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
122 #include <sys/spa.h>
123 #include <sys/zio.h>
124 #include <sys/zio_compress.h>
125 #include <sys/zfs_context.h>
126 #include <sys/arc.h>
127 #include <sys/refcount.h>
128 #include <sys/vdev.h>
129 #include <sys/vdev_impl.h>
130 #include <sys/dsl_pool.h>
131 #ifdef _KERNEL
132 #include <sys/vmsystm.h>
133 #include <vm/anon.h>
134 #include <sys/fs/swapnode.h>
135 #include <sys/dnlc.h>
136 #endif
137 #include <sys/callb.h>
138 #include <sys/kstat.h>
139 #include <zfs_fletcher.h>
140
141 #ifndef _KERNEL
142 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
143 boolean_t arc_watch = B_FALSE;
144 int arc_procfd;
145 #endif
146
147 static kmutex_t arc_reclaim_thr_lock;
148 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
149 static uint8_t arc_thread_exit;
150
151 #define ARC_REDUCE_DNLC_PERCENT 3
152 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
153
154 typedef enum arc_reclaim_strategy {
155 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
156 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
157 } arc_reclaim_strategy_t;
158
159 /*
160 * The number of iterations through arc_evict_*() before we
161 * drop & reacquire the lock.
162 */
163 int arc_evict_iterations = 100;
164
165 /* number of seconds before growing cache again */
166 static int arc_grow_retry = 60;
167
168 /* shift of arc_c for calculating both min and max arc_p */
169 static int arc_p_min_shift = 4;
170
171 /* log2(fraction of arc to reclaim) */
172 static int arc_shrink_shift = 5;
173
174 /*
175 * minimum lifespan of a prefetch block in clock ticks
176 * (initialized in arc_init())
177 */
178 static int arc_min_prefetch_lifespan;
179
180 /*
181 * If this percent of memory is free, don't throttle.
182 */
183 int arc_lotsfree_percent = 10;
184
185 static int arc_dead;
186
187 /*
188 * The arc has filled available memory and has now warmed up.
189 */
190 static boolean_t arc_warm;
191
192 /*
193 * These tunables are for performance analysis.
194 */
195 uint64_t zfs_arc_max;
196 uint64_t zfs_arc_min;
197 uint64_t zfs_arc_meta_limit = 0;
198 int zfs_arc_grow_retry = 0;
199 int zfs_arc_shrink_shift = 0;
200 int zfs_arc_p_min_shift = 0;
201 int zfs_disable_dup_eviction = 0;
202
203 /*
204 * Note that buffers can be in one of 6 states:
205 * ARC_anon - anonymous (discussed below)
206 * ARC_mru - recently used, currently cached
207 * ARC_mru_ghost - recentely used, no longer in cache
208 * ARC_mfu - frequently used, currently cached
209 * ARC_mfu_ghost - frequently used, no longer in cache
210 * ARC_l2c_only - exists in L2ARC but not other states
211 * When there are no active references to the buffer, they are
212 * are linked onto a list in one of these arc states. These are
213 * the only buffers that can be evicted or deleted. Within each
214 * state there are multiple lists, one for meta-data and one for
215 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
216 * etc.) is tracked separately so that it can be managed more
217 * explicitly: favored over data, limited explicitly.
218 *
219 * Anonymous buffers are buffers that are not associated with
220 * a DVA. These are buffers that hold dirty block copies
221 * before they are written to stable storage. By definition,
222 * they are "ref'd" and are considered part of arc_mru
223 * that cannot be freed. Generally, they will aquire a DVA
224 * as they are written and migrate onto the arc_mru list.
225 *
226 * The ARC_l2c_only state is for buffers that are in the second
227 * level ARC but no longer in any of the ARC_m* lists. The second
228 * level ARC itself may also contain buffers that are in any of
229 * the ARC_m* states - meaning that a buffer can exist in two
230 * places. The reason for the ARC_l2c_only state is to keep the
231 * buffer header in the hash table, so that reads that hit the
232 * second level ARC benefit from these fast lookups.
233 */
234
235 typedef struct arc_state {
236 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
237 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
238 uint64_t arcs_size; /* total amount of data in this state */
239 kmutex_t arcs_mtx;
240 } arc_state_t;
241
242 /* The 6 states: */
243 static arc_state_t ARC_anon;
244 static arc_state_t ARC_mru;
245 static arc_state_t ARC_mru_ghost;
246 static arc_state_t ARC_mfu;
247 static arc_state_t ARC_mfu_ghost;
248 static arc_state_t ARC_l2c_only;
249
250 typedef struct arc_stats {
251 kstat_named_t arcstat_hits;
252 kstat_named_t arcstat_misses;
253 kstat_named_t arcstat_demand_data_hits;
254 kstat_named_t arcstat_demand_data_misses;
255 kstat_named_t arcstat_demand_metadata_hits;
256 kstat_named_t arcstat_demand_metadata_misses;
257 kstat_named_t arcstat_prefetch_data_hits;
258 kstat_named_t arcstat_prefetch_data_misses;
259 kstat_named_t arcstat_prefetch_metadata_hits;
260 kstat_named_t arcstat_prefetch_metadata_misses;
261 kstat_named_t arcstat_mru_hits;
262 kstat_named_t arcstat_mru_ghost_hits;
263 kstat_named_t arcstat_mfu_hits;
264 kstat_named_t arcstat_mfu_ghost_hits;
265 kstat_named_t arcstat_deleted;
266 kstat_named_t arcstat_recycle_miss;
267 /*
268 * Number of buffers that could not be evicted because the hash lock
269 * was held by another thread. The lock may not necessarily be held
270 * by something using the same buffer, since hash locks are shared
271 * by multiple buffers.
272 */
273 kstat_named_t arcstat_mutex_miss;
274 /*
275 * Number of buffers skipped because they have I/O in progress, are
276 * indrect prefetch buffers that have not lived long enough, or are
277 * not from the spa we're trying to evict from.
278 */
279 kstat_named_t arcstat_evict_skip;
280 kstat_named_t arcstat_evict_l2_cached;
281 kstat_named_t arcstat_evict_l2_eligible;
282 kstat_named_t arcstat_evict_l2_ineligible;
283 kstat_named_t arcstat_hash_elements;
284 kstat_named_t arcstat_hash_elements_max;
285 kstat_named_t arcstat_hash_collisions;
286 kstat_named_t arcstat_hash_chains;
287 kstat_named_t arcstat_hash_chain_max;
288 kstat_named_t arcstat_p;
289 kstat_named_t arcstat_c;
290 kstat_named_t arcstat_c_min;
291 kstat_named_t arcstat_c_max;
292 kstat_named_t arcstat_size;
293 kstat_named_t arcstat_hdr_size;
294 kstat_named_t arcstat_data_size;
295 kstat_named_t arcstat_other_size;
296 kstat_named_t arcstat_l2_hits;
297 kstat_named_t arcstat_l2_misses;
298 kstat_named_t arcstat_l2_feeds;
299 kstat_named_t arcstat_l2_rw_clash;
300 kstat_named_t arcstat_l2_read_bytes;
301 kstat_named_t arcstat_l2_write_bytes;
302 kstat_named_t arcstat_l2_writes_sent;
303 kstat_named_t arcstat_l2_writes_done;
304 kstat_named_t arcstat_l2_writes_error;
305 kstat_named_t arcstat_l2_writes_hdr_miss;
306 kstat_named_t arcstat_l2_evict_lock_retry;
307 kstat_named_t arcstat_l2_evict_reading;
308 kstat_named_t arcstat_l2_free_on_write;
309 kstat_named_t arcstat_l2_abort_lowmem;
310 kstat_named_t arcstat_l2_cksum_bad;
311 kstat_named_t arcstat_l2_io_error;
312 kstat_named_t arcstat_l2_size;
313 kstat_named_t arcstat_l2_asize;
314 kstat_named_t arcstat_l2_hdr_size;
315 kstat_named_t arcstat_l2_compress_successes;
316 kstat_named_t arcstat_l2_compress_zeros;
317 kstat_named_t arcstat_l2_compress_failures;
318 kstat_named_t arcstat_memory_throttle_count;
319 kstat_named_t arcstat_duplicate_buffers;
320 kstat_named_t arcstat_duplicate_buffers_size;
321 kstat_named_t arcstat_duplicate_reads;
322 kstat_named_t arcstat_meta_used;
323 kstat_named_t arcstat_meta_limit;
324 kstat_named_t arcstat_meta_max;
325 } arc_stats_t;
326
327 static arc_stats_t arc_stats = {
328 { "hits", KSTAT_DATA_UINT64 },
329 { "misses", KSTAT_DATA_UINT64 },
330 { "demand_data_hits", KSTAT_DATA_UINT64 },
331 { "demand_data_misses", KSTAT_DATA_UINT64 },
332 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
333 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
334 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
335 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
336 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
337 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
338 { "mru_hits", KSTAT_DATA_UINT64 },
339 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
340 { "mfu_hits", KSTAT_DATA_UINT64 },
341 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
342 { "deleted", KSTAT_DATA_UINT64 },
343 { "recycle_miss", KSTAT_DATA_UINT64 },
344 { "mutex_miss", KSTAT_DATA_UINT64 },
345 { "evict_skip", KSTAT_DATA_UINT64 },
346 { "evict_l2_cached", KSTAT_DATA_UINT64 },
347 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
348 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
349 { "hash_elements", KSTAT_DATA_UINT64 },
350 { "hash_elements_max", KSTAT_DATA_UINT64 },
351 { "hash_collisions", KSTAT_DATA_UINT64 },
352 { "hash_chains", KSTAT_DATA_UINT64 },
353 { "hash_chain_max", KSTAT_DATA_UINT64 },
354 { "p", KSTAT_DATA_UINT64 },
355 { "c", KSTAT_DATA_UINT64 },
356 { "c_min", KSTAT_DATA_UINT64 },
357 { "c_max", KSTAT_DATA_UINT64 },
358 { "size", KSTAT_DATA_UINT64 },
359 { "hdr_size", KSTAT_DATA_UINT64 },
360 { "data_size", KSTAT_DATA_UINT64 },
361 { "other_size", KSTAT_DATA_UINT64 },
362 { "l2_hits", KSTAT_DATA_UINT64 },
363 { "l2_misses", KSTAT_DATA_UINT64 },
364 { "l2_feeds", KSTAT_DATA_UINT64 },
365 { "l2_rw_clash", KSTAT_DATA_UINT64 },
366 { "l2_read_bytes", KSTAT_DATA_UINT64 },
367 { "l2_write_bytes", KSTAT_DATA_UINT64 },
368 { "l2_writes_sent", KSTAT_DATA_UINT64 },
369 { "l2_writes_done", KSTAT_DATA_UINT64 },
370 { "l2_writes_error", KSTAT_DATA_UINT64 },
371 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
372 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
373 { "l2_evict_reading", KSTAT_DATA_UINT64 },
374 { "l2_free_on_write", KSTAT_DATA_UINT64 },
375 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
376 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
377 { "l2_io_error", KSTAT_DATA_UINT64 },
378 { "l2_size", KSTAT_DATA_UINT64 },
379 { "l2_asize", KSTAT_DATA_UINT64 },
380 { "l2_hdr_size", KSTAT_DATA_UINT64 },
381 { "l2_compress_successes", KSTAT_DATA_UINT64 },
382 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
383 { "l2_compress_failures", KSTAT_DATA_UINT64 },
384 { "memory_throttle_count", KSTAT_DATA_UINT64 },
385 { "duplicate_buffers", KSTAT_DATA_UINT64 },
386 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
387 { "duplicate_reads", KSTAT_DATA_UINT64 },
388 { "arc_meta_used", KSTAT_DATA_UINT64 },
389 { "arc_meta_limit", KSTAT_DATA_UINT64 },
390 { "arc_meta_max", KSTAT_DATA_UINT64 }
391 };
392
393 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
394
395 #define ARCSTAT_INCR(stat, val) \
396 atomic_add_64(&arc_stats.stat.value.ui64, (val))
397
398 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
399 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
400
401 #define ARCSTAT_MAX(stat, val) { \
402 uint64_t m; \
403 while ((val) > (m = arc_stats.stat.value.ui64) && \
404 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
405 continue; \
406 }
407
408 #define ARCSTAT_MAXSTAT(stat) \
409 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
410
411 /*
412 * We define a macro to allow ARC hits/misses to be easily broken down by
413 * two separate conditions, giving a total of four different subtypes for
414 * each of hits and misses (so eight statistics total).
415 */
416 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
417 if (cond1) { \
418 if (cond2) { \
419 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
420 } else { \
421 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
422 } \
423 } else { \
424 if (cond2) { \
425 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
426 } else { \
427 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
428 } \
429 }
430
431 kstat_t *arc_ksp;
432 static arc_state_t *arc_anon;
433 static arc_state_t *arc_mru;
434 static arc_state_t *arc_mru_ghost;
435 static arc_state_t *arc_mfu;
436 static arc_state_t *arc_mfu_ghost;
437 static arc_state_t *arc_l2c_only;
438
439 /*
440 * There are several ARC variables that are critical to export as kstats --
441 * but we don't want to have to grovel around in the kstat whenever we wish to
442 * manipulate them. For these variables, we therefore define them to be in
443 * terms of the statistic variable. This assures that we are not introducing
444 * the possibility of inconsistency by having shadow copies of the variables,
445 * while still allowing the code to be readable.
446 */
447 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
448 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
449 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
450 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
451 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
452 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
453 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
454 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
455
456 #define L2ARC_IS_VALID_COMPRESS(_c_) \
457 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
458
459 static int arc_no_grow; /* Don't try to grow cache size */
460 static uint64_t arc_tempreserve;
461 static uint64_t arc_loaned_bytes;
462
463 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
464
465 typedef struct arc_callback arc_callback_t;
466
467 struct arc_callback {
468 void *acb_private;
469 arc_done_func_t *acb_done;
470 arc_buf_t *acb_buf;
471 zio_t *acb_zio_dummy;
472 arc_callback_t *acb_next;
473 };
474
475 typedef struct arc_write_callback arc_write_callback_t;
476
477 struct arc_write_callback {
478 void *awcb_private;
479 arc_done_func_t *awcb_ready;
480 arc_done_func_t *awcb_physdone;
481 arc_done_func_t *awcb_done;
482 arc_buf_t *awcb_buf;
483 };
484
485 struct arc_buf_hdr {
486 /* protected by hash lock */
487 dva_t b_dva;
488 uint64_t b_birth;
489 uint64_t b_cksum0;
490
491 kmutex_t b_freeze_lock;
492 zio_cksum_t *b_freeze_cksum;
493 void *b_thawed;
494
495 arc_buf_hdr_t *b_hash_next;
496 arc_buf_t *b_buf;
497 uint32_t b_flags;
498 uint32_t b_datacnt;
499
500 arc_callback_t *b_acb;
501 kcondvar_t b_cv;
502
503 /* immutable */
504 arc_buf_contents_t b_type;
505 uint64_t b_size;
506 uint64_t b_spa;
507
508 /* protected by arc state mutex */
509 arc_state_t *b_state;
510 list_node_t b_arc_node;
511
512 /* updated atomically */
513 clock_t b_arc_access;
514
515 /* self protecting */
516 refcount_t b_refcnt;
517
518 l2arc_buf_hdr_t *b_l2hdr;
519 list_node_t b_l2node;
520 };
521
522 static arc_buf_t *arc_eviction_list;
523 static kmutex_t arc_eviction_mtx;
524 static arc_buf_hdr_t arc_eviction_hdr;
525 static void arc_get_data_buf(arc_buf_t *buf);
526 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
527 static int arc_evict_needed(arc_buf_contents_t type);
528 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
529 static void arc_buf_watch(arc_buf_t *buf);
530
531 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
532
533 #define GHOST_STATE(state) \
534 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
535 (state) == arc_l2c_only)
536
537 /*
538 * Private ARC flags. These flags are private ARC only flags that will show up
539 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
540 * be passed in as arc_flags in things like arc_read. However, these flags
541 * should never be passed and should only be set by ARC code. When adding new
542 * public flags, make sure not to smash the private ones.
543 */
544
545 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
546 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
547 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
548 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
549 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
550 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
551 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
552 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
553 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
554 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
555
556 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
557 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
558 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
559 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
560 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
561 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
562 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
563 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
564 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
565 (hdr)->b_l2hdr != NULL)
566 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
567 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
568 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
569
570 /*
571 * Other sizes
572 */
573
574 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
575 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
576
577 /*
578 * Hash table routines
579 */
580
581 #define HT_LOCK_PAD 64
582
583 struct ht_lock {
584 kmutex_t ht_lock;
585 #ifdef _KERNEL
586 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
587 #endif
588 };
589
590 #define BUF_LOCKS 256
591 typedef struct buf_hash_table {
592 uint64_t ht_mask;
593 arc_buf_hdr_t **ht_table;
594 struct ht_lock ht_locks[BUF_LOCKS];
595 } buf_hash_table_t;
596
597 static buf_hash_table_t buf_hash_table;
598
599 #define BUF_HASH_INDEX(spa, dva, birth) \
600 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
601 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
602 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
603 #define HDR_LOCK(hdr) \
604 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
605
606 uint64_t zfs_crc64_table[256];
607
608 /*
609 * Level 2 ARC
610 */
611
612 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
613 #define L2ARC_HEADROOM 2 /* num of writes */
614 /*
615 * If we discover during ARC scan any buffers to be compressed, we boost
616 * our headroom for the next scanning cycle by this percentage multiple.
617 */
618 #define L2ARC_HEADROOM_BOOST 200
619 #define L2ARC_FEED_SECS 1 /* caching interval secs */
620 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
621
622 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
623 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
624
625 /* L2ARC Performance Tunables */
626 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
627 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
628 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
629 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
630 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
631 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
632 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
633 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
634 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
635
636 /*
637 * L2ARC Internals
638 */
639 typedef struct l2arc_dev {
640 vdev_t *l2ad_vdev; /* vdev */
641 spa_t *l2ad_spa; /* spa */
642 uint64_t l2ad_hand; /* next write location */
643 uint64_t l2ad_start; /* first addr on device */
644 uint64_t l2ad_end; /* last addr on device */
645 uint64_t l2ad_evict; /* last addr eviction reached */
646 boolean_t l2ad_first; /* first sweep through */
647 boolean_t l2ad_writing; /* currently writing */
648 list_t *l2ad_buflist; /* buffer list */
649 list_node_t l2ad_node; /* device list node */
650 } l2arc_dev_t;
651
652 static list_t L2ARC_dev_list; /* device list */
653 static list_t *l2arc_dev_list; /* device list pointer */
654 static kmutex_t l2arc_dev_mtx; /* device list mutex */
655 static l2arc_dev_t *l2arc_dev_last; /* last device used */
656 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
657 static list_t L2ARC_free_on_write; /* free after write buf list */
658 static list_t *l2arc_free_on_write; /* free after write list ptr */
659 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
660 static uint64_t l2arc_ndev; /* number of devices */
661
662 typedef struct l2arc_read_callback {
663 arc_buf_t *l2rcb_buf; /* read buffer */
664 spa_t *l2rcb_spa; /* spa */
665 blkptr_t l2rcb_bp; /* original blkptr */
666 zbookmark_t l2rcb_zb; /* original bookmark */
667 int l2rcb_flags; /* original flags */
668 enum zio_compress l2rcb_compress; /* applied compress */
669 } l2arc_read_callback_t;
670
671 typedef struct l2arc_write_callback {
672 l2arc_dev_t *l2wcb_dev; /* device info */
673 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
674 } l2arc_write_callback_t;
675
676 struct l2arc_buf_hdr {
677 /* protected by arc_buf_hdr mutex */
678 l2arc_dev_t *b_dev; /* L2ARC device */
679 uint64_t b_daddr; /* disk address, offset byte */
680 /* compression applied to buffer data */
681 enum zio_compress b_compress;
682 /* real alloc'd buffer size depending on b_compress applied */
683 int b_asize;
684 /* temporary buffer holder for in-flight compressed data */
685 void *b_tmp_cdata;
686 };
687
688 typedef struct l2arc_data_free {
689 /* protected by l2arc_free_on_write_mtx */
690 void *l2df_data;
691 size_t l2df_size;
692 void (*l2df_func)(void *, size_t);
693 list_node_t l2df_list_node;
694 } l2arc_data_free_t;
695
696 static kmutex_t l2arc_feed_thr_lock;
697 static kcondvar_t l2arc_feed_thr_cv;
698 static uint8_t l2arc_thread_exit;
699
700 static void l2arc_read_done(zio_t *zio);
701 static void l2arc_hdr_stat_add(void);
702 static void l2arc_hdr_stat_remove(void);
703
704 static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr);
705 static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr,
706 enum zio_compress c);
707 static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab);
708
709 static uint64_t
710 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
711 {
712 uint8_t *vdva = (uint8_t *)dva;
713 uint64_t crc = -1ULL;
714 int i;
715
716 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
717
718 for (i = 0; i < sizeof (dva_t); i++)
719 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
720
721 crc ^= (spa>>8) ^ birth;
722
723 return (crc);
724 }
725
726 #define BUF_EMPTY(buf) \
727 ((buf)->b_dva.dva_word[0] == 0 && \
728 (buf)->b_dva.dva_word[1] == 0 && \
729 (buf)->b_birth == 0)
730
731 #define BUF_EQUAL(spa, dva, birth, buf) \
732 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
733 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
734 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
735
736 static void
737 buf_discard_identity(arc_buf_hdr_t *hdr)
738 {
739 hdr->b_dva.dva_word[0] = 0;
740 hdr->b_dva.dva_word[1] = 0;
741 hdr->b_birth = 0;
742 hdr->b_cksum0 = 0;
743 }
744
745 static arc_buf_hdr_t *
746 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
747 {
748 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
749 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
750 arc_buf_hdr_t *buf;
751
752 mutex_enter(hash_lock);
753 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
754 buf = buf->b_hash_next) {
755 if (BUF_EQUAL(spa, dva, birth, buf)) {
756 *lockp = hash_lock;
757 return (buf);
758 }
759 }
760 mutex_exit(hash_lock);
761 *lockp = NULL;
762 return (NULL);
763 }
764
765 /*
766 * Insert an entry into the hash table. If there is already an element
767 * equal to elem in the hash table, then the already existing element
768 * will be returned and the new element will not be inserted.
