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