1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 */
27
28 /*
29 * DVA-based Adjustable Replacement Cache
30 *
31 * While much of the theory of operation used here is
32 * based on the self-tuning, low overhead replacement cache
33 * presented by Megiddo and Modha at FAST 2003, there are some
34 * significant differences:
35 *
36 * 1. The Megiddo and Modha model assumes any page is evictable.
37 * Pages in its cache cannot be "locked" into memory. This makes
38 * the eviction algorithm simple: evict the last page in the list.
39 * This also make the performance characteristics easy to reason
40 * about. Our cache is not so simple. At any given moment, some
41 * subset of the blocks in the cache are un-evictable because we
42 * have handed out a reference to them. Blocks are only evictable
43 * when there are no external references active. This makes
44 * eviction far more problematic: we choose to evict the evictable
45 * blocks that are the "lowest" in the list.
46 *
47 * There are times when it is not possible to evict the requested
48 * space. In these circumstances we are unable to adjust the cache
49 * size. To prevent the cache growing unbounded at these times we
50 * implement a "cache throttle" that slows the flow of new data
51 * into the cache until we can make space available.
52 *
53 * 2. The Megiddo and Modha model assumes a fixed cache size.
54 * Pages are evicted when the cache is full and there is a cache
55 * miss. Our model has a variable sized cache. It grows with
56 * high use, but also tries to react to memory pressure from the
57 * operating system: decreasing its size when system memory is
58 * tight.
59 *
60 * 3. The Megiddo and Modha model assumes a fixed page size. All
61 * elements of the cache are therefore exactly the same size. So
62 * when adjusting the cache size following a cache miss, its simply
63 * a matter of choosing a single page to evict. In our model, we
64 * have variable sized cache blocks (rangeing from 512 bytes to
65 * 128K bytes). We therefore choose a set of blocks to evict to make
66 * space for a cache miss that approximates as closely as possible
67 * the space used by the new block.
68 *
69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
70 * by N. Megiddo & D. Modha, FAST 2003
71 */
72
73 /*
74 * The locking model:
75 *
76 * A new reference to a cache buffer can be obtained in two
77 * ways: 1) via a hash table lookup using the DVA as a key,
78 * or 2) via one of the ARC lists. The arc_read() interface
79 * uses method 1, while the internal arc algorithms for
80 * adjusting the cache use method 2. We therefore provide two
81 * types of locks: 1) the hash table lock array, and 2) the
82 * arc list locks.
83 *
84 * Buffers do not have their own mutexes, rather they rely on the
85 * hash table mutexes for the bulk of their protection (i.e. most
86 * fields in the arc_buf_hdr_t are protected by these mutexes).
87 *
88 * buf_hash_find() returns the appropriate mutex (held) when it
89 * locates the requested buffer in the hash table. It returns
90 * NULL for the mutex if the buffer was not in the table.
91 *
92 * buf_hash_remove() expects the appropriate hash mutex to be
93 * already held before it is invoked.
94 *
95 * Each arc state also has a mutex which is used to protect the
96 * buffer list associated with the state. When attempting to
97 * obtain a hash table lock while holding an arc list lock you
98 * must use: mutex_tryenter() to avoid deadlock. Also note that
99 * the active state mutex must be held before the ghost state mutex.
100 *
101 * Arc buffers may have an associated eviction callback function.
102 * This function will be invoked prior to removing the buffer (e.g.
103 * in arc_do_user_evicts()). Note however that the data associated
104 * with the buffer may be evicted prior to the callback. The callback
105 * must be made with *no locks held* (to prevent deadlock). Additionally,
106 * the users of callbacks must ensure that their private data is
107 * protected from simultaneous callbacks from arc_buf_evict()
108 * and arc_do_user_evicts().
109 *
110 * Note that the majority of the performance stats are manipulated
111 * with atomic operations.