769 * Otherwise returns NULL.
770 */
771 static arc_buf_hdr_t *
772 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
773 {
774 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
775 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
776 arc_buf_hdr_t *fbuf;
777 uint32_t i;
778
779 ASSERT(!HDR_IN_HASH_TABLE(buf));
780 *lockp = hash_lock;
781 mutex_enter(hash_lock);
782 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
783 fbuf = fbuf->b_hash_next, i++) {
784 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
785 return (fbuf);
786 }
787
788 buf->b_hash_next = buf_hash_table.ht_table[idx];
789 buf_hash_table.ht_table[idx] = buf;
790 buf->b_flags |= ARC_IN_HASH_TABLE;
791
792 /* collect some hash table performance data */
793 if (i > 0) {
794 ARCSTAT_BUMP(arcstat_hash_collisions);
795 if (i == 1)
796 ARCSTAT_BUMP(arcstat_hash_chains);
797
798 ARCSTAT_MAX(arcstat_hash_chain_max, i);
799 }
800
801 ARCSTAT_BUMP(arcstat_hash_elements);
802 ARCSTAT_MAXSTAT(arcstat_hash_elements);
803
804 return (NULL);
805 }
806
807 static void
808 buf_hash_remove(arc_buf_hdr_t *buf)
809 {
810 arc_buf_hdr_t *fbuf, **bufp;
811 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
812
813 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
814 ASSERT(HDR_IN_HASH_TABLE(buf));
815
816 bufp = &buf_hash_table.ht_table[idx];
817 while ((fbuf = *bufp) != buf) {
818 ASSERT(fbuf != NULL);
819 bufp = &fbuf->b_hash_next;
820 }
821 *bufp = buf->b_hash_next;
822 buf->b_hash_next = NULL;
823 buf->b_flags &= ~ARC_IN_HASH_TABLE;
824
825 /* collect some hash table performance data */
826 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
827
828 if (buf_hash_table.ht_table[idx] &&
829 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
830 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
831 }
832
833 /*
834 * Global data structures and functions for the buf kmem cache.
835 */
836 static kmem_cache_t *hdr_cache;
837 static kmem_cache_t *buf_cache;
838
839 static void
840 buf_fini(void)
841 {
842 int i;
843
844 kmem_free(buf_hash_table.ht_table,
845 (buf_hash_table.ht_mask + 1) * sizeof (void *));
846 for (i = 0; i < BUF_LOCKS; i++)
847 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
848 kmem_cache_destroy(hdr_cache);
849 kmem_cache_destroy(buf_cache);
850 }
851
852 /*
853 * Constructor callback - called when the cache is empty
854 * and a new buf is requested.
855 */
856 /* ARGSUSED */
857 static int
858 hdr_cons(void *vbuf, void *unused, int kmflag)
859 {
860 arc_buf_hdr_t *buf = vbuf;
861
862 bzero(buf, sizeof (arc_buf_hdr_t));
863 refcount_create(&buf->b_refcnt);
864 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
865 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
866 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
867
868 return (0);
869 }
870
871 /* ARGSUSED */
872 static int
873 buf_cons(void *vbuf, void *unused, int kmflag)
874 {
875 arc_buf_t *buf = vbuf;
876
877 bzero(buf, sizeof (arc_buf_t));
878 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
879 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
880
881 return (0);
882 }
883
884 /*
885 * Destructor callback - called when a cached buf is
886 * no longer required.
887 */
888 /* ARGSUSED */
889 static void
890 hdr_dest(void *vbuf, void *unused)
891 {
892 arc_buf_hdr_t *buf = vbuf;
893
894 ASSERT(BUF_EMPTY(buf));
895 refcount_destroy(&buf->b_refcnt);
896 cv_destroy(&buf->b_cv);
897 mutex_destroy(&buf->b_freeze_lock);
898 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
899 }
900
901 /* ARGSUSED */
902 static void
903 buf_dest(void *vbuf, void *unused)
904 {
905 arc_buf_t *buf = vbuf;
906
907 mutex_destroy(&buf->b_evict_lock);
908 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
909 }
910
911 /*
912 * Reclaim callback -- invoked when memory is low.
913 */
914 /* ARGSUSED */
915 static void
916 hdr_recl(void *unused)
917 {
918 dprintf("hdr_recl called\n");
919 /*
920 * umem calls the reclaim func when we destroy the buf cache,
921 * which is after we do arc_fini().
922 */
923 if (!arc_dead)
924 cv_signal(&arc_reclaim_thr_cv);
925 }
926
927 static void
928 buf_init(void)
929 {
930 uint64_t *ct;
931 uint64_t hsize = 1ULL << 12;
932 int i, j;
933
934 /*
935 * The hash table is big enough to fill all of physical memory
936 * with an average 64K block size. The table will take up
937 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
938 */
939 while (hsize * 65536 < physmem * PAGESIZE)
940 hsize <<= 1;
941 retry:
942 buf_hash_table.ht_mask = hsize - 1;
943 buf_hash_table.ht_table =
944 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
945 if (buf_hash_table.ht_table == NULL) {
946 ASSERT(hsize > (1ULL << 8));
947 hsize >>= 1;
948 goto retry;
949 }
950
951 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
952 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
953 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
954 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
955
956 for (i = 0; i < 256; i++)
957 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
958 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
959
960 for (i = 0; i < BUF_LOCKS; i++) {
961 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
962 NULL, MUTEX_DEFAULT, NULL);
963 }
964 }
965
966 #define ARC_MINTIME (hz>>4) /* 62 ms */
967
968 static void
969 arc_cksum_verify(arc_buf_t *buf)
970 {
971 zio_cksum_t zc;
972
973 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
974 return;
975
976 mutex_enter(&buf->b_hdr->b_freeze_lock);
977 if (buf->b_hdr->b_freeze_cksum == NULL ||
978 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
979 mutex_exit(&buf->b_hdr->b_freeze_lock);
980 return;
981 }
982 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
983 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
984 panic("buffer modified while frozen!");
985 mutex_exit(&buf->b_hdr->b_freeze_lock);
986 }
987
988 static int
989 arc_cksum_equal(arc_buf_t *buf)
990 {
991 zio_cksum_t zc;
992 int equal;
993
994 mutex_enter(&buf->b_hdr->b_freeze_lock);
995 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
996 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
997 mutex_exit(&buf->b_hdr->b_freeze_lock);
998
999 return (equal);
1000 }
1001
1002 static void
1003 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1004 {
1005 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1006 return;
1007
1008 mutex_enter(&buf->b_hdr->b_freeze_lock);
1009 if (buf->b_hdr->b_freeze_cksum != NULL) {
1010 mutex_exit(&buf->b_hdr->b_freeze_lock);
1011 return;
1012 }
1013 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1014 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1015 buf->b_hdr->b_freeze_cksum);
1016 mutex_exit(&buf->b_hdr->b_freeze_lock);
1017 arc_buf_watch(buf);
1018 }
1019
1020 #ifndef _KERNEL
1021 typedef struct procctl {
1022 long cmd;
1023 prwatch_t prwatch;
1024 } procctl_t;
1025 #endif
1026
1027 /* ARGSUSED */
1028 static void
1029 arc_buf_unwatch(arc_buf_t *buf)
1030 {
1031 #ifndef _KERNEL
1032 if (arc_watch) {
1033 int result;
1034 procctl_t ctl;
1035 ctl.cmd = PCWATCH;
1036 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1037 ctl.prwatch.pr_size = 0;
1038 ctl.prwatch.pr_wflags = 0;
1039 result = write(arc_procfd, &ctl, sizeof (ctl));
1040 ASSERT3U(result, ==, sizeof (ctl));
1041 }
1042 #endif
1043 }
1044
1045 /* ARGSUSED */
1046 static void
1047 arc_buf_watch(arc_buf_t *buf)
1048 {
1049 #ifndef _KERNEL
1050 if (arc_watch) {
1051 int result;
1052 procctl_t ctl;
1053 ctl.cmd = PCWATCH;
1054 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1055 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1056 ctl.prwatch.pr_wflags = WA_WRITE;
1057 result = write(arc_procfd, &ctl, sizeof (ctl));
1058 ASSERT3U(result, ==, sizeof (ctl));
1059 }
1060 #endif
1061 }
1062
1063 void
1064 arc_buf_thaw(arc_buf_t *buf)
1065 {
1066 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1067 if (buf->b_hdr->b_state != arc_anon)
1068 panic("modifying non-anon buffer!");
1069 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1070 panic("modifying buffer while i/o in progress!");
1071 arc_cksum_verify(buf);
1072 }
1073
1074 mutex_enter(&buf->b_hdr->b_freeze_lock);
1075 if (buf->b_hdr->b_freeze_cksum != NULL) {
1076 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1077 buf->b_hdr->b_freeze_cksum = NULL;
1078 }
1079
1080 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1081 if (buf->b_hdr->b_thawed)
1082 kmem_free(buf->b_hdr->b_thawed, 1);
1083 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1084 }
1085
1086 mutex_exit(&buf->b_hdr->b_freeze_lock);
1087
1088 arc_buf_unwatch(buf);
1089 }
1090
1091 void
1092 arc_buf_freeze(arc_buf_t *buf)
1093 {
1094 kmutex_t *hash_lock;
1095
1096 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1097 return;
1098
1099 hash_lock = HDR_LOCK(buf->b_hdr);
1100 mutex_enter(hash_lock);
1101
1102 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1103 buf->b_hdr->b_state == arc_anon);
1104 arc_cksum_compute(buf, B_FALSE);
1105 mutex_exit(hash_lock);
1106
1107 }
1108
1109 static void
1110 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1111 {
1112 ASSERT(MUTEX_HELD(hash_lock));
1113
1114 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1115 (ab->b_state != arc_anon)) {
1116 uint64_t delta = ab->b_size * ab->b_datacnt;
1117 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1118 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1119
1120 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1121 mutex_enter(&ab->b_state->arcs_mtx);
1122 ASSERT(list_link_active(&ab->b_arc_node));
1123 list_remove(list, ab);
1124 if (GHOST_STATE(ab->b_state)) {
1125 ASSERT0(ab->b_datacnt);
1126 ASSERT3P(ab->b_buf, ==, NULL);
1127 delta = ab->b_size;
1128 }
1129 ASSERT(delta > 0);
1130 ASSERT3U(*size, >=, delta);
1131 atomic_add_64(size, -delta);
1132 mutex_exit(&ab->b_state->arcs_mtx);
1133 /* remove the prefetch flag if we get a reference */
1134 if (ab->b_flags & ARC_PREFETCH)
1135 ab->b_flags &= ~ARC_PREFETCH;
1136 }
1137 }
1138
1139 static int
1140 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1141 {
1142 int cnt;
1143 arc_state_t *state = ab->b_state;
1144
1145 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1146 ASSERT(!GHOST_STATE(state));
1147
1148 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1149 (state != arc_anon)) {
1150 uint64_t *size = &state->arcs_lsize[ab->b_type];
1151
1152 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1153 mutex_enter(&state->arcs_mtx);
1154 ASSERT(!list_link_active(&ab->b_arc_node));
1155 list_insert_head(&state->arcs_list[ab->b_type], ab);
1156 ASSERT(ab->b_datacnt > 0);
1157 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1158 mutex_exit(&state->arcs_mtx);
1159 }
1160 return (cnt);
1161 }
1162
1163 /*
1164 * Move the supplied buffer to the indicated state. The mutex
1165 * for the buffer must be held by the caller.
1166 */
1167 static void
1168 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1169 {
1170 arc_state_t *old_state = ab->b_state;
1171 int64_t refcnt = refcount_count(&ab->b_refcnt);
1172 uint64_t from_delta, to_delta;
1173
1174 ASSERT(MUTEX_HELD(hash_lock));
1175 ASSERT3P(new_state, !=, old_state);
1176 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1177 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1178 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1179
1180 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1181
1182 /*
1183 * If this buffer is evictable, transfer it from the
1184 * old state list to the new state list.
1185 */
1186 if (refcnt == 0) {
1187 if (old_state != arc_anon) {
1188 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1189 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1190
1191 if (use_mutex)
1192 mutex_enter(&old_state->arcs_mtx);
1193
1194 ASSERT(list_link_active(&ab->b_arc_node));
1195 list_remove(&old_state->arcs_list[ab->b_type], ab);
1196
1197 /*
1198 * If prefetching out of the ghost cache,
1199 * we will have a non-zero datacnt.
1200 */
1201 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1202 /* ghost elements have a ghost size */
1203 ASSERT(ab->b_buf == NULL);
1204 from_delta = ab->b_size;
1205 }
1206 ASSERT3U(*size, >=, from_delta);
1207 atomic_add_64(size, -from_delta);
1208
1209 if (use_mutex)
1210 mutex_exit(&old_state->arcs_mtx);
1211 }
1212 if (new_state != arc_anon) {
1213 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1214 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1215
1216 if (use_mutex)
1217 mutex_enter(&new_state->arcs_mtx);
1218
1219 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1220
1221 /* ghost elements have a ghost size */
1222 if (GHOST_STATE(new_state)) {
1223 ASSERT(ab->b_datacnt == 0);
1224 ASSERT(ab->b_buf == NULL);
1225 to_delta = ab->b_size;
1226 }
1227 atomic_add_64(size, to_delta);
1228
1229 if (use_mutex)
1230 mutex_exit(&new_state->arcs_mtx);
1231 }
1232 }
1233
1234 ASSERT(!BUF_EMPTY(ab));
1235 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1236 buf_hash_remove(ab);
1237
1238 /* adjust state sizes */
1239 if (to_delta)
1240 atomic_add_64(&new_state->arcs_size, to_delta);
1241 if (from_delta) {
1242 ASSERT3U(old_state->arcs_size, >=, from_delta);
1243 atomic_add_64(&old_state->arcs_size, -from_delta);
1244 }
1245 ab->b_state = new_state;
1246
1247 /* adjust l2arc hdr stats */
1248 if (new_state == arc_l2c_only)
1249 l2arc_hdr_stat_add();
1250 else if (old_state == arc_l2c_only)
1251 l2arc_hdr_stat_remove();
1252 }
1253
1254 void
1255 arc_space_consume(uint64_t space, arc_space_type_t type)
1256 {
1257 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1258
1259 switch (type) {
1260 case ARC_SPACE_DATA:
1261 ARCSTAT_INCR(arcstat_data_size, space);
1262 break;
1263 case ARC_SPACE_OTHER:
1264 ARCSTAT_INCR(arcstat_other_size, space);
1265 break;
1266 case ARC_SPACE_HDRS:
1267 ARCSTAT_INCR(arcstat_hdr_size, space);
1268 break;
1269 case ARC_SPACE_L2HDRS:
1270 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1271 break;
1272 }
1273
1274 ARCSTAT_INCR(arcstat_meta_used, space);
1275 atomic_add_64(&arc_size, space);
1276 }
1277
1278 void
1279 arc_space_return(uint64_t space, arc_space_type_t type)
1280 {
1281 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1282
1283 switch (type) {
1284 case ARC_SPACE_DATA:
1285 ARCSTAT_INCR(arcstat_data_size, -space);
1286 break;
1287 case ARC_SPACE_OTHER:
1288 ARCSTAT_INCR(arcstat_other_size, -space);
1289 break;
1290 case ARC_SPACE_HDRS:
1291 ARCSTAT_INCR(arcstat_hdr_size, -space);
1292 break;
1293 case ARC_SPACE_L2HDRS:
1294 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1295 break;
1296 }
1297
1298 ASSERT(arc_meta_used >= space);
1299 if (arc_meta_max < arc_meta_used)
1300 arc_meta_max = arc_meta_used;
1301 ARCSTAT_INCR(arcstat_meta_used, -space);
1302 ASSERT(arc_size >= space);
1303 atomic_add_64(&arc_size, -space);
1304 }
1305
1306 void *
1307 arc_data_buf_alloc(uint64_t size)
1308 {
1309 if (arc_evict_needed(ARC_BUFC_DATA))
1310 cv_signal(&arc_reclaim_thr_cv);
1311 atomic_add_64(&arc_size, size);
1312 return (zio_data_buf_alloc(size));
1313 }
1314
1315 void
1316 arc_data_buf_free(void *buf, uint64_t size)
1317 {
1318 zio_data_buf_free(buf, size);
1319 ASSERT(arc_size >= size);
1320 atomic_add_64(&arc_size, -size);
1321 }
1322
1323 arc_buf_t *
1324 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1325 {
1326 arc_buf_hdr_t *hdr;
1327 arc_buf_t *buf;
1328
1329 ASSERT3U(size, >, 0);
1330 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1331 ASSERT(BUF_EMPTY(hdr));
1332 hdr->b_size = size;
1333 hdr->b_type = type;
1334 hdr->b_spa = spa_load_guid(spa);
1335 hdr->b_state = arc_anon;
1336 hdr->b_arc_access = 0;
1337 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1338 buf->b_hdr = hdr;
1339 buf->b_data = NULL;
1340 buf->b_efunc = NULL;
1341 buf->b_private = NULL;
1342 buf->b_next = NULL;
1343 hdr->b_buf = buf;
1344 arc_get_data_buf(buf);
1345 hdr->b_datacnt = 1;
1346 hdr->b_flags = 0;
1347 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1348 (void) refcount_add(&hdr->b_refcnt, tag);
1349
1350 return (buf);
1351 }
1352
1353 static char *arc_onloan_tag = "onloan";
1354
1355 /*
1356 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1357 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1358 * buffers must be returned to the arc before they can be used by the DMU or
1359 * freed.
1360 */
1361 arc_buf_t *
1362 arc_loan_buf(spa_t *spa, int size)
1363 {
1364 arc_buf_t *buf;
1365
1366 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1367
1368 atomic_add_64(&arc_loaned_bytes, size);
1369 return (buf);
1370 }
1371
1372 /*
1373 * Return a loaned arc buffer to the arc.
1374 */
1375 void
1376 arc_return_buf(arc_buf_t *buf, void *tag)
1377 {
1378 arc_buf_hdr_t *hdr = buf->b_hdr;
1379
1380 ASSERT(buf->b_data != NULL);
1381 (void) refcount_add(&hdr->b_refcnt, tag);
1382 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1383
1384 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1385 }
1386
1387 /* Detach an arc_buf from a dbuf (tag) */
1388 void
1389 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1390 {
1391 arc_buf_hdr_t *hdr;
1392
1393 ASSERT(buf->b_data != NULL);
1394 hdr = buf->b_hdr;
1395 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1396 (void) refcount_remove(&hdr->b_refcnt, tag);
1397 buf->b_efunc = NULL;
1398 buf->b_private = NULL;
1399
1400 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1401 }
1402
1403 static arc_buf_t *
1404 arc_buf_clone(arc_buf_t *from)
1405 {
1406 arc_buf_t *buf;
1407 arc_buf_hdr_t *hdr = from->b_hdr;
1408 uint64_t size = hdr->b_size;
1409
1410 ASSERT(hdr->b_state != arc_anon);
1411
1412 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1413 buf->b_hdr = hdr;
1414 buf->b_data = NULL;
1415 buf->b_efunc = NULL;
1416 buf->b_private = NULL;
1417 buf->b_next = hdr->b_buf;
1418 hdr->b_buf = buf;
1419 arc_get_data_buf(buf);
1420 bcopy(from->b_data, buf->b_data, size);
1421
1422 /*
1423 * This buffer already exists in the arc so create a duplicate
1424 * copy for the caller. If the buffer is associated with user data
1425 * then track the size and number of duplicates. These stats will be
1426 * updated as duplicate buffers are created and destroyed.
