&defer_write_list, dwe);
4691 ab = dwe->dwe_buf;
4692
4693 /*
4694 * Accessing ab->b_l2hdr without locking is safe here because
4695 * we're holding the l2arc_buflist_mtx and no other thread will
4696 * ever directly modify the L2 fields. In particular ab->b_buf
4697 * may be invalid by now due to ARC eviction.
4698 */
4699 l2hdr = ab->b_l2hdr;
4700 l2hdr->b_daddr = dev->l2ad_hand;
4701
4702 if ((ab->b_flags & ARC_L2COMPRESS) &&
4703 l2hdr->b_asize >= buf_compress_minsz &&
4704 l2arc_compress_buf(dwe->dwe_orig_data, dwe->dwe_orig_size,
4705 &buf_data, &buf_sz, &l2hdr->b_compress)) {
4706 /*
4707 * If compression succeeded, enable headroom
4708 * boost on the next scan cycle.
4709 */
4710 *headroom_boost = B_TRUE;
4711 l2hdr->b_asize = buf_sz;
4712 } else {
4713 buf_data = dwe->dwe_orig_data;
4714 buf_sz = dwe->dwe_orig_size;
4715 l2hdr->b_asize = dwe->dwe_orig_size;
4716 }
4717
4718 /* Compression may have squashed the buffer to zero length. */
4719 if (buf_sz != 0) {
4720 uint64_t buf_p_sz;
4721
4722 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4723 dev->l2ad_hand, l2hdr->b_asize, buf_data,
4724 ZIO_CHECKSUM_OFF, NULL, NULL,
4725 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL,
4726 B_FALSE);
4727
4728 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4729 zio_t *, wzio);
4730 (void) zio_nowait(wzio);
4731
4732 write_asize += l2hdr->b_asize;
4733 /*
4734 * Keep the clock hand suitably device-aligned.
4735 */
4736 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4737 write_psize += buf_p_sz;
4738 dev->l2ad_hand += buf_p_sz;
4739 }
4740
4741 list_remove(&defer_write_list, dwe);
4742 kmem_free(dwe, sizeof (*dwe));
4743 }
4744
4745 list_destroy(&defer_write_list);
4746
4747 ASSERT3U(write_asize, <=, target_sz);
4748 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4749 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
4750 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4751 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
4752 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
4753
4754 /*
4755 * Bump device hand to the device start if it is approaching the end.
4756 * l2arc_evict() will already have evicted ahead for this case.
4757 */
4758 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4759 vdev_space_update(dev->l2ad_vdev,
4760 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4761 dev->l2ad_hand = dev->l2ad_start;
4762 dev->l2ad_evict = dev->l2ad_start;
4763 dev->l2ad_first = B_FALSE;
4764 }
4765
4766 dev->l2ad_writing = B_TRUE;
4767 (void) zio_wait(pio);
4768 dev->l2ad_writing = B_FALSE;
4769
4770 return (write_asize);
4771 }
4772
4773 /*
4774 * Compresses an L2ARC buffer.
4775 * The data to be compressed is in in_data and its size in in_sz. This routine
4776 * tries to compress the data and depending on the compression result there
4777 * are three possible outcomes:
4778 * *) The buffer was incompressible. The function returns with B_FALSE and
4779 * does nothing else.
4780 * *) The buffer was all-zeros, so there is no need to write it to an L2
4781 * device. To indicate this situation, the *out_data is set to NULL,
4782 * *out_sz is set to zero, *compress is set to ZIO_COMPRESS_EMPTY and
4783 * the function returns B_TRUE.
4784 * *) Compression succeeded and *out_data was set to point to a buffer holding
4785 * the compressed data buffer, *out_sz was set to indicate the output size,
4786 * *compress was set to the appropriate compression algorithm and B_TRUE is
4787 * returned. Once writing is done the buffer will be automatically freed by
4788 * l2arc_do_free_on_write().