112 *
113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
114 *
115 * - L2ARC buflist creation
116 * - L2ARC buflist eviction
117 * - L2ARC write completion, which walks L2ARC buflists
118 * - ARC header destruction, as it removes from L2ARC buflists
119 * - ARC header release, as it removes from L2ARC buflists
120 */
121
122 #include <sys/spa.h>
123 #include <sys/zio.h>
124 #include <sys/zio_compress.h>
125 #include <sys/zfs_context.h>
126 #include <sys/arc.h>
127 #include <sys/refcount.h>
128 #include <sys/vdev.h>
129 #include <sys/vdev_impl.h>
130 #ifdef _KERNEL
131 #include <sys/vmsystm.h>
132 #include <vm/anon.h>
133 #include <sys/fs/swapnode.h>
134 #include <sys/dnlc.h>
135 #endif
136 #include <sys/callb.h>
137 #include <sys/kstat.h>
138 #include <zfs_fletcher.h>
139
140 #ifndef _KERNEL
141 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
142 boolean_t arc_watch = B_FALSE;
143 int arc_procfd;
144 #endif
145
146 static kmutex_t arc_reclaim_thr_lock;
147 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
148 static uint8_t arc_thread_exit;
149
150 extern int zfs_write_limit_shift;
151 extern uint64_t zfs_write_limit_max;
152 extern kmutex_t zfs_write_limit_lock;
153
154 #define ARC_REDUCE_DNLC_PERCENT 3
155 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
156
157 typedef enum arc_reclaim_strategy {
158 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
159 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
160 } arc_reclaim_strategy_t;
161
162 /* number of seconds before growing cache again */
163 static int arc_grow_retry = 60;
164
165 /* shift of arc_c for calculating both min and max arc_p */
166 static int arc_p_min_shift = 4;
167
168 /* log2(fraction of arc to reclaim) */
169 static int arc_shrink_shift = 5;
170
171 /*
172 * minimum lifespan of a prefetch block in clock ticks
173 * (initialized in arc_init())
174 */
175 static int arc_min_prefetch_lifespan;
176
177 static int arc_dead;
178
179 /*
180 * The arc has filled available memory and has now warmed up.
181 */
182 static boolean_t arc_warm;
183
184 /*
185 * These tunables are for performance analysis.
186 */
187 uint64_t zfs_arc_max;
188 uint64_t zfs_arc_min;
189 uint64_t zfs_arc_meta_limit = 0;
190 int zfs_arc_grow_retry = 0;
191 int zfs_arc_shrink_shift = 0;
192 int zfs_arc_p_min_shift = 0;
193 int zfs_disable_dup_eviction = 0;
194
195 /*
196 * Note that buffers can be in one of 6 states:
197 * ARC_anon - anonymous (discussed below)
198 * ARC_mru - recently used, currently cached
199 * ARC_mru_ghost - recentely used, no longer in cache
200 * ARC_mfu - frequently used, currently cached
201 * ARC_mfu_ghost - frequently used, no longer in cache
202 * ARC_l2c_only - exists in L2ARC but not other states
203 * When there are no active references to the buffer, they are
204 * are linked onto a list in one of these arc states. These are
205 * the only buffers that can be evicted or deleted. Within each
206 * state there are multiple lists, one for meta-data and one for
207 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
208 * etc.) is tracked separately so that it can be managed more
209 * explicitly: favored over data, limited explicitly.
210 *
211 * Anonymous buffers are buffers that are not associated with
212 * a DVA. These are buffers that hold dirty block copies
213 * before they are written to stable storage. By definition,
214 * they are "ref'd" and are considered part of arc_mru
215 * that cannot be freed. Generally, they will aquire a DVA
216 * as they are written and migrate onto the arc_mru list.
217 *
218 * The ARC_l2c_only state is for buffers that are in the second
219 * level ARC but no longer in any of the ARC_m* lists. The second
220 * level ARC itself may also contain buffers that are in any of
221 * the ARC_m* states - meaning that a buffer can exist in two
222 * places. The reason for the ARC_l2c_only state is to keep the
223 * buffer header in the hash table, so that reads that hit the
224 * second level ARC benefit from these fast lookups.