1427 */
1428 if (hdr->b_type == ARC_BUFC_DATA) {
1429 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1430 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1431 }
1432 hdr->b_datacnt += 1;
1433 return (buf);
1434 }
1435
1436 void
1437 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1438 {
1439 arc_buf_hdr_t *hdr;
1440 kmutex_t *hash_lock;
1441
1442 /*
1443 * Check to see if this buffer is evicted. Callers
1444 * must verify b_data != NULL to know if the add_ref
1445 * was successful.
1446 */
1447 mutex_enter(&buf->b_evict_lock);
1448 if (buf->b_data == NULL) {
1449 mutex_exit(&buf->b_evict_lock);
1450 return;
1451 }
1452 hash_lock = HDR_LOCK(buf->b_hdr);
1453 mutex_enter(hash_lock);
1454 hdr = buf->b_hdr;
1455 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1456 mutex_exit(&buf->b_evict_lock);
1457
1458 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1459 add_reference(hdr, hash_lock, tag);
1460 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1461 arc_access(hdr, hash_lock);
1462 mutex_exit(hash_lock);
1463 ARCSTAT_BUMP(arcstat_hits);
1464 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1465 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1466 data, metadata, hits);
1467 }
1468
1469 /*
1470 * Free the arc data buffer. If it is an l2arc write in progress,
1471 * the buffer is placed on l2arc_free_on_write to be freed later.
1472 */
1473 static void
1474 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1475 {
1476 arc_buf_hdr_t *hdr = buf->b_hdr;
1477
1478 if (HDR_L2_WRITING(hdr)) {
1479 l2arc_data_free_t *df;
1480 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1481 df->l2df_data = buf->b_data;
1482 df->l2df_size = hdr->b_size;
1483 df->l2df_func = free_func;
1484 mutex_enter(&l2arc_free_on_write_mtx);
1485 list_insert_head(l2arc_free_on_write, df);
1486 mutex_exit(&l2arc_free_on_write_mtx);
1487 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1488 } else {
1489 free_func(buf->b_data, hdr->b_size);
1490 }
1491 }
1492
1493 static void
1494 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1495 {
1496 arc_buf_t **bufp;
1497
1498 /* free up data associated with the buf */
1499 if (buf->b_data) {
1500 arc_state_t *state = buf->b_hdr->b_state;
1501 uint64_t size = buf->b_hdr->b_size;
1502 arc_buf_contents_t type = buf->b_hdr->b_type;
1503
1504 arc_cksum_verify(buf);
1505 arc_buf_unwatch(buf);
1506
1507 if (!recycle) {
1508 if (type == ARC_BUFC_METADATA) {
1509 arc_buf_data_free(buf, zio_buf_free);
1510 arc_space_return(size, ARC_SPACE_DATA);
1511 } else {
1512 ASSERT(type == ARC_BUFC_DATA);
1513 arc_buf_data_free(buf, zio_data_buf_free);
1514 ARCSTAT_INCR(arcstat_data_size, -size);
1515 atomic_add_64(&arc_size, -size);
1516 }
1517 }
1518 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1519 uint64_t *cnt = &state->arcs_lsize[type];
1520
1521 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1522 ASSERT(state != arc_anon);
1523
1524 ASSERT3U(*cnt, >=, size);
1525 atomic_add_64(cnt, -size);
1526 }
1527 ASSERT3U(state->arcs_size, >=, size);
1528 atomic_add_64(&state->arcs_size, -size);
1529 buf->b_data = NULL;
1530
1531 /*
1532 * If we're destroying a duplicate buffer make sure
1533 * that the appropriate statistics are updated.
1534 */
1535 if (buf->b_hdr->b_datacnt > 1 &&
1536 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1537 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1538 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1539 }
1540 ASSERT(buf->b_hdr->b_datacnt > 0);
1541 buf->b_hdr->b_datacnt -= 1;
1542 }
1543
1544 /* only remove the buf if requested */
1545 if (!all)
1546 return;
1547
1548 /* remove the buf from the hdr list */
1549 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1550 continue;
1551 *bufp = buf->b_next;
1552 buf->b_next = NULL;
1553
1554 ASSERT(buf->b_efunc == NULL);
1555
1556 /* clean up the buf */
1557 buf->b_hdr = NULL;
1558 kmem_cache_free(buf_cache, buf);
1559 }
1560
1561 static void
1562 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1563 {
1564 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1565 ASSERT3P(hdr->b_state, ==, arc_anon);
1566 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1567 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1568
1569 if (l2hdr != NULL) {
1570 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1571 /*
1572 * To prevent arc_free() and l2arc_evict() from
1573 * attempting to free the same buffer at the same time,
1574 * a FREE_IN_PROGRESS flag is given to arc_free() to
1575 * give it priority. l2arc_evict() can't destroy this
1576 * header while we are waiting on l2arc_buflist_mtx.
1577 *
1578 * The hdr may be removed from l2ad_buflist before we
1579 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1580 */
1581 if (!buflist_held) {
1582 mutex_enter(&l2arc_buflist_mtx);
1583 l2hdr = hdr->b_l2hdr;
1584 }
1585
1586 if (l2hdr != NULL) {
1587 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1588 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1589 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
1590 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1591 if (hdr->b_state == arc_l2c_only)
1592 l2arc_hdr_stat_remove();
1593 hdr->b_l2hdr = NULL;
1594 }
1595
1596 if (!buflist_held)
1597 mutex_exit(&l2arc_buflist_mtx);
1598 }
1599
1600 if (!BUF_EMPTY(hdr)) {
1601 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1602 buf_discard_identity(hdr);
1603 }
1604 while (hdr->b_buf) {
1605 arc_buf_t *buf = hdr->b_buf;
1606
1607 if (buf->b_efunc) {
1608 mutex_enter(&arc_eviction_mtx);
1609 mutex_enter(&buf->b_evict_lock);
1610 ASSERT(buf->b_hdr != NULL);
1611 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1612 hdr->b_buf = buf->b_next;
1613 buf->b_hdr = &arc_eviction_hdr;
1614 buf->b_next = arc_eviction_list;
1615 arc_eviction_list = buf;
1616 mutex_exit(&buf->b_evict_lock);
1617 mutex_exit(&arc_eviction_mtx);
1618 } else {
1619 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1620 }
1621 }
1622 if (hdr->b_freeze_cksum != NULL) {
1623 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1624 hdr->b_freeze_cksum = NULL;
1625 }
1626 if (hdr->b_thawed) {
1627 kmem_free(hdr->b_thawed, 1);
1628 hdr->b_thawed = NULL;
1629 }
1630
1631 ASSERT(!list_link_active(&hdr->b_arc_node));
1632 ASSERT3P(hdr->b_hash_next, ==, NULL);
1633 ASSERT3P(hdr->b_acb, ==, NULL);
1634 kmem_cache_free(hdr_cache, hdr);
1635 }
1636
1637 void
1638 arc_buf_free(arc_buf_t *buf, void *tag)
1639 {
1640 arc_buf_hdr_t *hdr = buf->b_hdr;
1641 int hashed = hdr->b_state != arc_anon;
1642
1643 ASSERT(buf->b_efunc == NULL);
1644 ASSERT(buf->b_data != NULL);
1645
1646 if (hashed) {
1647 kmutex_t *hash_lock = HDR_LOCK(hdr);
1648
1649 mutex_enter(hash_lock);
1650 hdr = buf->b_hdr;
1651 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1652
1653 (void) remove_reference(hdr, hash_lock, tag);
1654 if (hdr->b_datacnt > 1) {
1655 arc_buf_destroy(buf, FALSE, TRUE);
1656 } else {
1657 ASSERT(buf == hdr->b_buf);
1658 ASSERT(buf->b_efunc == NULL);
1659 hdr->b_flags |= ARC_BUF_AVAILABLE;
1660 }
1661 mutex_exit(hash_lock);
1662 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1663 int destroy_hdr;
1664 /*
1665 * We are in the middle of an async write. Don't destroy
1666 * this buffer unless the write completes before we finish
1667 * decrementing the reference count.
1668 */
1669 mutex_enter(&arc_eviction_mtx);
1670 (void) remove_reference(hdr, NULL, tag);
1671 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1672 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1673 mutex_exit(&arc_eviction_mtx);
1674 if (destroy_hdr)
1675 arc_hdr_destroy(hdr);
1676 } else {
1677 if (remove_reference(hdr, NULL, tag) > 0)
1678 arc_buf_destroy(buf, FALSE, TRUE);
1679 else
1680 arc_hdr_destroy(hdr);
1681 }
1682 }
1683
1684 boolean_t
1685 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1686 {
1687 arc_buf_hdr_t *hdr = buf->b_hdr;
1688 kmutex_t *hash_lock = HDR_LOCK(hdr);
1689 boolean_t no_callback = (buf->b_efunc == NULL);
1690
1691 if (hdr->b_state == arc_anon) {
1692 ASSERT(hdr->b_datacnt == 1);
1693 arc_buf_free(buf, tag);
1694 return (no_callback);
1695 }
1696
1697 mutex_enter(hash_lock);
1698 hdr = buf->b_hdr;
1699 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1700 ASSERT(hdr->b_state != arc_anon);
1701 ASSERT(buf->b_data != NULL);
1702
1703 (void) remove_reference(hdr, hash_lock, tag);
1704 if (hdr->b_datacnt > 1) {
1705 if (no_callback)
1706 arc_buf_destroy(buf, FALSE, TRUE);
1707 } else if (no_callback) {
1708 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1709 ASSERT(buf->b_efunc == NULL);
1710 hdr->b_flags |= ARC_BUF_AVAILABLE;
1711 }
1712 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1713 refcount_is_zero(&hdr->b_refcnt));
1714 mutex_exit(hash_lock);
1715 return (no_callback);
1716 }
1717
1718 int
1719 arc_buf_size(arc_buf_t *buf)
1720 {
1721 return (buf->b_hdr->b_size);
1722 }
1723
1724 /*
1725 * Called from the DMU to determine if the current buffer should be
1726 * evicted. In order to ensure proper locking, the eviction must be initiated
1727 * from the DMU. Return true if the buffer is associated with user data and
1728 * duplicate buffers still exist.
1729 */
1730 boolean_t
1731 arc_buf_eviction_needed(arc_buf_t *buf)
1732 {
1733 arc_buf_hdr_t *hdr;
1734 boolean_t evict_needed = B_FALSE;
1735
1736 if (zfs_disable_dup_eviction)
1737 return (B_FALSE);
1738
1739 mutex_enter(&buf->b_evict_lock);
1740 hdr = buf->b_hdr;
1741 if (hdr == NULL) {
1742 /*
1743 * We are in arc_do_user_evicts(); let that function
1744 * perform the eviction.
1745 */
1746 ASSERT(buf->b_data == NULL);
1747 mutex_exit(&buf->b_evict_lock);
1748 return (B_FALSE);
1749 } else if (buf->b_data == NULL) {
1750 /*
1751 * We have already been added to the arc eviction list;
1752 * recommend eviction.
1753 */
1754 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1755 mutex_exit(&buf->b_evict_lock);
1756 return (B_TRUE);
1757 }
1758
1759 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1760 evict_needed = B_TRUE;
1761
1762 mutex_exit(&buf->b_evict_lock);
1763 return (evict_needed);
1764 }
1765
1766 /*
1767 * Evict buffers from list until we've removed the specified number of
1768 * bytes. Move the removed buffers to the appropriate evict state.
1769 * If the recycle flag is set, then attempt to "recycle" a buffer:
1770 * - look for a buffer to evict that is `bytes' long.
1771 * - return the data block from this buffer rather than freeing it.
1772 * This flag is used by callers that are trying to make space for a
1773 * new buffer in a full arc cache.
1774 *
1775 * This function makes a "best effort". It skips over any buffers
1776 * it can't get a hash_lock on, and so may not catch all candidates.
1777 * It may also return without evicting as much space as requested.
1778 */
1779 static void *
1780 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1781 arc_buf_contents_t type)
1782 {
1783 arc_state_t *evicted_state;
1784 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1785 arc_buf_hdr_t *ab, *ab_prev = NULL;
1786 list_t *list = &state->arcs_list[type];
1787 kmutex_t *hash_lock;
1788 boolean_t have_lock;
1789 void *stolen = NULL;
1790 arc_buf_hdr_t marker = { 0 };
1791 int count = 0;
1792
1793 ASSERT(state == arc_mru || state == arc_mfu);
1794
1795 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1796
1797 mutex_enter(&state->arcs_mtx);
1798 mutex_enter(&evicted_state->arcs_mtx);
1799
1800 for (ab = list_tail(list); ab; ab = ab_prev) {
1801 ab_prev = list_prev(list, ab);
1802 /* prefetch buffers have a minimum lifespan */
1803 if (HDR_IO_IN_PROGRESS(ab) ||
1804 (spa && ab->b_spa != spa) ||
1805 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1806 ddi_get_lbolt() - ab->b_arc_access <
1807 arc_min_prefetch_lifespan)) {
1808 skipped++;
1809 continue;
1810 }
1811 /* "lookahead" for better eviction candidate */
1812 if (recycle && ab->b_size != bytes &&
1813 ab_prev && ab_prev->b_size == bytes)
1814 continue;
1815
1816 /* ignore markers */
1817 if (ab->b_spa == 0)
1818 continue;
1819
1820 /*
1821 * It may take a long time to evict all the bufs requested.
1822 * To avoid blocking all arc activity, periodically drop
1823 * the arcs_mtx and give other threads a chance to run
1824 * before reacquiring the lock.
1825 *
1826 * If we are looking for a buffer to recycle, we are in
1827 * the hot code path, so don't sleep.
1828 */
1829 if (!recycle && count++ > arc_evict_iterations) {
1830 list_insert_after(list, ab, &marker);
1831 mutex_exit(&evicted_state->arcs_mtx);
1832 mutex_exit(&state->arcs_mtx);
1833 kpreempt(KPREEMPT_SYNC);
1834 mutex_enter(&state->arcs_mtx);
1835 mutex_enter(&evicted_state->arcs_mtx);
1836 ab_prev = list_prev(list, &marker);
1837 list_remove(list, &marker);
1838 count = 0;
1839 continue;
1840 }
1841
1842 hash_lock = HDR_LOCK(ab);
1843 have_lock = MUTEX_HELD(hash_lock);
1844 if (have_lock || mutex_tryenter(hash_lock)) {
1845 ASSERT0(refcount_count(&ab->b_refcnt));
1846 ASSERT(ab->b_datacnt > 0);
1847 while (ab->b_buf) {
1848 arc_buf_t *buf = ab->b_buf;
1849 if (!mutex_tryenter(&buf->b_evict_lock)) {
1850 missed += 1;
1851 break;
1852 }
1853 if (buf->b_data) {
1854 bytes_evicted += ab->b_size;
1855 if (recycle && ab->b_type == type &&
1856 ab->b_size == bytes &&
1857 !HDR_L2_WRITING(ab)) {
1858 stolen = buf->b_data;
1859 recycle = FALSE;
1860 }
1861 }
1862 if (buf->b_efunc) {
1863 mutex_enter(&arc_eviction_mtx);
1864 arc_buf_destroy(buf,
1865 buf->b_data == stolen, FALSE);
1866 ab->b_buf = buf->b_next;
1867 buf->b_hdr = &arc_eviction_hdr;
1868 buf->b_next = arc_eviction_list;
1869 arc_eviction_list = buf;
1870 mutex_exit(&arc_eviction_mtx);
1871 mutex_exit(&buf->b_evict_lock);
1872 } else {
1873 mutex_exit(&buf->b_evict_lock);
1874 arc_buf_destroy(buf,
1875 buf->b_data == stolen, TRUE);
1876 }
1877 }
1878
1879 if (ab->b_l2hdr) {
1880 ARCSTAT_INCR(arcstat_evict_l2_cached,
1881 ab->b_size);
1882 } else {
1883 if (l2arc_write_eligible(ab->b_spa, ab)) {
1884 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1885 ab->b_size);
1886 } else {
1887 ARCSTAT_INCR(
1888 arcstat_evict_l2_ineligible,
1889 ab->b_size);
1890 }
1891 }
1892
1893 if (ab->b_datacnt == 0) {
1894 arc_change_state(evicted_state, ab, hash_lock);
1895 ASSERT(HDR_IN_HASH_TABLE(ab));
1896 ab->b_flags |= ARC_IN_HASH_TABLE;
1897 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1898 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1899 }
1900 if (!have_lock)
1901 mutex_exit(hash_lock);
1902 if (bytes >= 0 && bytes_evicted >= bytes)
1903 break;
1904 } else {
1905 missed += 1;
1906 }
1907 }
1908
1909 mutex_exit(&evicted_state->arcs_mtx);
1910 mutex_exit(&state->arcs_mtx);
1911
1912 if (bytes_evicted < bytes)
1913 dprintf("only evicted %lld bytes from %x",
1914 (longlong_t)bytes_evicted, state);
1915
1916 if (skipped)
1917 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1918
1919 if (missed)
1920 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1921
1922 /*
1923 * Note: we have just evicted some data into the ghost state,
1924 * potentially putting the ghost size over the desired size. Rather
1925 * that evicting from the ghost list in this hot code path, leave
1926 * this chore to the arc_reclaim_thread().
1927 */
1928
1929 return (stolen);
1930 }
1931
1932 /*
1933 * Remove buffers from list until we've removed the specified number of
1934 * bytes. Destroy the buffers that are removed.
1935 */
1936 static void
1937 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1938 {
1939 arc_buf_hdr_t *ab, *ab_prev;
1940 arc_buf_hdr_t marker = { 0 };
1941 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1942 kmutex_t *hash_lock;
1943 uint64_t bytes_deleted = 0;
1944 uint64_t bufs_skipped = 0;
1945 int count = 0;
1946
1947 ASSERT(GHOST_STATE(state));
1948 top:
1949 mutex_enter(&state->arcs_mtx);
1950 for (ab = list_tail(list); ab; ab = ab_prev) {
1951 ab_prev = list_prev(list, ab);
1952 if (ab->b_type > ARC_BUFC_NUMTYPES)
1953 panic("invalid ab=%p", (void *)ab);
1954 if (spa && ab->b_spa != spa)
1955 continue;
1956
1957 /* ignore markers */
1958 if (ab->b_spa == 0)
1959 continue;
1960
1961 hash_lock = HDR_LOCK(ab);
1962 /* caller may be trying to modify this buffer, skip it */
1963 if (MUTEX_HELD(hash_lock))
1964 continue;
1965
1966 /*
1967 * It may take a long time to evict all the bufs requested.
1968 * To avoid blocking all arc activity, periodically drop
1969 * the arcs_mtx and give other threads a chance to run
1970 * before reacquiring the lock.
1971 */
1972 if (count++ > arc_evict_iterations) {
1973 list_insert_after(list, ab, &marker);
1974 mutex_exit(&state->arcs_mtx);
1975 kpreempt(KPREEMPT_SYNC);
1976 mutex_enter(&state->arcs_mtx);
1977 ab_prev = list_prev(list, &marker);
1978 list_remove(list, &marker);
1979 count = 0;
1980 continue;
1981 }
1982 if (mutex_tryenter(hash_lock)) {
1983 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1984 ASSERT(ab->b_buf == NULL);
1985 ARCSTAT_BUMP(arcstat_deleted);
1986 bytes_deleted += ab->b_size;
1987
1988 if (ab->b_l2hdr != NULL) {
1989 /*
1990 * This buffer is cached on the 2nd Level ARC;
1991 * don't destroy the header.
