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