225 */
226
227 typedef struct arc_state {
228 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
229 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
230 uint64_t arcs_size; /* total amount of data in this state */
231 kmutex_t arcs_mtx;
232 } arc_state_t;
233
234 /* The 6 states: */
235 static arc_state_t ARC_anon;
236 static arc_state_t ARC_mru;
237 static arc_state_t ARC_mru_ghost;
238 static arc_state_t ARC_mfu;
239 static arc_state_t ARC_mfu_ghost;
240 static arc_state_t ARC_l2c_only;
241
242 typedef struct arc_stats {
243 kstat_named_t arcstat_hits;
244 kstat_named_t arcstat_misses;
245 kstat_named_t arcstat_demand_data_hits;
246 kstat_named_t arcstat_demand_data_misses;
247 kstat_named_t arcstat_demand_metadata_hits;
248 kstat_named_t arcstat_demand_metadata_misses;
249 kstat_named_t arcstat_prefetch_data_hits;
250 kstat_named_t arcstat_prefetch_data_misses;
251 kstat_named_t arcstat_prefetch_metadata_hits;
252 kstat_named_t arcstat_prefetch_metadata_misses;
253 kstat_named_t arcstat_mru_hits;
254 kstat_named_t arcstat_mru_ghost_hits;
255 kstat_named_t arcstat_mfu_hits;
256 kstat_named_t arcstat_mfu_ghost_hits;
257 kstat_named_t arcstat_deleted;
258 kstat_named_t arcstat_recycle_miss;
259 /*
260 * Number of buffers that could not be evicted because the hash lock
261 * was held by another thread. The lock may not necessarily be held
262 * by something using the same buffer, since hash locks are shared
263 * by multiple buffers.
264 */
265 kstat_named_t arcstat_mutex_miss;
266 /*
267 * Number of buffers skipped because they have I/O in progress, are
268 * indrect prefetch buffers that have not lived long enough, or are
269 * not from the spa we're trying to evict from.
270 */
271 kstat_named_t arcstat_evict_skip;
272 kstat_named_t arcstat_evict_l2_cached;
273 kstat_named_t arcstat_evict_l2_eligible;
274 kstat_named_t arcstat_evict_l2_ineligible;
275 kstat_named_t arcstat_hash_elements;
276 kstat_named_t arcstat_hash_elements_max;
277 kstat_named_t arcstat_hash_collisions;
278 kstat_named_t arcstat_hash_chains;
279 kstat_named_t arcstat_hash_chain_max;
280 kstat_named_t arcstat_p;
281 kstat_named_t arcstat_c;
282 kstat_named_t arcstat_c_min;
283 kstat_named_t arcstat_c_max;
284 kstat_named_t arcstat_size;
285 kstat_named_t arcstat_hdr_size;
286 kstat_named_t arcstat_data_size;
287 kstat_named_t arcstat_other_size;
288 kstat_named_t arcstat_l2_hits;
289 kstat_named_t arcstat_l2_misses;
290 kstat_named_t arcstat_l2_feeds;
291 kstat_named_t arcstat_l2_rw_clash;
292 kstat_named_t arcstat_l2_read_bytes;
293 kstat_named_t arcstat_l2_write_bytes;
294 kstat_named_t arcstat_l2_writes_sent;
295 kstat_named_t arcstat_l2_writes_done;
296 kstat_named_t arcstat_l2_writes_error;
297 kstat_named_t arcstat_l2_writes_hdr_miss;
298 kstat_named_t arcstat_l2_evict_lock_retry;
299 kstat_named_t arcstat_l2_evict_reading;
300 kstat_named_t arcstat_l2_free_on_write;
301 kstat_named_t arcstat_l2_abort_lowmem;
302 kstat_named_t arcstat_l2_cksum_bad;
303 kstat_named_t arcstat_l2_io_error;
304 kstat_named_t arcstat_l2_size;
305 kstat_named_t arcstat_l2_asize;
306 kstat_named_t arcstat_l2_hdr_size;
307 kstat_named_t arcstat_l2_compress_successes;
308 kstat_named_t arcstat_l2_compress_zeros;
309 kstat_named_t arcstat_l2_compress_failures;
310 kstat_named_t arcstat_memory_throttle_count;
311 kstat_named_t arcstat_duplicate_buffers;
312 kstat_named_t arcstat_duplicate_buffers_size;
313 kstat_named_t arcstat_duplicate_reads;
314 kstat_named_t arcstat_meta_used;
315 kstat_named_t arcstat_meta_limit;
316 kstat_named_t arcstat_meta_max;
317 } arc_stats_t;
318
319 static arc_stats_t arc_stats = {
320 { "hits", KSTAT_DATA_UINT64 },