1992 */
1993 arc_change_state(arc_l2c_only, ab, hash_lock);
1994 mutex_exit(hash_lock);
1995 } else {
1996 arc_change_state(arc_anon, ab, hash_lock);
1997 mutex_exit(hash_lock);
1998 arc_hdr_destroy(ab);
1999 }
2000
2001 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
2002 if (bytes >= 0 && bytes_deleted >= bytes)
2003 break;
2004 } else if (bytes < 0) {
2005 /*
2006 * Insert a list marker and then wait for the
2007 * hash lock to become available. Once its
2008 * available, restart from where we left off.
2009 */
2010 list_insert_after(list, ab, &marker);
2011 mutex_exit(&state->arcs_mtx);
2012 mutex_enter(hash_lock);
2013 mutex_exit(hash_lock);
2014 mutex_enter(&state->arcs_mtx);
2015 ab_prev = list_prev(list, &marker);
2016 list_remove(list, &marker);
2017 } else {
2018 bufs_skipped += 1;
2019 }
2020
2021 }
2022 mutex_exit(&state->arcs_mtx);
2023
2024 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
2025 (bytes < 0 || bytes_deleted < bytes)) {
2026 list = &state->arcs_list[ARC_BUFC_METADATA];
2027 goto top;
2028 }
2029
2030 if (bufs_skipped) {
2031 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2032 ASSERT(bytes >= 0);
2033 }
2034
2035 if (bytes_deleted < bytes)
2036 dprintf("only deleted %lld bytes from %p",
2037 (longlong_t)bytes_deleted, state);
2038 }
2039
2040 static void
2041 arc_adjust(void)
2042 {
2043 int64_t adjustment, delta;
2044
2045 /*
2046 * Adjust MRU size
2047 */
2048
2049 adjustment = MIN((int64_t)(arc_size - arc_c),
2050 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2051 arc_p));
2052
2053 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2054 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2055 (void) arc_evict(arc_mru, NULL, delta, FALSE, ARC_BUFC_DATA);
2056 adjustment -= delta;
2057 }
2058
2059 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2060 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2061 (void) arc_evict(arc_mru, NULL, delta, FALSE,
2062 ARC_BUFC_METADATA);
2063 }
2064
2065 /*
2066 * Adjust MFU size
2067 */
2068
2069 adjustment = arc_size - arc_c;
2070
2071 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2072 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2073 (void) arc_evict(arc_mfu, NULL, delta, FALSE, ARC_BUFC_DATA);
2074 adjustment -= delta;
2075 }
2076
2077 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2078 int64_t delta = MIN(adjustment,
2079 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2080 (void) arc_evict(arc_mfu, NULL, delta, FALSE,
2081 ARC_BUFC_METADATA);
2082 }
2083
2084 /*
2085 * Adjust ghost lists
2086 */
2087
2088 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2089
2090 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2091 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2092 arc_evict_ghost(arc_mru_ghost, NULL, delta);
2093 }
2094
2095 adjustment =
2096 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2097
2098 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2099 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2100 arc_evict_ghost(arc_mfu_ghost, NULL, delta);
2101 }
2102 }
2103
2104 static void
2105 arc_do_user_evicts(void)
2106 {
2107 mutex_enter(&arc_eviction_mtx);
2108 while (arc_eviction_list != NULL) {
2109 arc_buf_t *buf = arc_eviction_list;
2110 arc_eviction_list = buf->b_next;
2111 mutex_enter(&buf->b_evict_lock);
2112 buf->b_hdr = NULL;
2113 mutex_exit(&buf->b_evict_lock);
2114 mutex_exit(&arc_eviction_mtx);
2115
2116 if (buf->b_efunc != NULL)
2117 VERIFY(buf->b_efunc(buf) == 0);
2118
2119 buf->b_efunc = NULL;
2120 buf->b_private = NULL;
2121 kmem_cache_free(buf_cache, buf);
2122 mutex_enter(&arc_eviction_mtx);
2123 }
2124 mutex_exit(&arc_eviction_mtx);
2125 }
2126
2127 /*
2128 * Flush all *evictable* data from the cache for the given spa.
2129 * NOTE: this will not touch "active" (i.e. referenced) data.
2130 */
2131 void
2132 arc_flush(spa_t *spa)
2133 {
2134 uint64_t guid = 0;
2135
2136 if (spa)
2137 guid = spa_load_guid(spa);
2138
2139 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2140 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2141 if (spa)
2142 break;
2143 }
2144 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2145 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2146 if (spa)
2147 break;
2148 }
2149 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2150 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2151 if (spa)
2152 break;
2153 }
2154 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2155 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2156 if (spa)
2157 break;
2158 }
2159
2160 arc_evict_ghost(arc_mru_ghost, guid, -1);
2161 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2162
2163 mutex_enter(&arc_reclaim_thr_lock);
2164 arc_do_user_evicts();
2165 mutex_exit(&arc_reclaim_thr_lock);
2166 ASSERT(spa || arc_eviction_list == NULL);
2167 }
2168
2169 void
2170 arc_shrink(void)
2171 {
2172 if (arc_c > arc_c_min) {
2173 uint64_t to_free;
2174
2175 #ifdef _KERNEL
2176 to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
2177 #else
2178 to_free = arc_c >> arc_shrink_shift;
2179 #endif
2180 if (arc_c > arc_c_min + to_free)
2181 atomic_add_64(&arc_c, -to_free);
2182 else
2183 arc_c = arc_c_min;
2184
2185 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2186 if (arc_c > arc_size)
2187 arc_c = MAX(arc_size, arc_c_min);
2188 if (arc_p > arc_c)
2189 arc_p = (arc_c >> 1);
2190 ASSERT(arc_c >= arc_c_min);
2191 ASSERT((int64_t)arc_p >= 0);
2192 }
2193
2194 if (arc_size > arc_c)
2195 arc_adjust();
2196 }
2197
2198 /*
2199 * Determine if the system is under memory pressure and is asking
2200 * to reclaim memory. A return value of 1 indicates that the system
2201 * is under memory pressure and that the arc should adjust accordingly.
2202 */
2203 static int
2204 arc_reclaim_needed(void)
2205 {
2206 uint64_t extra;
2207
2208 #ifdef _KERNEL
2209
2210 if (needfree)
2211 return (1);
2212
2213 /*
2214 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2215 */
2216 extra = desfree;
2217
2218 /*
2219 * check that we're out of range of the pageout scanner. It starts to
2220 * schedule paging if freemem is less than lotsfree and needfree.
2221 * lotsfree is the high-water mark for pageout, and needfree is the
2222 * number of needed free pages. We add extra pages here to make sure
2223 * the scanner doesn't start up while we're freeing memory.
2224 */
2225 if (freemem < lotsfree + needfree + extra)
2226 return (1);
2227
2228 /*
2229 * check to make sure that swapfs has enough space so that anon
2230 * reservations can still succeed. anon_resvmem() checks that the
2231 * availrmem is greater than swapfs_minfree, and the number of reserved
2232 * swap pages. We also add a bit of extra here just to prevent
2233 * circumstances from getting really dire.
2234 */
2235 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2236 return (1);
2237
2238 #if defined(__i386)
2239 /*
2240 * If we're on an i386 platform, it's possible that we'll exhaust the
2241 * kernel heap space before we ever run out of available physical
2242 * memory. Most checks of the size of the heap_area compare against
2243 * tune.t_minarmem, which is the minimum available real memory that we
2244 * can have in the system. However, this is generally fixed at 25 pages
2245 * which is so low that it's useless. In this comparison, we seek to
2246 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2247 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2248 * free)
2249 */
2250 if (vmem_size(heap_arena, VMEM_FREE) <
2251 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2))
2252 return (1);
2253 #endif
2254
2255 /*
2256 * If zio data pages are being allocated out of a separate heap segment,
2257 * then enforce that the size of available vmem for this arena remains
2258 * above about 1/16th free.
2259 *
2260 * Note: The 1/16th arena free requirement was put in place
2261 * to aggressively evict memory from the arc in order to avoid
2262 * memory fragmentation issues.
2263 */
2264 if (zio_arena != NULL &&
2265 vmem_size(zio_arena, VMEM_FREE) <
2266 (vmem_size(zio_arena, VMEM_ALLOC) >> 4))
2267 return (1);
2268 #else
2269 if (spa_get_random(100) == 0)
2270 return (1);
2271 #endif
2272 return (0);
2273 }
2274
2275 static void
2276 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2277 {
2278 size_t i;
2279 kmem_cache_t *prev_cache = NULL;
2280 kmem_cache_t *prev_data_cache = NULL;
2281 extern kmem_cache_t *zio_buf_cache[];
2282 extern kmem_cache_t *zio_data_buf_cache[];
2283
2284 #ifdef _KERNEL
2285 if (arc_meta_used >= arc_meta_limit) {
2286 /*
2287 * We are exceeding our meta-data cache limit.
2288 * Purge some DNLC entries to release holds on meta-data.
2289 */
2290 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2291 }
2292 #if defined(__i386)
2293 /*
2294 * Reclaim unused memory from all kmem caches.
2295 */
2296 kmem_reap();
2297 #endif
2298 #endif
2299
2300 /*
2301 * An aggressive reclamation will shrink the cache size as well as
2302 * reap free buffers from the arc kmem caches.
2303 */
2304 if (strat == ARC_RECLAIM_AGGR)
2305 arc_shrink();
2306
2307 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2308 if (zio_buf_cache[i] != prev_cache) {
2309 prev_cache = zio_buf_cache[i];
2310 kmem_cache_reap_now(zio_buf_cache[i]);
2311 }
2312 if (zio_data_buf_cache[i] != prev_data_cache) {
2313 prev_data_cache = zio_data_buf_cache[i];
2314 kmem_cache_reap_now(zio_data_buf_cache[i]);
2315 }
2316 }
2317 kmem_cache_reap_now(buf_cache);
2318 kmem_cache_reap_now(hdr_cache);
2319
2320 /*
2321 * Ask the vmem areana to reclaim unused memory from its
2322 * quantum caches.
2323 */
2324 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR)
2325 vmem_qcache_reap(zio_arena);
2326 }
2327
2328 static void
2329 arc_reclaim_thread(void)
2330 {
2331 clock_t growtime = 0;
2332 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2333 callb_cpr_t cpr;
2334
2335 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2336
2337 mutex_enter(&arc_reclaim_thr_lock);
2338 while (arc_thread_exit == 0) {
2339 if (arc_reclaim_needed()) {
2340
2341 if (arc_no_grow) {
2342 if (last_reclaim == ARC_RECLAIM_CONS) {
2343 last_reclaim = ARC_RECLAIM_AGGR;
2344 } else {
2345 last_reclaim = ARC_RECLAIM_CONS;
2346 }
2347 } else {
2348 arc_no_grow = TRUE;
2349 last_reclaim = ARC_RECLAIM_AGGR;
2350 membar_producer();
2351 }
2352
2353 /* reset the growth delay for every reclaim */
2354 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2355
2356 arc_kmem_reap_now(last_reclaim);
2357 arc_warm = B_TRUE;
2358
2359 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2360 arc_no_grow = FALSE;
2361 }
2362
2363 arc_adjust();
2364
2365 if (arc_eviction_list != NULL)
2366 arc_do_user_evicts();
2367
2368 /* block until needed, or one second, whichever is shorter */
2369 CALLB_CPR_SAFE_BEGIN(&cpr);
2370 (void) cv_timedwait(&arc_reclaim_thr_cv,
2371 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2372 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2373 }
2374
2375 arc_thread_exit = 0;
2376 cv_broadcast(&arc_reclaim_thr_cv);
2377 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2378 thread_exit();
2379 }
2380
2381 /*
2382 * Adapt arc info given the number of bytes we are trying to add and
2383 * the state that we are comming from. This function is only called
2384 * when we are adding new content to the cache.
2385 */
2386 static void
2387 arc_adapt(int bytes, arc_state_t *state)
2388 {
2389 int mult;
2390 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2391
2392 if (state == arc_l2c_only)
2393 return;
2394
2395 ASSERT(bytes > 0);
2396 /*
2397 * Adapt the target size of the MRU list:
2398 * - if we just hit in the MRU ghost list, then increase
2399 * the target size of the MRU list.
2400 * - if we just hit in the MFU ghost list, then increase
2401 * the target size of the MFU list by decreasing the
2402 * target size of the MRU list.
2403 */
2404 if (state == arc_mru_ghost) {
2405 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2406 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2407 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2408
2409 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2410 } else if (state == arc_mfu_ghost) {
2411 uint64_t delta;
2412
2413 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2414 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2415 mult = MIN(mult, 10);
2416
2417 delta = MIN(bytes * mult, arc_p);
2418 arc_p = MAX(arc_p_min, arc_p - delta);
2419 }
2420 ASSERT((int64_t)arc_p >= 0);
2421
2422 if (arc_reclaim_needed()) {
2423 cv_signal(&arc_reclaim_thr_cv);
2424 return;
2425 }
2426
2427 if (arc_no_grow)
2428 return;
2429
2430 if (arc_c >= arc_c_max)
2431 return;
2432
2433 /*
2434 * If we're within (2 * maxblocksize) bytes of the target
2435 * cache size, increment the target cache size
2436 */
2437 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2438 atomic_add_64(&arc_c, (int64_t)bytes);
2439 if (arc_c > arc_c_max)
2440 arc_c = arc_c_max;
2441 else if (state == arc_anon)
2442 atomic_add_64(&arc_p, (int64_t)bytes);
2443 if (arc_p > arc_c)
2444 arc_p = arc_c;
2445 }
2446 ASSERT((int64_t)arc_p >= 0);
2447 }
2448
2449 /*
2450 * Check if the cache has reached its limits and eviction is required
2451 * prior to insert.
2452 */
2453 static int
2454 arc_evict_needed(arc_buf_contents_t type)
2455 {
2456 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2457 return (1);
2458
2459 if (arc_reclaim_needed())
2460 return (1);
2461
2462 return (arc_size > arc_c);
2463 }
2464
2465 /*
2466 * The buffer, supplied as the first argument, needs a data block.
2467 * So, if we are at cache max, determine which cache should be victimized.
2468 * We have the following cases:
2469 *
2470 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2471 * In this situation if we're out of space, but the resident size of the MFU is
2472 * under the limit, victimize the MFU cache to satisfy this insertion request.
2473 *
2474 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2475 * Here, we've used up all of the available space for the MRU, so we need to
2476 * evict from our own cache instead. Evict from the set of resident MRU
2477 * entries.
2478 *
2479 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2480 * c minus p represents the MFU space in the cache, since p is the size of the
2481 * cache that is dedicated to the MRU. In this situation there's still space on
2482 * the MFU side, so the MRU side needs to be victimized.
2483 *
2484 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2485 * MFU's resident set is consuming more space than it has been allotted. In
2486 * this situation, we must victimize our own cache, the MFU, for this insertion.
2487 */
2488 static void
2489 arc_get_data_buf(arc_buf_t *buf)
2490 {
2491 arc_state_t *state = buf->b_hdr->b_state;
2492 uint64_t size = buf->b_hdr->b_size;
2493 arc_buf_contents_t type = buf->b_hdr->b_type;
2494
2495 arc_adapt(size, state);
2496
2497 /*
2498 * We have not yet reached cache maximum size,
2499 * just allocate a new buffer.
2500 */
2501 if (!arc_evict_needed(type)) {
2502 if (type == ARC_BUFC_METADATA) {
2503 buf->b_data = zio_buf_alloc(size);
2504 arc_space_consume(size, ARC_SPACE_DATA);
2505 } else {
2506 ASSERT(type == ARC_BUFC_DATA);
2507 buf->b_data = zio_data_buf_alloc(size);
2508 ARCSTAT_INCR(arcstat_data_size, size);
2509 atomic_add_64(&arc_size, size);
2510 }
2511 goto out;
2512 }
2513
2514 /*
2515 * If we are prefetching from the mfu ghost list, this buffer
2516 * will end up on the mru list; so steal space from there.
2517 */
2518 if (state == arc_mfu_ghost)
2519 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2520 else if (state == arc_mru_ghost)
2521 state = arc_mru;
2522
2523 if (state == arc_mru || state == arc_anon) {
2524 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2525 state = (arc_mfu->arcs_lsize[type] >= size &&
2526 arc_p > mru_used) ? arc_mfu : arc_mru;
2527 } else {
2528 /* MFU cases */
2529 uint64_t mfu_space = arc_c - arc_p;
2530 state = (arc_mru->arcs_lsize[type] >= size &&
2531 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2532 }
2533 if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) {
2534 if (type == ARC_BUFC_METADATA) {
2535 buf->b_data = zio_buf_alloc(size);
2536 arc_space_consume(size, ARC_SPACE_DATA);
2537 } else {
2538 ASSERT(type == ARC_BUFC_DATA);
2539 buf->b_data = zio_data_buf_alloc(size);
2540 ARCSTAT_INCR(arcstat_data_size, size);
2541 atomic_add_64(&arc_size, size);
2542 }
2543 ARCSTAT_BUMP(arcstat_recycle_miss);
2544 }
2545 ASSERT(buf->b_data != NULL);
2546 out:
2547 /*
2548 * Update the state size. Note that ghost states have a
2549 * "ghost size" and so don't need to be updated.
2550 */
2551 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2552 arc_buf_hdr_t *hdr = buf->b_hdr;
2553
2554 atomic_add_64(&hdr->b_state->arcs_size, size);
2555 if (list_link_active(&hdr->b_arc_node)) {
2556 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2557 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2558 }
2559 /*
2560 * If we are growing the cache, and we are adding anonymous
2561 * data, and we have outgrown arc_p, update arc_p
2562 */
2563 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2564 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2565 arc_p = MIN(arc_c, arc_p + size);
2566 }
2567 }
2568
2569 /*
2570 * This routine is called whenever a buffer is accessed.
2571 * NOTE: the hash lock is dropped in this function.
2572 */
2573 static void
2574 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2575 {
2576 clock_t now;
2577
2578 ASSERT(MUTEX_HELD(hash_lock));
2579
2580 if (buf->b_state == arc_anon) {
2581 /*
2582 * This buffer is not in the cache, and does not
2583 * appear in our "ghost" list. Add the new buffer
2584 * to the MRU state.
2585 */
2586
2587 ASSERT(buf->b_arc_access == 0);
2588 buf->b_arc_access = ddi_get_lbolt();
2589 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2590 arc_change_state(arc_mru, buf, hash_lock);
2591
2592 } else if (buf->b_state == arc_mru) {
2593 now = ddi_get_lbolt();
2594
2595 /*
2596 * If this buffer is here because of a prefetch, then either:
2597 * - clear the flag if this is a "referencing" read
2598 * (any subsequent access will bump this into the MFU state).
2599 * or
2600 * - move the buffer to the head of the list if this is
2601 * another prefetch (to make it less likely to be evicted).
2602 */
2603 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2604 if (refcount_count(&buf->b_refcnt) == 0) {
2605 ASSERT(list_link_active(&buf->b_arc_node));
2606 } else {
2607 buf->b_flags &= ~ARC_PREFETCH;
2608 ARCSTAT_BUMP(arcstat_mru_hits);
2609 }
2610 buf->b_arc_access = now;
2611 return;
2612 }
2613
2614 /*
2615 * This buffer has been "accessed" only once so far,
2616 * but it is still in the cache. Move it to the MFU
2617 * state.
2618 */
2619 if (now > buf->b_arc_access + ARC_MINTIME) {
2620 /*
2621 * More than 125ms have passed since we
2622 * instantiated this buffer. Move it to the
2623 * most frequently used state.
2624 */
2625 buf->b_arc_access = now;
2626 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2627 arc_change_state(arc_mfu, buf, hash_lock);
2628 }
2629 ARCSTAT_BUMP(arcstat_mru_hits);
2630 } else if (buf->b_state == arc_mru_ghost) {
2631 arc_state_t *new_state;
2632 /*
2633 * This buffer has been "accessed" recently, but
2634 * was evicted from the cache. Move it to the
2635 * MFU state.
2636 */
2637
2638 if (buf->b_flags & ARC_PREFETCH) {
2639 new_state = arc_mru;
2640 if (refcount_count(&buf->b_refcnt) > 0)
2641 buf->b_flags &= ~ARC_PREFETCH;
2642 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2643 } else {
2644 new_state = arc_mfu;
2645 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2646 }
2647
2648 buf->b_arc_access = ddi_get_lbolt();
2649 arc_change_state(new_state, buf, hash_lock);
2650
2651 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2652 } else if (buf->b_state == arc_mfu) {
2653 /*
2654 * This buffer has been accessed more than once and is
2655 * still in the cache. Keep it in the MFU state.
2656 *
2657 * NOTE: an add_reference() that occurred when we did
2658 * the arc_read() will have kicked this off the list.
2659 * If it was a prefetch, we will explicitly move it to
2660 * the head of the list now.
