1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2014 Nexenta Systems, Inc. All rights reserved.
27 */
28
29 /*
30 * DVA-based Adjustable Replacement Cache
31 *
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
36 *
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
47 *
48 * There are times when it is not possible to evict the requested
49 * space. In these circumstances we are unable to adjust the cache
50 * size. To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
53 *
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss. Our model has a variable sized cache. It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
59 * tight.
60 *
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size. So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict. In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
69 *
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
72 */
73
74 /*
75 * The locking model:
76 *
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists. The arc_read() interface
80 * uses method 1, while the internal arc algorithms for
81 * adjusting the cache use method 2. We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
83 * arc list locks.
84 *
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
88 *
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table. It returns
91 * NULL for the mutex if the buffer was not in the table.
92 *
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
95 *
96 * Each arc state also has a mutex which is used to protect the
97 * buffer list associated with the state. When attempting to
98 * obtain a hash table lock while holding an arc list lock you
99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 * the active state mutex must be held before the ghost state mutex.
101 *
102 * Arc buffers may have an associated eviction callback function.
103 * This function will be invoked prior to removing the buffer (e.g.
104 * in arc_do_user_evicts()). Note however that the data associated
105 * with the buffer may be evicted prior to the callback. The callback
106 * must be made with *no locks held* (to prevent deadlock). Additionally,
107 * the users of callbacks must ensure that their private data is
108 * protected from simultaneous callbacks from arc_clear_callback()
109 * and arc_do_user_evicts().
110 *
111 * Note that the majority of the performance stats are manipulated
112 * with atomic operations.
113 *
114 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
115 *
116 * - L2ARC buflist creation
117 * - L2ARC buflist eviction
118 * - L2ARC write completion, which walks L2ARC buflists
119 * - ARC header destruction, as it removes from L2ARC buflists
120 * - ARC header release, as it removes from L2ARC buflists
121 */
122
123 #include <sys/spa.h>
124 #include <sys/zio.h>
125 #include <sys/zio_compress.h>
126 #include <sys/zfs_context.h>
127 #include <sys/arc.h>
128 #include <sys/refcount.h>
129 #include <sys/vdev.h>
130 #include <sys/vdev_impl.h>
131 #include <sys/dsl_pool.h>
132 #ifdef _KERNEL
133 #include <sys/vmsystm.h>
134 #include <vm/anon.h>
135 #include <sys/fs/swapnode.h>
136 #include <sys/dnlc.h>
137 #endif
138 #include <sys/callb.h>
139 #include <sys/kstat.h>
140 #include <zfs_fletcher.h>
141
142 #ifndef _KERNEL
143 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
144 boolean_t arc_watch = B_FALSE;
145 int arc_procfd;
146 #endif
147
148 static kmutex_t arc_reclaim_thr_lock;
149 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
150 static uint8_t arc_thread_exit;
151
152 #define ARC_REDUCE_DNLC_PERCENT 3
153 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
154
155 typedef enum arc_reclaim_strategy {
156 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
157 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
158 } arc_reclaim_strategy_t;
159
160 /*
161 * The number of iterations through arc_evict_*() before we
162 * drop & reacquire the lock.
163 */
164 int arc_evict_iterations = 100;
165
166 /* number of seconds before growing cache again */
167 static int arc_grow_retry = 60;
168
169 /* shift of arc_c for calculating both min and max arc_p */
170 static int arc_p_min_shift = 4;
171
172 /* log2(fraction of arc to reclaim) */
173 static int arc_shrink_shift = 5;
174
175 /*
176 * minimum lifespan of a prefetch block in clock ticks
177 * (initialized in arc_init())
178 */
179 static int arc_min_prefetch_lifespan;
180
181 /*
182 * If this percent of memory is free, don't throttle.
183 */
184 int arc_lotsfree_percent = 10;
185
186 static int arc_dead;
187
188 /*
189 * The arc has filled available memory and has now warmed up.
190 */
191 static boolean_t arc_warm;
192
193 /*
194 * These tunables are for performance analysis.
195 */
196 uint64_t zfs_arc_max;
197 uint64_t zfs_arc_min;
198 uint64_t zfs_arc_meta_limit = 0;
199 int zfs_arc_grow_retry = 0;
200 int zfs_arc_shrink_shift = 0;
201 int zfs_arc_p_min_shift = 0;
202 int zfs_disable_dup_eviction = 0;
203 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
204
205 /*
206 * Note that buffers can be in one of 6 states:
207 * ARC_anon - anonymous (discussed below)
208 * ARC_mru - recently used, currently cached
209 * ARC_mru_ghost - recentely used, no longer in cache
210 * ARC_mfu - frequently used, currently cached
211 * ARC_mfu_ghost - frequently used, no longer in cache
212 * ARC_l2c_only - exists in L2ARC but not other states
213 * When there are no active references to the buffer, they are
214 * are linked onto a list in one of these arc states. These are
215 * the only buffers that can be evicted or deleted. Within each
216 * state there are multiple lists, one for meta-data and one for
217 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
218 * etc.) is tracked separately so that it can be managed more
219 * explicitly: favored over data, limited explicitly.
220 *
221 * Anonymous buffers are buffers that are not associated with
222 * a DVA. These are buffers that hold dirty block copies
223 * before they are written to stable storage. By definition,
224 * they are "ref'd" and are considered part of arc_mru
225 * that cannot be freed. Generally, they will aquire a DVA
226 * as they are written and migrate onto the arc_mru list.
227 *
228 * The ARC_l2c_only state is for buffers that are in the second
229 * level ARC but no longer in any of the ARC_m* lists. The second
230 * level ARC itself may also contain buffers that are in any of
231 * the ARC_m* states - meaning that a buffer can exist in two
232 * places. The reason for the ARC_l2c_only state is to keep the
233 * buffer header in the hash table, so that reads that hit the
234 * second level ARC benefit from these fast lookups.
235 */
236
237 typedef struct arc_state {
238 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
239 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
240 uint64_t arcs_size; /* total amount of data in this state */
241 kmutex_t arcs_mtx;
242 } arc_state_t;
243
244 /* The 6 states: */
245 static arc_state_t ARC_anon;
246 static arc_state_t ARC_mru;
247 static arc_state_t ARC_mru_ghost;
248 static arc_state_t ARC_mfu;
249 static arc_state_t ARC_mfu_ghost;
250 static arc_state_t ARC_l2c_only;
251
252 typedef struct arc_stats {
253 kstat_named_t arcstat_hits;
254 kstat_named_t arcstat_misses;
255 kstat_named_t arcstat_demand_data_hits;
256 kstat_named_t arcstat_demand_data_misses;
257 kstat_named_t arcstat_demand_metadata_hits;
258 kstat_named_t arcstat_demand_metadata_misses;
259 kstat_named_t arcstat_prefetch_data_hits;
260 kstat_named_t arcstat_prefetch_data_misses;
261 kstat_named_t arcstat_prefetch_metadata_hits;
262 kstat_named_t arcstat_prefetch_metadata_misses;
263 kstat_named_t arcstat_mru_hits;
264 kstat_named_t arcstat_mru_ghost_hits;
265 kstat_named_t arcstat_mfu_hits;
266 kstat_named_t arcstat_mfu_ghost_hits;
267 kstat_named_t arcstat_deleted;
268 kstat_named_t arcstat_recycle_miss;
269 /*
270 * Number of buffers that could not be evicted because the hash lock
271 * was held by another thread. The lock may not necessarily be held
272 * by something using the same buffer, since hash locks are shared
273 * by multiple buffers.
274 */
275 kstat_named_t arcstat_mutex_miss;
276 /*
277 * Number of buffers skipped because they have I/O in progress, are
278 * indrect prefetch buffers that have not lived long enough, or are
279 * not from the spa we're trying to evict from.
280 */
281 kstat_named_t arcstat_evict_skip;
282 kstat_named_t arcstat_evict_l2_cached;
283 kstat_named_t arcstat_evict_l2_eligible;
284 kstat_named_t arcstat_evict_l2_ineligible;
285 kstat_named_t arcstat_hash_elements;
286 kstat_named_t arcstat_hash_elements_max;
287 kstat_named_t arcstat_hash_collisions;
288 kstat_named_t arcstat_hash_chains;
289 kstat_named_t arcstat_hash_chain_max;
290 kstat_named_t arcstat_p;
291 kstat_named_t arcstat_c;
292 kstat_named_t arcstat_c_min;
293 kstat_named_t arcstat_c_max;
294 kstat_named_t arcstat_size;
295 kstat_named_t arcstat_hdr_size;
296 kstat_named_t arcstat_data_size;
297 kstat_named_t arcstat_other_size;
298 kstat_named_t arcstat_l2_hits;
299 kstat_named_t arcstat_l2_misses;
300 kstat_named_t arcstat_l2_feeds;
301 kstat_named_t arcstat_l2_rw_clash;
302 kstat_named_t arcstat_l2_read_bytes;
303 kstat_named_t arcstat_l2_write_bytes;
304 kstat_named_t arcstat_l2_writes_sent;
305 kstat_named_t arcstat_l2_writes_done;
306 kstat_named_t arcstat_l2_writes_error;
307 kstat_named_t arcstat_l2_writes_hdr_miss;
308 kstat_named_t arcstat_l2_evict_lock_retry;
309 kstat_named_t arcstat_l2_evict_reading;
310 kstat_named_t arcstat_l2_free_on_write;
311 kstat_named_t arcstat_l2_cdata_free_on_write;
312 kstat_named_t arcstat_l2_abort_lowmem;
313 kstat_named_t arcstat_l2_cksum_bad;
314 kstat_named_t arcstat_l2_io_error;
315 kstat_named_t arcstat_l2_size;
316 kstat_named_t arcstat_l2_asize;
317 kstat_named_t arcstat_l2_hdr_size;
318 kstat_named_t arcstat_l2_compress_successes;
319 kstat_named_t arcstat_l2_compress_zeros;
320 kstat_named_t arcstat_l2_compress_failures;
321 kstat_named_t arcstat_memory_throttle_count;
322 kstat_named_t arcstat_duplicate_buffers;
323 kstat_named_t arcstat_duplicate_buffers_size;
324 kstat_named_t arcstat_duplicate_reads;
325 kstat_named_t arcstat_meta_used;
326 kstat_named_t arcstat_meta_limit;
327 kstat_named_t arcstat_meta_max;
328 } arc_stats_t;
329
330 static arc_stats_t arc_stats = {
331 { "hits", KSTAT_DATA_UINT64 },
332 { "misses", KSTAT_DATA_UINT64 },
333 { "demand_data_hits", KSTAT_DATA_UINT64 },
334 { "demand_data_misses", KSTAT_DATA_UINT64 },
335 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
336 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
337 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
338 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
339 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
340 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
341 { "mru_hits", KSTAT_DATA_UINT64 },
342 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
343 { "mfu_hits", KSTAT_DATA_UINT64 },
344 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
345 { "deleted", KSTAT_DATA_UINT64 },
346 { "recycle_miss", KSTAT_DATA_UINT64 },
347 { "mutex_miss", KSTAT_DATA_UINT64 },
348 { "evict_skip", KSTAT_DATA_UINT64 },
349 { "evict_l2_cached", KSTAT_DATA_UINT64 },
350 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
351 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
352 { "hash_elements", KSTAT_DATA_UINT64 },
353 { "hash_elements_max", KSTAT_DATA_UINT64 },
354 { "hash_collisions", KSTAT_DATA_UINT64 },
355 { "hash_chains", KSTAT_DATA_UINT64 },
356 { "hash_chain_max", KSTAT_DATA_UINT64 },
357 { "p", KSTAT_DATA_UINT64 },
358 { "c", KSTAT_DATA_UINT64 },
359 { "c_min", KSTAT_DATA_UINT64 },
360 { "c_max", KSTAT_DATA_UINT64 },
361 { "size", KSTAT_DATA_UINT64 },
362 { "hdr_size", KSTAT_DATA_UINT64 },
363 { "data_size", KSTAT_DATA_UINT64 },
364 { "other_size", KSTAT_DATA_UINT64 },
365 { "l2_hits", KSTAT_DATA_UINT64 },
366 { "l2_misses", KSTAT_DATA_UINT64 },
367 { "l2_feeds", KSTAT_DATA_UINT64 },
368 { "l2_rw_clash", KSTAT_DATA_UINT64 },
369 { "l2_read_bytes", KSTAT_DATA_UINT64 },
370 { "l2_write_bytes", KSTAT_DATA_UINT64 },
371 { "l2_writes_sent", KSTAT_DATA_UINT64 },
372 { "l2_writes_done", KSTAT_DATA_UINT64 },
373 { "l2_writes_error", KSTAT_DATA_UINT64 },
374 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
375 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
376 { "l2_evict_reading", KSTAT_DATA_UINT64 },
377 { "l2_free_on_write", KSTAT_DATA_UINT64 },
378 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 },
379 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
380 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
381 { "l2_io_error", KSTAT_DATA_UINT64 },
382 { "l2_size", KSTAT_DATA_UINT64 },
383 { "l2_asize", KSTAT_DATA_UINT64 },
384 { "l2_hdr_size", KSTAT_DATA_UINT64 },
385 { "l2_compress_successes", KSTAT_DATA_UINT64 },
386 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
387 { "l2_compress_failures", KSTAT_DATA_UINT64 },
388 { "memory_throttle_count", KSTAT_DATA_UINT64 },
389 { "duplicate_buffers", KSTAT_DATA_UINT64 },
390 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
391 { "duplicate_reads", KSTAT_DATA_UINT64 },
392 { "arc_meta_used", KSTAT_DATA_UINT64 },
393 { "arc_meta_limit", KSTAT_DATA_UINT64 },
394 { "arc_meta_max", KSTAT_DATA_UINT64 }
395 };
396
397 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
398
399 #define ARCSTAT_INCR(stat, val) \
400 atomic_add_64(&arc_stats.stat.value.ui64, (val))
401
402 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
403 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
404
405 #define ARCSTAT_MAX(stat, val) { \
406 uint64_t m; \
407 while ((val) > (m = arc_stats.stat.value.ui64) && \
408 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
409 continue; \
410 }
411
412 #define ARCSTAT_MAXSTAT(stat) \
413 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
414
415 /*
416 * We define a macro to allow ARC hits/misses to be easily broken down by
417 * two separate conditions, giving a total of four different subtypes for
418 * each of hits and misses (so eight statistics total).
419 */
420 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
421 if (cond1) { \
422 if (cond2) { \
423 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
424 } else { \
425 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
426 } \
427 } else { \
428 if (cond2) { \
429 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
430 } else { \
431 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
432 } \
433 }
434
435 kstat_t *arc_ksp;
436 static arc_state_t *arc_anon;
437 static arc_state_t *arc_mru;
438 static arc_state_t *arc_mru_ghost;
439 static arc_state_t *arc_mfu;
440 static arc_state_t *arc_mfu_ghost;
441 static arc_state_t *arc_l2c_only;
442
443 /*
444 * There are several ARC variables that are critical to export as kstats --
445 * but we don't want to have to grovel around in the kstat whenever we wish to
446 * manipulate them. For these variables, we therefore define them to be in
447 * terms of the statistic variable. This assures that we are not introducing
448 * the possibility of inconsistency by having shadow copies of the variables,
449 * while still allowing the code to be readable.
450 */
451 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
452 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
453 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
454 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
455 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
456 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
457 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
458 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
459
460 #define L2ARC_IS_VALID_COMPRESS(_c_) \
461 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
462
463 static int arc_no_grow; /* Don't try to grow cache size */
464 static uint64_t arc_tempreserve;
465 static uint64_t arc_loaned_bytes;
466
467 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
468
469 typedef struct arc_callback arc_callback_t;
470
471 struct arc_callback {
472 void *acb_private;
473 arc_done_func_t *acb_done;
474 arc_buf_t *acb_buf;
475 zio_t *acb_zio_dummy;
476 arc_callback_t *acb_next;
477 };
478
479 typedef struct arc_write_callback arc_write_callback_t;
480
481 struct arc_write_callback {
482 void *awcb_private;
483 arc_done_func_t *awcb_ready;
484 arc_done_func_t *awcb_physdone;
485 arc_done_func_t *awcb_done;
486 arc_buf_t *awcb_buf;
487 };
488
489 struct arc_buf_hdr {
490 /* protected by hash lock */
491 dva_t b_dva;
492 uint64_t b_birth;
493 uint64_t b_cksum0;
494
495 kmutex_t b_freeze_lock;
496 zio_cksum_t *b_freeze_cksum;
497 void *b_thawed;
498
499 arc_buf_hdr_t *b_hash_next;
500 arc_buf_t *b_buf;
501 uint32_t b_flags;
502 uint32_t b_datacnt;
503
504 arc_callback_t *b_acb;
505 kcondvar_t b_cv;
506
507 /* immutable */
508 arc_buf_contents_t b_type;
509 uint64_t b_size;
510 uint64_t b_spa;
511
512 /* protected by arc state mutex */
513 arc_state_t *b_state;
514 list_node_t b_arc_node;
515
516 /* updated atomically */
517 clock_t b_arc_access;
518
519 /* self protecting */
520 refcount_t b_refcnt;
521
522 l2arc_buf_hdr_t *b_l2hdr;
523 list_node_t b_l2node;
524 };
525
526 static arc_buf_t *arc_eviction_list;
527 static kmutex_t arc_eviction_mtx;
528 static arc_buf_hdr_t arc_eviction_hdr;
529 static void arc_get_data_buf(arc_buf_t *buf);
530 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
531 static int arc_evict_needed(arc_buf_contents_t type);
532 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
533 static void arc_buf_watch(arc_buf_t *buf);
534
535 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
536
537 #define GHOST_STATE(state) \
538 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
539 (state) == arc_l2c_only)
540
541 /*
542 * Private ARC flags. These flags are private ARC only flags that will show up
543 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
544 * be passed in as arc_flags in things like arc_read. However, these flags
545 * should never be passed and should only be set by ARC code. When adding new
546 * public flags, make sure not to smash the private ones.
547 */
548
549 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
550 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
551 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
552 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
553 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
554 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
555 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
556 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
557 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
558 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
559
560 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
561 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
562 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
563 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
564 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
565 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
566 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
567 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
568 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
569 (hdr)->b_l2hdr != NULL)
570 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
571 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
572 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
573
574 /*
575 * Other sizes
576 */
577
578 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
579 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
580
581 /*
582 * Hash table routines
583 */
584
585 #define HT_LOCK_PAD 64
586
587 struct ht_lock {
588 kmutex_t ht_lock;
589 #ifdef _KERNEL
590 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
591 #endif
592 };
593
594 #define BUF_LOCKS 256
595 typedef struct buf_hash_table {
596 uint64_t ht_mask;
597 arc_buf_hdr_t **ht_table;
598 struct ht_lock ht_locks[BUF_LOCKS];
599 } buf_hash_table_t;
600
601 static buf_hash_table_t buf_hash_table;
602
603 #define BUF_HASH_INDEX(spa, dva, birth) \
604 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
605 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
606 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
607 #define HDR_LOCK(hdr) \
608 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
609
610 uint64_t zfs_crc64_table[256];
611
612 /*
613 * Level 2 ARC
614 */
615
616 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
617 #define L2ARC_HEADROOM 2 /* num of writes */
618 /*
619 * If we discover during ARC scan any buffers to be compressed, we boost
620 * our headroom for the next scanning cycle by this percentage multiple.