4789 */
4790 static boolean_t
4791 l2arc_compress_buf(void *in_data, uint64_t in_sz, void **out_data,
4792 uint64_t *out_sz, enum zio_compress *compress)
4793 {
4794 void *cdata;
4795
4796 cdata = zio_data_buf_alloc(in_sz);
4797 *out_sz = zio_compress_data(ZIO_COMPRESS_LZ4, in_data, cdata, in_sz);
4798
4799 if (*out_sz == 0) {
4800 /* Zero block, indicate that there's nothing to write. */
4801 zio_data_buf_free(cdata, in_sz);
4802 *compress = ZIO_COMPRESS_EMPTY;
4803 *out_data = NULL;
4804 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
4805 return (B_TRUE);
4806 } else if (*out_sz > 0 && *out_sz < in_sz) {
4807 /*
4808 * Compression succeeded, we'll keep the cdata around for
4809 * writing and release it after writing.
4810 */
4811 l2arc_data_free_t *df;
4812
4813 *compress = ZIO_COMPRESS_LZ4;
4814 *out_data = cdata;
4815
4816 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
4817 df->l2df_data = cdata;
4818 df->l2df_size = *out_sz;
4819 df->l2df_func = zio_data_buf_free;
4820 mutex_enter(&l2arc_free_on_write_mtx);
4821 list_insert_head(l2arc_free_on_write, df);
4822 mutex_exit(&l2arc_free_on_write_mtx);
4823
4824 ARCSTAT_BUMP(arcstat_l2_compress_successes);
4825 ARCSTAT_BUMP(arcstat_l2_free_on_write);
4826 return (B_TRUE);
4827 } else {
4828 /*
4829 * Compression failed, release the compressed buffer.
4830 */
4831 zio_data_buf_free(cdata, in_sz);
4832 ARCSTAT_BUMP(arcstat_l2_compress_failures);
4833 return (B_FALSE);
4834 }
4835 }
4836
4837 /*
4838 * Decompresses a zio read back from an l2arc device. On success, the
4839 * underlying zio's io_data buffer is overwritten by the uncompressed
4840 * version. On decompression error (corrupt compressed stream), the
4841 * zio->io_error value is set to signal an I/O error.
4842 *
4843 * Please note that the compressed data stream is not checksummed, so
4844 * if the underlying device is experiencing data corruption, we may feed
4845 * corrupt data to the decompressor, so the decompressor needs to be
4846 * able to handle this situation (LZ4 does).
4847 */
4848 static void
4849 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
4850 {
4851 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
4852
4853 if (zio->io_error != 0) {
4854 /*
4855 * An io error has occured, just restore the original io
4856 * size in preparation for a main pool read.
4857 */
4858 zio->io_orig_size = zio->io_size = hdr->b_size;
4859 return;
4860 }
4861
4862 if (c == ZIO_COMPRESS_EMPTY) {
4863 /*
4864 * An empty buffer results in a null zio, which means we
4865 * need to fill its io_data after we're done restoring the
4866 * buffer's contents.
4867 */
4868 ASSERT(hdr->b_buf != NULL);
4869 bzero(hdr->b_buf->b_data, hdr->b_size);
4870 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data;
4871 } else {
4872 ASSERT(zio->io_data != NULL);
4873 /*
4874 * We copy the compressed data from the start of the arc buffer
4875 * (the zio_read will have pulled in only what we need, the
4876 * rest is garbage which we will overwrite at decompression)
4877 * and then decompress back to the ARC data buffer. This way we
4878 * can minimize copying by simply decompressing back over the
4879 * original compressed data (rather than decompressing to an
4880 * aux buffer and then copying back the uncompressed buffer,
4881 * which is likely to be much larger).
4882 */
4883 uint64_t csize;
4884 void *cdata;
4885
4886 csize = zio->io_size;
4887 cdata = zio_data_buf_alloc(csize);
4888 bcopy(zio->io_data, cdata, csize);
4889 if (zio_decompress_data(c, cdata, zio->io_data, csize,
4890 hdr->b_size) != 0)
4891 zio->io_error = EIO;
4892 zio_data_buf_free(cdata, csize);
4893 }
4894
4895 /* Restore the expected uncompressed IO size. */
4896 zio->io_orig_size = zio->io_size = hdr->b_size;
4897 }
4898
4899 /*
4900 * This thread feeds the L2ARC at regular intervals. This is the beating
4901 * heart of the L2ARC.