2661 */
2662 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2663 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2664 ASSERT(list_link_active(&buf->b_arc_node));
2665 }
2666 ARCSTAT_BUMP(arcstat_mfu_hits);
2667 buf->b_arc_access = ddi_get_lbolt();
2668 } else if (buf->b_state == arc_mfu_ghost) {
2669 arc_state_t *new_state = arc_mfu;
2670 /*
2671 * This buffer has been accessed more than once but has
2672 * been evicted from the cache. Move it back to the
2673 * MFU state.
2674 */
2675
2676 if (buf->b_flags & ARC_PREFETCH) {
2677 /*
2678 * This is a prefetch access...
2679 * move this block back to the MRU state.
2680 */
2681 ASSERT0(refcount_count(&buf->b_refcnt));
2682 new_state = arc_mru;
2683 }
2684
2685 buf->b_arc_access = ddi_get_lbolt();
2686 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2687 arc_change_state(new_state, buf, hash_lock);
2688
2689 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2690 } else if (buf->b_state == arc_l2c_only) {
2691 /*
2692 * This buffer is on the 2nd Level ARC.
2693 */
2694
2695 buf->b_arc_access = ddi_get_lbolt();
2696 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2697 arc_change_state(arc_mfu, buf, hash_lock);
2698 } else {
2699 ASSERT(!"invalid arc state");
2700 }
2701 }
2702
2703 /* a generic arc_done_func_t which you can use */
2704 /* ARGSUSED */
2705 void
2706 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2707 {
2708 if (zio == NULL || zio->io_error == 0)
2709 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2710 VERIFY(arc_buf_remove_ref(buf, arg));
2711 }
2712
2713 /* a generic arc_done_func_t */
2714 void
2715 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2716 {
2717 arc_buf_t **bufp = arg;
2718 if (zio && zio->io_error) {
2719 VERIFY(arc_buf_remove_ref(buf, arg));
2720 *bufp = NULL;
2721 } else {
2722 *bufp = buf;
2723 ASSERT(buf->b_data);
2724 }
2725 }
2726
2727 static void
2728 arc_read_done(zio_t *zio)
2729 {
2730 arc_buf_hdr_t *hdr, *found;
2731 arc_buf_t *buf;
2732 arc_buf_t *abuf; /* buffer we're assigning to callback */
2733 kmutex_t *hash_lock;
2734 arc_callback_t *callback_list, *acb;
2735 int freeable = FALSE;
2736
2737 buf = zio->io_private;
2738 hdr = buf->b_hdr;
2739
2740 /*
2741 * The hdr was inserted into hash-table and removed from lists
2742 * prior to starting I/O. We should find this header, since
2743 * it's in the hash table, and it should be legit since it's
2744 * not possible to evict it during the I/O. The only possible
2745 * reason for it not to be found is if we were freed during the
2746 * read.
2747 */
2748 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2749 &hash_lock);
2750
2751 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2752 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2753 (found == hdr && HDR_L2_READING(hdr)));
2754
2755 hdr->b_flags &= ~ARC_L2_EVICTED;
2756 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2757 hdr->b_flags &= ~ARC_L2CACHE;
2758
2759 /* byteswap if necessary */
2760 callback_list = hdr->b_acb;
2761 ASSERT(callback_list != NULL);
2762 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2763 dmu_object_byteswap_t bswap =
2764 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2765 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2766 byteswap_uint64_array :
2767 dmu_ot_byteswap[bswap].ob_func;
2768 func(buf->b_data, hdr->b_size);
2769 }
2770
2771 arc_cksum_compute(buf, B_FALSE);
2772 arc_buf_watch(buf);
2773
2774 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2775 /*
2776 * Only call arc_access on anonymous buffers. This is because
2777 * if we've issued an I/O for an evicted buffer, we've already
2778 * called arc_access (to prevent any simultaneous readers from
2779 * getting confused).
2780 */
2781 arc_access(hdr, hash_lock);
2782 }
2783
2784 /* create copies of the data buffer for the callers */
2785 abuf = buf;
2786 for (acb = callback_list; acb; acb = acb->acb_next) {
2787 if (acb->acb_done) {
2788 if (abuf == NULL) {
2789 ARCSTAT_BUMP(arcstat_duplicate_reads);
2790 abuf = arc_buf_clone(buf);
2791 }
2792 acb->acb_buf = abuf;
2793 abuf = NULL;
2794 }
2795 }
2796 hdr->b_acb = NULL;
2797 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2798 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2799 if (abuf == buf) {
2800 ASSERT(buf->b_efunc == NULL);
2801 ASSERT(hdr->b_datacnt == 1);
2802 hdr->b_flags |= ARC_BUF_AVAILABLE;
2803 }
2804
2805 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2806
2807 if (zio->io_error != 0) {
2808 hdr->b_flags |= ARC_IO_ERROR;
2809 if (hdr->b_state != arc_anon)
2810 arc_change_state(arc_anon, hdr, hash_lock);
2811 if (HDR_IN_HASH_TABLE(hdr))
2812 buf_hash_remove(hdr);
2813 freeable = refcount_is_zero(&hdr->b_refcnt);
2814 }
2815
2816 /*
2817 * Broadcast before we drop the hash_lock to avoid the possibility
2818 * that the hdr (and hence the cv) might be freed before we get to
2819 * the cv_broadcast().
2820 */
2821 cv_broadcast(&hdr->b_cv);
2822
2823 if (hash_lock) {
2824 mutex_exit(hash_lock);
2825 } else {
2826 /*
2827 * This block was freed while we waited for the read to
2828 * complete. It has been removed from the hash table and
2829 * moved to the anonymous state (so that it won't show up
2830 * in the cache).
2831 */
2832 ASSERT3P(hdr->b_state, ==, arc_anon);
2833 freeable = refcount_is_zero(&hdr->b_refcnt);
2834 }
2835
2836 /* execute each callback and free its structure */
2837 while ((acb = callback_list) != NULL) {
2838 if (acb->acb_done)
2839 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2840
2841 if (acb->acb_zio_dummy != NULL) {
2842 acb->acb_zio_dummy->io_error = zio->io_error;
2843 zio_nowait(acb->acb_zio_dummy);
2844 }
2845
2846 callback_list = acb->acb_next;
2847 kmem_free(acb, sizeof (arc_callback_t));
2848 }
2849
2850 if (freeable)
2851 arc_hdr_destroy(hdr);
2852 }
2853
2854 /*
2855 * "Read" the block at the specified DVA (in bp) via the
2856 * cache. If the block is found in the cache, invoke the provided
2857 * callback immediately and return. Note that the `zio' parameter
2858 * in the callback will be NULL in this case, since no IO was
2859 * required. If the block is not in the cache pass the read request
2860 * on to the spa with a substitute callback function, so that the
2861 * requested block will be added to the cache.
2862 *
2863 * If a read request arrives for a block that has a read in-progress,
2864 * either wait for the in-progress read to complete (and return the
2865 * results); or, if this is a read with a "done" func, add a record
2866 * to the read to invoke the "done" func when the read completes,
2867 * and return; or just return.
2868 *
2869 * arc_read_done() will invoke all the requested "done" functions
2870 * for readers of this block.
2871 */
2872 int
2873 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
2874 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags,
2875 const zbookmark_t *zb)
2876 {
2877 arc_buf_hdr_t *hdr;
2878 arc_buf_t *buf = NULL;
2879 kmutex_t *hash_lock;
2880 zio_t *rzio;
2881 uint64_t guid = spa_load_guid(spa);
2882
2883 top:
2884 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2885 &hash_lock);
2886 if (hdr && hdr->b_datacnt > 0) {
2887
2888 *arc_flags |= ARC_CACHED;
2889
2890 if (HDR_IO_IN_PROGRESS(hdr)) {
2891
2892 if (*arc_flags & ARC_WAIT) {
2893 cv_wait(&hdr->b_cv, hash_lock);
2894 mutex_exit(hash_lock);
2895 goto top;
2896 }
2897 ASSERT(*arc_flags & ARC_NOWAIT);
2898
2899 if (done) {
2900 arc_callback_t *acb = NULL;
2901
2902 acb = kmem_zalloc(sizeof (arc_callback_t),
2903 KM_SLEEP);
2904 acb->acb_done = done;
2905 acb->acb_private = private;
2906 if (pio != NULL)
2907 acb->acb_zio_dummy = zio_null(pio,
2908 spa, NULL, NULL, NULL, zio_flags);
2909
2910 ASSERT(acb->acb_done != NULL);
2911 acb->acb_next = hdr->b_acb;
2912 hdr->b_acb = acb;
2913 add_reference(hdr, hash_lock, private);
2914 mutex_exit(hash_lock);
2915 return (0);
2916 }
2917 mutex_exit(hash_lock);
2918 return (0);
2919 }
2920
2921 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2922
2923 if (done) {
2924 add_reference(hdr, hash_lock, private);
2925 /*
2926 * If this block is already in use, create a new
2927 * copy of the data so that we will be guaranteed
2928 * that arc_release() will always succeed.
2929 */
2930 buf = hdr->b_buf;
2931 ASSERT(buf);
2932 ASSERT(buf->b_data);
2933 if (HDR_BUF_AVAILABLE(hdr)) {
2934 ASSERT(buf->b_efunc == NULL);
2935 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2936 } else {
2937 buf = arc_buf_clone(buf);
2938 }
2939
2940 } else if (*arc_flags & ARC_PREFETCH &&
2941 refcount_count(&hdr->b_refcnt) == 0) {
2942 hdr->b_flags |= ARC_PREFETCH;
2943 }
2944 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2945 arc_access(hdr, hash_lock);
2946 if (*arc_flags & ARC_L2CACHE)
2947 hdr->b_flags |= ARC_L2CACHE;
2948 if (*arc_flags & ARC_L2COMPRESS)
2949 hdr->b_flags |= ARC_L2COMPRESS;
2950 mutex_exit(hash_lock);
2951 ARCSTAT_BUMP(arcstat_hits);
2952 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2953 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2954 data, metadata, hits);
2955
2956 if (done)
2957 done(NULL, buf, private);
2958 } else {
2959 uint64_t size = BP_GET_LSIZE(bp);
2960 arc_callback_t *acb;
2961 vdev_t *vd = NULL;
2962 uint64_t addr = 0;
2963 boolean_t devw = B_FALSE;
2964
2965 if (hdr == NULL) {
2966 /* this block is not in the cache */
2967 arc_buf_hdr_t *exists;
2968 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2969 buf = arc_buf_alloc(spa, size, private, type);
2970 hdr = buf->b_hdr;
2971 hdr->b_dva = *BP_IDENTITY(bp);
2972 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
2973 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2974 exists = buf_hash_insert(hdr, &hash_lock);
2975 if (exists) {
2976 /* somebody beat us to the hash insert */
2977 mutex_exit(hash_lock);
2978 buf_discard_identity(hdr);
2979 (void) arc_buf_remove_ref(buf, private);
2980 goto top; /* restart the IO request */
2981 }
2982 /* if this is a prefetch, we don't have a reference */
2983 if (*arc_flags & ARC_PREFETCH) {
2984 (void) remove_reference(hdr, hash_lock,
2985 private);
2986 hdr->b_flags |= ARC_PREFETCH;
2987 }
2988 if (*arc_flags & ARC_L2CACHE)
2989 hdr->b_flags |= ARC_L2CACHE;
2990 if (*arc_flags & ARC_L2COMPRESS)
2991 hdr->b_flags |= ARC_L2COMPRESS;
2992 if (BP_GET_LEVEL(bp) > 0)
2993 hdr->b_flags |= ARC_INDIRECT;
2994 } else {
2995 /* this block is in the ghost cache */
2996 ASSERT(GHOST_STATE(hdr->b_state));
2997 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2998 ASSERT0(refcount_count(&hdr->b_refcnt));
2999 ASSERT(hdr->b_buf == NULL);
3000
3001 /* if this is a prefetch, we don't have a reference */
3002 if (*arc_flags & ARC_PREFETCH)
3003 hdr->b_flags |= ARC_PREFETCH;
3004 else
3005 add_reference(hdr, hash_lock, private);
3006 if (*arc_flags & ARC_L2CACHE)
3007 hdr->b_flags |= ARC_L2CACHE;
3008 if (*arc_flags & ARC_L2COMPRESS)
3009 hdr->b_flags |= ARC_L2COMPRESS;
3010 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3011 buf->b_hdr = hdr;
3012 buf->b_data = NULL;
3013 buf->b_efunc = NULL;
3014 buf->b_private = NULL;
3015 buf->b_next = NULL;
3016 hdr->b_buf = buf;
3017 ASSERT(hdr->b_datacnt == 0);
3018 hdr->b_datacnt = 1;
3019 arc_get_data_buf(buf);
3020 arc_access(hdr, hash_lock);
3021 }
3022
3023 ASSERT(!GHOST_STATE(hdr->b_state));
3024
3025 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
3026 acb->acb_done = done;
3027 acb->acb_private = private;
3028
3029 ASSERT(hdr->b_acb == NULL);
3030 hdr->b_acb = acb;
3031 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3032
3033 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
3034 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3035 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3036 addr = hdr->b_l2hdr->b_daddr;
3037 /*
3038 * Lock out device removal.
3039 */
3040 if (vdev_is_dead(vd) ||
3041 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3042 vd = NULL;
3043 }
3044
3045 mutex_exit(hash_lock);
3046
3047 /*
3048 * At this point, we have a level 1 cache miss. Try again in
3049 * L2ARC if possible.
3050 */
3051 ASSERT3U(hdr->b_size, ==, size);
3052 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3053 uint64_t, size, zbookmark_t *, zb);
3054 ARCSTAT_BUMP(arcstat_misses);
3055 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3056 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3057 data, metadata, misses);
3058
3059 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3060 /*
3061 * Read from the L2ARC if the following are true:
3062 * 1. The L2ARC vdev was previously cached.
3063 * 2. This buffer still has L2ARC metadata.
3064 * 3. This buffer isn't currently writing to the L2ARC.
3065 * 4. The L2ARC entry wasn't evicted, which may
3066 * also have invalidated the vdev.
3067 * 5. This isn't prefetch and l2arc_noprefetch is set.
3068 */
3069 if (hdr->b_l2hdr != NULL &&
3070 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3071 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3072 l2arc_read_callback_t *cb;
3073
3074 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3075 ARCSTAT_BUMP(arcstat_l2_hits);
3076
3077 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3078 KM_SLEEP);
3079 cb->l2rcb_buf = buf;
3080 cb->l2rcb_spa = spa;
3081 cb->l2rcb_bp = *bp;
3082 cb->l2rcb_zb = *zb;
3083 cb->l2rcb_flags = zio_flags;
3084 cb->l2rcb_compress = hdr->b_l2hdr->b_compress;
3085
3086 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
3087 addr + size < vd->vdev_psize -
3088 VDEV_LABEL_END_SIZE);
3089
3090 /*
3091 * l2arc read. The SCL_L2ARC lock will be
3092 * released by l2arc_read_done().
3093 * Issue a null zio if the underlying buffer
3094 * was squashed to zero size by compression.
3095 */
3096 if (hdr->b_l2hdr->b_compress ==
3097 ZIO_COMPRESS_EMPTY) {
3098 rzio = zio_null(pio, spa, vd,
3099 l2arc_read_done, cb,
3100 zio_flags | ZIO_FLAG_DONT_CACHE |
3101 ZIO_FLAG_CANFAIL |
3102 ZIO_FLAG_DONT_PROPAGATE |
3103 ZIO_FLAG_DONT_RETRY);
3104 } else {
3105 rzio = zio_read_phys(pio, vd, addr,
3106 hdr->b_l2hdr->b_asize,
3107 buf->b_data, ZIO_CHECKSUM_OFF,
3108 l2arc_read_done, cb, priority,
3109 zio_flags | ZIO_FLAG_DONT_CACHE |
3110 ZIO_FLAG_CANFAIL |
3111 ZIO_FLAG_DONT_PROPAGATE |
3112 ZIO_FLAG_DONT_RETRY, B_FALSE);
3113 }
3114 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3115 zio_t *, rzio);
3116 ARCSTAT_INCR(arcstat_l2_read_bytes,
3117 hdr->b_l2hdr->b_asize);
3118
3119 if (*arc_flags & ARC_NOWAIT) {
3120 zio_nowait(rzio);
3121 return (0);
3122 }
3123
3124 ASSERT(*arc_flags & ARC_WAIT);
3125 if (zio_wait(rzio) == 0)
3126 return (0);
3127
3128 /* l2arc read error; goto zio_read() */
3129 } else {
3130 DTRACE_PROBE1(l2arc__miss,
3131 arc_buf_hdr_t *, hdr);
3132 ARCSTAT_BUMP(arcstat_l2_misses);
3133 if (HDR_L2_WRITING(hdr))
3134 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3135 spa_config_exit(spa, SCL_L2ARC, vd);
3136 }
3137 } else {
3138 if (vd != NULL)
3139 spa_config_exit(spa, SCL_L2ARC, vd);
3140 if (l2arc_ndev != 0) {
3141 DTRACE_PROBE1(l2arc__miss,
3142 arc_buf_hdr_t *, hdr);
3143 ARCSTAT_BUMP(arcstat_l2_misses);
3144 }
3145 }
3146
3147 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3148 arc_read_done, buf, priority, zio_flags, zb);
3149
3150 if (*arc_flags & ARC_WAIT)
3151 return (zio_wait(rzio));
3152
3153 ASSERT(*arc_flags & ARC_NOWAIT);
3154 zio_nowait(rzio);
3155 }
3156 return (0);
3157 }
3158
3159 void
3160 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3161 {
3162 ASSERT(buf->b_hdr != NULL);
3163 ASSERT(buf->b_hdr->b_state != arc_anon);
3164 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3165 ASSERT(buf->b_efunc == NULL);
3166 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3167
3168 buf->b_efunc = func;
3169 buf->b_private = private;
3170 }
3171
3172 /*
3173 * Notify the arc that a block was freed, and thus will never be used again.
3174 */
3175 void
3176 arc_freed(spa_t *spa, const blkptr_t *bp)
3177 {
3178 arc_buf_hdr_t *hdr;
3179 kmutex_t *hash_lock;
3180 uint64_t guid = spa_load_guid(spa);
3181
3182 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3183 &hash_lock);
3184 if (hdr == NULL)
3185 return;
3186 if (HDR_BUF_AVAILABLE(hdr)) {
3187 arc_buf_t *buf = hdr->b_buf;
3188 add_reference(hdr, hash_lock, FTAG);
3189 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3190 mutex_exit(hash_lock);
3191
3192 arc_release(buf, FTAG);
3193 (void) arc_buf_remove_ref(buf, FTAG);
3194 } else {
3195 mutex_exit(hash_lock);
3196 }
3197
3198 }
3199
3200 /*
3201 * This is used by the DMU to let the ARC know that a buffer is
3202 * being evicted, so the ARC should clean up. If this arc buf
3203 * is not yet in the evicted state, it will be put there.
3204 */
3205 int
3206 arc_buf_evict(arc_buf_t *buf)
3207 {
3208 arc_buf_hdr_t *hdr;
3209 kmutex_t *hash_lock;
3210 arc_buf_t **bufp;
3211
3212 mutex_enter(&buf->b_evict_lock);
3213 hdr = buf->b_hdr;
3214 if (hdr == NULL) {
3215 /*
3216 * We are in arc_do_user_evicts().
3217 */
3218 ASSERT(buf->b_data == NULL);
3219 mutex_exit(&buf->b_evict_lock);
3220 return (0);
3221 } else if (buf->b_data == NULL) {
3222 arc_buf_t copy = *buf; /* structure assignment */
3223 /*
3224 * We are on the eviction list; process this buffer now
3225 * but let arc_do_user_evicts() do the reaping.