621 */
622 #define L2ARC_HEADROOM_BOOST 200
623 #define L2ARC_FEED_SECS 1 /* caching interval secs */
624 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
625
626 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
627 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
628
629 /* L2ARC Performance Tunables */
630 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
631 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
632 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
633 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
634 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
635 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
636 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
637 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
638 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
639
640 /*
641 * L2ARC Internals
642 */
643 typedef struct l2arc_dev {
644 vdev_t *l2ad_vdev; /* vdev */
645 spa_t *l2ad_spa; /* spa */
646 uint64_t l2ad_hand; /* next write location */
647 uint64_t l2ad_start; /* first addr on device */
648 uint64_t l2ad_end; /* last addr on device */
649 uint64_t l2ad_evict; /* last addr eviction reached */
650 boolean_t l2ad_first; /* first sweep through */
651 boolean_t l2ad_writing; /* currently writing */
652 list_t *l2ad_buflist; /* buffer list */
653 list_node_t l2ad_node; /* device list node */
654 } l2arc_dev_t;
655
656 static list_t L2ARC_dev_list; /* device list */
657 static list_t *l2arc_dev_list; /* device list pointer */
658 static kmutex_t l2arc_dev_mtx; /* device list mutex */
659 static l2arc_dev_t *l2arc_dev_last; /* last device used */
660 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
661 static list_t L2ARC_free_on_write; /* free after write buf list */
662 static list_t *l2arc_free_on_write; /* free after write list ptr */
663 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
664 static uint64_t l2arc_ndev; /* number of devices */
665
666 typedef struct l2arc_read_callback {
667 arc_buf_t *l2rcb_buf; /* read buffer */
668 spa_t *l2rcb_spa; /* spa */
669 blkptr_t l2rcb_bp; /* original blkptr */
670 zbookmark_phys_t l2rcb_zb; /* original bookmark */
671 int l2rcb_flags; /* original flags */
672 enum zio_compress l2rcb_compress; /* applied compress */
673 } l2arc_read_callback_t;
674
675 typedef struct l2arc_write_callback {
676 l2arc_dev_t *l2wcb_dev; /* device info */
677 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
678 } l2arc_write_callback_t;
679
680 struct l2arc_buf_hdr {
681 /* protected by arc_buf_hdr mutex */
682 l2arc_dev_t *b_dev; /* L2ARC device */
683 uint64_t b_daddr; /* disk address, offset byte */
684 /* compression applied to buffer data */
685 enum zio_compress b_compress;
686 /* real alloc'd buffer size depending on b_compress applied */
687 int b_asize;
688 /* temporary buffer holder for in-flight compressed data */
689 void *b_tmp_cdata;
690 };
691
692 typedef struct l2arc_data_free {
693 /* protected by l2arc_free_on_write_mtx */
694 void *l2df_data;
695 size_t l2df_size;
696 void (*l2df_func)(void *, size_t);
697 list_node_t l2df_list_node;
698 } l2arc_data_free_t;
699
700 static kmutex_t l2arc_feed_thr_lock;
701 static kcondvar_t l2arc_feed_thr_cv;
702 static uint8_t l2arc_thread_exit;
703
704 static void l2arc_read_done(zio_t *zio);
705 static void l2arc_hdr_stat_add(void);
706 static void l2arc_hdr_stat_remove(void);
707
708 static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr);
709 static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr,
710 enum zio_compress c);
711 static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab);
712
713 static uint64_t
714 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
715 {
716 uint8_t *vdva = (uint8_t *)dva;
717 uint64_t crc = -1ULL;
718 int i;
719
720 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
721
722 for (i = 0; i < sizeof (dva_t); i++)
723 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
724
725 crc ^= (spa>>8) ^ birth;
726
727 return (crc);
728 }
729
730 #define BUF_EMPTY(buf) \
731 ((buf)->b_dva.dva_word[0] == 0 && \
732 (buf)->b_dva.dva_word[1] == 0 && \
733 (buf)->b_cksum0 == 0)
734
735 #define BUF_EQUAL(spa, dva, birth, buf) \
736 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
737 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
738 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
739
740 static void
741 buf_discard_identity(arc_buf_hdr_t *hdr)
742 {
743 hdr->b_dva.dva_word[0] = 0;
744 hdr->b_dva.dva_word[1] = 0;
745 hdr->b_birth = 0;
746 hdr->b_cksum0 = 0;
747 }
748
749 static arc_buf_hdr_t *
750 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
751 {
752 const dva_t *dva = BP_IDENTITY(bp);
753 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
754 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
755 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
756 arc_buf_hdr_t *buf;
757
758 mutex_enter(hash_lock);
759 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
760 buf = buf->b_hash_next) {
761 if (BUF_EQUAL(spa, dva, birth, buf)) {
762 *lockp = hash_lock;
763 return (buf);
764 }
765 }
766 mutex_exit(hash_lock);
767 *lockp = NULL;
768 return (NULL);
769 }
770
771 /*
772 * Insert an entry into the hash table. If there is already an element
773 * equal to elem in the hash table, then the already existing element
774 * will be returned and the new element will not be inserted.
775 * Otherwise returns NULL.
776 */
777 static arc_buf_hdr_t *
778 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
779 {
780 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
781 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
782 arc_buf_hdr_t *fbuf;
783 uint32_t i;
784
785 ASSERT(!DVA_IS_EMPTY(&buf->b_dva));
786 ASSERT(buf->b_birth != 0);
787 ASSERT(!HDR_IN_HASH_TABLE(buf));
788 *lockp = hash_lock;
789 mutex_enter(hash_lock);
790 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
791 fbuf = fbuf->b_hash_next, i++) {
792 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
793 return (fbuf);
794 }
795
796 buf->b_hash_next = buf_hash_table.ht_table[idx];
797 buf_hash_table.ht_table[idx] = buf;
798 buf->b_flags |= ARC_IN_HASH_TABLE;
799
800 /* collect some hash table performance data */
801 if (i > 0) {
802 ARCSTAT_BUMP(arcstat_hash_collisions);
803 if (i == 1)
804 ARCSTAT_BUMP(arcstat_hash_chains);
805
806 ARCSTAT_MAX(arcstat_hash_chain_max, i);
807 }
808
809 ARCSTAT_BUMP(arcstat_hash_elements);
810 ARCSTAT_MAXSTAT(arcstat_hash_elements);
811
812 return (NULL);
813 }
814
815 static void
816 buf_hash_remove(arc_buf_hdr_t *buf)
817 {
818 arc_buf_hdr_t *fbuf, **bufp;
819 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
820
821 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
822 ASSERT(HDR_IN_HASH_TABLE(buf));
823
824 bufp = &buf_hash_table.ht_table[idx];
825 while ((fbuf = *bufp) != buf) {
826 ASSERT(fbuf != NULL);
827 bufp = &fbuf->b_hash_next;
828 }
829 *bufp = buf->b_hash_next;
830 buf->b_hash_next = NULL;
831 buf->b_flags &= ~ARC_IN_HASH_TABLE;
832
833 /* collect some hash table performance data */
834 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
835
836 if (buf_hash_table.ht_table[idx] &&
837 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
838 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
839 }
840
841 /*
842 * Global data structures and functions for the buf kmem cache.
843 */
844 static kmem_cache_t *hdr_cache;
845 static kmem_cache_t *buf_cache;
846
847 static void
848 buf_fini(void)
849 {
850 int i;
851
852 kmem_free(buf_hash_table.ht_table,
853 (buf_hash_table.ht_mask + 1) * sizeof (void *));
854 for (i = 0; i < BUF_LOCKS; i++)
855 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
856 kmem_cache_destroy(hdr_cache);
857 kmem_cache_destroy(buf_cache);
858 }
859
860 /*
861 * Constructor callback - called when the cache is empty
862 * and a new buf is requested.
863 */
864 /* ARGSUSED */
865 static int
866 hdr_cons(void *vbuf, void *unused, int kmflag)
867 {
868 arc_buf_hdr_t *buf = vbuf;
869
870 bzero(buf, sizeof (arc_buf_hdr_t));
871 refcount_create(&buf->b_refcnt);
872 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
873 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
874 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
875
876 return (0);
877 }
878
879 /* ARGSUSED */
880 static int
881 buf_cons(void *vbuf, void *unused, int kmflag)
882 {
883 arc_buf_t *buf = vbuf;
884
885 bzero(buf, sizeof (arc_buf_t));
886 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
887 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
888
889 return (0);
890 }
891
892 /*
893 * Destructor callback - called when a cached buf is
894 * no longer required.
895 */
896 /* ARGSUSED */
897 static void
898 hdr_dest(void *vbuf, void *unused)
899 {
900 arc_buf_hdr_t *buf = vbuf;
901
902 ASSERT(BUF_EMPTY(buf));
903 refcount_destroy(&buf->b_refcnt);
904 cv_destroy(&buf->b_cv);
905 mutex_destroy(&buf->b_freeze_lock);
906 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
907 }
908
909 /* ARGSUSED */
910 static void
911 buf_dest(void *vbuf, void *unused)
912 {
913 arc_buf_t *buf = vbuf;
914
915 mutex_destroy(&buf->b_evict_lock);
916 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
917 }
918
919 /*
920 * Reclaim callback -- invoked when memory is low.
921 */
922 /* ARGSUSED */
923 static void
924 hdr_recl(void *unused)
925 {
926 dprintf("hdr_recl called\n");
927 /*
928 * umem calls the reclaim func when we destroy the buf cache,
929 * which is after we do arc_fini().
930 */
931 if (!arc_dead)
932 cv_signal(&arc_reclaim_thr_cv);
933 }
934
935 static void
936 buf_init(void)
937 {
938 uint64_t *ct;
939 uint64_t hsize = 1ULL << 12;
940 int i, j;
941
942 /*
943 * The hash table is big enough to fill all of physical memory
944 * with an average block size of zfs_arc_average_blocksize (default 8K).
945 * By default, the table will take up
946 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
947 */
948 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
949 hsize <<= 1;
950 retry:
951 buf_hash_table.ht_mask = hsize - 1;
952 buf_hash_table.ht_table =
953 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
954 if (buf_hash_table.ht_table == NULL) {
955 ASSERT(hsize > (1ULL << 8));
956 hsize >>= 1;
957 goto retry;
958 }
959
960 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
961 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
962 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
963 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
964
965 for (i = 0; i < 256; i++)
966 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
967 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
968
969 for (i = 0; i < BUF_LOCKS; i++) {
970 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
971 NULL, MUTEX_DEFAULT, NULL);
972 }
973 }
974
975 #define ARC_MINTIME (hz>>4) /* 62 ms */
976
977 static void
978 arc_cksum_verify(arc_buf_t *buf)
979 {
980 zio_cksum_t zc;
981
982 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
983 return;
984
985 mutex_enter(&buf->b_hdr->b_freeze_lock);
986 if (buf->b_hdr->b_freeze_cksum == NULL ||
987 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
988 mutex_exit(&buf->b_hdr->b_freeze_lock);
989 return;
990 }
991 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
992 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
993 panic("buffer modified while frozen!");
994 mutex_exit(&buf->b_hdr->b_freeze_lock);
995 }
996
997 static int
998 arc_cksum_equal(arc_buf_t *buf)
999 {
1000 zio_cksum_t zc;
1001 int equal;
1002
1003 mutex_enter(&buf->b_hdr->b_freeze_lock);
1004 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1005 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1006 mutex_exit(&buf->b_hdr->b_freeze_lock);
1007
1008 return (equal);
1009 }
1010
1011 static void
1012 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1013 {
1014 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1015 return;
1016
1017 mutex_enter(&buf->b_hdr->b_freeze_lock);
1018 if (buf->b_hdr->b_freeze_cksum != NULL) {
1019 mutex_exit(&buf->b_hdr->b_freeze_lock);
1020 return;
1021 }
1022 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1023 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1024 buf->b_hdr->b_freeze_cksum);
1025 mutex_exit(&buf->b_hdr->b_freeze_lock);
1026 arc_buf_watch(buf);
1027 }
1028
1029 #ifndef _KERNEL
1030 typedef struct procctl {
1031 long cmd;
1032 prwatch_t prwatch;
1033 } procctl_t;
1034 #endif
1035
1036 /* ARGSUSED */
1037 static void
1038 arc_buf_unwatch(arc_buf_t *buf)
1039 {
1040 #ifndef _KERNEL
1041 if (arc_watch) {
1042 int result;
1043 procctl_t ctl;
1044 ctl.cmd = PCWATCH;
1045 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1046 ctl.prwatch.pr_size = 0;
1047 ctl.prwatch.pr_wflags = 0;
1048 result = write(arc_procfd, &ctl, sizeof (ctl));
1049 ASSERT3U(result, ==, sizeof (ctl));
1050 }
1051 #endif
1052 }
1053
1054 /* ARGSUSED */
1055 static void
1056 arc_buf_watch(arc_buf_t *buf)
1057 {
1058 #ifndef _KERNEL
1059 if (arc_watch) {
1060 int result;
1061 procctl_t ctl;
1062 ctl.cmd = PCWATCH;
1063 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1064 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1065 ctl.prwatch.pr_wflags = WA_WRITE;
1066 result = write(arc_procfd, &ctl, sizeof (ctl));
1067 ASSERT3U(result, ==, sizeof (ctl));
1068 }
1069 #endif
1070 }
1071
1072 void
1073 arc_buf_thaw(arc_buf_t *buf)
1074 {
1075 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1076 if (buf->b_hdr->b_state != arc_anon)
1077 panic("modifying non-anon buffer!");
1078 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1079 panic("modifying buffer while i/o in progress!");
1080 arc_cksum_verify(buf);
1081 }
1082
1083 mutex_enter(&buf->b_hdr->b_freeze_lock);
1084 if (buf->b_hdr->b_freeze_cksum != NULL) {
1085 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1086 buf->b_hdr->b_freeze_cksum = NULL;
1087 }
1088
1089 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1090 if (buf->b_hdr->b_thawed)
1091 kmem_free(buf->b_hdr->b_thawed, 1);
1092 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1093 }
1094
1095 mutex_exit(&buf->b_hdr->b_freeze_lock);
1096
1097 arc_buf_unwatch(buf);
1098 }
1099
1100 void
1101 arc_buf_freeze(arc_buf_t *buf)
1102 {
1103 kmutex_t *hash_lock;
1104
1105 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1106 return;
1107
1108 hash_lock = HDR_LOCK(buf->b_hdr);
1109 mutex_enter(hash_lock);
1110
1111 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1112 buf->b_hdr->b_state == arc_anon);
1113 arc_cksum_compute(buf, B_FALSE);
1114 mutex_exit(hash_lock);
1115
1116 }
1117
1118 static void
1119 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1120 {
1121 ASSERT(MUTEX_HELD(hash_lock));
1122
1123 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1124 (ab->b_state != arc_anon)) {
1125 uint64_t delta = ab->b_size * ab->b_datacnt;
1126 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1127 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1128
1129 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1130 mutex_enter(&ab->b_state->arcs_mtx);
1131 ASSERT(list_link_active(&ab->b_arc_node));
1132 list_remove(list, ab);
1133 if (GHOST_STATE(ab->b_state)) {
1134 ASSERT0(ab->b_datacnt);
1135 ASSERT3P(ab->b_buf, ==, NULL);
1136 delta = ab->b_size;
1137 }
1138 ASSERT(delta > 0);
1139 ASSERT3U(*size, >=, delta);
1140 atomic_add_64(size, -delta);
1141 mutex_exit(&ab->b_state->arcs_mtx);
1142 /* remove the prefetch flag if we get a reference */
1143 if (ab->b_flags & ARC_PREFETCH)
1144 ab->b_flags &= ~ARC_PREFETCH;
1145 }
1146 }
1147
1148 static int
1149 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1150 {
1151 int cnt;
1152 arc_state_t *state = ab->b_state;
1153
1154 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1155 ASSERT(!GHOST_STATE(state));
1156
1157 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1158 (state != arc_anon)) {
1159 uint64_t *size = &state->arcs_lsize[ab->b_type];
1160
1161 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1162 mutex_enter(&state->arcs_mtx);
1163 ASSERT(!list_link_active(&ab->b_arc_node));
1164 list_insert_head(&state->arcs_list[ab->b_type], ab);
1165 ASSERT(ab->b_datacnt > 0);
1166 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1167 mutex_exit(&state->arcs_mtx);
1168 }
1169 return (cnt);
1170 }
1171
1172 /*
1173 * Move the supplied buffer to the indicated state. The mutex
1174 * for the buffer must be held by the caller.
1175 */
1176 static void
1177 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1178 {
1179 arc_state_t *old_state = ab->b_state;
1180 int64_t refcnt = refcount_count(&ab->b_refcnt);
1181 uint64_t from_delta, to_delta;
1182
1183 ASSERT(MUTEX_HELD(hash_lock));
1184 ASSERT3P(new_state, !=, old_state);
1185 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1186 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1187 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1188
1189 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1190
1191 /*
1192 * If this buffer is evictable, transfer it from the
1193 * old state list to the new state list.
1194 */
1195 if (refcnt == 0) {
1196 if (old_state != arc_anon) {
1197 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1198 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1199
1200 if (use_mutex)
1201 mutex_enter(&old_state->arcs_mtx);
1202
1203 ASSERT(list_link_active(&ab->b_arc_node));
1204 list_remove(&old_state->arcs_list[ab->b_type], ab);
1205
1206 /*
1207 * If prefetching out of the ghost cache,
1208 * we will have a non-zero datacnt.
1209 */
1210 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1211 /* ghost elements have a ghost size */
1212 ASSERT(ab->b_buf == NULL);
1213 from_delta = ab->b_size;
1214 }
1215 ASSERT3U(*size, >=, from_delta);
1216 atomic_add_64(size, -from_delta);
1217
1218 if (use_mutex)
1219 mutex_exit(&old_state->arcs_mtx);
1220 }
1221 if (new_state != arc_anon) {
1222 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1223 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1224
1225 if (use_mutex)
1226 mutex_enter(&new_state->arcs_mtx);
1227
1228 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1229
1230 /* ghost elements have a ghost size */
1231 if (GHOST_STATE(new_state)) {
1232 ASSERT(ab->b_datacnt == 0);
1233 ASSERT(ab->b_buf == NULL);
1234 to_delta = ab->b_size;
1235 }
1236 atomic_add_64(size, to_delta);
1237
1238 if (use_mutex)
1239 mutex_exit(&new_state->arcs_mtx);
1240 }
1241 }
1242
1243 ASSERT(!BUF_EMPTY(ab));
1244 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1245 buf_hash_remove(ab);
1246
1247 /* adjust state sizes */
1248 if (to_delta)
1249 atomic_add_64(&new_state->arcs_size, to_delta);
1250 if (from_delta) {
1251 ASSERT3U(old_state->arcs_size, >=, from_delta);
1252 atomic_add_64(&old_state->arcs_size, -from_delta);
1253 }
1254 ab->b_state = new_state;
1255
1256 /* adjust l2arc hdr stats */
1257 if (new_state == arc_l2c_only)
1258 l2arc_hdr_stat_add();
1259 else if (old_state == arc_l2c_only)
1260 l2arc_hdr_stat_remove();
1261 }
1262
1263 void
1264 arc_space_consume(uint64_t space, arc_space_type_t type)
1265 {
1266 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1267
1268 switch (type) {
1269 case ARC_SPACE_DATA:
1270 ARCSTAT_INCR(arcstat_data_size, space);
1271 break;
1272 case ARC_SPACE_OTHER:
1273 ARCSTAT_INCR(arcstat_other_size, space);
1274 break;
1275 case ARC_SPACE_HDRS:
1276 ARCSTAT_INCR(arcstat_hdr_size, space);
1277 break;
1278 case ARC_SPACE_L2HDRS:
1279 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1280 break;
1281 }
1282
1283 ARCSTAT_INCR(arcstat_meta_used, space);
1284 atomic_add_64(&arc_size, space);
1285 }
1286
1287 void
1288 arc_space_return(uint64_t space, arc_space_type_t type)
1289 {
1290 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1291
1292 switch (type) {
1293 case ARC_SPACE_DATA:
1294 ARCSTAT_INCR(arcstat_data_size, -space);
1295 break;
1296 case ARC_SPACE_OTHER:
1297 ARCSTAT_INCR(arcstat_other_size, -space);
1298 break;
1299 case ARC_SPACE_HDRS:
1300 ARCSTAT_INCR(arcstat_hdr_size, -space);
1301 break;
1302 case ARC_SPACE_L2HDRS:
1303 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1304 break;
1305 }
1306
1307 ASSERT(arc_meta_used >= space);
1308 if (arc_meta_max < arc_meta_used)
1309 arc_meta_max = arc_meta_used;
1310 ARCSTAT_INCR(arcstat_meta_used, -space);
1311 ASSERT(arc_size >= space);
1312 atomic_add_64(&arc_size, -space);
1313 }
1314
1315 void *
1316 arc_data_buf_alloc(uint64_t size)
1317 {
1318 if (arc_evict_needed(ARC_BUFC_DATA))
1319 cv_signal(&arc_reclaim_thr_cv);
1320 atomic_add_64(&arc_size, size);
1321 return (zio_data_buf_alloc(size));
1322 }
1323
1324 void
1325 arc_data_buf_free(void *buf, uint64_t size)
1326 {
1327 zio_data_buf_free(buf, size);
1328 ASSERT(arc_size >= size);
1329 atomic_add_64(&arc_size, -size);
1330 }
1331
1332 arc_buf_t *
1333 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1334 {
1335 arc_buf_hdr_t *hdr;
1336 arc_buf_t *buf;
1337
1338 ASSERT3U(size, >, 0);
1339 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1340 ASSERT(BUF_EMPTY(hdr));
1341 hdr->b_size = size;
1342 hdr->b_type = type;
1343 hdr->b_spa = spa_load_guid(spa);
1344 hdr->b_state = arc_anon;
1345 hdr->b_arc_access = 0;
1346 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1347 buf->b_hdr = hdr;
1348 buf->b_data = NULL;
1349 buf->b_efunc = NULL;
1350 buf->b_private = NULL;
1351 buf->b_next = NULL;
1352 hdr->b_buf = buf;
1353 arc_get_data_buf(buf);
1354 hdr->b_datacnt = 1;
1355 hdr->b_flags = 0;
1356 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1357 (void) refcount_add(&hdr->b_refcnt, tag);
1358
1359 return (buf);
1360 }
1361
1362 static char *arc_onloan_tag = "onloan";
1363
1364 /*
1365 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1366 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1367 * buffers must be returned to the arc before they can be used by the DMU or
1368 * freed.
1369 */
1370 arc_buf_t *
1371 arc_loan_buf(spa_t *spa, int size)
1372 {
1373 arc_buf_t *buf;
1374
1375 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1376
1377 atomic_add_64(&arc_loaned_bytes, size);
1378 return (buf);
1379 }
1380
1381 /*
1382 * Return a loaned arc buffer to the arc.