321 { "misses", KSTAT_DATA_UINT64 },
322 { "demand_data_hits", KSTAT_DATA_UINT64 },
323 { "demand_data_misses", KSTAT_DATA_UINT64 },
324 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
325 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
326 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
327 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
328 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
329 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
330 { "mru_hits", KSTAT_DATA_UINT64 },
331 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
332 { "mfu_hits", KSTAT_DATA_UINT64 },
333 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
334 { "deleted", KSTAT_DATA_UINT64 },
335 { "recycle_miss", KSTAT_DATA_UINT64 },
336 { "mutex_miss", KSTAT_DATA_UINT64 },
337 { "evict_skip", KSTAT_DATA_UINT64 },
338 { "evict_l2_cached", KSTAT_DATA_UINT64 },
339 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
340 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
341 { "hash_elements", KSTAT_DATA_UINT64 },
342 { "hash_elements_max", KSTAT_DATA_UINT64 },
343 { "hash_collisions", KSTAT_DATA_UINT64 },
344 { "hash_chains", KSTAT_DATA_UINT64 },
345 { "hash_chain_max", KSTAT_DATA_UINT64 },
346 { "p", KSTAT_DATA_UINT64 },
347 { "c", KSTAT_DATA_UINT64 },
348 { "c_min", KSTAT_DATA_UINT64 },
349 { "c_max", KSTAT_DATA_UINT64 },
350 { "size", KSTAT_DATA_UINT64 },
351 { "hdr_size", KSTAT_DATA_UINT64 },
352 { "data_size", KSTAT_DATA_UINT64 },
353 { "other_size", KSTAT_DATA_UINT64 },
354 { "l2_hits", KSTAT_DATA_UINT64 },
355 { "l2_misses", KSTAT_DATA_UINT64 },
356 { "l2_feeds", KSTAT_DATA_UINT64 },
357 { "l2_rw_clash", KSTAT_DATA_UINT64 },
358 { "l2_read_bytes", KSTAT_DATA_UINT64 },
359 { "l2_write_bytes", KSTAT_DATA_UINT64 },
360 { "l2_writes_sent", KSTAT_DATA_UINT64 },
361 { "l2_writes_done", KSTAT_DATA_UINT64 },
362 { "l2_writes_error", KSTAT_DATA_UINT64 },
363 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
364 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
365 { "l2_evict_reading", KSTAT_DATA_UINT64 },
366 { "l2_free_on_write", KSTAT_DATA_UINT64 },
367 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
368 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
369 { "l2_io_error", KSTAT_DATA_UINT64 },
370 { "l2_size", KSTAT_DATA_UINT64 },
371 { "l2_asize", KSTAT_DATA_UINT64 },
372 { "l2_hdr_size", KSTAT_DATA_UINT64 },
373 { "l2_compress_successes", KSTAT_DATA_UINT64 },
374 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
375 { "l2_compress_failures", KSTAT_DATA_UINT64 },
376 { "memory_throttle_count", KSTAT_DATA_UINT64 },
377 { "duplicate_buffers", KSTAT_DATA_UINT64 },
378 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
379 { "duplicate_reads", KSTAT_DATA_UINT64 },
380 { "arc_meta_used", KSTAT_DATA_UINT64 },
381 { "arc_meta_limit", KSTAT_DATA_UINT64 },
382 { "arc_meta_max", KSTAT_DATA_UINT64 }
383 };
384
385 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
386
387 #define ARCSTAT_INCR(stat, val) \
388 atomic_add_64(&arc_stats.stat.value.ui64, (val))
389
390 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
391 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
392
393 #define ARCSTAT_MAX(stat, val) { \
394 uint64_t m; \
395 while ((val) > (m = arc_stats.stat.value.ui64) && \
396 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
397 continue; \
398 }
399
400 #define ARCSTAT_MAXSTAT(stat) \
401 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
402
403 /*
404 * We define a macro to allow ARC hits/misses to be easily broken down by
405 * two separate conditions, giving a total of four different subtypes for
406 * each of hits and misses (so eight statistics total).