3226 */
3227 buf->b_efunc = NULL;
3228 mutex_exit(&buf->b_evict_lock);
3229 VERIFY(copy.b_efunc(©) == 0);
3230 return (1);
3231 }
3232 hash_lock = HDR_LOCK(hdr);
3233 mutex_enter(hash_lock);
3234 hdr = buf->b_hdr;
3235 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3236
3237 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3238 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3239
3240 /*
3241 * Pull this buffer off of the hdr
3242 */
3243 bufp = &hdr->b_buf;
3244 while (*bufp != buf)
3245 bufp = &(*bufp)->b_next;
3246 *bufp = buf->b_next;
3247
3248 ASSERT(buf->b_data != NULL);
3249 arc_buf_destroy(buf, FALSE, FALSE);
3250
3251 if (hdr->b_datacnt == 0) {
3252 arc_state_t *old_state = hdr->b_state;
3253 arc_state_t *evicted_state;
3254
3255 ASSERT(hdr->b_buf == NULL);
3256 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3257
3258 evicted_state =
3259 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3260
3261 mutex_enter(&old_state->arcs_mtx);
3262 mutex_enter(&evicted_state->arcs_mtx);
3263
3264 arc_change_state(evicted_state, hdr, hash_lock);
3265 ASSERT(HDR_IN_HASH_TABLE(hdr));
3266 hdr->b_flags |= ARC_IN_HASH_TABLE;
3267 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3268
3269 mutex_exit(&evicted_state->arcs_mtx);
3270 mutex_exit(&old_state->arcs_mtx);
3271 }
3272 mutex_exit(hash_lock);
3273 mutex_exit(&buf->b_evict_lock);
3274
3275 VERIFY(buf->b_efunc(buf) == 0);
3276 buf->b_efunc = NULL;
3277 buf->b_private = NULL;
3278 buf->b_hdr = NULL;
3279 buf->b_next = NULL;
3280 kmem_cache_free(buf_cache, buf);
3281 return (1);
3282 }
3283
3284 /*
3285 * Release this buffer from the cache, making it an anonymous buffer. This
3286 * must be done after a read and prior to modifying the buffer contents.
3287 * If the buffer has more than one reference, we must make
3288 * a new hdr for the buffer.
3289 */
3290 void
3291 arc_release(arc_buf_t *buf, void *tag)
3292 {
3293 arc_buf_hdr_t *hdr;
3294 kmutex_t *hash_lock = NULL;
3295 l2arc_buf_hdr_t *l2hdr;
3296 uint64_t buf_size;
3297
3298 /*
3299 * It would be nice to assert that if it's DMU metadata (level >
3300 * 0 || it's the dnode file), then it must be syncing context.
3301 * But we don't know that information at this level.
3302 */
3303
3304 mutex_enter(&buf->b_evict_lock);
3305 hdr = buf->b_hdr;
3306
3307 /* this buffer is not on any list */
3308 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3309
3310 if (hdr->b_state == arc_anon) {
3311 /* this buffer is already released */
3312 ASSERT(buf->b_efunc == NULL);
3313 } else {
3314 hash_lock = HDR_LOCK(hdr);
3315 mutex_enter(hash_lock);
3316 hdr = buf->b_hdr;
3317 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3318 }
3319
3320 l2hdr = hdr->b_l2hdr;
3321 if (l2hdr) {
3322 mutex_enter(&l2arc_buflist_mtx);
3323 hdr->b_l2hdr = NULL;
3324 }
3325 buf_size = hdr->b_size;
3326
3327 /*
3328 * Do we have more than one buf?
3329 */
3330 if (hdr->b_datacnt > 1) {
3331 arc_buf_hdr_t *nhdr;
3332 arc_buf_t **bufp;
3333 uint64_t blksz = hdr->b_size;
3334 uint64_t spa = hdr->b_spa;
3335 arc_buf_contents_t type = hdr->b_type;
3336 uint32_t flags = hdr->b_flags;
3337
3338 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3339 /*
3340 * Pull the data off of this hdr and attach it to
3341 * a new anonymous hdr.
3342 */
3343 (void) remove_reference(hdr, hash_lock, tag);
3344 bufp = &hdr->b_buf;
3345 while (*bufp != buf)
3346 bufp = &(*bufp)->b_next;
3347 *bufp = buf->b_next;
3348 buf->b_next = NULL;
3349
3350 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3351 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3352 if (refcount_is_zero(&hdr->b_refcnt)) {
3353 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3354 ASSERT3U(*size, >=, hdr->b_size);
3355 atomic_add_64(size, -hdr->b_size);
3356 }
3357
3358 /*
3359 * We're releasing a duplicate user data buffer, update
3360 * our statistics accordingly.
3361 */
3362 if (hdr->b_type == ARC_BUFC_DATA) {
3363 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3364 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3365 -hdr->b_size);
3366 }
3367 hdr->b_datacnt -= 1;
3368 arc_cksum_verify(buf);
3369 arc_buf_unwatch(buf);
3370
3371 mutex_exit(hash_lock);
3372
3373 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3374 nhdr->b_size = blksz;
3375 nhdr->b_spa = spa;
3376 nhdr->b_type = type;
3377 nhdr->b_buf = buf;
3378 nhdr->b_state = arc_anon;
3379 nhdr->b_arc_access = 0;
3380 nhdr->b_flags = flags & ARC_L2_WRITING;
3381 nhdr->b_l2hdr = NULL;
3382 nhdr->b_datacnt = 1;
3383 nhdr->b_freeze_cksum = NULL;
3384 (void) refcount_add(&nhdr->b_refcnt, tag);
3385 buf->b_hdr = nhdr;
3386 mutex_exit(&buf->b_evict_lock);
3387 atomic_add_64(&arc_anon->arcs_size, blksz);
3388 } else {
3389 mutex_exit(&buf->b_evict_lock);
3390 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3391 ASSERT(!list_link_active(&hdr->b_arc_node));
3392 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3393 if (hdr->b_state != arc_anon)
3394 arc_change_state(arc_anon, hdr, hash_lock);
3395 hdr->b_arc_access = 0;
3396 if (hash_lock)
3397 mutex_exit(hash_lock);
3398
3399 buf_discard_identity(hdr);
3400 arc_buf_thaw(buf);
3401 }
3402 buf->b_efunc = NULL;
3403 buf->b_private = NULL;
3404
3405 if (l2hdr) {
3406 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
3407 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3408 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3409 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3410 mutex_exit(&l2arc_buflist_mtx);
3411 }
3412 }
3413
3414 int
3415 arc_released(arc_buf_t *buf)
3416 {
3417 int released;
3418
3419 mutex_enter(&buf->b_evict_lock);
3420 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3421 mutex_exit(&buf->b_evict_lock);
3422 return (released);
3423 }
3424
3425 int
3426 arc_has_callback(arc_buf_t *buf)
3427 {
3428 int callback;
3429
3430 mutex_enter(&buf->b_evict_lock);
3431 callback = (buf->b_efunc != NULL);
3432 mutex_exit(&buf->b_evict_lock);
3433 return (callback);
3434 }
3435
3436 #ifdef ZFS_DEBUG
3437 int
3438 arc_referenced(arc_buf_t *buf)
3439 {
3440 int referenced;
3441
3442 mutex_enter(&buf->b_evict_lock);
3443 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3444 mutex_exit(&buf->b_evict_lock);
3445 return (referenced);
3446 }
3447 #endif
3448
3449 static void
3450 arc_write_ready(zio_t *zio)
3451 {
3452 arc_write_callback_t *callback = zio->io_private;
3453 arc_buf_t *buf = callback->awcb_buf;
3454 arc_buf_hdr_t *hdr = buf->b_hdr;
3455
3456 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3457 callback->awcb_ready(zio, buf, callback->awcb_private);
3458
3459 /*
3460 * If the IO is already in progress, then this is a re-write
3461 * attempt, so we need to thaw and re-compute the cksum.
3462 * It is the responsibility of the callback to handle the
3463 * accounting for any re-write attempt.
3464 */
3465 if (HDR_IO_IN_PROGRESS(hdr)) {
3466 mutex_enter(&hdr->b_freeze_lock);
3467 if (hdr->b_freeze_cksum != NULL) {
3468 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3469 hdr->b_freeze_cksum = NULL;
3470 }
3471 mutex_exit(&hdr->b_freeze_lock);
3472 }
3473 arc_cksum_compute(buf, B_FALSE);
3474 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3475 }
3476
3477 /*
3478 * The SPA calls this callback for each physical write that happens on behalf
3479 * of a logical write. See the comment in dbuf_write_physdone() for details.
3480 */
3481 static void
3482 arc_write_physdone(zio_t *zio)
3483 {
3484 arc_write_callback_t *cb = zio->io_private;
3485 if (cb->awcb_physdone != NULL)
3486 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
3487 }
3488
3489 static void
3490 arc_write_done(zio_t *zio)
3491 {
3492 arc_write_callback_t *callback = zio->io_private;
3493 arc_buf_t *buf = callback->awcb_buf;
3494 arc_buf_hdr_t *hdr = buf->b_hdr;
3495
3496 ASSERT(hdr->b_acb == NULL);
3497
3498 if (zio->io_error == 0) {
3499 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3500 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3501 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3502 } else {
3503 ASSERT(BUF_EMPTY(hdr));
3504 }
3505
3506 /*
3507 * If the block to be written was all-zero, we may have
3508 * compressed it away. In this case no write was performed
3509 * so there will be no dva/birth/checksum. The buffer must
3510 * therefore remain anonymous (and uncached).
3511 */
3512 if (!BUF_EMPTY(hdr)) {
3513 arc_buf_hdr_t *exists;
3514 kmutex_t *hash_lock;
3515
3516 ASSERT(zio->io_error == 0);
3517
3518 arc_cksum_verify(buf);
3519
3520 exists = buf_hash_insert(hdr, &hash_lock);
3521 if (exists) {
3522 /*
3523 * This can only happen if we overwrite for
3524 * sync-to-convergence, because we remove
3525 * buffers from the hash table when we arc_free().
3526 */
3527 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3528 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3529 panic("bad overwrite, hdr=%p exists=%p",
3530 (void *)hdr, (void *)exists);
3531 ASSERT(refcount_is_zero(&exists->b_refcnt));
3532 arc_change_state(arc_anon, exists, hash_lock);
3533 mutex_exit(hash_lock);
3534 arc_hdr_destroy(exists);
3535 exists = buf_hash_insert(hdr, &hash_lock);
3536 ASSERT3P(exists, ==, NULL);
3537 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3538 /* nopwrite */
3539 ASSERT(zio->io_prop.zp_nopwrite);
3540 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3541 panic("bad nopwrite, hdr=%p exists=%p",
3542 (void *)hdr, (void *)exists);
3543 } else {
3544 /* Dedup */
3545 ASSERT(hdr->b_datacnt == 1);
3546 ASSERT(hdr->b_state == arc_anon);
3547 ASSERT(BP_GET_DEDUP(zio->io_bp));
3548 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3549 }
3550 }
3551 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3552 /* if it's not anon, we are doing a scrub */
3553 if (!exists && hdr->b_state == arc_anon)
3554 arc_access(hdr, hash_lock);
3555 mutex_exit(hash_lock);
3556 } else {
3557 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3558 }
3559
3560 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3561 callback->awcb_done(zio, buf, callback->awcb_private);
3562
3563 kmem_free(callback, sizeof (arc_write_callback_t));
3564 }
3565
3566 zio_t *
3567 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3568 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
3569 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
3570 arc_done_func_t *done, void *private, zio_priority_t priority,
3571 int zio_flags, const zbookmark_t *zb)
3572 {
3573 arc_buf_hdr_t *hdr = buf->b_hdr;
3574 arc_write_callback_t *callback;
3575 zio_t *zio;
3576
3577 ASSERT(ready != NULL);
3578 ASSERT(done != NULL);
3579 ASSERT(!HDR_IO_ERROR(hdr));
3580 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3581 ASSERT(hdr->b_acb == NULL);
3582 if (l2arc)
3583 hdr->b_flags |= ARC_L2CACHE;
3584 if (l2arc_compress)
3585 hdr->b_flags |= ARC_L2COMPRESS;
3586 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3587 callback->awcb_ready = ready;
3588 callback->awcb_physdone = physdone;
3589 callback->awcb_done = done;
3590 callback->awcb_private = private;
3591 callback->awcb_buf = buf;
3592
3593 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3594 arc_write_ready, arc_write_physdone, arc_write_done, callback,
3595 priority, zio_flags, zb);
3596
3597 return (zio);
3598 }
3599
3600 static int
3601 arc_memory_throttle(uint64_t reserve, uint64_t txg)
3602 {
3603 #ifdef _KERNEL
3604 uint64_t available_memory = ptob(freemem);
3605 static uint64_t page_load = 0;
3606 static uint64_t last_txg = 0;
3607
3608 #if defined(__i386)
3609 available_memory =
3610 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3611 #endif
3612
3613 if (freemem > physmem * arc_lotsfree_percent / 100)
3614 return (0);
3615
3616 if (txg > last_txg) {
3617 last_txg = txg;
3618 page_load = 0;
3619 }
3620 /*
3621 * If we are in pageout, we know that memory is already tight,
3622 * the arc is already going to be evicting, so we just want to
3623 * continue to let page writes occur as quickly as possible.
3624 */
3625 if (curproc == proc_pageout) {
3626 if (page_load > MAX(ptob(minfree), available_memory) / 4)
3627 return (SET_ERROR(ERESTART));
3628 /* Note: reserve is inflated, so we deflate */
3629 page_load += reserve / 8;
3630 return (0);
3631 } else if (page_load > 0 && arc_reclaim_needed()) {
3632 /* memory is low, delay before restarting */
3633 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3634 return (SET_ERROR(EAGAIN));
3635 }
3636 page_load = 0;
3637 #endif
3638 return (0);
3639 }
3640
3641 void
3642 arc_tempreserve_clear(uint64_t reserve)
3643 {
3644 atomic_add_64(&arc_tempreserve, -reserve);
3645 ASSERT((int64_t)arc_tempreserve >= 0);
3646 }
3647
3648 int
3649 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3650 {
3651 int error;
3652 uint64_t anon_size;
3653
3654 if (reserve > arc_c/4 && !arc_no_grow)
3655 arc_c = MIN(arc_c_max, reserve * 4);
3656 if (reserve > arc_c)
3657 return (SET_ERROR(ENOMEM));
3658
3659 /*
3660 * Don't count loaned bufs as in flight dirty data to prevent long
3661 * network delays from blocking transactions that are ready to be
3662 * assigned to a txg.
3663 */
3664 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3665
3666 /*
3667 * Writes will, almost always, require additional memory allocations
3668 * in order to compress/encrypt/etc the data. We therefore need to
3669 * make sure that there is sufficient available memory for this.
3670 */
3671 error = arc_memory_throttle(reserve, txg);
3672 if (error != 0)
3673 return (error);
3674
3675 /*
3676 * Throttle writes when the amount of dirty data in the cache
3677 * gets too large. We try to keep the cache less than half full
3678 * of dirty blocks so that our sync times don't grow too large.
3679 * Note: if two requests come in concurrently, we might let them
3680 * both succeed, when one of them should fail. Not a huge deal.
3681 */
3682
3683 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3684 anon_size > arc_c / 4) {
3685 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3686 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3687 arc_tempreserve>>10,
3688 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3689 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3690 reserve>>10, arc_c>>10);
3691 return (SET_ERROR(ERESTART));
3692 }
3693 atomic_add_64(&arc_tempreserve, reserve);
3694 return (0);
3695 }
3696
3697 void
3698 arc_init(void)
3699 {
3700 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3701 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3702
3703 /* Convert seconds to clock ticks */
3704 arc_min_prefetch_lifespan = 1 * hz;
3705
3706 /* Start out with 1/8 of all memory */
3707 arc_c = physmem * PAGESIZE / 8;
3708
3709 #ifdef _KERNEL
3710 /*
3711 * On architectures where the physical memory can be larger
3712 * than the addressable space (intel in 32-bit mode), we may
3713 * need to limit the cache to 1/8 of VM size.
3714 */
3715 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3716 #endif
3717
3718 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3719 arc_c_min = MAX(arc_c / 4, 64<<20);
3720 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
3721 if (arc_c * 8 >= 1<<30)
3722 arc_c_max = (arc_c * 8) - (1<<30);
3723 else
3724 arc_c_max = arc_c_min;
3725 arc_c_max = MAX(arc_c * 6, arc_c_max);
3726
3727 /*
3728 * Allow the tunables to override our calculations if they are
3729 * reasonable (ie. over 64MB)
3730 */
3731 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3732 arc_c_max = zfs_arc_max;
3733 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3734 arc_c_min = zfs_arc_min;
3735
3736 arc_c = arc_c_max;
3737 arc_p = (arc_c >> 1);
3738
3739 /* limit meta-data to 1/4 of the arc capacity */
3740 arc_meta_limit = arc_c_max / 4;
3741
3742 /* Allow the tunable to override if it is reasonable */
3743 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3744 arc_meta_limit = zfs_arc_meta_limit;
3745
3746 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3747 arc_c_min = arc_meta_limit / 2;
3748
3749 if (zfs_arc_grow_retry > 0)
3750 arc_grow_retry = zfs_arc_grow_retry;
3751
3752 if (zfs_arc_shrink_shift > 0)
3753 arc_shrink_shift = zfs_arc_shrink_shift;
3754
3755 if (zfs_arc_p_min_shift > 0)
3756 arc_p_min_shift = zfs_arc_p_min_shift;
3757
3758 /* if kmem_flags are set, lets try to use less memory */
3759 if (kmem_debugging())
3760 arc_c = arc_c / 2;
3761 if (arc_c < arc_c_min)
3762 arc_c = arc_c_min;
3763
3764 arc_anon = &ARC_anon;
3765 arc_mru = &ARC_mru;
3766 arc_mru_ghost = &ARC_mru_ghost;
3767 arc_mfu = &ARC_mfu;
3768 arc_mfu_ghost = &ARC_mfu_ghost;
3769 arc_l2c_only = &ARC_l2c_only;
3770 arc_size = 0;
3771
3772 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3773 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3774 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3775 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3776 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3777 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3778
3779 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3780 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3781 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3782 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3783 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3784 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3785 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3786 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3787 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3788 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3789 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3790 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3791 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3792 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3793 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3794 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3795 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3796 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3797 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3798 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3799
3800 buf_init();
3801
3802 arc_thread_exit = 0;
3803 arc_eviction_list = NULL;
3804 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3805 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3806
3807 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3808 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3809
3810 if (arc_ksp != NULL) {
3811 arc_ksp->ks_data = &arc_stats;
3812 kstat_install(arc_ksp);
3813 }
3814
3815 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3816 TS_RUN, minclsyspri);
3817
3818 arc_dead = FALSE;
3819 arc_warm = B_FALSE;
3820
3821 /*
3822 * Calculate maximum amount of dirty data per pool.
3823 *
3824 * If it has been set by /etc/system, take that.
3825 * Otherwise, use a percentage of physical memory defined by
3826 * zfs_dirty_data_max_percent (default 10%) with a cap at
3827 * zfs_dirty_data_max_max (default 4GB).
3828 */
3829 if (zfs_dirty_data_max == 0) {
3830 zfs_dirty_data_max = physmem * PAGESIZE *
3831 zfs_dirty_data_max_percent / 100;
3832 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
3833 zfs_dirty_data_max_max);
3834 }
3835 }
3836
3837 void
3838 arc_fini(void)
3839 {
3840 mutex_enter(&arc_reclaim_thr_lock);
3841 arc_thread_exit = 1;
3842 while (arc_thread_exit != 0)
3843 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3844 mutex_exit(&arc_reclaim_thr_lock);
3845
3846 arc_flush(NULL);
3847
3848 arc_dead = TRUE;
3849
3850 if (arc_ksp != NULL) {
3851 kstat_delete(arc_ksp);
3852 arc_ksp = NULL;
3853 }
3854
3855 mutex_destroy(&arc_eviction_mtx);
3856 mutex_destroy(&arc_reclaim_thr_lock);
3857 cv_destroy(&arc_reclaim_thr_cv);
3858
3859 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3860 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3861 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3862 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3863 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3864 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3865 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3866 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3867
3868 mutex_destroy(&arc_anon->arcs_mtx);
3869 mutex_destroy(&arc_mru->arcs_mtx);
3870 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3871 mutex_destroy(&arc_mfu->arcs_mtx);
3872 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3873 mutex_destroy(&arc_l2c_only->arcs_mtx);
3874
3875 buf_fini();
3876
3877 ASSERT(arc_loaned_bytes == 0);
3878 }
3879
3880 /*
3881 * Level 2 ARC
3882 *
3883 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3884 * It uses dedicated storage devices to hold cached data, which are populated
3885 * using large infrequent writes. The main role of this cache is to boost
3886 * the performance of random read workloads. The intended L2ARC devices
3887 * include short-stroked disks, solid state disks, and other media with
3888 * substantially faster read latency than disk.
3889 *
3890 * +-----------------------+
3891 * | ARC |
3892 * +-----------------------+
3893 * | ^ ^
3894 * | | |
3895 * l2arc_feed_thread() arc_read()
3896 * | | |
3897 * | l2arc read |
3898 * V | |
3899 * +---------------+ |
3900 * | L2ARC | |
3901 * +---------------+ |
3902 * | ^ |
3903 * l2arc_write() | |
3904 * | | |
3905 * V | |
3906 * +-------+ +-------+
3907 * | vdev | | vdev |
3908 * | cache | | cache |
3909 * +-------+ +-------+
3910 * +=========+ .-----.