4902 */
4903 static void
4904 l2arc_feed_thread(void)
4905 {
4906 callb_cpr_t cpr;
4907 l2arc_dev_t *dev;
4908 spa_t *spa;
4909 uint64_t size, wrote;
4910 clock_t begin, next = ddi_get_lbolt();
4911 boolean_t headroom_boost = B_FALSE;
4912
4913 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4914
4915 mutex_enter(&l2arc_feed_thr_lock);
4916
4917 while (l2arc_thread_exit == 0) {
4918 CALLB_CPR_SAFE_BEGIN(&cpr);
4919 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4920 next);
4921 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4922 next = ddi_get_lbolt() + hz;
4923
4924 /*
4925 * Quick check for L2ARC devices.
4926 */
4927 mutex_enter(&l2arc_dev_mtx);
4928 if (l2arc_ndev == 0) {
4929 mutex_exit(&l2arc_dev_mtx);
4930 continue;
4931 }
4932 mutex_exit(&l2arc_dev_mtx);
4933 begin = ddi_get_lbolt();
4934
4935 /*
4936 * This selects the next l2arc device to write to, and in
4937 * doing so the next spa to feed from: dev->l2ad_spa. This
4938 * will return NULL if there are now no l2arc devices or if
4939 * they are all faulted.
4940 *
4941 * If a device is returned, its spa's config lock is also
4942 * held to prevent device removal. l2arc_dev_get_next()
4943 * will grab and release l2arc_dev_mtx.
4944 */
4945 if ((dev = l2arc_dev_get_next()) == NULL)
4946 continue;
4947
4948 spa = dev->l2ad_spa;
4949 ASSERT(spa != NULL);
4950
4951 /*
4952 * If the pool is read-only then force the feed thread to
4953 * sleep a little longer.
4954 */
4955 if (!spa_writeable(spa)) {
4956 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4957 spa_config_exit(spa, SCL_L2ARC, dev);
4958 continue;
4959 }
4960
4961 /*
4962 * Avoid contributing to memory pressure.
4963 */
4964 if (arc_reclaim_needed()) {
4965 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4966 spa_config_exit(spa, SCL_L2ARC, dev);
4967 continue;
4968 }
4969
4970 ARCSTAT_BUMP(arcstat_l2_feeds);
4971
4972 size = l2arc_write_size();
4973
4974 /*
4975 * Evict L2ARC buffers that will be overwritten.
4976 */
4977 l2arc_evict(dev, size, B_FALSE);
4978
4979 /*
4980 * Write ARC buffers.
4981 */
4982 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
4983
4984 /*
4985 * Calculate interval between writes.
4986 */
4987 next = l2arc_write_interval(begin, size, wrote);
4988 spa_config_exit(spa, SCL_L2ARC, dev);
4989 }
4990
4991 l2arc_thread_exit = 0;
4992 cv_broadcast(&l2arc_feed_thr_cv);
4993 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4994 thread_exit();
4995 }
4996
4997 boolean_t
4998 l2arc_vdev_present(vdev_t *vd)
4999 {
5000 l2arc_dev_t *dev;
5001
5002 mutex_enter(&l2arc_dev_mtx);
5003 for (dev = list_head(l2arc_dev_list); dev != NULL;
5004 dev = list_next(l2arc_dev_list, dev)) {
5005 if (dev->l2ad_vdev == vd)
5006 break;
5007 }
5008 mutex_exit(&l2arc_dev_mtx);
5009
5010 return (dev != NULL);
5011 }
5012
5013 /*
5014 * Add a vdev for use by the L2ARC. By this point the spa has already
5015 * validated the vdev and opened it.
5016 */
5017 void
5018 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
5019 {
5020 l2arc_dev_t *adddev;
5021
5022 ASSERT(!l2arc_vdev_present(vd));
5023
5024 /*
5025 * Create a new l2arc device entry.