407 */
408 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
409 if (cond1) { \
410 if (cond2) { \
411 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
412 } else { \
413 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
414 } \
415 } else { \
416 if (cond2) { \
417 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
418 } else { \
419 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
420 } \
421 }
422
423 kstat_t *arc_ksp;
424 static arc_state_t *arc_anon;
425 static arc_state_t *arc_mru;
426 static arc_state_t *arc_mru_ghost;
427 static arc_state_t *arc_mfu;
428 static arc_state_t *arc_mfu_ghost;
429 static arc_state_t *arc_l2c_only;
430
431 /*
432 * There are several ARC variables that are critical to export as kstats --
433 * but we don't want to have to grovel around in the kstat whenever we wish to
434 * manipulate them. For these variables, we therefore define them to be in
435 * terms of the statistic variable. This assures that we are not introducing
436 * the possibility of inconsistency by having shadow copies of the variables,
437 * while still allowing the code to be readable.
438 */
439 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
440 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
441 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
442 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
443 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
444 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
445 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
446 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
447
448 #define L2ARC_IS_VALID_COMPRESS(_c_) \
449 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
450
451 static int arc_no_grow; /* Don't try to grow cache size */
452 static uint64_t arc_tempreserve;
453 static uint64_t arc_loaned_bytes;
454
455 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
456
457 typedef struct arc_callback arc_callback_t;
458
459 struct arc_callback {
460 void *acb_private;
461 arc_done_func_t *acb_done;
462 arc_buf_t *acb_buf;
463 zio_t *acb_zio_dummy;
464 arc_callback_t *acb_next;
465 };
466
467 typedef struct arc_write_callback arc_write_callback_t;
468
469 struct arc_write_callback {
470 void *awcb_private;
471 arc_done_func_t *awcb_ready;
472 arc_done_func_t *awcb_done;
473 arc_buf_t *awcb_buf;
474 };
475
476 struct arc_buf_hdr {
477 /* protected by hash lock */
478 dva_t b_dva;
479 uint64_t b_birth;
480 uint64_t b_cksum0;
481
482 kmutex_t b_freeze_lock;
483 zio_cksum_t *b_freeze_cksum;
484 void *b_thawed;
485
486 arc_buf_hdr_t *b_hash_next;
487 arc_buf_t *b_buf;
488 uint32_t b_flags;
489 uint32_t b_datacnt;
490
491 arc_callback_t *b_acb;
492 kcondvar_t b_cv;
493
494 /* immutable */
495 arc_buf_contents_t b_type;
496 uint64_t b_size;
497 uint64_t b_spa;
498
499 /* protected by arc state mutex */
500 arc_state_t *b_state;
501 list_node_t b_arc_node;
502
503 /* updated atomically */
504 clock_t b_arc_access;
505
506 /* self protecting */
507 refcount_t b_refcnt;
508
509 l2arc_buf_hdr_t *b_l2hdr;
510 list_node_t b_l2node;
511 };
512
513 static arc_buf_t *arc_eviction_list;
514 static kmutex_t arc_eviction_mtx;
515 static arc_buf_hdr_t arc_eviction_hdr;
516 static void arc_get_data_buf(arc_buf_t *buf);
517 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
518 static int arc_evict_needed(arc_buf_contents_t type);
519 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
520 static void arc_buf_watch(arc_buf_t *buf);
521
522 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
523
524 #define GHOST_STATE(state) \
525 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
526 (state) == arc_l2c_only)
527
528 /*
529 * Private ARC flags. These flags are private ARC only flags that will show up
530 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
531 * be passed in as arc_flags in things like arc_read. However, these flags
532 * should never be passed and should only be set by ARC code. When adding new
533 * public flags, make sure not to smash the private ones.