1383 */
1384 void
1385 arc_return_buf(arc_buf_t *buf, void *tag)
1386 {
1387 arc_buf_hdr_t *hdr = buf->b_hdr;
1388
1389 ASSERT(buf->b_data != NULL);
1390 (void) refcount_add(&hdr->b_refcnt, tag);
1391 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1392
1393 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1394 }
1395
1396 /* Detach an arc_buf from a dbuf (tag) */
1397 void
1398 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1399 {
1400 arc_buf_hdr_t *hdr;
1401
1402 ASSERT(buf->b_data != NULL);
1403 hdr = buf->b_hdr;
1404 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1405 (void) refcount_remove(&hdr->b_refcnt, tag);
1406 buf->b_efunc = NULL;
1407 buf->b_private = NULL;
1408
1409 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1410 }
1411
1412 static arc_buf_t *
1413 arc_buf_clone(arc_buf_t *from)
1414 {
1415 arc_buf_t *buf;
1416 arc_buf_hdr_t *hdr = from->b_hdr;
1417 uint64_t size = hdr->b_size;
1418
1419 ASSERT(hdr->b_state != arc_anon);
1420
1421 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1422 buf->b_hdr = hdr;
1423 buf->b_data = NULL;
1424 buf->b_efunc = NULL;
1425 buf->b_private = NULL;
1426 buf->b_next = hdr->b_buf;
1427 hdr->b_buf = buf;
1428 arc_get_data_buf(buf);
1429 bcopy(from->b_data, buf->b_data, size);
1430
1431 /*
1432 * This buffer already exists in the arc so create a duplicate
1433 * copy for the caller. If the buffer is associated with user data
1434 * then track the size and number of duplicates. These stats will be
1435 * updated as duplicate buffers are created and destroyed.
1436 */
1437 if (hdr->b_type == ARC_BUFC_DATA) {
1438 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1439 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1440 }
1441 hdr->b_datacnt += 1;
1442 return (buf);
1443 }
1444
1445 void
1446 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1447 {
1448 arc_buf_hdr_t *hdr;
1449 kmutex_t *hash_lock;
1450
1451 /*
1452 * Check to see if this buffer is evicted. Callers
1453 * must verify b_data != NULL to know if the add_ref
1454 * was successful.
1455 */
1456 mutex_enter(&buf->b_evict_lock);
1457 if (buf->b_data == NULL) {
1458 mutex_exit(&buf->b_evict_lock);
1459 return;
1460 }
1461 hash_lock = HDR_LOCK(buf->b_hdr);
1462 mutex_enter(hash_lock);
1463 hdr = buf->b_hdr;
1464 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1465 mutex_exit(&buf->b_evict_lock);
1466
1467 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1468 add_reference(hdr, hash_lock, tag);
1469 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1470 arc_access(hdr, hash_lock);
1471 mutex_exit(hash_lock);
1472 ARCSTAT_BUMP(arcstat_hits);
1473 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1474 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1475 data, metadata, hits);
1476 }
1477
1478 static void
1479 arc_buf_free_on_write(void *data, size_t size,
1480 void (*free_func)(void *, size_t))
1481 {
1482 l2arc_data_free_t *df;
1483
1484 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1485 df->l2df_data = data;
1486 df->l2df_size = size;
1487 df->l2df_func = free_func;
1488 mutex_enter(&l2arc_free_on_write_mtx);
1489 list_insert_head(l2arc_free_on_write, df);
1490 mutex_exit(&l2arc_free_on_write_mtx);
1491 }
1492
1493 /*
1494 * Free the arc data buffer. If it is an l2arc write in progress,
1495 * the buffer is placed on l2arc_free_on_write to be freed later.
1496 */
1497 static void
1498 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1499 {
1500 arc_buf_hdr_t *hdr = buf->b_hdr;
1501
1502 if (HDR_L2_WRITING(hdr)) {
1503 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func);
1504 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1505 } else {
1506 free_func(buf->b_data, hdr->b_size);
1507 }
1508 }
1509
1510 /*
1511 * Free up buf->b_data and if 'remove' is set, then pull the
1512 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
1513 */
1514 static void
1515 arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr)
1516 {
1517 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1518
1519 ASSERT(MUTEX_HELD(&l2arc_buflist_mtx));
1520
1521 if (l2hdr->b_tmp_cdata == NULL)
1522 return;
1523
1524 ASSERT(HDR_L2_WRITING(hdr));
1525 arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size,
1526 zio_data_buf_free);
1527 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write);
1528 l2hdr->b_tmp_cdata = NULL;
1529 }
1530
1531 static void
1532 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove)
1533 {
1534 arc_buf_t **bufp;
1535
1536 /* free up data associated with the buf */
1537 if (buf->b_data) {
1538 arc_state_t *state = buf->b_hdr->b_state;
1539 uint64_t size = buf->b_hdr->b_size;
1540 arc_buf_contents_t type = buf->b_hdr->b_type;
1541
1542 arc_cksum_verify(buf);
1543 arc_buf_unwatch(buf);
1544
1545 if (!recycle) {
1546 if (type == ARC_BUFC_METADATA) {
1547 arc_buf_data_free(buf, zio_buf_free);
1548 arc_space_return(size, ARC_SPACE_DATA);
1549 } else {
1550 ASSERT(type == ARC_BUFC_DATA);
1551 arc_buf_data_free(buf, zio_data_buf_free);
1552 ARCSTAT_INCR(arcstat_data_size, -size);
1553 atomic_add_64(&arc_size, -size);
1554 }
1555 }
1556 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1557 uint64_t *cnt = &state->arcs_lsize[type];
1558
1559 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1560 ASSERT(state != arc_anon);
1561
1562 ASSERT3U(*cnt, >=, size);
1563 atomic_add_64(cnt, -size);
1564 }
1565 ASSERT3U(state->arcs_size, >=, size);
1566 atomic_add_64(&state->arcs_size, -size);
1567 buf->b_data = NULL;
1568
1569 /*
1570 * If we're destroying a duplicate buffer make sure
1571 * that the appropriate statistics are updated.
1572 */
1573 if (buf->b_hdr->b_datacnt > 1 &&
1574 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1575 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1576 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1577 }
1578 ASSERT(buf->b_hdr->b_datacnt > 0);
1579 buf->b_hdr->b_datacnt -= 1;
1580 }
1581
1582 /* only remove the buf if requested */
1583 if (!remove)
1584 return;
1585
1586 /* remove the buf from the hdr list */
1587 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1588 continue;
1589 *bufp = buf->b_next;
1590 buf->b_next = NULL;
1591
1592 ASSERT(buf->b_efunc == NULL);
1593
1594 /* clean up the buf */
1595 buf->b_hdr = NULL;
1596 kmem_cache_free(buf_cache, buf);
1597 }
1598
1599 static void
1600 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1601 {
1602 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1603 ASSERT3P(hdr->b_state, ==, arc_anon);
1604 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1605 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1606
1607 if (l2hdr != NULL) {
1608 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1609 /*
1610 * To prevent arc_free() and l2arc_evict() from
1611 * attempting to free the same buffer at the same time,
1612 * a FREE_IN_PROGRESS flag is given to arc_free() to
1613 * give it priority. l2arc_evict() can't destroy this
1614 * header while we are waiting on l2arc_buflist_mtx.
1615 *
1616 * The hdr may be removed from l2ad_buflist before we
1617 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1618 */
1619 if (!buflist_held) {
1620 mutex_enter(&l2arc_buflist_mtx);
1621 l2hdr = hdr->b_l2hdr;
1622 }
1623
1624 if (l2hdr != NULL) {
1625 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1626 arc_buf_l2_cdata_free(hdr);
1627 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1628 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
1629 vdev_space_update(l2hdr->b_dev->l2ad_vdev,
1630 -l2hdr->b_asize, 0, 0);
1631 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1632 if (hdr->b_state == arc_l2c_only)
1633 l2arc_hdr_stat_remove();
1634 hdr->b_l2hdr = NULL;
1635 }
1636
1637 if (!buflist_held)
1638 mutex_exit(&l2arc_buflist_mtx);
1639 }
1640
1641 if (!BUF_EMPTY(hdr)) {
1642 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1643 buf_discard_identity(hdr);
1644 }
1645 while (hdr->b_buf) {
1646 arc_buf_t *buf = hdr->b_buf;
1647
1648 if (buf->b_efunc) {
1649 mutex_enter(&arc_eviction_mtx);
1650 mutex_enter(&buf->b_evict_lock);
1651 ASSERT(buf->b_hdr != NULL);
1652 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1653 hdr->b_buf = buf->b_next;
1654 buf->b_hdr = &arc_eviction_hdr;
1655 buf->b_next = arc_eviction_list;
1656 arc_eviction_list = buf;
1657 mutex_exit(&buf->b_evict_lock);
1658 mutex_exit(&arc_eviction_mtx);
1659 } else {
1660 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1661 }
1662 }
1663 if (hdr->b_freeze_cksum != NULL) {
1664 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1665 hdr->b_freeze_cksum = NULL;
1666 }
1667 if (hdr->b_thawed) {
1668 kmem_free(hdr->b_thawed, 1);
1669 hdr->b_thawed = NULL;
1670 }
1671
1672 ASSERT(!list_link_active(&hdr->b_arc_node));
1673 ASSERT3P(hdr->b_hash_next, ==, NULL);
1674 ASSERT3P(hdr->b_acb, ==, NULL);
1675 kmem_cache_free(hdr_cache, hdr);
1676 }
1677
1678 void
1679 arc_buf_free(arc_buf_t *buf, void *tag)
1680 {
1681 arc_buf_hdr_t *hdr = buf->b_hdr;
1682 int hashed = hdr->b_state != arc_anon;
1683
1684 ASSERT(buf->b_efunc == NULL);
1685 ASSERT(buf->b_data != NULL);
1686
1687 if (hashed) {
1688 kmutex_t *hash_lock = HDR_LOCK(hdr);
1689
1690 mutex_enter(hash_lock);
1691 hdr = buf->b_hdr;
1692 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1693
1694 (void) remove_reference(hdr, hash_lock, tag);
1695 if (hdr->b_datacnt > 1) {
1696 arc_buf_destroy(buf, FALSE, TRUE);
1697 } else {
1698 ASSERT(buf == hdr->b_buf);
1699 ASSERT(buf->b_efunc == NULL);
1700 hdr->b_flags |= ARC_BUF_AVAILABLE;
1701 }
1702 mutex_exit(hash_lock);
1703 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1704 int destroy_hdr;
1705 /*
1706 * We are in the middle of an async write. Don't destroy
1707 * this buffer unless the write completes before we finish
1708 * decrementing the reference count.
1709 */
1710 mutex_enter(&arc_eviction_mtx);
1711 (void) remove_reference(hdr, NULL, tag);
1712 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1713 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1714 mutex_exit(&arc_eviction_mtx);
1715 if (destroy_hdr)
1716 arc_hdr_destroy(hdr);
1717 } else {
1718 if (remove_reference(hdr, NULL, tag) > 0)
1719 arc_buf_destroy(buf, FALSE, TRUE);
1720 else
1721 arc_hdr_destroy(hdr);
1722 }
1723 }
1724
1725 boolean_t
1726 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1727 {
1728 arc_buf_hdr_t *hdr = buf->b_hdr;
1729 kmutex_t *hash_lock = HDR_LOCK(hdr);
1730 boolean_t no_callback = (buf->b_efunc == NULL);
1731
1732 if (hdr->b_state == arc_anon) {
1733 ASSERT(hdr->b_datacnt == 1);
1734 arc_buf_free(buf, tag);
1735 return (no_callback);
1736 }
1737
1738 mutex_enter(hash_lock);
1739 hdr = buf->b_hdr;
1740 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1741 ASSERT(hdr->b_state != arc_anon);
1742 ASSERT(buf->b_data != NULL);
1743
1744 (void) remove_reference(hdr, hash_lock, tag);
1745 if (hdr->b_datacnt > 1) {
1746 if (no_callback)
1747 arc_buf_destroy(buf, FALSE, TRUE);
1748 } else if (no_callback) {
1749 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1750 ASSERT(buf->b_efunc == NULL);
1751 hdr->b_flags |= ARC_BUF_AVAILABLE;
1752 }
1753 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1754 refcount_is_zero(&hdr->b_refcnt));
1755 mutex_exit(hash_lock);
1756 return (no_callback);
1757 }
1758
1759 int
1760 arc_buf_size(arc_buf_t *buf)
1761 {
1762 return (buf->b_hdr->b_size);
1763 }
1764
1765 /*
1766 * Called from the DMU to determine if the current buffer should be
1767 * evicted. In order to ensure proper locking, the eviction must be initiated
1768 * from the DMU. Return true if the buffer is associated with user data and
1769 * duplicate buffers still exist.
1770 */
1771 boolean_t
1772 arc_buf_eviction_needed(arc_buf_t *buf)
1773 {
1774 arc_buf_hdr_t *hdr;
1775 boolean_t evict_needed = B_FALSE;
1776
1777 if (zfs_disable_dup_eviction)
1778 return (B_FALSE);
1779
1780 mutex_enter(&buf->b_evict_lock);
1781 hdr = buf->b_hdr;
1782 if (hdr == NULL) {
1783 /*
1784 * We are in arc_do_user_evicts(); let that function
1785 * perform the eviction.
1786 */
1787 ASSERT(buf->b_data == NULL);
1788 mutex_exit(&buf->b_evict_lock);
1789 return (B_FALSE);
1790 } else if (buf->b_data == NULL) {
1791 /*
1792 * We have already been added to the arc eviction list;
1793 * recommend eviction.
1794 */
1795 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1796 mutex_exit(&buf->b_evict_lock);
1797 return (B_TRUE);
1798 }
1799
1800 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1801 evict_needed = B_TRUE;
1802
1803 mutex_exit(&buf->b_evict_lock);
1804 return (evict_needed);
1805 }
1806
1807 /*
1808 * Evict buffers from list until we've removed the specified number of
1809 * bytes. Move the removed buffers to the appropriate evict state.
1810 * If the recycle flag is set, then attempt to "recycle" a buffer:
1811 * - look for a buffer to evict that is `bytes' long.
1812 * - return the data block from this buffer rather than freeing it.
1813 * This flag is used by callers that are trying to make space for a
1814 * new buffer in a full arc cache.
1815 *
1816 * This function makes a "best effort". It skips over any buffers
1817 * it can't get a hash_lock on, and so may not catch all candidates.
1818 * It may also return without evicting as much space as requested.
1819 */
1820 static void *
1821 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1822 arc_buf_contents_t type)
1823 {
1824 arc_state_t *evicted_state;
1825 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1826 arc_buf_hdr_t *ab, *ab_prev = NULL;
1827 list_t *list = &state->arcs_list[type];
1828 kmutex_t *hash_lock;
1829 boolean_t have_lock;
1830 void *stolen = NULL;
1831 arc_buf_hdr_t marker = { 0 };
1832 int count = 0;
1833
1834 ASSERT(state == arc_mru || state == arc_mfu);
1835
1836 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1837
1838 mutex_enter(&state->arcs_mtx);
1839 mutex_enter(&evicted_state->arcs_mtx);
1840
1841 for (ab = list_tail(list); ab; ab = ab_prev) {
1842 ab_prev = list_prev(list, ab);
1843 /* prefetch buffers have a minimum lifespan */
1844 if (HDR_IO_IN_PROGRESS(ab) ||
1845 (spa && ab->b_spa != spa) ||
1846 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1847 ddi_get_lbolt() - ab->b_arc_access <
1848 arc_min_prefetch_lifespan)) {
1849 skipped++;
1850 continue;
1851 }
1852 /* "lookahead" for better eviction candidate */
1853 if (recycle && ab->b_size != bytes &&
1854 ab_prev && ab_prev->b_size == bytes)
1855 continue;
1856
1857 /* ignore markers */
1858 if (ab->b_spa == 0)
1859 continue;
1860
1861 /*
1862 * It may take a long time to evict all the bufs requested.
1863 * To avoid blocking all arc activity, periodically drop
1864 * the arcs_mtx and give other threads a chance to run
1865 * before reacquiring the lock.
1866 *
1867 * If we are looking for a buffer to recycle, we are in
1868 * the hot code path, so don't sleep.
1869 */
1870 if (!recycle && count++ > arc_evict_iterations) {
1871 list_insert_after(list, ab, &marker);
1872 mutex_exit(&evicted_state->arcs_mtx);
1873 mutex_exit(&state->arcs_mtx);
1874 kpreempt(KPREEMPT_SYNC);
1875 mutex_enter(&state->arcs_mtx);
1876 mutex_enter(&evicted_state->arcs_mtx);
1877 ab_prev = list_prev(list, &marker);
1878 list_remove(list, &marker);
1879 count = 0;
1880 continue;
1881 }
1882
1883 hash_lock = HDR_LOCK(ab);
1884 have_lock = MUTEX_HELD(hash_lock);
1885 if (have_lock || mutex_tryenter(hash_lock)) {
1886 ASSERT0(refcount_count(&ab->b_refcnt));
1887 ASSERT(ab->b_datacnt > 0);
1888 while (ab->b_buf) {
1889 arc_buf_t *buf = ab->b_buf;
1890 if (!mutex_tryenter(&buf->b_evict_lock)) {
1891 missed += 1;
1892 break;
1893 }
1894 if (buf->b_data) {
1895 bytes_evicted += ab->b_size;
1896 if (recycle && ab->b_type == type &&
1897 ab->b_size == bytes &&
1898 !HDR_L2_WRITING(ab)) {
1899 stolen = buf->b_data;
1900 recycle = FALSE;
1901 }
1902 }
1903 if (buf->b_efunc) {
1904 mutex_enter(&arc_eviction_mtx);
1905 arc_buf_destroy(buf,
1906 buf->b_data == stolen, FALSE);
1907 ab->b_buf = buf->b_next;
1908 buf->b_hdr = &arc_eviction_hdr;
1909 buf->b_next = arc_eviction_list;
1910 arc_eviction_list = buf;
1911 mutex_exit(&arc_eviction_mtx);
1912 mutex_exit(&buf->b_evict_lock);
1913 } else {
1914 mutex_exit(&buf->b_evict_lock);
1915 arc_buf_destroy(buf,
1916 buf->b_data == stolen, TRUE);
1917 }
1918 }
1919
1920 if (ab->b_l2hdr) {
1921 ARCSTAT_INCR(arcstat_evict_l2_cached,
1922 ab->b_size);
1923 } else {
1924 if (l2arc_write_eligible(ab->b_spa, ab)) {
1925 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1926 ab->b_size);
1927 } else {
1928 ARCSTAT_INCR(
1929 arcstat_evict_l2_ineligible,
1930 ab->b_size);
1931 }
1932 }
1933
1934 if (ab->b_datacnt == 0) {
1935 arc_change_state(evicted_state, ab, hash_lock);
1936 ASSERT(HDR_IN_HASH_TABLE(ab));
1937 ab->b_flags |= ARC_IN_HASH_TABLE;
1938 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1939 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1940 }
1941 if (!have_lock)
1942 mutex_exit(hash_lock);
1943 if (bytes >= 0 && bytes_evicted >= bytes)
1944 break;
1945 } else {
1946 missed += 1;
1947 }
1948 }
1949
1950 mutex_exit(&evicted_state->arcs_mtx);
1951 mutex_exit(&state->arcs_mtx);
1952
1953 if (bytes_evicted < bytes)
1954 dprintf("only evicted %lld bytes from %x",
1955 (longlong_t)bytes_evicted, state);
1956
1957 if (skipped)
1958 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1959
1960 if (missed)
1961 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1962
1963 /*
1964 * Note: we have just evicted some data into the ghost state,
1965 * potentially putting the ghost size over the desired size. Rather
1966 * that evicting from the ghost list in this hot code path, leave
1967 * this chore to the arc_reclaim_thread().
1968 */
1969
1970 return (stolen);
1971 }
1972
1973 /*
1974 * Remove buffers from list until we've removed the specified number of
1975 * bytes. Destroy the buffers that are removed.
1976 */
1977 static void
1978 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1979 {
1980 arc_buf_hdr_t *ab, *ab_prev;
1981 arc_buf_hdr_t marker = { 0 };
1982 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1983 kmutex_t *hash_lock;
1984 uint64_t bytes_deleted = 0;
1985 uint64_t bufs_skipped = 0;
1986 int count = 0;
1987
1988 ASSERT(GHOST_STATE(state));
1989 top:
1990 mutex_enter(&state->arcs_mtx);
1991 for (ab = list_tail(list); ab; ab = ab_prev) {
1992 ab_prev = list_prev(list, ab);
1993 if (ab->b_type > ARC_BUFC_NUMTYPES)
1994 panic("invalid ab=%p", (void *)ab);
1995 if (spa && ab->b_spa != spa)
1996 continue;
1997
1998 /* ignore markers */
1999 if (ab->b_spa == 0)
2000 continue;
2001
2002 hash_lock = HDR_LOCK(ab);
2003 /* caller may be trying to modify this buffer, skip it */
2004 if (MUTEX_HELD(hash_lock))
2005 continue;
2006
2007 /*
2008 * It may take a long time to evict all the bufs requested.