3911 * : L2ARC : |-_____-|
3912 * : devices : | Disks |
3913 * +=========+ `-_____-'
3914 *
3915 * Read requests are satisfied from the following sources, in order:
3916 *
3917 * 1) ARC
3918 * 2) vdev cache of L2ARC devices
3919 * 3) L2ARC devices
3920 * 4) vdev cache of disks
3921 * 5) disks
3922 *
3923 * Some L2ARC device types exhibit extremely slow write performance.
3924 * To accommodate for this there are some significant differences between
3925 * the L2ARC and traditional cache design:
3926 *
3927 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3928 * the ARC behave as usual, freeing buffers and placing headers on ghost
3929 * lists. The ARC does not send buffers to the L2ARC during eviction as
3930 * this would add inflated write latencies for all ARC memory pressure.
3931 *
3932 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3933 * It does this by periodically scanning buffers from the eviction-end of
3934 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3935 * not already there. It scans until a headroom of buffers is satisfied,
3936 * which itself is a buffer for ARC eviction. If a compressible buffer is
3937 * found during scanning and selected for writing to an L2ARC device, we
3938 * temporarily boost scanning headroom during the next scan cycle to make
3939 * sure we adapt to compression effects (which might significantly reduce
3940 * the data volume we write to L2ARC). The thread that does this is
3941 * l2arc_feed_thread(), illustrated below; example sizes are included to
3942 * provide a better sense of ratio than this diagram:
3943 *
3944 * head --> tail
3945 * +---------------------+----------+
3946 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
3947 * +---------------------+----------+ | o L2ARC eligible
3948 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
3949 * +---------------------+----------+ |
3950 * 15.9 Gbytes ^ 32 Mbytes |
3951 * headroom |
3952 * l2arc_feed_thread()
3953 * |
3954 * l2arc write hand <--[oooo]--'
3955 * | 8 Mbyte
3956 * | write max
3957 * V
3958 * +==============================+
3959 * L2ARC dev |####|#|###|###| |####| ... |
3960 * +==============================+
3961 * 32 Gbytes
3962 *
3963 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
3964 * evicted, then the L2ARC has cached a buffer much sooner than it probably
3965 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
3966 * safe to say that this is an uncommon case, since buffers at the end of
3967 * the ARC lists have moved there due to inactivity.
3968 *
3969 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
3970 * then the L2ARC simply misses copying some buffers. This serves as a
3971 * pressure valve to prevent heavy read workloads from both stalling the ARC
3972 * with waits and clogging the L2ARC with writes. This also helps prevent
3973 * the potential for the L2ARC to churn if it attempts to cache content too
3974 * quickly, such as during backups of the entire pool.
3975 *
3976 * 5. After system boot and before the ARC has filled main memory, there are
3977 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
3978 * lists can remain mostly static. Instead of searching from tail of these
3979 * lists as pictured, the l2arc_feed_thread() will search from the list heads
3980 * for eligible buffers, greatly increasing its chance of finding them.
3981 *
3982 * The L2ARC device write speed is also boosted during this time so that
3983 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
3984 * there are no L2ARC reads, and no fear of degrading read performance
3985 * through increased writes.
3986 *
3987 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
3988 * the vdev queue can aggregate them into larger and fewer writes. Each
3989 * device is written to in a rotor fashion, sweeping writes through
3990 * available space then repeating.
3991 *
3992 * 7. The L2ARC does not store dirty content. It never needs to flush
3993 * write buffers back to disk based storage.
3994 *
3995 * 8. If an ARC buffer is written (and dirtied) which also exists in the
3996 * L2ARC, the now stale L2ARC buffer is immediately dropped.
3997 *
3998 * The performance of the L2ARC can be tweaked by a number of tunables, which
3999 * may be necessary for different workloads:
4000 *
4001 * l2arc_write_max max write bytes per interval
4002 * l2arc_write_boost extra write bytes during device warmup
4003 * l2arc_noprefetch skip caching prefetched buffers
4004 * l2arc_headroom number of max device writes to precache
4005 * l2arc_headroom_boost when we find compressed buffers during ARC
4006 * scanning, we multiply headroom by this
4007 * percentage factor for the next scan cycle,
4008 * since more compressed buffers are likely to
4009 * be present
4010 * l2arc_feed_secs seconds between L2ARC writing
4011 *
4012 * Tunables may be removed or added as future performance improvements are
4013 * integrated, and also may become zpool properties.
4014 *
4015 * There are three key functions that control how the L2ARC warms up:
4016 *
4017 * l2arc_write_eligible() check if a buffer is eligible to cache
4018 * l2arc_write_size() calculate how much to write
4019 * l2arc_write_interval() calculate sleep delay between writes
4020 *
4021 * These three functions determine what to write, how much, and how quickly
4022 * to send writes.
4023 */
4024
4025 static boolean_t
4026 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4027 {
4028 /*
4029 * A buffer is *not* eligible for the L2ARC if it:
4030 * 1. belongs to a different spa.
4031 * 2. is already cached on the L2ARC.
4032 * 3. has an I/O in progress (it may be an incomplete read).
4033 * 4. is flagged not eligible (zfs property).
4034 */
4035 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
4036 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
4037 return (B_FALSE);
4038
4039 return (B_TRUE);
4040 }
4041
4042 static uint64_t
4043 l2arc_write_size(void)
4044 {
4045 uint64_t size;
4046
4047 /*
4048 * Make sure our globals have meaningful values in case the user
4049 * altered them.
4050 */
4051 size = l2arc_write_max;
4052 if (size == 0) {
4053 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
4054 "be greater than zero, resetting it to the default (%d)",
4055 L2ARC_WRITE_SIZE);
4056 size = l2arc_write_max = L2ARC_WRITE_SIZE;
4057 }
4058
4059 if (arc_warm == B_FALSE)
4060 size += l2arc_write_boost;
4061
4062 return (size);
4063
4064 }
4065
4066 static clock_t
4067 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4068 {
4069 clock_t interval, next, now;
4070
4071 /*
4072 * If the ARC lists are busy, increase our write rate; if the
4073 * lists are stale, idle back. This is achieved by checking
4074 * how much we previously wrote - if it was more than half of
4075 * what we wanted, schedule the next write much sooner.
4076 */
4077 if (l2arc_feed_again && wrote > (wanted / 2))
4078 interval = (hz * l2arc_feed_min_ms) / 1000;
4079 else
4080 interval = hz * l2arc_feed_secs;
4081
4082 now = ddi_get_lbolt();
4083 next = MAX(now, MIN(now + interval, began + interval));
4084
4085 return (next);
4086 }
4087
4088 static void
4089 l2arc_hdr_stat_add(void)
4090 {
4091 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4092 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4093 }
4094
4095 static void
4096 l2arc_hdr_stat_remove(void)
4097 {
4098 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4099 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4100 }
4101
4102 /*
4103 * Cycle through L2ARC devices. This is how L2ARC load balances.
4104 * If a device is returned, this also returns holding the spa config lock.
4105 */
4106 static l2arc_dev_t *
4107 l2arc_dev_get_next(void)
4108 {
4109 l2arc_dev_t *first, *next = NULL;
4110
4111 /*
4112 * Lock out the removal of spas (spa_namespace_lock), then removal
4113 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4114 * both locks will be dropped and a spa config lock held instead.
4115 */
4116 mutex_enter(&spa_namespace_lock);
4117 mutex_enter(&l2arc_dev_mtx);
4118
4119 /* if there are no vdevs, there is nothing to do */
4120 if (l2arc_ndev == 0)
4121 goto out;
4122
4123 first = NULL;
4124 next = l2arc_dev_last;
4125 do {
4126 /* loop around the list looking for a non-faulted vdev */
4127 if (next == NULL) {
4128 next = list_head(l2arc_dev_list);
4129 } else {
4130 next = list_next(l2arc_dev_list, next);
4131 if (next == NULL)
4132 next = list_head(l2arc_dev_list);
4133 }
4134
4135 /* if we have come back to the start, bail out */
4136 if (first == NULL)
4137 first = next;
4138 else if (next == first)
4139 break;
4140
4141 } while (vdev_is_dead(next->l2ad_vdev));
4142
4143 /* if we were unable to find any usable vdevs, return NULL */
4144 if (vdev_is_dead(next->l2ad_vdev))
4145 next = NULL;
4146
4147 l2arc_dev_last = next;
4148
4149 out:
4150 mutex_exit(&l2arc_dev_mtx);
4151
4152 /*
4153 * Grab the config lock to prevent the 'next' device from being
4154 * removed while we are writing to it.
4155 */
4156 if (next != NULL)
4157 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4158 mutex_exit(&spa_namespace_lock);
4159
4160 return (next);
4161 }
4162
4163 /*
4164 * Free buffers that were tagged for destruction.
4165 */
4166 static void
4167 l2arc_do_free_on_write()
4168 {
4169 list_t *buflist;
4170 l2arc_data_free_t *df, *df_prev;
4171
4172 mutex_enter(&l2arc_free_on_write_mtx);
4173 buflist = l2arc_free_on_write;
4174
4175 for (df = list_tail(buflist); df; df = df_prev) {
4176 df_prev = list_prev(buflist, df);
4177 ASSERT(df->l2df_data != NULL);
4178 ASSERT(df->l2df_func != NULL);
4179 df->l2df_func(df->l2df_data, df->l2df_size);
4180 list_remove(buflist, df);
4181 kmem_free(df, sizeof (l2arc_data_free_t));
4182 }
4183
4184 mutex_exit(&l2arc_free_on_write_mtx);
4185 }
4186
4187 /*
4188 * A write to a cache device has completed. Update all headers to allow
4189 * reads from these buffers to begin.
4190 */
4191 static void
4192 l2arc_write_done(zio_t *zio)
4193 {
4194 l2arc_write_callback_t *cb;
4195 l2arc_dev_t *dev;
4196 list_t *buflist;
4197 arc_buf_hdr_t *head, *ab, *ab_prev;
4198 l2arc_buf_hdr_t *abl2;
4199 kmutex_t *hash_lock;
4200
4201 cb = zio->io_private;
4202 ASSERT(cb != NULL);
4203 dev = cb->l2wcb_dev;
4204 ASSERT(dev != NULL);
4205 head = cb->l2wcb_head;
4206 ASSERT(head != NULL);
4207 buflist = dev->l2ad_buflist;
4208 ASSERT(buflist != NULL);
4209 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4210 l2arc_write_callback_t *, cb);
4211
4212 if (zio->io_error != 0)
4213 ARCSTAT_BUMP(arcstat_l2_writes_error);
4214
4215 mutex_enter(&l2arc_buflist_mtx);
4216
4217 /*
4218 * All writes completed, or an error was hit.
4219 */
4220 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4221 ab_prev = list_prev(buflist, ab);
4222
4223 hash_lock = HDR_LOCK(ab);
4224 if (!mutex_tryenter(hash_lock)) {
4225 /*
4226 * This buffer misses out. It may be in a stage
4227 * of eviction. Its ARC_L2_WRITING flag will be
4228 * left set, denying reads to this buffer.
4229 */
4230 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4231 continue;
4232 }
4233
4234 abl2 = ab->b_l2hdr;
4235
4236 /*
4237 * Release the temporary compressed buffer as soon as possible.
4238 */
4239 if (abl2->b_compress != ZIO_COMPRESS_OFF)
4240 l2arc_release_cdata_buf(ab);
4241
4242 if (zio->io_error != 0) {
4243 /*
4244 * Error - drop L2ARC entry.
4245 */
4246 list_remove(buflist, ab);
4247 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4248 ab->b_l2hdr = NULL;
4249 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4250 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4251 }
4252
4253 /*
4254 * Allow ARC to begin reads to this L2ARC entry.
4255 */
4256 ab->b_flags &= ~ARC_L2_WRITING;
4257
4258 mutex_exit(hash_lock);
4259 }
4260
4261 atomic_inc_64(&l2arc_writes_done);
4262 list_remove(buflist, head);
4263 kmem_cache_free(hdr_cache, head);
4264 mutex_exit(&l2arc_buflist_mtx);
4265
4266 l2arc_do_free_on_write();
4267
4268 kmem_free(cb, sizeof (l2arc_write_callback_t));
4269 }
4270
4271 /*
4272 * A read to a cache device completed. Validate buffer contents before
4273 * handing over to the regular ARC routines.
4274 */
4275 static void
4276 l2arc_read_done(zio_t *zio)
4277 {
4278 l2arc_read_callback_t *cb;
4279 arc_buf_hdr_t *hdr;
4280 arc_buf_t *buf;
4281 kmutex_t *hash_lock;
4282 int equal;
4283
4284 ASSERT(zio->io_vd != NULL);
4285 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4286
4287 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4288
4289 cb = zio->io_private;
4290 ASSERT(cb != NULL);
4291 buf = cb->l2rcb_buf;
4292 ASSERT(buf != NULL);
4293
4294 hash_lock = HDR_LOCK(buf->b_hdr);
4295 mutex_enter(hash_lock);
4296 hdr = buf->b_hdr;
4297 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4298
4299 /*
4300 * If the buffer was compressed, decompress it first.
4301 */
4302 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
4303 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
4304 ASSERT(zio->io_data != NULL);
4305
4306 /*
4307 * Check this survived the L2ARC journey.
4308 */
4309 equal = arc_cksum_equal(buf);
4310 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4311 mutex_exit(hash_lock);
4312 zio->io_private = buf;
4313 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4314 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4315 arc_read_done(zio);
4316 } else {
4317 mutex_exit(hash_lock);
4318 /*
4319 * Buffer didn't survive caching. Increment stats and
4320 * reissue to the original storage device.
4321 */
4322 if (zio->io_error != 0) {
4323 ARCSTAT_BUMP(arcstat_l2_io_error);
4324 } else {
4325 zio->io_error = SET_ERROR(EIO);
4326 }
4327 if (!equal)
4328 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4329
4330 /*
4331 * If there's no waiter, issue an async i/o to the primary
4332 * storage now. If there *is* a waiter, the caller must
4333 * issue the i/o in a context where it's OK to block.
4334 */
4335 if (zio->io_waiter == NULL) {
4336 zio_t *pio = zio_unique_parent(zio);
4337
4338 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4339
4340 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4341 buf->b_data, zio->io_size, arc_read_done, buf,
4342 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4343 }
4344 }
4345
4346 kmem_free(cb, sizeof (l2arc_read_callback_t));
4347 }
4348
4349 /*
4350 * This is the list priority from which the L2ARC will search for pages to
4351 * cache. This is used within loops (0..3) to cycle through lists in the
4352 * desired order. This order can have a significant effect on cache
4353 * performance.
4354 *
4355 * Currently the metadata lists are hit first, MFU then MRU, followed by
4356 * the data lists. This function returns a locked list, and also returns
4357 * the lock pointer.
4358 */
4359 static list_t *
4360 l2arc_list_locked(int list_num, kmutex_t **lock)
4361 {
4362 list_t *list = NULL;
4363
4364 ASSERT(list_num >= 0 && list_num <= 3);
4365
4366 switch (list_num) {
4367 case 0:
4368 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4369 *lock = &arc_mfu->arcs_mtx;
4370 break;
4371 case 1:
4372 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4373 *lock = &arc_mru->arcs_mtx;
4374 break;
4375 case 2:
4376 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4377 *lock = &arc_mfu->arcs_mtx;
4378 break;
4379 case 3:
4380 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4381 *lock = &arc_mru->arcs_mtx;
4382 break;
4383 }
4384
4385 ASSERT(!(MUTEX_HELD(*lock)));
4386 mutex_enter(*lock);
4387 return (list);
4388 }
4389
4390 /*
4391 * Evict buffers from the device write hand to the distance specified in
4392 * bytes. This distance may span populated buffers, it may span nothing.
4393 * This is clearing a region on the L2ARC device ready for writing.
4394 * If the 'all' boolean is set, every buffer is evicted.
4395 */
4396 static void
4397 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4398 {
4399 list_t *buflist;
4400 l2arc_buf_hdr_t *abl2;
4401 arc_buf_hdr_t *ab, *ab_prev;
4402 kmutex_t *hash_lock;
4403 uint64_t taddr;
4404
4405 buflist = dev->l2ad_buflist;
4406
4407 if (buflist == NULL)
4408 return;
4409
4410 if (!all && dev->l2ad_first) {
4411 /*
4412 * This is the first sweep through the device. There is
4413 * nothing to evict.
4414 */
4415 return;
4416 }
4417
4418 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4419 /*
4420 * When nearing the end of the device, evict to the end
4421 * before the device write hand jumps to the start.
4422 */
4423 taddr = dev->l2ad_end;
4424 } else {
4425 taddr = dev->l2ad_hand + distance;
4426 }
4427 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4428 uint64_t, taddr, boolean_t, all);
4429
4430 top:
4431 mutex_enter(&l2arc_buflist_mtx);
4432 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4433 ab_prev = list_prev(buflist, ab);
4434
4435 hash_lock = HDR_LOCK(ab);
4436 if (!mutex_tryenter(hash_lock)) {
4437 /*
4438 * Missed the hash lock. Retry.
4439 */
4440 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4441 mutex_exit(&l2arc_buflist_mtx);
4442 mutex_enter(hash_lock);
4443 mutex_exit(hash_lock);
4444 goto top;
4445 }
4446
4447 if (HDR_L2_WRITE_HEAD(ab)) {
4448 /*
4449 * We hit a write head node. Leave it for
4450 * l2arc_write_done().
4451 */
4452 list_remove(buflist, ab);
4453 mutex_exit(hash_lock);
4454 continue;
4455 }
4456
4457 if (!all && ab->b_l2hdr != NULL &&
4458 (ab->b_l2hdr->b_daddr > taddr ||
4459 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4460 /*
4461 * We've evicted to the target address,
4462 * or the end of the device.
4463 */
4464 mutex_exit(hash_lock);
4465 break;
4466 }
4467
4468 if (HDR_FREE_IN_PROGRESS(ab)) {
4469 /*
4470 * Already on the path to destruction.
4471 */
4472 mutex_exit(hash_lock);
4473 continue;
4474 }
4475
4476 if (ab->b_state == arc_l2c_only) {
4477 ASSERT(!HDR_L2_READING(ab));
4478 /*
4479 * This doesn't exist in the ARC. Destroy.
4480 * arc_hdr_destroy() will call list_remove()
4481 * and decrement arcstat_l2_size.
4482 */
4483 arc_change_state(arc_anon, ab, hash_lock);
4484 arc_hdr_destroy(ab);
4485 } else {
4486 /*
4487 * Invalidate issued or about to be issued
4488 * reads, since we may be about to write
4489 * over this location.
4490 */
4491 if (HDR_L2_READING(ab)) {
4492 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4493 ab->b_flags |= ARC_L2_EVICTED;
4494 }
4495
4496 /*
4497 * Tell ARC this no longer exists in L2ARC.
4498 */
4499 if (ab->b_l2hdr != NULL) {
4500 abl2 = ab->b_l2hdr;
4501 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4502 ab->b_l2hdr = NULL;
4503 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4504 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4505 }
4506 list_remove(buflist, ab);
4507
4508 /*
4509 * This may have been leftover after a
4510 * failed write.
4511 */
4512 ab->b_flags &= ~ARC_L2_WRITING;
4513 }
4514 mutex_exit(hash_lock);
4515 }
4516 mutex_exit(&l2arc_buflist_mtx);
4517
4518 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4519 dev->l2ad_evict = taddr;
4520 }
4521
4522 /*
4523 * Find and write ARC buffers to the L2ARC device.
4524 *
4525 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4526 * for reading until they have completed writing.
4527 * The headroom_boost is an in-out parameter used to maintain headroom boost
4528 * state between calls to this function.
4529 *
4530 * Returns the number of bytes actually written (which may be smaller than
4531 * the delta by which the device hand has changed due to alignment).