534 */
535
536 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
537 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
538 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
539 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
540 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
541 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
542 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
543 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
544 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
545 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
546
547 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
548 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
549 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
550 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
551 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
552 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
553 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
554 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
555 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
556 (hdr)->b_l2hdr != NULL)
557 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
558 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
559 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
560
561 /*
562 * Other sizes
563 */
564
565 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
566 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
567
568 /*
569 * Hash table routines
570 */
571
572 #define HT_LOCK_PAD 64
573
574 struct ht_lock {
575 kmutex_t ht_lock;
576 #ifdef _KERNEL
577 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
578 #endif
579 };
580
581 #define BUF_LOCKS 256
582 typedef struct buf_hash_table {
583 uint64_t ht_mask;
584 arc_buf_hdr_t **ht_table;
585 struct ht_lock ht_locks[BUF_LOCKS];
586 } buf_hash_table_t;
587
588 static buf_hash_table_t buf_hash_table;
589
590 #define BUF_HASH_INDEX(spa, dva, birth) \
591 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
592 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
593 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
594 #define HDR_LOCK(hdr) \
595 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
596
597 uint64_t zfs_crc64_table[256];
598
599 /*
600 * Level 2 ARC
601 */
602
603 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
604 #define L2ARC_HEADROOM 2 /* num of writes */
605 /*
606 * If we discover during ARC scan any buffers to be compressed, we boost
607 * our headroom for the next scanning cycle by this percentage multiple.
608 */
609 #define L2ARC_HEADROOM_BOOST 200
610 #define L2ARC_FEED_SECS 1 /* caching interval secs */
611 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
612
613 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
614 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
615
616 /* L2ARC Performance Tunables */
617 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
618 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
619 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
620 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
621 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
622 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
623 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
624 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
625 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
626
627 /*
628 * L2ARC Internals
629 */
630 typedef struct l2arc_dev {
631 vdev_t *l2ad_vdev; /* vdev */
632 spa_t *l2ad_spa; /* spa */
633 uint64_t l2ad_hand; /* next write location */
634 uint64_t l2ad_start; /* first addr on device */
635 uint64_t l2ad_end; /* last addr on device */
636 uint64_t l2ad_evict; /* last addr eviction reached */
637 boolean_t l2ad_first; /* first sweep through */
638 boolean_t l2ad_writing; /* currently writing */
639 list_t *l2ad_buflist; /* buffer list */
640 list_node_t l2ad_node; /* device list node */
641 } l2arc_dev_t;
642
643 static list_t L2ARC_dev_list; /* device list */
644 static list_t *l2arc_dev_list; /* device list pointer */
645 static kmutex_t l2arc_dev_mtx; /* device list mutex */
646 static l2arc_dev_t *l2arc_dev_last; /* last device used */
647 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
648 static list_t L2ARC_free_on_write; /* free after write buf list */
649 static list_t *l2arc_free_on_write; /* free after write list ptr */
650 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
651 static uint64_t l2arc_ndev; /* number of devices */
652
653 typedef struct l2arc_read_callback {
654 arc_buf_t *l2rcb_buf; /* read buffer */
655 spa_t *l2rcb_spa; /* spa */
656 blkptr_t l2rcb_bp; /* original blkptr */
657 zbookmark_t l2rcb_zb; /* original bookmark */
658 int l2rcb_flags; /* original flags */
659 enum zio_compress l2rcb_compress; /* applied compress */
660 } l2arc_read_callback_t;
661