2009 * To avoid blocking all arc activity, periodically drop
2010 * the arcs_mtx and give other threads a chance to run
2011 * before reacquiring the lock.
2012 */
2013 if (count++ > arc_evict_iterations) {
2014 list_insert_after(list, ab, &marker);
2015 mutex_exit(&state->arcs_mtx);
2016 kpreempt(KPREEMPT_SYNC);
2017 mutex_enter(&state->arcs_mtx);
2018 ab_prev = list_prev(list, &marker);
2019 list_remove(list, &marker);
2020 count = 0;
2021 continue;
2022 }
2023 if (mutex_tryenter(hash_lock)) {
2024 ASSERT(!HDR_IO_IN_PROGRESS(ab));
2025 ASSERT(ab->b_buf == NULL);
2026 ARCSTAT_BUMP(arcstat_deleted);
2027 bytes_deleted += ab->b_size;
2028
2029 if (ab->b_l2hdr != NULL) {
2030 /*
2031 * This buffer is cached on the 2nd Level ARC;
2032 * don't destroy the header.
2033 */
2034 arc_change_state(arc_l2c_only, ab, hash_lock);
2035 mutex_exit(hash_lock);
2036 } else {
2037 arc_change_state(arc_anon, ab, hash_lock);
2038 mutex_exit(hash_lock);
2039 arc_hdr_destroy(ab);
2040 }
2041
2042 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
2043 if (bytes >= 0 && bytes_deleted >= bytes)
2044 break;
2045 } else if (bytes < 0) {
2046 /*
2047 * Insert a list marker and then wait for the
2048 * hash lock to become available. Once its
2049 * available, restart from where we left off.
2050 */
2051 list_insert_after(list, ab, &marker);
2052 mutex_exit(&state->arcs_mtx);
2053 mutex_enter(hash_lock);
2054 mutex_exit(hash_lock);
2055 mutex_enter(&state->arcs_mtx);
2056 ab_prev = list_prev(list, &marker);
2057 list_remove(list, &marker);
2058 } else {
2059 bufs_skipped += 1;
2060 }
2061
2062 }
2063 mutex_exit(&state->arcs_mtx);
2064
2065 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
2066 (bytes < 0 || bytes_deleted < bytes)) {
2067 list = &state->arcs_list[ARC_BUFC_METADATA];
2068 goto top;
2069 }
2070
2071 if (bufs_skipped) {
2072 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
2073 ASSERT(bytes >= 0);
2074 }
2075
2076 if (bytes_deleted < bytes)
2077 dprintf("only deleted %lld bytes from %p",
2078 (longlong_t)bytes_deleted, state);
2079 }
2080
2081 static void
2082 arc_adjust(void)
2083 {
2084 int64_t adjustment, delta;
2085
2086 /*
2087 * Adjust MRU size
2088 */
2089
2090 adjustment = MIN((int64_t)(arc_size - arc_c),
2091 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2092 arc_p));
2093
2094 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2095 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2096 (void) arc_evict(arc_mru, NULL, delta, FALSE, ARC_BUFC_DATA);
2097 adjustment -= delta;
2098 }
2099
2100 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2101 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2102 (void) arc_evict(arc_mru, NULL, delta, FALSE,
2103 ARC_BUFC_METADATA);
2104 }
2105
2106 /*
2107 * Adjust MFU size
2108 */
2109
2110 adjustment = arc_size - arc_c;
2111
2112 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2113 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2114 (void) arc_evict(arc_mfu, NULL, delta, FALSE, ARC_BUFC_DATA);
2115 adjustment -= delta;
2116 }
2117
2118 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2119 int64_t delta = MIN(adjustment,
2120 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2121 (void) arc_evict(arc_mfu, NULL, delta, FALSE,
2122 ARC_BUFC_METADATA);
2123 }
2124
2125 /*
2126 * Adjust ghost lists
2127 */
2128
2129 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2130
2131 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2132 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2133 arc_evict_ghost(arc_mru_ghost, NULL, delta);
2134 }
2135
2136 adjustment =
2137 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2138
2139 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2140 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2141 arc_evict_ghost(arc_mfu_ghost, NULL, delta);
2142 }
2143 }
2144
2145 static void
2146 arc_do_user_evicts(void)
2147 {
2148 mutex_enter(&arc_eviction_mtx);
2149 while (arc_eviction_list != NULL) {
2150 arc_buf_t *buf = arc_eviction_list;
2151 arc_eviction_list = buf->b_next;
2152 mutex_enter(&buf->b_evict_lock);
2153 buf->b_hdr = NULL;
2154 mutex_exit(&buf->b_evict_lock);
2155 mutex_exit(&arc_eviction_mtx);
2156
2157 if (buf->b_efunc != NULL)
2158 VERIFY0(buf->b_efunc(buf->b_private));
2159
2160 buf->b_efunc = NULL;
2161 buf->b_private = NULL;
2162 kmem_cache_free(buf_cache, buf);
2163 mutex_enter(&arc_eviction_mtx);
2164 }
2165 mutex_exit(&arc_eviction_mtx);
2166 }
2167
2168 /*
2169 * Flush all *evictable* data from the cache for the given spa.
2170 * NOTE: this will not touch "active" (i.e. referenced) data.
2171 */
2172 void
2173 arc_flush(spa_t *spa)
2174 {
2175 uint64_t guid = 0;
2176
2177 if (spa)
2178 guid = spa_load_guid(spa);
2179
2180 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2181 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2182 if (spa)
2183 break;
2184 }
2185 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2186 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2187 if (spa)
2188 break;
2189 }
2190 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2191 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2192 if (spa)
2193 break;
2194 }
2195 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2196 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2197 if (spa)
2198 break;
2199 }
2200
2201 arc_evict_ghost(arc_mru_ghost, guid, -1);
2202 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2203
2204 mutex_enter(&arc_reclaim_thr_lock);
2205 arc_do_user_evicts();
2206 mutex_exit(&arc_reclaim_thr_lock);
2207 ASSERT(spa || arc_eviction_list == NULL);
2208 }
2209
2210 void
2211 arc_shrink(void)
2212 {
2213 if (arc_c > arc_c_min) {
2214 uint64_t to_free;
2215
2216 #ifdef _KERNEL
2217 to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
2218 #else
2219 to_free = arc_c >> arc_shrink_shift;
2220 #endif
2221 if (arc_c > arc_c_min + to_free)
2222 atomic_add_64(&arc_c, -to_free);
2223 else
2224 arc_c = arc_c_min;
2225
2226 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2227 if (arc_c > arc_size)
2228 arc_c = MAX(arc_size, arc_c_min);
2229 if (arc_p > arc_c)
2230 arc_p = (arc_c >> 1);
2231 ASSERT(arc_c >= arc_c_min);
2232 ASSERT((int64_t)arc_p >= 0);
2233 }
2234
2235 if (arc_size > arc_c)
2236 arc_adjust();
2237 }
2238
2239 /*
2240 * Determine if the system is under memory pressure and is asking
2241 * to reclaim memory. A return value of 1 indicates that the system
2242 * is under memory pressure and that the arc should adjust accordingly.
2243 */
2244 static int
2245 arc_reclaim_needed(void)
2246 {
2247 uint64_t extra;
2248
2249 #ifdef _KERNEL
2250
2251 if (needfree)
2252 return (1);
2253
2254 /*
2255 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2256 */
2257 extra = desfree;
2258
2259 /*
2260 * check that we're out of range of the pageout scanner. It starts to
2261 * schedule paging if freemem is less than lotsfree and needfree.
2262 * lotsfree is the high-water mark for pageout, and needfree is the
2263 * number of needed free pages. We add extra pages here to make sure
2264 * the scanner doesn't start up while we're freeing memory.
2265 */
2266 if (freemem < lotsfree + needfree + extra)
2267 return (1);
2268
2269 /*
2270 * check to make sure that swapfs has enough space so that anon
2271 * reservations can still succeed. anon_resvmem() checks that the
2272 * availrmem is greater than swapfs_minfree, and the number of reserved
2273 * swap pages. We also add a bit of extra here just to prevent
2274 * circumstances from getting really dire.
2275 */
2276 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2277 return (1);
2278
2279 /*
2280 * Check that we have enough availrmem that memory locking (e.g., via
2281 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
2282 * stores the number of pages that cannot be locked; when availrmem
2283 * drops below pages_pp_maximum, page locking mechanisms such as
2284 * page_pp_lock() will fail.)
2285 */
2286 if (availrmem <= pages_pp_maximum)
2287 return (1);
2288
2289 #if defined(__i386)
2290 /*
2291 * If we're on an i386 platform, it's possible that we'll exhaust the
2292 * kernel heap space before we ever run out of available physical
2293 * memory. Most checks of the size of the heap_area compare against
2294 * tune.t_minarmem, which is the minimum available real memory that we
2295 * can have in the system. However, this is generally fixed at 25 pages
2296 * which is so low that it's useless. In this comparison, we seek to
2297 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2298 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2299 * free)
2300 */
2301 if (vmem_size(heap_arena, VMEM_FREE) <
2302 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2))
2303 return (1);
2304 #endif
2305
2306 /*
2307 * If zio data pages are being allocated out of a separate heap segment,
2308 * then enforce that the size of available vmem for this arena remains
2309 * above about 1/16th free.
2310 *
2311 * Note: The 1/16th arena free requirement was put in place
2312 * to aggressively evict memory from the arc in order to avoid
2313 * memory fragmentation issues.
2314 */
2315 if (zio_arena != NULL &&
2316 vmem_size(zio_arena, VMEM_FREE) <
2317 (vmem_size(zio_arena, VMEM_ALLOC) >> 4))
2318 return (1);
2319 #else
2320 if (spa_get_random(100) == 0)
2321 return (1);
2322 #endif
2323 return (0);
2324 }
2325
2326 static void
2327 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2328 {
2329 size_t i;
2330 kmem_cache_t *prev_cache = NULL;
2331 kmem_cache_t *prev_data_cache = NULL;
2332 extern kmem_cache_t *zio_buf_cache[];
2333 extern kmem_cache_t *zio_data_buf_cache[];
2334 extern kmem_cache_t *range_seg_cache;
2335
2336 #ifdef _KERNEL
2337 if (arc_meta_used >= arc_meta_limit) {
2338 /*
2339 * We are exceeding our meta-data cache limit.
2340 * Purge some DNLC entries to release holds on meta-data.
2341 */
2342 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2343 }
2344 #if defined(__i386)
2345 /*
2346 * Reclaim unused memory from all kmem caches.
2347 */
2348 kmem_reap();
2349 #endif
2350 #endif
2351
2352 /*
2353 * An aggressive reclamation will shrink the cache size as well as
2354 * reap free buffers from the arc kmem caches.
2355 */
2356 if (strat == ARC_RECLAIM_AGGR)
2357 arc_shrink();
2358
2359 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2360 if (zio_buf_cache[i] != prev_cache) {
2361 prev_cache = zio_buf_cache[i];
2362 kmem_cache_reap_now(zio_buf_cache[i]);
2363 }
2364 if (zio_data_buf_cache[i] != prev_data_cache) {
2365 prev_data_cache = zio_data_buf_cache[i];
2366 kmem_cache_reap_now(zio_data_buf_cache[i]);
2367 }
2368 }
2369 kmem_cache_reap_now(buf_cache);
2370 kmem_cache_reap_now(hdr_cache);
2371 kmem_cache_reap_now(range_seg_cache);
2372
2373 /*
2374 * Ask the vmem areana to reclaim unused memory from its
2375 * quantum caches.
2376 */
2377 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR)
2378 vmem_qcache_reap(zio_arena);
2379 }
2380
2381 static void
2382 arc_reclaim_thread(void)
2383 {
2384 clock_t growtime = 0;
2385 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2386 callb_cpr_t cpr;
2387
2388 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2389
2390 mutex_enter(&arc_reclaim_thr_lock);
2391 while (arc_thread_exit == 0) {
2392 if (arc_reclaim_needed()) {
2393
2394 if (arc_no_grow) {
2395 if (last_reclaim == ARC_RECLAIM_CONS) {
2396 last_reclaim = ARC_RECLAIM_AGGR;
2397 } else {
2398 last_reclaim = ARC_RECLAIM_CONS;
2399 }
2400 } else {
2401 arc_no_grow = TRUE;
2402 last_reclaim = ARC_RECLAIM_AGGR;
2403 membar_producer();
2404 }
2405
2406 /* reset the growth delay for every reclaim */
2407 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2408
2409 arc_kmem_reap_now(last_reclaim);
2410 arc_warm = B_TRUE;
2411
2412 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2413 arc_no_grow = FALSE;
2414 }
2415
2416 arc_adjust();
2417
2418 if (arc_eviction_list != NULL)
2419 arc_do_user_evicts();
2420
2421 /* block until needed, or one second, whichever is shorter */
2422 CALLB_CPR_SAFE_BEGIN(&cpr);
2423 (void) cv_timedwait(&arc_reclaim_thr_cv,
2424 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2425 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2426 }
2427
2428 arc_thread_exit = 0;
2429 cv_broadcast(&arc_reclaim_thr_cv);
2430 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2431 thread_exit();
2432 }
2433
2434 /*
2435 * Adapt arc info given the number of bytes we are trying to add and
2436 * the state that we are comming from. This function is only called
2437 * when we are adding new content to the cache.
2438 */
2439 static void
2440 arc_adapt(int bytes, arc_state_t *state)
2441 {
2442 int mult;
2443 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2444
2445 if (state == arc_l2c_only)
2446 return;
2447
2448 ASSERT(bytes > 0);
2449 /*
2450 * Adapt the target size of the MRU list:
2451 * - if we just hit in the MRU ghost list, then increase
2452 * the target size of the MRU list.
2453 * - if we just hit in the MFU ghost list, then increase
2454 * the target size of the MFU list by decreasing the
2455 * target size of the MRU list.
2456 */
2457 if (state == arc_mru_ghost) {
2458 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2459 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2460 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2461
2462 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2463 } else if (state == arc_mfu_ghost) {
2464 uint64_t delta;
2465
2466 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2467 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2468 mult = MIN(mult, 10);
2469
2470 delta = MIN(bytes * mult, arc_p);
2471 arc_p = MAX(arc_p_min, arc_p - delta);
2472 }
2473 ASSERT((int64_t)arc_p >= 0);
2474
2475 if (arc_reclaim_needed()) {
2476 cv_signal(&arc_reclaim_thr_cv);
2477 return;
2478 }
2479
2480 if (arc_no_grow)
2481 return;
2482
2483 if (arc_c >= arc_c_max)
2484 return;
2485
2486 /*
2487 * If we're within (2 * maxblocksize) bytes of the target
2488 * cache size, increment the target cache size
2489 */
2490 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2491 atomic_add_64(&arc_c, (int64_t)bytes);
2492 if (arc_c > arc_c_max)
2493 arc_c = arc_c_max;
2494 else if (state == arc_anon)
2495 atomic_add_64(&arc_p, (int64_t)bytes);
2496 if (arc_p > arc_c)
2497 arc_p = arc_c;
2498 }
2499 ASSERT((int64_t)arc_p >= 0);
2500 }
2501
2502 /*
2503 * Check if the cache has reached its limits and eviction is required
2504 * prior to insert.
2505 */
2506 static int
2507 arc_evict_needed(arc_buf_contents_t type)
2508 {
2509 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2510 return (1);
2511
2512 if (arc_reclaim_needed())
2513 return (1);
2514
2515 return (arc_size > arc_c);
2516 }
2517
2518 /*
2519 * The buffer, supplied as the first argument, needs a data block.
2520 * So, if we are at cache max, determine which cache should be victimized.
2521 * We have the following cases:
2522 *
2523 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2524 * In this situation if we're out of space, but the resident size of the MFU is
2525 * under the limit, victimize the MFU cache to satisfy this insertion request.
2526 *
2527 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2528 * Here, we've used up all of the available space for the MRU, so we need to
2529 * evict from our own cache instead. Evict from the set of resident MRU
2530 * entries.
2531 *
2532 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2533 * c minus p represents the MFU space in the cache, since p is the size of the
2534 * cache that is dedicated to the MRU. In this situation there's still space on
2535 * the MFU side, so the MRU side needs to be victimized.
2536 *
2537 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2538 * MFU's resident set is consuming more space than it has been allotted. In
2539 * this situation, we must victimize our own cache, the MFU, for this insertion.
2540 */
2541 static void
2542 arc_get_data_buf(arc_buf_t *buf)
2543 {
2544 arc_state_t *state = buf->b_hdr->b_state;
2545 uint64_t size = buf->b_hdr->b_size;
2546 arc_buf_contents_t type = buf->b_hdr->b_type;
2547
2548 arc_adapt(size, state);
2549
2550 /*
2551 * We have not yet reached cache maximum size,
2552 * just allocate a new buffer.
2553 */
2554 if (!arc_evict_needed(type)) {
2555 if (type == ARC_BUFC_METADATA) {
2556 buf->b_data = zio_buf_alloc(size);
2557 arc_space_consume(size, ARC_SPACE_DATA);
2558 } else {
2559 ASSERT(type == ARC_BUFC_DATA);
2560 buf->b_data = zio_data_buf_alloc(size);
2561 ARCSTAT_INCR(arcstat_data_size, size);
2562 atomic_add_64(&arc_size, size);
2563 }
2564 goto out;
2565 }
2566
2567 /*
2568 * If we are prefetching from the mfu ghost list, this buffer
2569 * will end up on the mru list; so steal space from there.
2570 */
2571 if (state == arc_mfu_ghost)
2572 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2573 else if (state == arc_mru_ghost)
2574 state = arc_mru;
2575
2576 if (state == arc_mru || state == arc_anon) {
2577 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2578 state = (arc_mfu->arcs_lsize[type] >= size &&
2579 arc_p > mru_used) ? arc_mfu : arc_mru;
2580 } else {
2581 /* MFU cases */
2582 uint64_t mfu_space = arc_c - arc_p;
2583 state = (arc_mru->arcs_lsize[type] >= size &&
2584 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2585 }
2586 if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) {
2587 if (type == ARC_BUFC_METADATA) {
2588 buf->b_data = zio_buf_alloc(size);
2589 arc_space_consume(size, ARC_SPACE_DATA);
2590 } else {
2591 ASSERT(type == ARC_BUFC_DATA);
2592 buf->b_data = zio_data_buf_alloc(size);
2593 ARCSTAT_INCR(arcstat_data_size, size);
2594 atomic_add_64(&arc_size, size);
2595 }
2596 ARCSTAT_BUMP(arcstat_recycle_miss);
2597 }
2598 ASSERT(buf->b_data != NULL);
2599 out:
2600 /*
2601 * Update the state size. Note that ghost states have a
2602 * "ghost size" and so don't need to be updated.
2603 */
2604 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2605 arc_buf_hdr_t *hdr = buf->b_hdr;
2606
2607 atomic_add_64(&hdr->b_state->arcs_size, size);
2608 if (list_link_active(&hdr->b_arc_node)) {
2609 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2610 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2611 }
2612 /*
2613 * If we are growing the cache, and we are adding anonymous
2614 * data, and we have outgrown arc_p, update arc_p
2615 */
2616 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2617 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2618 arc_p = MIN(arc_c, arc_p + size);
2619 }
2620 }
2621
2622 /*
2623 * This routine is called whenever a buffer is accessed.
2624 * NOTE: the hash lock is dropped in this function.
2625 */
2626 static void
2627 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2628 {
2629 clock_t now;
2630
2631 ASSERT(MUTEX_HELD(hash_lock));
2632
2633 if (buf->b_state == arc_anon) {
2634 /*
2635 * This buffer is not in the cache, and does not
2636 * appear in our "ghost" list. Add the new buffer
2637 * to the MRU state.
2638 */
2639
2640 ASSERT(buf->b_arc_access == 0);
2641 buf->b_arc_access = ddi_get_lbolt();
2642 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2643 arc_change_state(arc_mru, buf, hash_lock);
2644
2645 } else if (buf->b_state == arc_mru) {
2646 now = ddi_get_lbolt();
2647
2648 /*
2649 * If this buffer is here because of a prefetch, then either:
2650 * - clear the flag if this is a "referencing" read
2651 * (any subsequent access will bump this into the MFU state).
2652 * or
2653 * - move the buffer to the head of the list if this is
2654 * another prefetch (to make it less likely to be evicted).
2655 */
2656 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2657 if (refcount_count(&buf->b_refcnt) == 0) {
2658 ASSERT(list_link_active(&buf->b_arc_node));
2659 } else {
2660 buf->b_flags &= ~ARC_PREFETCH;
2661 ARCSTAT_BUMP(arcstat_mru_hits);
2662 }
2663 buf->b_arc_access = now;
2664 return;
2665 }
2666
2667 /*
2668 * This buffer has been "accessed" only once so far,
2669 * but it is still in the cache. Move it to the MFU
2670 * state.
2671 */
2672 if (now > buf->b_arc_access + ARC_MINTIME) {
2673 /*
2674 * More than 125ms have passed since we
2675 * instantiated this buffer. Move it to the
2676 * most frequently used state.