4532 */
4533 static uint64_t
4534 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
4535 boolean_t *headroom_boost)
4536 {
4537 arc_buf_hdr_t *ab, *ab_prev, *head;
4538 list_t *list;
4539 uint64_t write_asize, write_psize, write_sz, headroom,
4540 buf_compress_minsz;
4541 void *buf_data;
4542 kmutex_t *list_lock;
4543 boolean_t full;
4544 l2arc_write_callback_t *cb;
4545 zio_t *pio, *wzio;
4546 uint64_t guid = spa_load_guid(spa);
4547 const boolean_t do_headroom_boost = *headroom_boost;
4548
4549 ASSERT(dev->l2ad_vdev != NULL);
4550
4551 /* Lower the flag now, we might want to raise it again later. */
4552 *headroom_boost = B_FALSE;
4553
4554 pio = NULL;
4555 write_sz = write_asize = write_psize = 0;
4556 full = B_FALSE;
4557 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4558 head->b_flags |= ARC_L2_WRITE_HEAD;
4559
4560 /*
4561 * We will want to try to compress buffers that are at least 2x the
4562 * device sector size.
4563 */
4564 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
4565
4566 /*
4567 * Copy buffers for L2ARC writing.
4568 */
4569 mutex_enter(&l2arc_buflist_mtx);
4570 for (int try = 0; try <= 3; try++) {
4571 uint64_t passed_sz = 0;
4572
4573 list = l2arc_list_locked(try, &list_lock);
4574
4575 /*
4576 * L2ARC fast warmup.
4577 *
4578 * Until the ARC is warm and starts to evict, read from the
4579 * head of the ARC lists rather than the tail.
4580 */
4581 if (arc_warm == B_FALSE)
4582 ab = list_head(list);
4583 else
4584 ab = list_tail(list);
4585
4586 headroom = target_sz * l2arc_headroom;
4587 if (do_headroom_boost)
4588 headroom = (headroom * l2arc_headroom_boost) / 100;
4589
4590 for (; ab; ab = ab_prev) {
4591 l2arc_buf_hdr_t *l2hdr;
4592 kmutex_t *hash_lock;
4593 uint64_t buf_sz;
4594
4595 if (arc_warm == B_FALSE)
4596 ab_prev = list_next(list, ab);
4597 else
4598 ab_prev = list_prev(list, ab);
4599
4600 hash_lock = HDR_LOCK(ab);
4601 if (!mutex_tryenter(hash_lock)) {
4602 /*
4603 * Skip this buffer rather than waiting.
4604 */
4605 continue;
4606 }
4607
4608 passed_sz += ab->b_size;
4609 if (passed_sz > headroom) {
4610 /*
4611 * Searched too far.
4612 */
4613 mutex_exit(hash_lock);
4614 break;
4615 }
4616
4617 if (!l2arc_write_eligible(guid, ab)) {
4618 mutex_exit(hash_lock);
4619 continue;
4620 }
4621
4622 if ((write_sz + ab->b_size) > target_sz) {
4623 full = B_TRUE;
4624 mutex_exit(hash_lock);
4625 break;
4626 }
4627
4628 if (pio == NULL) {
4629 /*
4630 * Insert a dummy header on the buflist so
4631 * l2arc_write_done() can find where the
4632 * write buffers begin without searching.
4633 */
4634 list_insert_head(dev->l2ad_buflist, head);
4635
4636 cb = kmem_alloc(
4637 sizeof (l2arc_write_callback_t), KM_SLEEP);
4638 cb->l2wcb_dev = dev;
4639 cb->l2wcb_head = head;
4640 pio = zio_root(spa, l2arc_write_done, cb,
4641 ZIO_FLAG_CANFAIL);
4642 }
4643
4644 /*
4645 * Create and add a new L2ARC header.
4646 */
4647 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4648 l2hdr->b_dev = dev;
4649 ab->b_flags |= ARC_L2_WRITING;
4650
4651 /*
4652 * Temporarily stash the data buffer in b_tmp_cdata.
4653 * The subsequent write step will pick it up from
4654 * there. This is because can't access ab->b_buf
4655 * without holding the hash_lock, which we in turn
4656 * can't access without holding the ARC list locks
4657 * (which we want to avoid during compression/writing).
4658 */
4659 l2hdr->b_compress = ZIO_COMPRESS_OFF;
4660 l2hdr->b_asize = ab->b_size;
4661 l2hdr->b_tmp_cdata = ab->b_buf->b_data;
4662
4663 buf_sz = ab->b_size;
4664 ab->b_l2hdr = l2hdr;
4665
4666 list_insert_head(dev->l2ad_buflist, ab);
4667
4668 /*
4669 * Compute and store the buffer cksum before
4670 * writing. On debug the cksum is verified first.
4671 */
4672 arc_cksum_verify(ab->b_buf);
4673 arc_cksum_compute(ab->b_buf, B_TRUE);
4674
4675 mutex_exit(hash_lock);
4676
4677 write_sz += buf_sz;
4678 }
4679
4680 mutex_exit(list_lock);
4681
4682 if (full == B_TRUE)
4683 break;
4684 }
4685
4686 /* No buffers selected for writing? */
4687 if (pio == NULL) {
4688 ASSERT0(write_sz);
4689 mutex_exit(&l2arc_buflist_mtx);
4690 kmem_cache_free(hdr_cache, head);
4691 return (0);
4692 }
4693
4694 /*
4695 * Now start writing the buffers. We're starting at the write head
4696 * and work backwards, retracing the course of the buffer selector
4697 * loop above.
4698 */
4699 for (ab = list_prev(dev->l2ad_buflist, head); ab;
4700 ab = list_prev(dev->l2ad_buflist, ab)) {
4701 l2arc_buf_hdr_t *l2hdr;
4702 uint64_t buf_sz;
4703
4704 /*
4705 * We shouldn't need to lock the buffer here, since we flagged
4706 * it as ARC_L2_WRITING in the previous step, but we must take
4707 * care to only access its L2 cache parameters. In particular,
4708 * ab->b_buf may be invalid by now due to ARC eviction.
4709 */
4710 l2hdr = ab->b_l2hdr;
4711 l2hdr->b_daddr = dev->l2ad_hand;
4712
4713 if ((ab->b_flags & ARC_L2COMPRESS) &&
4714 l2hdr->b_asize >= buf_compress_minsz) {
4715 if (l2arc_compress_buf(l2hdr)) {
4716 /*
4717 * If compression succeeded, enable headroom
4718 * boost on the next scan cycle.
4719 */
4720 *headroom_boost = B_TRUE;
4721 }
4722 }
4723
4724 /*
4725 * Pick up the buffer data we had previously stashed away
4726 * (and now potentially also compressed).
4727 */
4728 buf_data = l2hdr->b_tmp_cdata;
4729 buf_sz = l2hdr->b_asize;
4730
4731 /* Compression may have squashed the buffer to zero length. */
4732 if (buf_sz != 0) {
4733 uint64_t buf_p_sz;
4734
4735 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4736 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4737 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4738 ZIO_FLAG_CANFAIL, B_FALSE);
4739
4740 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4741 zio_t *, wzio);
4742 (void) zio_nowait(wzio);
4743
4744 write_asize += buf_sz;
4745 /*
4746 * Keep the clock hand suitably device-aligned.
4747 */
4748 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4749 write_psize += buf_p_sz;
4750 dev->l2ad_hand += buf_p_sz;
4751 }
4752 }
4753
4754 mutex_exit(&l2arc_buflist_mtx);
4755
4756 ASSERT3U(write_asize, <=, target_sz);
4757 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4758 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
4759 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4760 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
4761 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
4762
4763 /*
4764 * Bump device hand to the device start if it is approaching the end.
4765 * l2arc_evict() will already have evicted ahead for this case.
4766 */
4767 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4768 vdev_space_update(dev->l2ad_vdev,
4769 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4770 dev->l2ad_hand = dev->l2ad_start;
4771 dev->l2ad_evict = dev->l2ad_start;
4772 dev->l2ad_first = B_FALSE;
4773 }
4774
4775 dev->l2ad_writing = B_TRUE;
4776 (void) zio_wait(pio);
4777 dev->l2ad_writing = B_FALSE;
4778
4779 return (write_asize);
4780 }
4781
4782 /*
4783 * Compresses an L2ARC buffer.
4784 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its
4785 * size in l2hdr->b_asize. This routine tries to compress the data and
4786 * depending on the compression result there are three possible outcomes:
4787 * *) The buffer was incompressible. The original l2hdr contents were left
4788 * untouched and are ready for writing to an L2 device.
4789 * *) The buffer was all-zeros, so there is no need to write it to an L2
4790 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
4791 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
4792 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
4793 * data buffer which holds the compressed data to be written, and b_asize
4794 * tells us how much data there is. b_compress is set to the appropriate
4795 * compression algorithm. Once writing is done, invoke
4796 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
4797 *
4798 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
4799 * buffer was incompressible).
4800 */
4801 static boolean_t
4802 l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr)
4803 {
4804 void *cdata;
4805 size_t csize, len;
4806
4807 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF);
4808 ASSERT(l2hdr->b_tmp_cdata != NULL);
4809
4810 len = l2hdr->b_asize;
4811 cdata = zio_data_buf_alloc(len);
4812 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata,
4813 cdata, l2hdr->b_asize);
4814
4815 if (csize == 0) {
4816 /* zero block, indicate that there's nothing to write */
4817 zio_data_buf_free(cdata, len);
4818 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
4819 l2hdr->b_asize = 0;
4820 l2hdr->b_tmp_cdata = NULL;
4821 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
4822 return (B_TRUE);
4823 } else if (csize > 0 && csize < len) {
4824 /*
4825 * Compression succeeded, we'll keep the cdata around for
4826 * writing and release it afterwards.
4827 */
4828 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
4829 l2hdr->b_asize = csize;
4830 l2hdr->b_tmp_cdata = cdata;
4831 ARCSTAT_BUMP(arcstat_l2_compress_successes);
4832 return (B_TRUE);
4833 } else {
4834 /*
4835 * Compression failed, release the compressed buffer.
4836 * l2hdr will be left unmodified.
4837 */
4838 zio_data_buf_free(cdata, len);
4839 ARCSTAT_BUMP(arcstat_l2_compress_failures);
4840 return (B_FALSE);
4841 }
4842 }
4843
4844 /*
4845 * Decompresses a zio read back from an l2arc device. On success, the
4846 * underlying zio's io_data buffer is overwritten by the uncompressed
4847 * version. On decompression error (corrupt compressed stream), the
4848 * zio->io_error value is set to signal an I/O error.
4849 *
4850 * Please note that the compressed data stream is not checksummed, so
4851 * if the underlying device is experiencing data corruption, we may feed
4852 * corrupt data to the decompressor, so the decompressor needs to be
4853 * able to handle this situation (LZ4 does).
4854 */
4855 static void
4856 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
4857 {
4858 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
4859
4860 if (zio->io_error != 0) {
4861 /*
4862 * An io error has occured, just restore the original io
4863 * size in preparation for a main pool read.
4864 */
4865 zio->io_orig_size = zio->io_size = hdr->b_size;
4866 return;
4867 }
4868
4869 if (c == ZIO_COMPRESS_EMPTY) {
4870 /*
4871 * An empty buffer results in a null zio, which means we
4872 * need to fill its io_data after we're done restoring the
4873 * buffer's contents.
4874 */
4875 ASSERT(hdr->b_buf != NULL);
4876 bzero(hdr->b_buf->b_data, hdr->b_size);
4877 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data;
4878 } else {
4879 ASSERT(zio->io_data != NULL);
4880 /*
4881 * We copy the compressed data from the start of the arc buffer
4882 * (the zio_read will have pulled in only what we need, the
4883 * rest is garbage which we will overwrite at decompression)
4884 * and then decompress back to the ARC data buffer. This way we
4885 * can minimize copying by simply decompressing back over the
4886 * original compressed data (rather than decompressing to an
4887 * aux buffer and then copying back the uncompressed buffer,
4888 * which is likely to be much larger).
4889 */
4890 uint64_t csize;
4891 void *cdata;
4892
4893 csize = zio->io_size;
4894 cdata = zio_data_buf_alloc(csize);
4895 bcopy(zio->io_data, cdata, csize);
4896 if (zio_decompress_data(c, cdata, zio->io_data, csize,
4897 hdr->b_size) != 0)
4898 zio->io_error = EIO;
4899 zio_data_buf_free(cdata, csize);
4900 }
4901
4902 /* Restore the expected uncompressed IO size. */
4903 zio->io_orig_size = zio->io_size = hdr->b_size;
4904 }
4905
4906 /*
4907 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
4908 * This buffer serves as a temporary holder of compressed data while
4909 * the buffer entry is being written to an l2arc device. Once that is
4910 * done, we can dispose of it.
4911 */
4912 static void
4913 l2arc_release_cdata_buf(arc_buf_hdr_t *ab)
4914 {
4915 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr;
4916
4917 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) {
4918 /*
4919 * If the data was compressed, then we've allocated a
4920 * temporary buffer for it, so now we need to release it.
4921 */
4922 ASSERT(l2hdr->b_tmp_cdata != NULL);
4923 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size);
4924 }
4925 l2hdr->b_tmp_cdata = NULL;
4926 }
4927
4928 /*
4929 * This thread feeds the L2ARC at regular intervals. This is the beating
4930 * heart of the L2ARC.
4931 */
4932 static void
4933 l2arc_feed_thread(void)
4934 {
4935 callb_cpr_t cpr;
4936 l2arc_dev_t *dev;
4937 spa_t *spa;
4938 uint64_t size, wrote;
4939 clock_t begin, next = ddi_get_lbolt();
4940 boolean_t headroom_boost = B_FALSE;
4941
4942 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4943
4944 mutex_enter(&l2arc_feed_thr_lock);
4945
4946 while (l2arc_thread_exit == 0) {
4947 CALLB_CPR_SAFE_BEGIN(&cpr);
4948 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4949 next);
4950 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4951 next = ddi_get_lbolt() + hz;
4952
4953 /*
4954 * Quick check for L2ARC devices.
4955 */
4956 mutex_enter(&l2arc_dev_mtx);
4957 if (l2arc_ndev == 0) {
4958 mutex_exit(&l2arc_dev_mtx);
4959 continue;
4960 }
4961 mutex_exit(&l2arc_dev_mtx);
4962 begin = ddi_get_lbolt();
4963
4964 /*
4965 * This selects the next l2arc device to write to, and in
4966 * doing so the next spa to feed from: dev->l2ad_spa. This
4967 * will return NULL if there are now no l2arc devices or if
4968 * they are all faulted.
4969 *
4970 * If a device is returned, its spa's config lock is also
4971 * held to prevent device removal. l2arc_dev_get_next()
4972 * will grab and release l2arc_dev_mtx.
4973 */
4974 if ((dev = l2arc_dev_get_next()) == NULL)
4975 continue;
4976
4977 spa = dev->l2ad_spa;
4978 ASSERT(spa != NULL);
4979
4980 /*
4981 * If the pool is read-only then force the feed thread to
4982 * sleep a little longer.
4983 */
4984 if (!spa_writeable(spa)) {
4985 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4986 spa_config_exit(spa, SCL_L2ARC, dev);
4987 continue;
4988 }
4989
4990 /*
4991 * Avoid contributing to memory pressure.
4992 */
4993 if (arc_reclaim_needed()) {
4994 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4995 spa_config_exit(spa, SCL_L2ARC, dev);
4996 continue;
4997 }
4998
4999 ARCSTAT_BUMP(arcstat_l2_feeds);
5000
5001 size = l2arc_write_size();
5002
5003 /*
5004 * Evict L2ARC buffers that will be overwritten.
5005 */
5006 l2arc_evict(dev, size, B_FALSE);
5007
5008 /*
5009 * Write ARC buffers.
5010 */
5011 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
5012
5013 /*
5014 * Calculate interval between writes.
5015 */
5016 next = l2arc_write_interval(begin, size, wrote);
5017 spa_config_exit(spa, SCL_L2ARC, dev);
5018 }
5019
5020 l2arc_thread_exit = 0;
5021 cv_broadcast(&l2arc_feed_thr_cv);
5022 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
5023 thread_exit();
5024 }
5025
5026 boolean_t
5027 l2arc_vdev_present(vdev_t *vd)
5028 {
5029 l2arc_dev_t *dev;
5030
5031 mutex_enter(&l2arc_dev_mtx);
5032 for (dev = list_head(l2arc_dev_list); dev != NULL;
5033 dev = list_next(l2arc_dev_list, dev)) {
5034 if (dev->l2ad_vdev == vd)
5035 break;
5036 }
5037 mutex_exit(&l2arc_dev_mtx);
5038
5039 return (dev != NULL);
5040 }
5041
5042 /*
5043 * Add a vdev for use by the L2ARC. By this point the spa has already
5044 * validated the vdev and opened it.
5045 */
5046 void
5047 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
5048 {
5049 l2arc_dev_t *adddev;
5050
5051 ASSERT(!l2arc_vdev_present(vd));
5052
5053 /*
5054 * Create a new l2arc device entry.
5055 */
5056 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
5057 adddev->l2ad_spa = spa;
5058 adddev->l2ad_vdev = vd;
5059 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
5060 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
5061 adddev->l2ad_hand = adddev->l2ad_start;
5062 adddev->l2ad_evict = adddev->l2ad_start;
5063 adddev->l2ad_first = B_TRUE;
5064 adddev->l2ad_writing = B_FALSE;
5065
5066 /*
5067 * This is a list of all ARC buffers that are still valid on the
5068 * device.
5069 */
5070 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
5071 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
5072 offsetof(arc_buf_hdr_t, b_l2node));
5073
5074 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
5075
5076 /*
5077 * Add device to global list
5078 */
5079 mutex_enter(&l2arc_dev_mtx);
5080 list_insert_head(l2arc_dev_list, adddev);
5081 atomic_inc_64(&l2arc_ndev);
5082 mutex_exit(&l2arc_dev_mtx);
5083 }
5084
5085 /*
5086 * Remove a vdev from the L2ARC.
5087 */
5088 void
5089 l2arc_remove_vdev(vdev_t *vd)
5090 {
5091 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
5092
5093 /*
5094 * Find the device by vdev
5095 */
5096 mutex_enter(&l2arc_dev_mtx);
5097 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
5098 nextdev = list_next(l2arc_dev_list, dev);
5099 if (vd == dev->l2ad_vdev) {
5100 remdev = dev;
5101 break;
5102 }
5103 }
5104 ASSERT(remdev != NULL);
5105
5106 /*
5107 * Remove device from global list
5108 */
5109 list_remove(l2arc_dev_list, remdev);
5110 l2arc_dev_last = NULL; /* may have been invalidated */
5111 atomic_dec_64(&l2arc_ndev);
5112 mutex_exit(&l2arc_dev_mtx);
5113
5114 /*
5115 * Clear all buflists and ARC references. L2ARC device flush.
5116 */
5117 l2arc_evict(remdev, 0, B_TRUE);
5118 list_destroy(remdev->l2ad_buflist);
5119 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5120 kmem_free(remdev, sizeof (l2arc_dev_t));
5121 }
5122
5123 void
5124 l2arc_init(void)
5125 {
5126 l2arc_thread_exit = 0;
5127 l2arc_ndev = 0;
5128 l2arc_writes_sent = 0;
5129 l2arc_writes_done = 0;
5130
5131 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5132 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5133 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5134 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5135 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5136
5137 l2arc_dev_list = &L2ARC_dev_list;
5138 l2arc_free_on_write = &L2ARC_free_on_write;
5139 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5140 offsetof(l2arc_dev_t, l2ad_node));
5141 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5142 offsetof(l2arc_data_free_t, l2df_list_node));
5143 }
5144
5145 void
5146 l2arc_fini(void)
5147 {
5148 /*
5149 * This is called from dmu_fini(), which is called from spa_fini();
5150 * Because of this, we can assume that all l2arc devices have
5151 * already been removed when the pools themselves were removed.
5152 */
5153
5154 l2arc_do_free_on_write();
5155
5156 mutex_destroy(&l2arc_feed_thr_lock);
5157 cv_destroy(&l2arc_feed_thr_cv);
5158 mutex_destroy(&l2arc_dev_mtx);
5159 mutex_destroy(&l2arc_buflist_mtx);
5160 mutex_destroy(&l2arc_free_on_write_mtx);
5161
5162 list_destroy(l2arc_dev_list);
5163 list_destroy(l2arc_free_on_write);
5164 }
5165
5166 void
5167 l2arc_start(void)
5168 {
5169 if (!(spa_mode_global & FWRITE))
5170 return;
5171
5172 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5173 TS_RUN, minclsyspri);
5174 }
5175
5176 void
5177 l2arc_stop(void)
5178 {
5179 if (!(spa_mode_global & FWRITE))
5180 return;
5181
5182 mutex_enter(&l2arc_feed_thr_lock);
5183 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5184 l2arc_thread_exit = 1;
5185 while (l2arc_thread_exit != 0)
5186 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5187 mutex_exit(&l2arc_feed_thr_lock);
5188 }