662 typedef struct l2arc_write_callback {
663 l2arc_dev_t *l2wcb_dev; /* device info */
664 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
665 } l2arc_write_callback_t;
666
667 struct l2arc_buf_hdr {
668 /* protected by arc_buf_hdr mutex */
669 l2arc_dev_t *b_dev; /* L2ARC device */
670 uint64_t b_daddr; /* disk address, offset byte */
671 /* compression applied to buffer data */
672 enum zio_compress b_compress;
673 /* real alloc'd buffer size depending on b_compress applied */
674 int b_asize;
675 /* temporary buffer holder for in-flight compressed data */
676 void *b_tmp_cdata;
677 };
678
679 typedef struct l2arc_data_free {
680 /* protected by l2arc_free_on_write_mtx */
681 void *l2df_data;
682 size_t l2df_size;
683 void (*l2df_func)(void *, size_t);
684 list_node_t l2df_list_node;
685 } l2arc_data_free_t;
686
687 static kmutex_t l2arc_feed_thr_lock;
688 static kcondvar_t l2arc_feed_thr_cv;
689 static uint8_t l2arc_thread_exit;
690
691 static void l2arc_read_done(zio_t *zio);
692 static void l2arc_hdr_stat_add(void);
693 static void l2arc_hdr_stat_remove(void);
694
695 static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr);
696 static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr,
697 enum zio_compress c);
698 static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab);
699
700 static uint64_t
701 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
702 {
703 uint8_t *vdva = (uint8_t *)dva;
704 uint64_t crc = -1ULL;
705 int i;
706
707 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
708
709 for (i = 0; i < sizeof (dva_t); i++)
710 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
711
712 crc ^= (spa>>8) ^ birth;
713
714 return (crc);
715 }
716
717 #define BUF_EMPTY(buf) \
718 ((buf)->b_dva.dva_word[0] == 0 && \
719 (buf)->b_dva.dva_word[1] == 0 && \
720 (buf)->b_birth == 0)
721
722 #define BUF_EQUAL(spa, dva, birth, buf) \
723 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
724 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
725 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
726
727 static void
728 buf_discard_identity(arc_buf_hdr_t *hdr)
729 {
730 hdr->b_dva.dva_word[0] = 0;
731 hdr->b_dva.dva_word[1] = 0;
732 hdr->b_birth = 0;
733 hdr->b_cksum0 = 0;
734 }
735
736 static arc_buf_hdr_t *
737 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
738 {
739 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
740 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
741 arc_buf_hdr_t *buf;
742
743 mutex_enter(hash_lock);
744 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
745 buf = buf->b_hash_next) {
746 if (BUF_EQUAL(spa, dva, birth, buf)) {
747 *lockp = hash_lock;
748 return (buf);
749 }
750 }
751 mutex_exit(hash_lock);
752 *lockp = NULL;
753 return (NULL);
754 }
755
756 /*
757 * Insert an entry into the hash table. If there is already an element
758 * equal to elem in the hash table, then the already existing element
759 * will be returned and the new element will not be inserted.
760 * Otherwise returns NULL.
761 */
762 static arc_buf_hdr_t *
763 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
764 {
765 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
766 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
767 arc_buf_hdr_t *fbuf;
768 uint32_t i;
769
770 ASSERT(!HDR_IN_HASH_TABLE(buf));
771 *lockp = hash_lock;
772 mutex_enter(hash_lock);
773 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
774 fbuf = fbuf->b_hash_next, i++) {
775 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
776 return (fbuf);
777 }
778
779 buf->b_hash_next = buf_hash_table.ht_table[idx];
780 buf_hash_table.ht_table[idx] = buf;
781 buf->b_flags |= ARC_IN_HASH_TABLE;
782
783 /* collect some hash table performance data */
784 if (i > 0) {
785 ARCSTAT_BUMP(arcstat_hash_collisions);
786 if (i == 1)
787 ARCSTAT_BUMP(arcstat_hash_chains);
788
789 ARCSTAT_MAX(arcstat_hash_chain_max, i);
790 }
791
792 ARCSTAT_BUMP(arcstat_hash_elements);
793 ARCSTAT_MAXSTAT(arcstat_hash_elements);
794
795 return (NULL);
796 }
797
798 static void
799 buf_hash_remove(arc_buf_hdr_t *buf)
800 {
801 arc_buf_hdr_t *fbuf, **bufp;
802 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
803
804 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
805 ASSERT(HDR_IN_HASH_TABLE(buf));
806
807 bufp = &buf_hash_table.ht_table[idx];
808 while ((fbuf = *bufp) != buf) {
809 ASSERT(fbuf != NULL);
810 bufp = &fbuf->b_hash_next;
811 }
812 *bufp = buf->b_hash_next;
813 buf->b_hash_next = NULL;
814 buf->b_flags &= ~ARC_IN_HASH_TABLE;
815
816 /* collect some hash table performance data */
817 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
818
819 if (buf_hash_table.ht_table[idx] &&
820 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
821 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
822 }
823
824 /*
825 * Global data structures and functions for the buf kmem cache.