2677 */
2678 buf->b_arc_access = now;
2679 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2680 arc_change_state(arc_mfu, buf, hash_lock);
2681 }
2682 ARCSTAT_BUMP(arcstat_mru_hits);
2683 } else if (buf->b_state == arc_mru_ghost) {
2684 arc_state_t *new_state;
2685 /*
2686 * This buffer has been "accessed" recently, but
2687 * was evicted from the cache. Move it to the
2688 * MFU state.
2689 */
2690
2691 if (buf->b_flags & ARC_PREFETCH) {
2692 new_state = arc_mru;
2693 if (refcount_count(&buf->b_refcnt) > 0)
2694 buf->b_flags &= ~ARC_PREFETCH;
2695 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2696 } else {
2697 new_state = arc_mfu;
2698 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2699 }
2700
2701 buf->b_arc_access = ddi_get_lbolt();
2702 arc_change_state(new_state, buf, hash_lock);
2703
2704 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2705 } else if (buf->b_state == arc_mfu) {
2706 /*
2707 * This buffer has been accessed more than once and is
2708 * still in the cache. Keep it in the MFU state.
2709 *
2710 * NOTE: an add_reference() that occurred when we did
2711 * the arc_read() will have kicked this off the list.
2712 * If it was a prefetch, we will explicitly move it to
2713 * the head of the list now.
2714 */
2715 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2716 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2717 ASSERT(list_link_active(&buf->b_arc_node));
2718 }
2719 ARCSTAT_BUMP(arcstat_mfu_hits);
2720 buf->b_arc_access = ddi_get_lbolt();
2721 } else if (buf->b_state == arc_mfu_ghost) {
2722 arc_state_t *new_state = arc_mfu;
2723 /*
2724 * This buffer has been accessed more than once but has
2725 * been evicted from the cache. Move it back to the
2726 * MFU state.
2727 */
2728
2729 if (buf->b_flags & ARC_PREFETCH) {
2730 /*
2731 * This is a prefetch access...
2732 * move this block back to the MRU state.
2733 */
2734 ASSERT0(refcount_count(&buf->b_refcnt));
2735 new_state = arc_mru;
2736 }
2737
2738 buf->b_arc_access = ddi_get_lbolt();
2739 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2740 arc_change_state(new_state, buf, hash_lock);
2741
2742 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2743 } else if (buf->b_state == arc_l2c_only) {
2744 /*
2745 * This buffer is on the 2nd Level ARC.
2746 */
2747
2748 buf->b_arc_access = ddi_get_lbolt();
2749 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2750 arc_change_state(arc_mfu, buf, hash_lock);
2751 } else {
2752 ASSERT(!"invalid arc state");
2753 }
2754 }
2755
2756 /* a generic arc_done_func_t which you can use */
2757 /* ARGSUSED */
2758 void
2759 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2760 {
2761 if (zio == NULL || zio->io_error == 0)
2762 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2763 VERIFY(arc_buf_remove_ref(buf, arg));
2764 }
2765
2766 /* a generic arc_done_func_t */
2767 void
2768 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2769 {
2770 arc_buf_t **bufp = arg;
2771 if (zio && zio->io_error) {
2772 VERIFY(arc_buf_remove_ref(buf, arg));
2773 *bufp = NULL;
2774 } else {
2775 *bufp = buf;
2776 ASSERT(buf->b_data);
2777 }
2778 }
2779
2780 static void
2781 arc_read_done(zio_t *zio)
2782 {
2783 arc_buf_hdr_t *hdr;
2784 arc_buf_t *buf;
2785 arc_buf_t *abuf; /* buffer we're assigning to callback */
2786 kmutex_t *hash_lock = NULL;
2787 arc_callback_t *callback_list, *acb;
2788 int freeable = FALSE;
2789
2790 buf = zio->io_private;
2791 hdr = buf->b_hdr;
2792
2793 /*
2794 * The hdr was inserted into hash-table and removed from lists
2795 * prior to starting I/O. We should find this header, since
2796 * it's in the hash table, and it should be legit since it's
2797 * not possible to evict it during the I/O. The only possible
2798 * reason for it not to be found is if we were freed during the
2799 * read.
2800 */
2801 if (HDR_IN_HASH_TABLE(hdr)) {
2802 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
2803 ASSERT3U(hdr->b_dva.dva_word[0], ==,
2804 BP_IDENTITY(zio->io_bp)->dva_word[0]);
2805 ASSERT3U(hdr->b_dva.dva_word[1], ==,
2806 BP_IDENTITY(zio->io_bp)->dva_word[1]);
2807
2808 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
2809 &hash_lock);
2810
2811 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
2812 hash_lock == NULL) ||
2813 (found == hdr &&
2814 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2815 (found == hdr && HDR_L2_READING(hdr)));
2816 }
2817
2818 hdr->b_flags &= ~ARC_L2_EVICTED;
2819 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2820 hdr->b_flags &= ~ARC_L2CACHE;
2821
2822 /* byteswap if necessary */
2823 callback_list = hdr->b_acb;
2824 ASSERT(callback_list != NULL);
2825 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2826 dmu_object_byteswap_t bswap =
2827 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2828 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2829 byteswap_uint64_array :
2830 dmu_ot_byteswap[bswap].ob_func;
2831 func(buf->b_data, hdr->b_size);
2832 }
2833
2834 arc_cksum_compute(buf, B_FALSE);
2835 arc_buf_watch(buf);
2836
2837 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2838 /*
2839 * Only call arc_access on anonymous buffers. This is because
2840 * if we've issued an I/O for an evicted buffer, we've already
2841 * called arc_access (to prevent any simultaneous readers from
2842 * getting confused).
2843 */
2844 arc_access(hdr, hash_lock);
2845 }
2846
2847 /* create copies of the data buffer for the callers */
2848 abuf = buf;
2849 for (acb = callback_list; acb; acb = acb->acb_next) {
2850 if (acb->acb_done) {
2851 if (abuf == NULL) {
2852 ARCSTAT_BUMP(arcstat_duplicate_reads);
2853 abuf = arc_buf_clone(buf);
2854 }
2855 acb->acb_buf = abuf;
2856 abuf = NULL;
2857 }
2858 }
2859 hdr->b_acb = NULL;
2860 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2861 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2862 if (abuf == buf) {
2863 ASSERT(buf->b_efunc == NULL);
2864 ASSERT(hdr->b_datacnt == 1);
2865 hdr->b_flags |= ARC_BUF_AVAILABLE;
2866 }
2867
2868 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2869
2870 if (zio->io_error != 0) {
2871 hdr->b_flags |= ARC_IO_ERROR;
2872 if (hdr->b_state != arc_anon)
2873 arc_change_state(arc_anon, hdr, hash_lock);
2874 if (HDR_IN_HASH_TABLE(hdr))
2875 buf_hash_remove(hdr);
2876 freeable = refcount_is_zero(&hdr->b_refcnt);
2877 }
2878
2879 /*
2880 * Broadcast before we drop the hash_lock to avoid the possibility
2881 * that the hdr (and hence the cv) might be freed before we get to
2882 * the cv_broadcast().
2883 */
2884 cv_broadcast(&hdr->b_cv);
2885
2886 if (hash_lock) {
2887 mutex_exit(hash_lock);
2888 } else {
2889 /*
2890 * This block was freed while we waited for the read to
2891 * complete. It has been removed from the hash table and
2892 * moved to the anonymous state (so that it won't show up
2893 * in the cache).
2894 */
2895 ASSERT3P(hdr->b_state, ==, arc_anon);
2896 freeable = refcount_is_zero(&hdr->b_refcnt);
2897 }
2898
2899 /* execute each callback and free its structure */
2900 while ((acb = callback_list) != NULL) {
2901 if (acb->acb_done)
2902 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2903
2904 if (acb->acb_zio_dummy != NULL) {
2905 acb->acb_zio_dummy->io_error = zio->io_error;
2906 zio_nowait(acb->acb_zio_dummy);
2907 }
2908
2909 callback_list = acb->acb_next;
2910 kmem_free(acb, sizeof (arc_callback_t));
2911 }
2912
2913 if (freeable)
2914 arc_hdr_destroy(hdr);
2915 }
2916
2917 /*
2918 * "Read" the block at the specified DVA (in bp) via the
2919 * cache. If the block is found in the cache, invoke the provided
2920 * callback immediately and return. Note that the `zio' parameter
2921 * in the callback will be NULL in this case, since no IO was
2922 * required. If the block is not in the cache pass the read request
2923 * on to the spa with a substitute callback function, so that the
2924 * requested block will be added to the cache.
2925 *
2926 * If a read request arrives for a block that has a read in-progress,
2927 * either wait for the in-progress read to complete (and return the
2928 * results); or, if this is a read with a "done" func, add a record
2929 * to the read to invoke the "done" func when the read completes,
2930 * and return; or just return.
2931 *
2932 * arc_read_done() will invoke all the requested "done" functions
2933 * for readers of this block.
2934 */
2935 int
2936 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
2937 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags,
2938 const zbookmark_phys_t *zb)
2939 {
2940 arc_buf_hdr_t *hdr = NULL;
2941 arc_buf_t *buf = NULL;
2942 kmutex_t *hash_lock = NULL;
2943 zio_t *rzio;
2944 uint64_t guid = spa_load_guid(spa);
2945
2946 ASSERT(!BP_IS_EMBEDDED(bp) ||
2947 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
2948
2949 top:
2950 if (!BP_IS_EMBEDDED(bp)) {
2951 /*
2952 * Embedded BP's have no DVA and require no I/O to "read".
2953 * Create an anonymous arc buf to back it.
2954 */
2955 hdr = buf_hash_find(guid, bp, &hash_lock);
2956 }
2957
2958 if (hdr != NULL && hdr->b_datacnt > 0) {
2959
2960 *arc_flags |= ARC_CACHED;
2961
2962 if (HDR_IO_IN_PROGRESS(hdr)) {
2963
2964 if (*arc_flags & ARC_WAIT) {
2965 cv_wait(&hdr->b_cv, hash_lock);
2966 mutex_exit(hash_lock);
2967 goto top;
2968 }
2969 ASSERT(*arc_flags & ARC_NOWAIT);
2970
2971 if (done) {
2972 arc_callback_t *acb = NULL;
2973
2974 acb = kmem_zalloc(sizeof (arc_callback_t),
2975 KM_SLEEP);
2976 acb->acb_done = done;
2977 acb->acb_private = private;
2978 if (pio != NULL)
2979 acb->acb_zio_dummy = zio_null(pio,
2980 spa, NULL, NULL, NULL, zio_flags);
2981
2982 ASSERT(acb->acb_done != NULL);
2983 acb->acb_next = hdr->b_acb;
2984 hdr->b_acb = acb;
2985 add_reference(hdr, hash_lock, private);
2986 mutex_exit(hash_lock);
2987 return (0);
2988 }
2989 mutex_exit(hash_lock);
2990 return (0);
2991 }
2992
2993 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2994
2995 if (done) {
2996 add_reference(hdr, hash_lock, private);
2997 /*
2998 * If this block is already in use, create a new
2999 * copy of the data so that we will be guaranteed
3000 * that arc_release() will always succeed.
3001 */
3002 buf = hdr->b_buf;
3003 ASSERT(buf);
3004 ASSERT(buf->b_data);
3005 if (HDR_BUF_AVAILABLE(hdr)) {
3006 ASSERT(buf->b_efunc == NULL);
3007 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3008 } else {
3009 buf = arc_buf_clone(buf);
3010 }
3011
3012 } else if (*arc_flags & ARC_PREFETCH &&
3013 refcount_count(&hdr->b_refcnt) == 0) {
3014 hdr->b_flags |= ARC_PREFETCH;
3015 }
3016 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
3017 arc_access(hdr, hash_lock);
3018 if (*arc_flags & ARC_L2CACHE)
3019 hdr->b_flags |= ARC_L2CACHE;
3020 if (*arc_flags & ARC_L2COMPRESS)
3021 hdr->b_flags |= ARC_L2COMPRESS;
3022 mutex_exit(hash_lock);
3023 ARCSTAT_BUMP(arcstat_hits);
3024 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3025 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3026 data, metadata, hits);
3027
3028 if (done)
3029 done(NULL, buf, private);
3030 } else {
3031 uint64_t size = BP_GET_LSIZE(bp);
3032 arc_callback_t *acb;
3033 vdev_t *vd = NULL;
3034 uint64_t addr = 0;
3035 boolean_t devw = B_FALSE;
3036 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
3037 uint64_t b_asize = 0;
3038
3039 if (hdr == NULL) {
3040 /* this block is not in the cache */
3041 arc_buf_hdr_t *exists = NULL;
3042 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
3043 buf = arc_buf_alloc(spa, size, private, type);
3044 hdr = buf->b_hdr;
3045 if (!BP_IS_EMBEDDED(bp)) {
3046 hdr->b_dva = *BP_IDENTITY(bp);
3047 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
3048 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
3049 exists = buf_hash_insert(hdr, &hash_lock);
3050 }
3051 if (exists != NULL) {
3052 /* somebody beat us to the hash insert */
3053 mutex_exit(hash_lock);
3054 buf_discard_identity(hdr);
3055 (void) arc_buf_remove_ref(buf, private);
3056 goto top; /* restart the IO request */
3057 }
3058 /* if this is a prefetch, we don't have a reference */
3059 if (*arc_flags & ARC_PREFETCH) {
3060 (void) remove_reference(hdr, hash_lock,
3061 private);
3062 hdr->b_flags |= ARC_PREFETCH;
3063 }
3064 if (*arc_flags & ARC_L2CACHE)
3065 hdr->b_flags |= ARC_L2CACHE;
3066 if (*arc_flags & ARC_L2COMPRESS)
3067 hdr->b_flags |= ARC_L2COMPRESS;
3068 if (BP_GET_LEVEL(bp) > 0)
3069 hdr->b_flags |= ARC_INDIRECT;
3070 } else {
3071 /* this block is in the ghost cache */
3072 ASSERT(GHOST_STATE(hdr->b_state));
3073 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3074 ASSERT0(refcount_count(&hdr->b_refcnt));
3075 ASSERT(hdr->b_buf == NULL);
3076
3077 /* if this is a prefetch, we don't have a reference */
3078 if (*arc_flags & ARC_PREFETCH)
3079 hdr->b_flags |= ARC_PREFETCH;
3080 else
3081 add_reference(hdr, hash_lock, private);
3082 if (*arc_flags & ARC_L2CACHE)
3083 hdr->b_flags |= ARC_L2CACHE;
3084 if (*arc_flags & ARC_L2COMPRESS)
3085 hdr->b_flags |= ARC_L2COMPRESS;
3086 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
3087 buf->b_hdr = hdr;
3088 buf->b_data = NULL;
3089 buf->b_efunc = NULL;
3090 buf->b_private = NULL;
3091 buf->b_next = NULL;
3092 hdr->b_buf = buf;
3093 ASSERT(hdr->b_datacnt == 0);
3094 hdr->b_datacnt = 1;
3095 arc_get_data_buf(buf);
3096 arc_access(hdr, hash_lock);
3097 }
3098
3099 ASSERT(!GHOST_STATE(hdr->b_state));
3100
3101 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
3102 acb->acb_done = done;
3103 acb->acb_private = private;
3104
3105 ASSERT(hdr->b_acb == NULL);
3106 hdr->b_acb = acb;
3107 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3108
3109 if (hdr->b_l2hdr != NULL &&
3110 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
3111 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
3112 addr = hdr->b_l2hdr->b_daddr;
3113 b_compress = hdr->b_l2hdr->b_compress;
3114 b_asize = hdr->b_l2hdr->b_asize;
3115 /*
3116 * Lock out device removal.
3117 */
3118 if (vdev_is_dead(vd) ||
3119 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
3120 vd = NULL;
3121 }
3122
3123 if (hash_lock != NULL)
3124 mutex_exit(hash_lock);
3125
3126 /*
3127 * At this point, we have a level 1 cache miss. Try again in
3128 * L2ARC if possible.
3129 */
3130 ASSERT3U(hdr->b_size, ==, size);
3131 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3132 uint64_t, size, zbookmark_phys_t *, zb);
3133 ARCSTAT_BUMP(arcstat_misses);
3134 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3135 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3136 data, metadata, misses);
3137
3138 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3139 /*
3140 * Read from the L2ARC if the following are true:
3141 * 1. The L2ARC vdev was previously cached.
3142 * 2. This buffer still has L2ARC metadata.
3143 * 3. This buffer isn't currently writing to the L2ARC.
3144 * 4. The L2ARC entry wasn't evicted, which may
3145 * also have invalidated the vdev.
3146 * 5. This isn't prefetch and l2arc_noprefetch is set.
3147 */
3148 if (hdr->b_l2hdr != NULL &&
3149 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3150 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3151 l2arc_read_callback_t *cb;
3152
3153 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3154 ARCSTAT_BUMP(arcstat_l2_hits);
3155
3156 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3157 KM_SLEEP);
3158 cb->l2rcb_buf = buf;
3159 cb->l2rcb_spa = spa;
3160 cb->l2rcb_bp = *bp;
3161 cb->l2rcb_zb = *zb;
3162 cb->l2rcb_flags = zio_flags;
3163 cb->l2rcb_compress = b_compress;
3164
3165 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
3166 addr + size < vd->vdev_psize -
3167 VDEV_LABEL_END_SIZE);
3168
3169 /*
3170 * l2arc read. The SCL_L2ARC lock will be
3171 * released by l2arc_read_done().
3172 * Issue a null zio if the underlying buffer
3173 * was squashed to zero size by compression.
3174 */
3175 if (b_compress == ZIO_COMPRESS_EMPTY) {
3176 rzio = zio_null(pio, spa, vd,
3177 l2arc_read_done, cb,
3178 zio_flags | ZIO_FLAG_DONT_CACHE |
3179 ZIO_FLAG_CANFAIL |
3180 ZIO_FLAG_DONT_PROPAGATE |
3181 ZIO_FLAG_DONT_RETRY);
3182 } else {
3183 rzio = zio_read_phys(pio, vd, addr,
3184 b_asize, buf->b_data,
3185 ZIO_CHECKSUM_OFF,
3186 l2arc_read_done, cb, priority,
3187 zio_flags | ZIO_FLAG_DONT_CACHE |
3188 ZIO_FLAG_CANFAIL |
3189 ZIO_FLAG_DONT_PROPAGATE |
3190 ZIO_FLAG_DONT_RETRY, B_FALSE);
3191 }
3192 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3193 zio_t *, rzio);
3194 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
3195
3196 if (*arc_flags & ARC_NOWAIT) {
3197 zio_nowait(rzio);
3198 return (0);
3199 }
3200
3201 ASSERT(*arc_flags & ARC_WAIT);
3202 if (zio_wait(rzio) == 0)
3203 return (0);
3204
3205 /* l2arc read error; goto zio_read() */
3206 } else {
3207 DTRACE_PROBE1(l2arc__miss,
3208 arc_buf_hdr_t *, hdr);
3209 ARCSTAT_BUMP(arcstat_l2_misses);
3210 if (HDR_L2_WRITING(hdr))
3211 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3212 spa_config_exit(spa, SCL_L2ARC, vd);
3213 }
3214 } else {
3215 if (vd != NULL)
3216 spa_config_exit(spa, SCL_L2ARC, vd);
3217 if (l2arc_ndev != 0) {
3218 DTRACE_PROBE1(l2arc__miss,
3219 arc_buf_hdr_t *, hdr);
3220 ARCSTAT_BUMP(arcstat_l2_misses);
3221 }
3222 }
3223
3224 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3225 arc_read_done, buf, priority, zio_flags, zb);
3226
3227 if (*arc_flags & ARC_WAIT)
3228 return (zio_wait(rzio));
3229
3230 ASSERT(*arc_flags & ARC_NOWAIT);
3231 zio_nowait(rzio);
3232 }
3233 return (0);
3234 }
3235
3236 void
3237 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3238 {
3239 ASSERT(buf->b_hdr != NULL);
3240 ASSERT(buf->b_hdr->b_state != arc_anon);
3241 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3242 ASSERT(buf->b_efunc == NULL);
3243 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3244
3245 buf->b_efunc = func;
3246 buf->b_private = private;
3247 }
3248
3249 /*
3250 * Notify the arc that a block was freed, and thus will never be used again.
3251 */
3252 void
3253 arc_freed(spa_t *spa, const blkptr_t *bp)
3254 {
3255 arc_buf_hdr_t *hdr;
3256 kmutex_t *hash_lock;
3257 uint64_t guid = spa_load_guid(spa);
3258
3259 ASSERT(!BP_IS_EMBEDDED(bp));
3260
3261 hdr = buf_hash_find(guid, bp, &hash_lock);
3262 if (hdr == NULL)
3263 return;
3264 if (HDR_BUF_AVAILABLE(hdr)) {
3265 arc_buf_t *buf = hdr->b_buf;
3266 add_reference(hdr, hash_lock, FTAG);
3267 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3268 mutex_exit(hash_lock);
3269
3270 arc_release(buf, FTAG);
3271 (void) arc_buf_remove_ref(buf, FTAG);
3272 } else {
3273 mutex_exit(hash_lock);
3274 }
3275
3276 }
3277
3278 /*
3279 * Clear the user eviction callback set by arc_set_callback(), first calling
3280 * it if it exists. Because the presence of a callback keeps an arc_buf cached
3281 * clearing the callback may result in the arc_buf being destroyed. However,
3282 * it will not result in the *last* arc_buf being destroyed, hence the data
3283 * will remain cached in the ARC. We make a copy of the arc buffer here so
3284 * that we can process the callback without holding any locks.