5026 */
5027 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
5028 adddev->l2ad_spa = spa;
5029 adddev->l2ad_vdev = vd;
5030 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
5031 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
5032 adddev->l2ad_hand = adddev->l2ad_start;
5033 adddev->l2ad_evict = adddev->l2ad_start;
5034 adddev->l2ad_first = B_TRUE;
5035 adddev->l2ad_writing = B_FALSE;
5036
5037 /*
5038 * This is a list of all ARC buffers that are still valid on the
5039 * device.
5040 */
5041 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
5042 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
5043 offsetof(arc_buf_hdr_t, b_l2node));
5044
5045 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
5046
5047 /*
5048 * Add device to global list
5049 */
5050 mutex_enter(&l2arc_dev_mtx);
5051 list_insert_head(l2arc_dev_list, adddev);
5052 atomic_inc_64(&l2arc_ndev);
5053 mutex_exit(&l2arc_dev_mtx);
5054 }
5055
5056 /*
5057 * Remove a vdev from the L2ARC.
5058 */
5059 void
5060 l2arc_remove_vdev(vdev_t *vd)
5061 {
5062 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
5063
5064 /*
5065 * Find the device by vdev
5066 */
5067 mutex_enter(&l2arc_dev_mtx);
5068 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
5069 nextdev = list_next(l2arc_dev_list, dev);
5070 if (vd == dev->l2ad_vdev) {
5071 remdev = dev;
5072 break;
5073 }
5074 }
5075 ASSERT(remdev != NULL);
5076
5077 /*
5078 * Remove device from global list
5079 */
5080 list_remove(l2arc_dev_list, remdev);
5081 l2arc_dev_last = NULL; /* may have been invalidated */
5082 atomic_dec_64(&l2arc_ndev);
5083 mutex_exit(&l2arc_dev_mtx);
5084
5085 /*
5086 * Clear all buflists and ARC references. L2ARC device flush.
5087 */
5088 l2arc_evict(remdev, 0, B_TRUE);
5089 list_destroy(remdev->l2ad_buflist);
5090 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5091 kmem_free(remdev, sizeof (l2arc_dev_t));
5092 }
5093
5094 void
5095 l2arc_init(void)
5096 {
5097 l2arc_thread_exit = 0;
5098 l2arc_ndev = 0;
5099 l2arc_writes_sent = 0;
5100 l2arc_writes_done = 0;
5101
5102 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5103 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5104 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5105 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5106 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5107
5108 l2arc_dev_list = &L2ARC_dev_list;
5109 l2arc_free_on_write = &L2ARC_free_on_write;
5110 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5111 offsetof(l2arc_dev_t, l2ad_node));
5112 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5113 offsetof(l2arc_data_free_t, l2df_list_node));
5114 }
5115
5116 void
5117 l2arc_fini(void)
5118 {
5119 /*
5120 * This is called from dmu_fini(), which is called from spa_fini();
5121 * Because of this, we can assume that all l2arc devices have
5122 * already been removed when the pools themselves were removed.
5123 */
5124
5125 l2arc_do_free_on_write();
5126
5127 mutex_destroy(&l2arc_feed_thr_lock);
5128 cv_destroy(&l2arc_feed_thr_cv);
5129 mutex_destroy(&l2arc_dev_mtx);
5130 mutex_destroy(&l2arc_buflist_mtx);
5131 mutex_destroy(&l2arc_free_on_write_mtx);
5132
5133 list_destroy(l2arc_dev_list);
5134 list_destroy(l2arc_free_on_write);
5135 }
5136
5137 void
5138 l2arc_start(void)
5139 {
5140 if (!(spa_mode_global & FWRITE))
5141 return;
5142
5143 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5144 TS_RUN, minclsyspri);
5145 }
5146
5147 void
5148 l2arc_stop(void)
5149 {
5150 if (!(spa_mode_global & FWRITE))
5151 return;
5152
5153 mutex_enter(&l2arc_feed_thr_lock);
5154 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5155 l2arc_thread_exit = 1;
5156 while (l2arc_thread_exit != 0)
5157 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5158 mutex_exit(&l2arc_feed_thr_lock);
5159 }