3285 *
3286 * It's possible that the callback is already in the process of being cleared
3287 * by another thread. In this case we can not clear the callback.
3288 *
3289 * Returns B_TRUE if the callback was successfully called and cleared.
3290 */
3291 boolean_t
3292 arc_clear_callback(arc_buf_t *buf)
3293 {
3294 arc_buf_hdr_t *hdr;
3295 kmutex_t *hash_lock;
3296 arc_evict_func_t *efunc = buf->b_efunc;
3297 void *private = buf->b_private;
3298
3299 mutex_enter(&buf->b_evict_lock);
3300 hdr = buf->b_hdr;
3301 if (hdr == NULL) {
3302 /*
3303 * We are in arc_do_user_evicts().
3304 */
3305 ASSERT(buf->b_data == NULL);
3306 mutex_exit(&buf->b_evict_lock);
3307 return (B_FALSE);
3308 } else if (buf->b_data == NULL) {
3309 /*
3310 * We are on the eviction list; process this buffer now
3311 * but let arc_do_user_evicts() do the reaping.
3312 */
3313 buf->b_efunc = NULL;
3314 mutex_exit(&buf->b_evict_lock);
3315 VERIFY0(efunc(private));
3316 return (B_TRUE);
3317 }
3318 hash_lock = HDR_LOCK(hdr);
3319 mutex_enter(hash_lock);
3320 hdr = buf->b_hdr;
3321 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3322
3323 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3324 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3325
3326 buf->b_efunc = NULL;
3327 buf->b_private = NULL;
3328
3329 if (hdr->b_datacnt > 1) {
3330 mutex_exit(&buf->b_evict_lock);
3331 arc_buf_destroy(buf, FALSE, TRUE);
3332 } else {
3333 ASSERT(buf == hdr->b_buf);
3334 hdr->b_flags |= ARC_BUF_AVAILABLE;
3335 mutex_exit(&buf->b_evict_lock);
3336 }
3337
3338 mutex_exit(hash_lock);
3339 VERIFY0(efunc(private));
3340 return (B_TRUE);
3341 }
3342
3343 /*
3344 * Release this buffer from the cache, making it an anonymous buffer. This
3345 * must be done after a read and prior to modifying the buffer contents.
3346 * If the buffer has more than one reference, we must make
3347 * a new hdr for the buffer.
3348 */
3349 void
3350 arc_release(arc_buf_t *buf, void *tag)
3351 {
3352 arc_buf_hdr_t *hdr;
3353 kmutex_t *hash_lock = NULL;
3354 l2arc_buf_hdr_t *l2hdr;
3355 uint64_t buf_size;
3356
3357 /*
3358 * It would be nice to assert that if it's DMU metadata (level >
3359 * 0 || it's the dnode file), then it must be syncing context.
3360 * But we don't know that information at this level.
3361 */
3362
3363 mutex_enter(&buf->b_evict_lock);
3364 hdr = buf->b_hdr;
3365
3366 /* this buffer is not on any list */
3367 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3368
3369 if (hdr->b_state == arc_anon) {
3370 /* this buffer is already released */
3371 ASSERT(buf->b_efunc == NULL);
3372 } else {
3373 hash_lock = HDR_LOCK(hdr);
3374 mutex_enter(hash_lock);
3375 hdr = buf->b_hdr;
3376 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3377 }
3378
3379 l2hdr = hdr->b_l2hdr;
3380 if (l2hdr) {
3381 mutex_enter(&l2arc_buflist_mtx);
3382 arc_buf_l2_cdata_free(hdr);
3383 hdr->b_l2hdr = NULL;
3384 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3385 }
3386 buf_size = hdr->b_size;
3387
3388 /*
3389 * Do we have more than one buf?
3390 */
3391 if (hdr->b_datacnt > 1) {
3392 arc_buf_hdr_t *nhdr;
3393 arc_buf_t **bufp;
3394 uint64_t blksz = hdr->b_size;
3395 uint64_t spa = hdr->b_spa;
3396 arc_buf_contents_t type = hdr->b_type;
3397 uint32_t flags = hdr->b_flags;
3398
3399 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3400 /*
3401 * Pull the data off of this hdr and attach it to
3402 * a new anonymous hdr.
3403 */
3404 (void) remove_reference(hdr, hash_lock, tag);
3405 bufp = &hdr->b_buf;
3406 while (*bufp != buf)
3407 bufp = &(*bufp)->b_next;
3408 *bufp = buf->b_next;
3409 buf->b_next = NULL;
3410
3411 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3412 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3413 if (refcount_is_zero(&hdr->b_refcnt)) {
3414 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3415 ASSERT3U(*size, >=, hdr->b_size);
3416 atomic_add_64(size, -hdr->b_size);
3417 }
3418
3419 /*
3420 * We're releasing a duplicate user data buffer, update
3421 * our statistics accordingly.
3422 */
3423 if (hdr->b_type == ARC_BUFC_DATA) {
3424 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3425 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3426 -hdr->b_size);
3427 }
3428 hdr->b_datacnt -= 1;
3429 arc_cksum_verify(buf);
3430 arc_buf_unwatch(buf);
3431
3432 mutex_exit(hash_lock);
3433
3434 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3435 nhdr->b_size = blksz;
3436 nhdr->b_spa = spa;
3437 nhdr->b_type = type;
3438 nhdr->b_buf = buf;
3439 nhdr->b_state = arc_anon;
3440 nhdr->b_arc_access = 0;
3441 nhdr->b_flags = flags & ARC_L2_WRITING;
3442 nhdr->b_l2hdr = NULL;
3443 nhdr->b_datacnt = 1;
3444 nhdr->b_freeze_cksum = NULL;
3445 (void) refcount_add(&nhdr->b_refcnt, tag);
3446 buf->b_hdr = nhdr;
3447 mutex_exit(&buf->b_evict_lock);
3448 atomic_add_64(&arc_anon->arcs_size, blksz);
3449 } else {
3450 mutex_exit(&buf->b_evict_lock);
3451 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3452 ASSERT(!list_link_active(&hdr->b_arc_node));
3453 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3454 if (hdr->b_state != arc_anon)
3455 arc_change_state(arc_anon, hdr, hash_lock);
3456 hdr->b_arc_access = 0;
3457 if (hash_lock)
3458 mutex_exit(hash_lock);
3459
3460 buf_discard_identity(hdr);
3461 arc_buf_thaw(buf);
3462 }
3463 buf->b_efunc = NULL;
3464 buf->b_private = NULL;
3465
3466 if (l2hdr) {
3467 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
3468 vdev_space_update(l2hdr->b_dev->l2ad_vdev,
3469 -l2hdr->b_asize, 0, 0);
3470 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3471 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3472 mutex_exit(&l2arc_buflist_mtx);
3473 }
3474 }
3475
3476 int
3477 arc_released(arc_buf_t *buf)
3478 {
3479 int released;
3480
3481 mutex_enter(&buf->b_evict_lock);
3482 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3483 mutex_exit(&buf->b_evict_lock);
3484 return (released);
3485 }
3486
3487 #ifdef ZFS_DEBUG
3488 int
3489 arc_referenced(arc_buf_t *buf)
3490 {
3491 int referenced;
3492
3493 mutex_enter(&buf->b_evict_lock);
3494 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3495 mutex_exit(&buf->b_evict_lock);
3496 return (referenced);
3497 }
3498 #endif
3499
3500 static void
3501 arc_write_ready(zio_t *zio)
3502 {
3503 arc_write_callback_t *callback = zio->io_private;
3504 arc_buf_t *buf = callback->awcb_buf;
3505 arc_buf_hdr_t *hdr = buf->b_hdr;
3506
3507 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3508 callback->awcb_ready(zio, buf, callback->awcb_private);
3509
3510 /*
3511 * If the IO is already in progress, then this is a re-write
3512 * attempt, so we need to thaw and re-compute the cksum.
3513 * It is the responsibility of the callback to handle the
3514 * accounting for any re-write attempt.
3515 */
3516 if (HDR_IO_IN_PROGRESS(hdr)) {
3517 mutex_enter(&hdr->b_freeze_lock);
3518 if (hdr->b_freeze_cksum != NULL) {
3519 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3520 hdr->b_freeze_cksum = NULL;
3521 }
3522 mutex_exit(&hdr->b_freeze_lock);
3523 }
3524 arc_cksum_compute(buf, B_FALSE);
3525 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3526 }
3527
3528 /*
3529 * The SPA calls this callback for each physical write that happens on behalf
3530 * of a logical write. See the comment in dbuf_write_physdone() for details.
3531 */
3532 static void
3533 arc_write_physdone(zio_t *zio)
3534 {
3535 arc_write_callback_t *cb = zio->io_private;
3536 if (cb->awcb_physdone != NULL)
3537 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
3538 }
3539
3540 static void
3541 arc_write_done(zio_t *zio)
3542 {
3543 arc_write_callback_t *callback = zio->io_private;
3544 arc_buf_t *buf = callback->awcb_buf;
3545 arc_buf_hdr_t *hdr = buf->b_hdr;
3546
3547 ASSERT(hdr->b_acb == NULL);
3548
3549 if (zio->io_error == 0) {
3550 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
3551 buf_discard_identity(hdr);
3552 } else {
3553 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3554 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3555 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3556 }
3557 } else {
3558 ASSERT(BUF_EMPTY(hdr));
3559 }
3560
3561 /*
3562 * If the block to be written was all-zero or compressed enough to be
3563 * embedded in the BP, no write was performed so there will be no
3564 * dva/birth/checksum. The buffer must therefore remain anonymous
3565 * (and uncached).
3566 */
3567 if (!BUF_EMPTY(hdr)) {
3568 arc_buf_hdr_t *exists;
3569 kmutex_t *hash_lock;
3570
3571 ASSERT(zio->io_error == 0);
3572
3573 arc_cksum_verify(buf);
3574
3575 exists = buf_hash_insert(hdr, &hash_lock);
3576 if (exists) {
3577 /*
3578 * This can only happen if we overwrite for
3579 * sync-to-convergence, because we remove
3580 * buffers from the hash table when we arc_free().
3581 */
3582 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3583 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3584 panic("bad overwrite, hdr=%p exists=%p",
3585 (void *)hdr, (void *)exists);
3586 ASSERT(refcount_is_zero(&exists->b_refcnt));
3587 arc_change_state(arc_anon, exists, hash_lock);
3588 mutex_exit(hash_lock);
3589 arc_hdr_destroy(exists);
3590 exists = buf_hash_insert(hdr, &hash_lock);
3591 ASSERT3P(exists, ==, NULL);
3592 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3593 /* nopwrite */
3594 ASSERT(zio->io_prop.zp_nopwrite);
3595 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3596 panic("bad nopwrite, hdr=%p exists=%p",
3597 (void *)hdr, (void *)exists);
3598 } else {
3599 /* Dedup */
3600 ASSERT(hdr->b_datacnt == 1);
3601 ASSERT(hdr->b_state == arc_anon);
3602 ASSERT(BP_GET_DEDUP(zio->io_bp));
3603 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3604 }
3605 }
3606 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3607 /* if it's not anon, we are doing a scrub */
3608 if (!exists && hdr->b_state == arc_anon)
3609 arc_access(hdr, hash_lock);
3610 mutex_exit(hash_lock);
3611 } else {
3612 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3613 }
3614
3615 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3616 callback->awcb_done(zio, buf, callback->awcb_private);
3617
3618 kmem_free(callback, sizeof (arc_write_callback_t));
3619 }
3620
3621 zio_t *
3622 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3623 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
3624 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
3625 arc_done_func_t *done, void *private, zio_priority_t priority,
3626 int zio_flags, const zbookmark_phys_t *zb)
3627 {
3628 arc_buf_hdr_t *hdr = buf->b_hdr;
3629 arc_write_callback_t *callback;
3630 zio_t *zio;
3631
3632 ASSERT(ready != NULL);
3633 ASSERT(done != NULL);
3634 ASSERT(!HDR_IO_ERROR(hdr));
3635 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3636 ASSERT(hdr->b_acb == NULL);
3637 if (l2arc)
3638 hdr->b_flags |= ARC_L2CACHE;
3639 if (l2arc_compress)
3640 hdr->b_flags |= ARC_L2COMPRESS;
3641 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3642 callback->awcb_ready = ready;
3643 callback->awcb_physdone = physdone;
3644 callback->awcb_done = done;
3645 callback->awcb_private = private;
3646 callback->awcb_buf = buf;
3647
3648 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3649 arc_write_ready, arc_write_physdone, arc_write_done, callback,
3650 priority, zio_flags, zb);
3651
3652 return (zio);
3653 }
3654
3655 static int
3656 arc_memory_throttle(uint64_t reserve, uint64_t txg)
3657 {
3658 #ifdef _KERNEL
3659 uint64_t available_memory = ptob(freemem);
3660 static uint64_t page_load = 0;
3661 static uint64_t last_txg = 0;
3662
3663 #if defined(__i386)
3664 available_memory =
3665 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3666 #endif
3667
3668 if (freemem > physmem * arc_lotsfree_percent / 100)
3669 return (0);
3670
3671 if (txg > last_txg) {
3672 last_txg = txg;
3673 page_load = 0;
3674 }
3675 /*
3676 * If we are in pageout, we know that memory is already tight,
3677 * the arc is already going to be evicting, so we just want to
3678 * continue to let page writes occur as quickly as possible.
3679 */
3680 if (curproc == proc_pageout) {
3681 if (page_load > MAX(ptob(minfree), available_memory) / 4)
3682 return (SET_ERROR(ERESTART));
3683 /* Note: reserve is inflated, so we deflate */
3684 page_load += reserve / 8;
3685 return (0);
3686 } else if (page_load > 0 && arc_reclaim_needed()) {
3687 /* memory is low, delay before restarting */
3688 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3689 return (SET_ERROR(EAGAIN));
3690 }
3691 page_load = 0;
3692 #endif
3693 return (0);
3694 }
3695
3696 void
3697 arc_tempreserve_clear(uint64_t reserve)
3698 {
3699 atomic_add_64(&arc_tempreserve, -reserve);
3700 ASSERT((int64_t)arc_tempreserve >= 0);
3701 }
3702
3703 int
3704 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3705 {
3706 int error;
3707 uint64_t anon_size;
3708
3709 if (reserve > arc_c/4 && !arc_no_grow)
3710 arc_c = MIN(arc_c_max, reserve * 4);
3711 if (reserve > arc_c)
3712 return (SET_ERROR(ENOMEM));
3713
3714 /*
3715 * Don't count loaned bufs as in flight dirty data to prevent long
3716 * network delays from blocking transactions that are ready to be
3717 * assigned to a txg.
3718 */
3719 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3720
3721 /*
3722 * Writes will, almost always, require additional memory allocations
3723 * in order to compress/encrypt/etc the data. We therefore need to
3724 * make sure that there is sufficient available memory for this.
3725 */
3726 error = arc_memory_throttle(reserve, txg);
3727 if (error != 0)
3728 return (error);
3729
3730 /*
3731 * Throttle writes when the amount of dirty data in the cache
3732 * gets too large. We try to keep the cache less than half full
3733 * of dirty blocks so that our sync times don't grow too large.
3734 * Note: if two requests come in concurrently, we might let them
3735 * both succeed, when one of them should fail. Not a huge deal.
3736 */
3737
3738 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3739 anon_size > arc_c / 4) {
3740 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3741 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3742 arc_tempreserve>>10,
3743 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3744 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3745 reserve>>10, arc_c>>10);
3746 return (SET_ERROR(ERESTART));
3747 }
3748 atomic_add_64(&arc_tempreserve, reserve);
3749 return (0);
3750 }
3751
3752 void
3753 arc_init(void)
3754 {
3755 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3756 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3757
3758 /* Convert seconds to clock ticks */
3759 arc_min_prefetch_lifespan = 1 * hz;
3760
3761 /* Start out with 1/8 of all memory */
3762 arc_c = physmem * PAGESIZE / 8;
3763
3764 #ifdef _KERNEL
3765 /*
3766 * On architectures where the physical memory can be larger
3767 * than the addressable space (intel in 32-bit mode), we may
3768 * need to limit the cache to 1/8 of VM size.
3769 */
3770 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3771 #endif
3772
3773 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3774 arc_c_min = MAX(arc_c / 4, 64<<20);
3775 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
3776 if (arc_c * 8 >= 1<<30)
3777 arc_c_max = (arc_c * 8) - (1<<30);
3778 else
3779 arc_c_max = arc_c_min;
3780 arc_c_max = MAX(arc_c * 6, arc_c_max);
3781
3782 /*
3783 * Allow the tunables to override our calculations if they are
3784 * reasonable (ie. over 64MB)
3785 */
3786 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3787 arc_c_max = zfs_arc_max;
3788 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3789 arc_c_min = zfs_arc_min;
3790
3791 arc_c = arc_c_max;
3792 arc_p = (arc_c >> 1);
3793
3794 /* limit meta-data to 1/4 of the arc capacity */
3795 arc_meta_limit = arc_c_max / 4;
3796
3797 /* Allow the tunable to override if it is reasonable */
3798 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3799 arc_meta_limit = zfs_arc_meta_limit;
3800
3801 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3802 arc_c_min = arc_meta_limit / 2;
3803
3804 if (zfs_arc_grow_retry > 0)
3805 arc_grow_retry = zfs_arc_grow_retry;
3806
3807 if (zfs_arc_shrink_shift > 0)
3808 arc_shrink_shift = zfs_arc_shrink_shift;
3809
3810 if (zfs_arc_p_min_shift > 0)
3811 arc_p_min_shift = zfs_arc_p_min_shift;
3812
3813 /* if kmem_flags are set, lets try to use less memory */
3814 if (kmem_debugging())
3815 arc_c = arc_c / 2;
3816 if (arc_c < arc_c_min)
3817 arc_c = arc_c_min;
3818
3819 arc_anon = &ARC_anon;
3820 arc_mru = &ARC_mru;
3821 arc_mru_ghost = &ARC_mru_ghost;
3822 arc_mfu = &ARC_mfu;
3823 arc_mfu_ghost = &ARC_mfu_ghost;
3824 arc_l2c_only = &ARC_l2c_only;
3825 arc_size = 0;
3826
3827 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3828 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3829 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3830 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3831 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3832 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3833
3834 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3835 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3836 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3837 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3838 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3839 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3840 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3841 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3842 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3843 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3844 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3845 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3846 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3847 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3848 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3849 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3850 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3851 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3852 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3853 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3854
3855 buf_init();
3856
3857 arc_thread_exit = 0;
3858 arc_eviction_list = NULL;
3859 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3860 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3861
3862 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3863 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3864
3865 if (arc_ksp != NULL) {
3866 arc_ksp->ks_data = &arc_stats;
3867 kstat_install(arc_ksp);
3868 }
3869
3870 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3871 TS_RUN, minclsyspri);
3872
3873 arc_dead = FALSE;
3874 arc_warm = B_FALSE;
3875
3876 /*
3877 * Calculate maximum amount of dirty data per pool.
3878 *
3879 * If it has been set by /etc/system, take that.
3880 * Otherwise, use a percentage of physical memory defined by
3881 * zfs_dirty_data_max_percent (default 10%) with a cap at
3882 * zfs_dirty_data_max_max (default 4GB).
3883 */
3884 if (zfs_dirty_data_max == 0) {
3885 zfs_dirty_data_max = physmem * PAGESIZE *
3886 zfs_dirty_data_max_percent / 100;
3887 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
3888 zfs_dirty_data_max_max);
3889 }
3890 }
3891
3892 void
3893 arc_fini(void)
3894 {
3895 mutex_enter(&arc_reclaim_thr_lock);
3896 arc_thread_exit = 1;
3897 while (arc_thread_exit != 0)
3898 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3899 mutex_exit(&arc_reclaim_thr_lock);
3900
3901 arc_flush(NULL);
3902
3903 arc_dead = TRUE;
3904
3905 if (arc_ksp != NULL) {
3906 kstat_delete(arc_ksp);
3907 arc_ksp = NULL;
3908 }
3909
3910 mutex_destroy(&arc_eviction_mtx);
3911 mutex_destroy(&arc_reclaim_thr_lock);
3912 cv_destroy(&arc_reclaim_thr_cv);
3913
3914 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3915 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3916 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3917 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3918 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3919 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3920 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3921 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3922
3923 mutex_destroy(&arc_anon->arcs_mtx);
3924 mutex_destroy(&arc_mru->arcs_mtx);
3925 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3926 mutex_destroy(&arc_mfu->arcs_mtx);
3927 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3928 mutex_destroy(&arc_l2c_only->arcs_mtx);
3929
3930 buf_fini();
3931
3932 ASSERT(arc_loaned_bytes == 0);
3933 }
3934
3935 /*
3936 * Level 2 ARC
3937 *
3938 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3939 * It uses dedicated storage devices to hold cached data, which are populated
3940 * using large infrequent writes. The main role of this cache is to boost
3941 * the performance of random read workloads. The intended L2ARC devices
3942 * include short-stroked disks, solid state disks, and other media with
3943 * substantially faster read latency than disk.
3944 *
3945 * +-----------------------+
3946 * | ARC |
3947 * +-----------------------+
3948 * | ^ ^
3949 * | | |
3950 * l2arc_feed_thread() arc_read()
3951 * | | |
3952 * | l2arc read |
3953 * V | |
3954 * +---------------+ |
3955 * | L2ARC | |
3956 * +---------------+ |
3957 * | ^ |
3958 * l2arc_write() | |
3959 * | | |
3960 * V | |
3961 * +-------+ +-------+
3962 * | vdev | | vdev |
3963 * | cache | | cache |
3964 * +-------+ +-------+
3965 * +=========+ .-----.
3966 * : L2ARC : |-_____-|
3967 * : devices : | Disks |
3968 * +=========+ `-_____-'
3969 *
3970 * Read requests are satisfied from the following sources, in order:
3971 *
3972 * 1) ARC
3973 * 2) vdev cache of L2ARC devices
3974 * 3) L2ARC devices
3975 * 4) vdev cache of disks
3976 * 5) disks
3977 *
3978 * Some L2ARC device types exhibit extremely slow write performance.
3979 * To accommodate for this there are some significant differences between
3980 * the L2ARC and traditional cache design:
3981 *
3982 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3983 * the ARC behave as usual, freeing buffers and placing headers on ghost
3984 * lists. The ARC does not send buffers to the L2ARC during eviction as
3985 * this would add inflated write latencies for all ARC memory pressure.
3986 *
3987 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3988 * It does this by periodically scanning buffers from the eviction-end of
3989 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3990 * not already there. It scans until a headroom of buffers is satisfied,
3991 * which itself is a buffer for ARC eviction. If a compressible buffer is
3992 * found during scanning and selected for writing to an L2ARC device, we
3993 * temporarily boost scanning headroom during the next scan cycle to make
3994 * sure we adapt to compression effects (which might significantly reduce
3995 * the data volume we write to L2ARC). The thread that does this is
3996 * l2arc_feed_thread(), illustrated below; example sizes are included to
3997 * provide a better sense of ratio than this diagram:
3998 *
3999 * head --> tail
4000 * +---------------------+----------+
4001 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4002 * +---------------------+----------+ | o L2ARC eligible
4003 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4004 * +---------------------+----------+ |
4005 * 15.9 Gbytes ^ 32 Mbytes |
4006 * headroom |
4007 * l2arc_feed_thread()
4008 * |
4009 * l2arc write hand <--[oooo]--'
4010 * | 8 Mbyte
4011 * | write max
4012 * V
4013 * +==============================+
4014 * L2ARC dev |####|#|###|###| |####| ... |
4015 * +==============================+
4016 * 32 Gbytes
4017 *
4018 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4019 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4020 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4021 * safe to say that this is an uncommon case, since buffers at the end of
4022 * the ARC lists have moved there due to inactivity.
4023 *
4024 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4025 * then the L2ARC simply misses copying some buffers. This serves as a
4026 * pressure valve to prevent heavy read workloads from both stalling the ARC
4027 * with waits and clogging the L2ARC with writes. This also helps prevent
4028 * the potential for the L2ARC to churn if it attempts to cache content too
4029 * quickly, such as during backups of the entire pool.
4030 *
4031 * 5. After system boot and before the ARC has filled main memory, there are
4032 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4033 * lists can remain mostly static. Instead of searching from tail of these
4034 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4035 * for eligible buffers, greatly increasing its chance of finding them.
4036 *
4037 * The L2ARC device write speed is also boosted during this time so that
4038 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4039 * there are no L2ARC reads, and no fear of degrading read performance
4040 * through increased writes.
4041 *
4042 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4043 * the vdev queue can aggregate them into larger and fewer writes. Each
4044 * device is written to in a rotor fashion, sweeping writes through
4045 * available space then repeating.
4046 *
4047 * 7. The L2ARC does not store dirty content. It never needs to flush
4048 * write buffers back to disk based storage.
4049 *
4050 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4051 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4052 *
4053 * The performance of the L2ARC can be tweaked by a number of tunables, which
4054 * may be necessary for different workloads:
4055 *
4056 * l2arc_write_max max write bytes per interval
4057 * l2arc_write_boost extra write bytes during device warmup
4058 * l2arc_noprefetch skip caching prefetched buffers
4059 * l2arc_headroom number of max device writes to precache
4060 * l2arc_headroom_boost when we find compressed buffers during ARC
4061 * scanning, we multiply headroom by this
4062 * percentage factor for the next scan cycle,
4063 * since more compressed buffers are likely to
4064 * be present
4065 * l2arc_feed_secs seconds between L2ARC writing
4066 *
4067 * Tunables may be removed or added as future performance improvements are
4068 * integrated, and also may become zpool properties.
4069 *
4070 * There are three key functions that control how the L2ARC warms up:
4071 *
4072 * l2arc_write_eligible() check if a buffer is eligible to cache
4073 * l2arc_write_size() calculate how much to write
4074 * l2arc_write_interval() calculate sleep delay between writes
4075 *
4076 * These three functions determine what to write, how much, and how quickly
4077 * to send writes.
4078 */
4079
4080 static boolean_t
4081 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
4082 {
4083 /*
4084 * A buffer is *not* eligible for the L2ARC if it:
4085 * 1. belongs to a different spa.
4086 * 2. is already cached on the L2ARC.
4087 * 3. has an I/O in progress (it may be an incomplete read).
4088 * 4. is flagged not eligible (zfs property).
4089 */
4090 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
4091 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
4092 return (B_FALSE);
4093
4094 return (B_TRUE);
4095 }
4096
4097 static uint64_t
4098 l2arc_write_size(void)
4099 {
4100 uint64_t size;
4101
4102 /*
4103 * Make sure our globals have meaningful values in case the user
4104 * altered them.
4105 */
4106 size = l2arc_write_max;
4107 if (size == 0) {
4108 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
4109 "be greater than zero, resetting it to the default (%d)",
4110 L2ARC_WRITE_SIZE);
4111 size = l2arc_write_max = L2ARC_WRITE_SIZE;
4112 }
4113
4114 if (arc_warm == B_FALSE)
4115 size += l2arc_write_boost;
4116
4117 return (size);
4118
4119 }
4120
4121 static clock_t
4122 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4123 {
4124 clock_t interval, next, now;
4125
4126 /*
4127 * If the ARC lists are busy, increase our write rate; if the
4128 * lists are stale, idle back. This is achieved by checking
4129 * how much we previously wrote - if it was more than half of
4130 * what we wanted, schedule the next write much sooner.
4131 */
4132 if (l2arc_feed_again && wrote > (wanted / 2))
4133 interval = (hz * l2arc_feed_min_ms) / 1000;
4134 else
4135 interval = hz * l2arc_feed_secs;
4136
4137 now = ddi_get_lbolt();
4138 next = MAX(now, MIN(now + interval, began + interval));
4139
4140 return (next);
4141 }
4142
4143 static void
4144 l2arc_hdr_stat_add(void)
4145 {
4146 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4147 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4148 }
4149
4150 static void
4151 l2arc_hdr_stat_remove(void)
4152 {
4153 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4154 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4155 }
4156
4157 /*
4158 * Cycle through L2ARC devices. This is how L2ARC load balances.
4159 * If a device is returned, this also returns holding the spa config lock.
4160 */
4161 static l2arc_dev_t *
4162 l2arc_dev_get_next(void)
4163 {
4164 l2arc_dev_t *first, *next = NULL;
4165
4166 /*
4167 * Lock out the removal of spas (spa_namespace_lock), then removal
4168 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4169 * both locks will be dropped and a spa config lock held instead.
4170 */
4171 mutex_enter(&spa_namespace_lock);
4172 mutex_enter(&l2arc_dev_mtx);
4173
4174 /* if there are no vdevs, there is nothing to do */
4175 if (l2arc_ndev == 0)
4176 goto out;
4177
4178 first = NULL;
4179 next = l2arc_dev_last;
4180 do {
4181 /* loop around the list looking for a non-faulted vdev */
4182 if (next == NULL) {
4183 next = list_head(l2arc_dev_list);
4184 } else {
4185 next = list_next(l2arc_dev_list, next);
4186 if (next == NULL)
4187 next = list_head(l2arc_dev_list);
4188 }
4189
4190 /* if we have come back to the start, bail out */
4191 if (first == NULL)
4192 first = next;
4193 else if (next == first)
4194 break;
4195
4196 } while (vdev_is_dead(next->l2ad_vdev));
4197
4198 /* if we were unable to find any usable vdevs, return NULL */
4199 if (vdev_is_dead(next->l2ad_vdev))
4200 next = NULL;
4201
4202 l2arc_dev_last = next;
4203
4204 out:
4205 mutex_exit(&l2arc_dev_mtx);
4206
4207 /*
4208 * Grab the config lock to prevent the 'next' device from being
4209 * removed while we are writing to it.
4210 */
4211 if (next != NULL)
4212 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4213 mutex_exit(&spa_namespace_lock);
4214
4215 return (next);
4216 }
4217
4218 /*
4219 * Free buffers that were tagged for destruction.
4220 */
4221 static void
4222 l2arc_do_free_on_write()
4223 {
4224 list_t *buflist;
4225 l2arc_data_free_t *df, *df_prev;
4226
4227 mutex_enter(&l2arc_free_on_write_mtx);
4228 buflist = l2arc_free_on_write;
4229
4230 for (df = list_tail(buflist); df; df = df_prev) {
4231 df_prev = list_prev(buflist, df);
4232 ASSERT(df->l2df_data != NULL);
4233 ASSERT(df->l2df_func != NULL);
4234 df->l2df_func(df->l2df_data, df->l2df_size);
4235 list_remove(buflist, df);
4236 kmem_free(df, sizeof (l2arc_data_free_t));
4237 }
4238
4239 mutex_exit(&l2arc_free_on_write_mtx);
4240 }
4241
4242 /*
4243 * A write to a cache device has completed. Update all headers to allow
4244 * reads from these buffers to begin.
4245 */
4246 static void
4247 l2arc_write_done(zio_t *zio)
4248 {
4249 l2arc_write_callback_t *cb;
4250 l2arc_dev_t *dev;
4251 list_t *buflist;
4252 arc_buf_hdr_t *head, *ab, *ab_prev;
4253 l2arc_buf_hdr_t *abl2;
4254 kmutex_t *hash_lock;
4255 int64_t bytes_dropped = 0;
4256
4257 cb = zio->io_private;
4258 ASSERT(cb != NULL);
4259 dev = cb->l2wcb_dev;
4260 ASSERT(dev != NULL);
4261 head = cb->l2wcb_head;
4262 ASSERT(head != NULL);
4263 buflist = dev->l2ad_buflist;
4264 ASSERT(buflist != NULL);
4265 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4266 l2arc_write_callback_t *, cb);
4267
4268 if (zio->io_error != 0)
4269 ARCSTAT_BUMP(arcstat_l2_writes_error);
4270
4271 mutex_enter(&l2arc_buflist_mtx);
4272
4273 /*
4274 * All writes completed, or an error was hit.
4275 */
4276 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4277 ab_prev = list_prev(buflist, ab);
4278 abl2 = ab->b_l2hdr;
4279
4280 /*
4281 * Release the temporary compressed buffer as soon as possible.
4282 */
4283 if (abl2->b_compress != ZIO_COMPRESS_OFF)
4284 l2arc_release_cdata_buf(ab);
4285
4286 hash_lock = HDR_LOCK(ab);
4287 if (!mutex_tryenter(hash_lock)) {
4288 /*
4289 * This buffer misses out. It may be in a stage
4290 * of eviction. Its ARC_L2_WRITING flag will be
4291 * left set, denying reads to this buffer.
4292 */
4293 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4294 continue;
4295 }
4296
4297 if (zio->io_error != 0) {
4298 /*
4299 * Error - drop L2ARC entry.
4300 */
4301 list_remove(buflist, ab);
4302 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4303 bytes_dropped += abl2->b_asize;
4304 ab->b_l2hdr = NULL;
4305 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4306 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4307 }
4308
4309 /*
4310 * Allow ARC to begin reads to this L2ARC entry.
4311 */
4312 ab->b_flags &= ~ARC_L2_WRITING;
4313
4314 mutex_exit(hash_lock);
4315 }
4316
4317 atomic_inc_64(&l2arc_writes_done);
4318 list_remove(buflist, head);
4319 kmem_cache_free(hdr_cache, head);
4320 mutex_exit(&l2arc_buflist_mtx);
4321
4322 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
4323
4324 l2arc_do_free_on_write();
4325
4326 kmem_free(cb, sizeof (l2arc_write_callback_t));
4327 }
4328
4329 /*
4330 * A read to a cache device completed. Validate buffer contents before
4331 * handing over to the regular ARC routines.
4332 */
4333 static void
4334 l2arc_read_done(zio_t *zio)
4335 {
4336 l2arc_read_callback_t *cb;
4337 arc_buf_hdr_t *hdr;
4338 arc_buf_t *buf;
4339 kmutex_t *hash_lock;
4340 int equal;
4341
4342 ASSERT(zio->io_vd != NULL);
4343 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4344
4345 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4346
4347 cb = zio->io_private;
4348 ASSERT(cb != NULL);
4349 buf = cb->l2rcb_buf;
4350 ASSERT(buf != NULL);
4351
4352 hash_lock = HDR_LOCK(buf->b_hdr);
4353 mutex_enter(hash_lock);
4354 hdr = buf->b_hdr;
4355 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4356
4357 /*
4358 * If the buffer was compressed, decompress it first.
4359 */
4360 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
4361 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
4362 ASSERT(zio->io_data != NULL);
4363
4364 /*
4365 * Check this survived the L2ARC journey.
4366 */
4367 equal = arc_cksum_equal(buf);
4368 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4369 mutex_exit(hash_lock);
4370 zio->io_private = buf;
4371 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4372 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4373 arc_read_done(zio);
4374 } else {
4375 mutex_exit(hash_lock);
4376 /*
4377 * Buffer didn't survive caching. Increment stats and
4378 * reissue to the original storage device.
4379 */
4380 if (zio->io_error != 0) {
4381 ARCSTAT_BUMP(arcstat_l2_io_error);
4382 } else {
4383 zio->io_error = SET_ERROR(EIO);
4384 }
4385 if (!equal)
4386 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4387
4388 /*
4389 * If there's no waiter, issue an async i/o to the primary
4390 * storage now. If there *is* a waiter, the caller must
4391 * issue the i/o in a context where it's OK to block.
4392 */
4393 if (zio->io_waiter == NULL) {
4394 zio_t *pio = zio_unique_parent(zio);
4395
4396 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4397
4398 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4399 buf->b_data, zio->io_size, arc_read_done, buf,
4400 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4401 }
4402 }
4403
4404 kmem_free(cb, sizeof (l2arc_read_callback_t));
4405 }
4406
4407 /*
4408 * This is the list priority from which the L2ARC will search for pages to
4409 * cache. This is used within loops (0..3) to cycle through lists in the
4410 * desired order. This order can have a significant effect on cache
4411 * performance.
4412 *
4413 * Currently the metadata lists are hit first, MFU then MRU, followed by
4414 * the data lists. This function returns a locked list, and also returns
4415 * the lock pointer.
4416 */
4417 static list_t *
4418 l2arc_list_locked(int list_num, kmutex_t **lock)
4419 {
4420 list_t *list = NULL;
4421
4422 ASSERT(list_num >= 0 && list_num <= 3);
4423
4424 switch (list_num) {
4425 case 0:
4426 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4427 *lock = &arc_mfu->arcs_mtx;
4428 break;
4429 case 1:
4430 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4431 *lock = &arc_mru->arcs_mtx;
4432 break;
4433 case 2:
4434 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4435 *lock = &arc_mfu->arcs_mtx;
4436 break;
4437 case 3:
4438 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4439 *lock = &arc_mru->arcs_mtx;
4440 break;
4441 }
4442
4443 ASSERT(!(MUTEX_HELD(*lock)));
4444 mutex_enter(*lock);
4445 return (list);
4446 }
4447
4448 /*
4449 * Evict buffers from the device write hand to the distance specified in
4450 * bytes. This distance may span populated buffers, it may span nothing.
4451 * This is clearing a region on the L2ARC device ready for writing.
4452 * If the 'all' boolean is set, every buffer is evicted.
4453 */
4454 static void
4455 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4456 {
4457 list_t *buflist;
4458 l2arc_buf_hdr_t *abl2;
4459 arc_buf_hdr_t *ab, *ab_prev;
4460 kmutex_t *hash_lock;
4461 uint64_t taddr;
4462 int64_t bytes_evicted = 0;
4463
4464 buflist = dev->l2ad_buflist;
4465
4466 if (buflist == NULL)
4467 return;
4468
4469 if (!all && dev->l2ad_first) {
4470 /*
4471 * This is the first sweep through the device. There is
4472 * nothing to evict.
4473 */
4474 return;
4475 }
4476
4477 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4478 /*
4479 * When nearing the end of the device, evict to the end
4480 * before the device write hand jumps to the start.
4481 */
4482 taddr = dev->l2ad_end;
4483 } else {
4484 taddr = dev->l2ad_hand + distance;
4485 }
4486 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4487 uint64_t, taddr, boolean_t, all);
4488
4489 top:
4490 mutex_enter(&l2arc_buflist_mtx);
4491 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4492 ab_prev = list_prev(buflist, ab);
4493
4494 hash_lock = HDR_LOCK(ab);
4495 if (!mutex_tryenter(hash_lock)) {
4496 /*
4497 * Missed the hash lock. Retry.
4498 */
4499 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4500 mutex_exit(&l2arc_buflist_mtx);
4501 mutex_enter(hash_lock);
4502 mutex_exit(hash_lock);
4503 goto top;
4504 }
4505
4506 if (HDR_L2_WRITE_HEAD(ab)) {
4507 /*
4508 * We hit a write head node. Leave it for
4509 * l2arc_write_done().
4510 */
4511 list_remove(buflist, ab);
4512 mutex_exit(hash_lock);
4513 continue;
4514 }
4515
4516 if (!all && ab->b_l2hdr != NULL &&
4517 (ab->b_l2hdr->b_daddr > taddr ||
4518 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4519 /*
4520 * We've evicted to the target address,
4521 * or the end of the device.
4522 */
4523 mutex_exit(hash_lock);
4524 break;
4525 }
4526
4527 if (HDR_FREE_IN_PROGRESS(ab)) {
4528 /*
4529 * Already on the path to destruction.
4530 */
4531 mutex_exit(hash_lock);
4532 continue;
4533 }
4534
4535 if (ab->b_state == arc_l2c_only) {
4536 ASSERT(!HDR_L2_READING(ab));
4537 /*
4538 * This doesn't exist in the ARC. Destroy.
4539 * arc_hdr_destroy() will call list_remove()
4540 * and decrement arcstat_l2_size.
4541 */
4542 arc_change_state(arc_anon, ab, hash_lock);
4543 arc_hdr_destroy(ab);
4544 } else {
4545 /*
4546 * Invalidate issued or about to be issued
4547 * reads, since we may be about to write
4548 * over this location.
4549 */
4550 if (HDR_L2_READING(ab)) {
4551 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4552 ab->b_flags |= ARC_L2_EVICTED;
4553 }
4554
4555 /*
4556 * Tell ARC this no longer exists in L2ARC.
4557 */
4558 if (ab->b_l2hdr != NULL) {
4559 abl2 = ab->b_l2hdr;
4560 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4561 bytes_evicted += abl2->b_asize;
4562 ab->b_l2hdr = NULL;
4563 /*
4564 * We are destroying l2hdr, so ensure that
4565 * its compressed buffer, if any, is not leaked.
4566 */
4567 ASSERT(abl2->b_tmp_cdata == NULL);
4568 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4569 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4570 }
4571 list_remove(buflist, ab);
4572
4573 /*
4574 * This may have been leftover after a
4575 * failed write.
4576 */
4577 ab->b_flags &= ~ARC_L2_WRITING;
4578 }
4579 mutex_exit(hash_lock);
4580 }
4581 mutex_exit(&l2arc_buflist_mtx);
4582
4583 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0);
4584 dev->l2ad_evict = taddr;
4585 }
4586
4587 /*
4588 * Find and write ARC buffers to the L2ARC device.
4589 *
4590 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4591 * for reading until they have completed writing.
4592 * The headroom_boost is an in-out parameter used to maintain headroom boost
4593 * state between calls to this function.
4594 *
4595 * Returns the number of bytes actually written (which may be smaller than
4596 * the delta by which the device hand has changed due to alignment).
4597 */
4598 static uint64_t
4599 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
4600 boolean_t *headroom_boost)
4601 {
4602 arc_buf_hdr_t *ab, *ab_prev, *head;
4603 list_t *list;
4604 uint64_t write_asize, write_psize, write_sz, headroom,
4605 buf_compress_minsz;
4606 void *buf_data;
4607 kmutex_t *list_lock;
4608 boolean_t full;
4609 l2arc_write_callback_t *cb;
4610 zio_t *pio, *wzio;
4611 uint64_t guid = spa_load_guid(spa);
4612 const boolean_t do_headroom_boost = *headroom_boost;
4613
4614 ASSERT(dev->l2ad_vdev != NULL);
4615
4616 /* Lower the flag now, we might want to raise it again later. */
4617 *headroom_boost = B_FALSE;
4618
4619 pio = NULL;
4620 write_sz = write_asize = write_psize = 0;
4621 full = B_FALSE;
4622 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4623 head->b_flags |= ARC_L2_WRITE_HEAD;
4624
4625 /*
4626 * We will want to try to compress buffers that are at least 2x the
4627 * device sector size.
4628 */
4629 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
4630
4631 /*
4632 * Copy buffers for L2ARC writing.
4633 */
4634 mutex_enter(&l2arc_buflist_mtx);
4635 for (int try = 0; try <= 3; try++) {
4636 uint64_t passed_sz = 0;
4637
4638 list = l2arc_list_locked(try, &list_lock);
4639
4640 /*
4641 * L2ARC fast warmup.
4642 *
4643 * Until the ARC is warm and starts to evict, read from the
4644 * head of the ARC lists rather than the tail.
4645 */
4646 if (arc_warm == B_FALSE)
4647 ab = list_head(list);
4648 else
4649 ab = list_tail(list);
4650
4651 headroom = target_sz * l2arc_headroom;
4652 if (do_headroom_boost)
4653 headroom = (headroom * l2arc_headroom_boost) / 100;
4654
4655 for (; ab; ab = ab_prev) {
4656 l2arc_buf_hdr_t *l2hdr;
4657 kmutex_t *hash_lock;
4658 uint64_t buf_sz;
4659
4660 if (arc_warm == B_FALSE)
4661 ab_prev = list_next(list, ab);
4662 else
4663 ab_prev = list_prev(list, ab);
4664
4665 hash_lock = HDR_LOCK(ab);
4666 if (!mutex_tryenter(hash_lock)) {
4667 /*
4668 * Skip this buffer rather than waiting.
4669 */
4670 continue;
4671 }
4672
4673 passed_sz += ab->b_size;
4674 if (passed_sz > headroom) {
4675 /*
4676 * Searched too far.
4677 */
4678 mutex_exit(hash_lock);
4679 break;
4680 }
4681
4682 if (!l2arc_write_eligible(guid, ab)) {
4683 mutex_exit(hash_lock);
4684 continue;
4685 }
4686
4687 if ((write_sz + ab->b_size) > target_sz) {
4688 full = B_TRUE;
4689 mutex_exit(hash_lock);
4690 break;
4691 }
4692
4693 if (pio == NULL) {
4694 /*
4695 * Insert a dummy header on the buflist so
4696 * l2arc_write_done() can find where the
4697 * write buffers begin without searching.
4698 */
4699 list_insert_head(dev->l2ad_buflist, head);
4700
4701 cb = kmem_alloc(
4702 sizeof (l2arc_write_callback_t), KM_SLEEP);
4703 cb->l2wcb_dev = dev;
4704 cb->l2wcb_head = head;
4705 pio = zio_root(spa, l2arc_write_done, cb,
4706 ZIO_FLAG_CANFAIL);
4707 }
4708
4709 /*
4710 * Create and add a new L2ARC header.
4711 */
4712 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4713 l2hdr->b_dev = dev;
4714 ab->b_flags |= ARC_L2_WRITING;
4715
4716 /*
4717 * Temporarily stash the data buffer in b_tmp_cdata.
4718 * The subsequent write step will pick it up from
4719 * there. This is because can't access ab->b_buf
4720 * without holding the hash_lock, which we in turn
4721 * can't access without holding the ARC list locks
4722 * (which we want to avoid during compression/writing).
4723 */
4724 l2hdr->b_compress = ZIO_COMPRESS_OFF;
4725 l2hdr->b_asize = ab->b_size;
4726 l2hdr->b_tmp_cdata = ab->b_buf->b_data;
4727
4728 buf_sz = ab->b_size;
4729 ab->b_l2hdr = l2hdr;
4730
4731 list_insert_head(dev->l2ad_buflist, ab);
4732
4733 /*
4734 * Compute and store the buffer cksum before
4735 * writing. On debug the cksum is verified first.
4736 */
4737 arc_cksum_verify(ab->b_buf);
4738 arc_cksum_compute(ab->b_buf, B_TRUE);
4739
4740 mutex_exit(hash_lock);
4741
4742 write_sz += buf_sz;
4743 }
4744
4745 mutex_exit(list_lock);
4746
4747 if (full == B_TRUE)
4748 break;
4749 }
4750
4751 /* No buffers selected for writing? */
4752 if (pio == NULL) {
4753 ASSERT0(write_sz);
4754 mutex_exit(&l2arc_buflist_mtx);
4755 kmem_cache_free(hdr_cache, head);
4756 return (0);
4757 }
4758
4759 /*
4760 * Now start writing the buffers. We're starting at the write head
4761 * and work backwards, retracing the course of the buffer selector
4762 * loop above.
4763 */
4764 for (ab = list_prev(dev->l2ad_buflist, head); ab;
4765 ab = list_prev(dev->l2ad_buflist, ab)) {
4766 l2arc_buf_hdr_t *l2hdr;
4767 uint64_t buf_sz;
4768
4769 /*
4770 * We shouldn't need to lock the buffer here, since we flagged
4771 * it as ARC_L2_WRITING in the previous step, but we must take
4772 * care to only access its L2 cache parameters. In particular,
4773 * ab->b_buf may be invalid by now due to ARC eviction.
4774 */
4775 l2hdr = ab->b_l2hdr;
4776 l2hdr->b_daddr = dev->l2ad_hand;
4777
4778 if ((ab->b_flags & ARC_L2COMPRESS) &&
4779 l2hdr->b_asize >= buf_compress_minsz) {
4780 if (l2arc_compress_buf(l2hdr)) {
4781 /*
4782 * If compression succeeded, enable headroom
4783 * boost on the next scan cycle.
4784 */
4785 *headroom_boost = B_TRUE;
4786 }
4787 }
4788
4789 /*
4790 * Pick up the buffer data we had previously stashed away
4791 * (and now potentially also compressed).
4792 */
4793 buf_data = l2hdr->b_tmp_cdata;
4794 buf_sz = l2hdr->b_asize;
4795
4796 /*
4797 * If the data has not been compressed, then clear b_tmp_cdata
4798 * to make sure that it points only to a temporary compression
4799 * buffer.
4800 */
4801 if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress))
4802 l2hdr->b_tmp_cdata = NULL;
4803
4804 /* Compression may have squashed the buffer to zero length. */
4805 if (buf_sz != 0) {
4806 uint64_t buf_p_sz;
4807
4808 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4809 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4810 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4811 ZIO_FLAG_CANFAIL, B_FALSE);
4812
4813 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4814 zio_t *, wzio);
4815 (void) zio_nowait(wzio);
4816
4817 write_asize += buf_sz;
4818 /*
4819 * Keep the clock hand suitably device-aligned.
4820 */
4821 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4822 write_psize += buf_p_sz;
4823 dev->l2ad_hand += buf_p_sz;
4824 }
4825 }
4826
4827 mutex_exit(&l2arc_buflist_mtx);
4828
4829 ASSERT3U(write_asize, <=, target_sz);
4830 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4831 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
4832 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4833 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
4834 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
4835
4836 /*
4837 * Bump device hand to the device start if it is approaching the end.
4838 * l2arc_evict() will already have evicted ahead for this case.
4839 */
4840 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4841 dev->l2ad_hand = dev->l2ad_start;
4842 dev->l2ad_evict = dev->l2ad_start;
4843 dev->l2ad_first = B_FALSE;
4844 }
4845
4846 dev->l2ad_writing = B_TRUE;
4847 (void) zio_wait(pio);
4848 dev->l2ad_writing = B_FALSE;
4849
4850 return (write_asize);
4851 }
4852
4853 /*
4854 * Compresses an L2ARC buffer.
4855 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its
4856 * size in l2hdr->b_asize. This routine tries to compress the data and
4857 * depending on the compression result there are three possible outcomes:
4858 * *) The buffer was incompressible. The original l2hdr contents were left
4859 * untouched and are ready for writing to an L2 device.
4860 * *) The buffer was all-zeros, so there is no need to write it to an L2
4861 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
4862 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
4863 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
4864 * data buffer which holds the compressed data to be written, and b_asize
4865 * tells us how much data there is. b_compress is set to the appropriate
4866 * compression algorithm. Once writing is done, invoke
4867 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
4868 *
4869 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
4870 * buffer was incompressible).
4871 */
4872 static boolean_t
4873 l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr)
4874 {
4875 void *cdata;
4876 size_t csize, len, rounded;
4877
4878 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF);
4879 ASSERT(l2hdr->b_tmp_cdata != NULL);
4880
4881 len = l2hdr->b_asize;
4882 cdata = zio_data_buf_alloc(len);
4883 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata,
4884 cdata, l2hdr->b_asize);
4885
4886 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE);
4887 if (rounded > csize) {
4888 bzero((char *)cdata + csize, rounded - csize);
4889 csize = rounded;
4890 }
4891
4892 if (csize == 0) {
4893 /* zero block, indicate that there's nothing to write */
4894 zio_data_buf_free(cdata, len);
4895 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
4896 l2hdr->b_asize = 0;
4897 l2hdr->b_tmp_cdata = NULL;
4898 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
4899 return (B_TRUE);
4900 } else if (csize > 0 && csize < len) {
4901 /*
4902 * Compression succeeded, we'll keep the cdata around for
4903 * writing and release it afterwards.
4904 */
4905 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
4906 l2hdr->b_asize = csize;
4907 l2hdr->b_tmp_cdata = cdata;
4908 ARCSTAT_BUMP(arcstat_l2_compress_successes);
4909 return (B_TRUE);
4910 } else {
4911 /*
4912 * Compression failed, release the compressed buffer.
4913 * l2hdr will be left unmodified.
4914 */
4915 zio_data_buf_free(cdata, len);
4916 ARCSTAT_BUMP(arcstat_l2_compress_failures);
4917 return (B_FALSE);
4918 }
4919 }
4920
4921 /*
4922 * Decompresses a zio read back from an l2arc device. On success, the
4923 * underlying zio's io_data buffer is overwritten by the uncompressed
4924 * version. On decompression error (corrupt compressed stream), the
4925 * zio->io_error value is set to signal an I/O error.
4926 *
4927 * Please note that the compressed data stream is not checksummed, so
4928 * if the underlying device is experiencing data corruption, we may feed
4929 * corrupt data to the decompressor, so the decompressor needs to be
4930 * able to handle this situation (LZ4 does).
4931 */
4932 static void
4933 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
4934 {
4935 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
4936
4937 if (zio->io_error != 0) {
4938 /*
4939 * An io error has occured, just restore the original io
4940 * size in preparation for a main pool read.
4941 */
4942 zio->io_orig_size = zio->io_size = hdr->b_size;
4943 return;
4944 }
4945
4946 if (c == ZIO_COMPRESS_EMPTY) {
4947 /*
4948 * An empty buffer results in a null zio, which means we
4949 * need to fill its io_data after we're done restoring the
4950 * buffer's contents.
4951 */
4952 ASSERT(hdr->b_buf != NULL);
4953 bzero(hdr->b_buf->b_data, hdr->b_size);
4954 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data;
4955 } else {
4956 ASSERT(zio->io_data != NULL);
4957 /*
4958 * We copy the compressed data from the start of the arc buffer
4959 * (the zio_read will have pulled in only what we need, the
4960 * rest is garbage which we will overwrite at decompression)
4961 * and then decompress back to the ARC data buffer. This way we
4962 * can minimize copying by simply decompressing back over the
4963 * original compressed data (rather than decompressing to an
4964 * aux buffer and then copying back the uncompressed buffer,
4965 * which is likely to be much larger).
4966 */
4967 uint64_t csize;
4968 void *cdata;
4969
4970 csize = zio->io_size;
4971 cdata = zio_data_buf_alloc(csize);
4972 bcopy(zio->io_data, cdata, csize);
4973 if (zio_decompress_data(c, cdata, zio->io_data, csize,
4974 hdr->b_size) != 0)
4975 zio->io_error = EIO;
4976 zio_data_buf_free(cdata, csize);
4977 }
4978
4979 /* Restore the expected uncompressed IO size. */
4980 zio->io_orig_size = zio->io_size = hdr->b_size;
4981 }
4982
4983 /*
4984 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
4985 * This buffer serves as a temporary holder of compressed data while
4986 * the buffer entry is being written to an l2arc device. Once that is
4987 * done, we can dispose of it.
4988 */
4989 static void
4990 l2arc_release_cdata_buf(arc_buf_hdr_t *ab)
4991 {
4992 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr;
4993
4994 ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress));
4995 if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) {
4996 /*
4997 * If the data was compressed, then we've allocated a
4998 * temporary buffer for it, so now we need to release it.
4999 */
5000 ASSERT(l2hdr->b_tmp_cdata != NULL);
5001 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size);
5002 l2hdr->b_tmp_cdata = NULL;
5003 } else {
5004 ASSERT(l2hdr->b_tmp_cdata == NULL);
5005 }
5006 }
5007
5008 /*
5009 * This thread feeds the L2ARC at regular intervals. This is the beating
5010 * heart of the L2ARC.
5011 */
5012 static void
5013 l2arc_feed_thread(void)
5014 {
5015 callb_cpr_t cpr;
5016 l2arc_dev_t *dev;
5017 spa_t *spa;
5018 uint64_t size, wrote;
5019 clock_t begin, next = ddi_get_lbolt();
5020 boolean_t headroom_boost = B_FALSE;
5021
5022 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
5023
5024 mutex_enter(&l2arc_feed_thr_lock);
5025
5026 while (l2arc_thread_exit == 0) {
5027 CALLB_CPR_SAFE_BEGIN(&cpr);
5028 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
5029 next);
5030 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
5031 next = ddi_get_lbolt() + hz;
5032
5033 /*
5034 * Quick check for L2ARC devices.
5035 */
5036 mutex_enter(&l2arc_dev_mtx);
5037 if (l2arc_ndev == 0) {
5038 mutex_exit(&l2arc_dev_mtx);
5039 continue;
5040 }
5041 mutex_exit(&l2arc_dev_mtx);
5042 begin = ddi_get_lbolt();
5043
5044 /*
5045 * This selects the next l2arc device to write to, and in
5046 * doing so the next spa to feed from: dev->l2ad_spa. This
5047 * will return NULL if there are now no l2arc devices or if
5048 * they are all faulted.
5049 *
5050 * If a device is returned, its spa's config lock is also
5051 * held to prevent device removal. l2arc_dev_get_next()
5052 * will grab and release l2arc_dev_mtx.
5053 */
5054 if ((dev = l2arc_dev_get_next()) == NULL)
5055 continue;
5056
5057 spa = dev->l2ad_spa;
5058 ASSERT(spa != NULL);
5059
5060 /*
5061 * If the pool is read-only then force the feed thread to
5062 * sleep a little longer.
5063 */
5064 if (!spa_writeable(spa)) {
5065 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
5066 spa_config_exit(spa, SCL_L2ARC, dev);
5067 continue;
5068 }
5069
5070 /*
5071 * Avoid contributing to memory pressure.
5072 */
5073 if (arc_reclaim_needed()) {
5074 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
5075 spa_config_exit(spa, SCL_L2ARC, dev);
5076 continue;
5077 }
5078
5079 ARCSTAT_BUMP(arcstat_l2_feeds);
5080
5081 size = l2arc_write_size();
5082
5083 /*
5084 * Evict L2ARC buffers that will be overwritten.
5085 */
5086 l2arc_evict(dev, size, B_FALSE);
5087
5088 /*
5089 * Write ARC buffers.
5090 */
5091 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
5092
5093 /*
5094 * Calculate interval between writes.
5095 */
5096 next = l2arc_write_interval(begin, size, wrote);
5097 spa_config_exit(spa, SCL_L2ARC, dev);
5098 }
5099
5100 l2arc_thread_exit = 0;
5101 cv_broadcast(&l2arc_feed_thr_cv);
5102 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
5103 thread_exit();
5104 }
5105
5106 boolean_t
5107 l2arc_vdev_present(vdev_t *vd)
5108 {
5109 l2arc_dev_t *dev;
5110
5111 mutex_enter(&l2arc_dev_mtx);
5112 for (dev = list_head(l2arc_dev_list); dev != NULL;
5113 dev = list_next(l2arc_dev_list, dev)) {
5114 if (dev->l2ad_vdev == vd)
5115 break;
5116 }
5117 mutex_exit(&l2arc_dev_mtx);
5118
5119 return (dev != NULL);
5120 }
5121
5122 /*
5123 * Add a vdev for use by the L2ARC. By this point the spa has already
5124 * validated the vdev and opened it.
5125 */
5126 void
5127 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
5128 {
5129 l2arc_dev_t *adddev;
5130
5131 ASSERT(!l2arc_vdev_present(vd));
5132
5133 /*
5134 * Create a new l2arc device entry.
5135 */
5136 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
5137 adddev->l2ad_spa = spa;
5138 adddev->l2ad_vdev = vd;
5139 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
5140 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
5141 adddev->l2ad_hand = adddev->l2ad_start;
5142 adddev->l2ad_evict = adddev->l2ad_start;
5143 adddev->l2ad_first = B_TRUE;
5144 adddev->l2ad_writing = B_FALSE;
5145
5146 /*
5147 * This is a list of all ARC buffers that are still valid on the
5148 * device.
5149 */
5150 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
5151 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
5152 offsetof(arc_buf_hdr_t, b_l2node));
5153
5154 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
5155
5156 /*
5157 * Add device to global list
5158 */
5159 mutex_enter(&l2arc_dev_mtx);
5160 list_insert_head(l2arc_dev_list, adddev);
5161 atomic_inc_64(&l2arc_ndev);
5162 mutex_exit(&l2arc_dev_mtx);
5163 }
5164
5165 /*
5166 * Remove a vdev from the L2ARC.
5167 */
5168 void
5169 l2arc_remove_vdev(vdev_t *vd)
5170 {
5171 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
5172
5173 /*
5174 * Find the device by vdev
5175 */
5176 mutex_enter(&l2arc_dev_mtx);
5177 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
5178 nextdev = list_next(l2arc_dev_list, dev);
5179 if (vd == dev->l2ad_vdev) {
5180 remdev = dev;
5181 break;
5182 }
5183 }
5184 ASSERT(remdev != NULL);
5185
5186 /*
5187 * Remove device from global list
5188 */
5189 list_remove(l2arc_dev_list, remdev);
5190 l2arc_dev_last = NULL; /* may have been invalidated */
5191 atomic_dec_64(&l2arc_ndev);
5192 mutex_exit(&l2arc_dev_mtx);
5193
5194 /*
5195 * Clear all buflists and ARC references. L2ARC device flush.
5196 */
5197 l2arc_evict(remdev, 0, B_TRUE);
5198 list_destroy(remdev->l2ad_buflist);
5199 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5200 kmem_free(remdev, sizeof (l2arc_dev_t));
5201 }
5202
5203 void
5204 l2arc_init(void)
5205 {
5206 l2arc_thread_exit = 0;
5207 l2arc_ndev = 0;
5208 l2arc_writes_sent = 0;
5209 l2arc_writes_done = 0;
5210
5211 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5212 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5213 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5214 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5215 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5216
5217 l2arc_dev_list = &L2ARC_dev_list;
5218 l2arc_free_on_write = &L2ARC_free_on_write;
5219 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5220 offsetof(l2arc_dev_t, l2ad_node));
5221 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5222 offsetof(l2arc_data_free_t, l2df_list_node));
5223 }
5224
5225 void
5226 l2arc_fini(void)
5227 {
5228 /*
5229 * This is called from dmu_fini(), which is called from spa_fini();
5230 * Because of this, we can assume that all l2arc devices have
5231 * already been removed when the pools themselves were removed.
5232 */
5233
5234 l2arc_do_free_on_write();
5235
5236 mutex_destroy(&l2arc_feed_thr_lock);
5237 cv_destroy(&l2arc_feed_thr_cv);
5238 mutex_destroy(&l2arc_dev_mtx);
5239 mutex_destroy(&l2arc_buflist_mtx);
5240 mutex_destroy(&l2arc_free_on_write_mtx);
5241
5242 list_destroy(l2arc_dev_list);
5243 list_destroy(l2arc_free_on_write);
5244 }
5245
5246 void
5247 l2arc_start(void)
5248 {
5249 if (!(spa_mode_global & FWRITE))
5250 return;
5251
5252 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5253 TS_RUN, minclsyspri);
5254 }
5255
5256 void
5257 l2arc_stop(void)
5258 {
5259 if (!(spa_mode_global & FWRITE))
5260 return;
5261
5262 mutex_enter(&l2arc_feed_thr_lock);
5263 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5264 l2arc_thread_exit = 1;
5265 while (l2arc_thread_exit != 0)
5266 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5267 mutex_exit(&l2arc_feed_thr_lock);
5268 }