826 */
827 static kmem_cache_t *hdr_cache;
828 static kmem_cache_t *buf_cache;
829
830 static void
831 buf_fini(void)
832 {
833 int i;
834
835 kmem_free(buf_hash_table.ht_table,
836 (buf_hash_table.ht_mask + 1) * sizeof (void *));
837 for (i = 0; i < BUF_LOCKS; i++)
838 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
839 kmem_cache_destroy(hdr_cache);
840 kmem_cache_destroy(buf_cache);
841 }
842
843 /*
844 * Constructor callback - called when the cache is empty
845 * and a new buf is requested.
846 */
847 /* ARGSUSED */
848 static int
849 hdr_cons(void *vbuf, void *unused, int kmflag)
850 {
851 arc_buf_hdr_t *buf = vbuf;
852
853 bzero(buf, sizeof (arc_buf_hdr_t));
854 refcount_create(&buf->b_refcnt);
855 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
856 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
857 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
858
859 return (0);
860 }
861
862 /* ARGSUSED */
863 static int
864 buf_cons(void *vbuf, void *unused, int kmflag)
865 {
866 arc_buf_t *buf = vbuf;
867
868 bzero(buf, sizeof (arc_buf_t));
869 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
870 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
871
872 return (0);
873 }
874
875 /*
876 * Destructor callback - called when a cached buf is
877 * no longer required.
878 */
879 /* ARGSUSED */
880 static void
881 hdr_dest(void *vbuf, void *unused)
882 {
883 arc_buf_hdr_t *buf = vbuf;
884
885 ASSERT(BUF_EMPTY(buf));
886 refcount_destroy(&buf->b_refcnt);
887 cv_destroy(&buf->b_cv);
888 mutex_destroy(&buf->b_freeze_lock);
889 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
890 }
891
892 /* ARGSUSED */
893 static void
894 buf_dest(void *vbuf, void *unused)
895 {
896 arc_buf_t *buf = vbuf;
897
898 mutex_destroy(&buf->b_evict_lock);
899 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
900 }
901
902 /*
903 * Reclaim callback -- invoked when memory is low.
904 */
905 /* ARGSUSED */
906 static void
907 hdr_recl(void *unused)
908 {
909 dprintf("hdr_recl called\n");
910 /*
911 * umem calls the reclaim func when we destroy the buf cache,
912 * which is after we do arc_fini().
913 */
914 if (!arc_dead)
915 cv_signal(&arc_reclaim_thr_cv);
916 }
917
918 static void
919 buf_init(void)
920 {
921 uint64_t *ct;
922 uint64_t hsize = 1ULL << 12;
923 int i, j;
924
925 /*
926 * The hash table is big enough to fill all of physical memory
927 * with an average 64K block size. The table will take up
928 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
929 */
930 while (hsize * 65536 < physmem * PAGESIZE)
931 hsize <<= 1;
932 retry:
933 buf_hash_table.ht_mask = hsize - 1;
934 buf_hash_table.ht_table =
935 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
936 if (buf_hash_table.ht_table == NULL) {
937 ASSERT(hsize > (1ULL << 8));
938 hsize >>= 1;
939 goto retry;
940 }
941
942 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
943 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
944 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
945 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
946
947 for (i = 0; i < 256; i++)
948 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
949 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
950
951 for (i = 0; i < BUF_LOCKS; i++) {
952 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
953 NULL, MUTEX_DEFAULT, NULL);
954 }
955 }
956
957 #define ARC_MINTIME (hz>>4) /* 62 ms */
958
959 static void
960 arc_cksum_verify(arc_buf_t *buf)
961 {
962 zio_cksum_t zc;
963
964 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
965 return;
966
967 mutex_enter(&buf->b_hdr->b_freeze_lock);
968 if (buf->b_hdr->b_freeze_cksum == NULL ||
969 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
970 mutex_exit(&buf->b_hdr->b_freeze_lock);
971 return;
972 }
973 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
974 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
975 panic("buffer modified while frozen!");
976 mutex_exit(&buf->b_hdr->b_freeze_lock);
977 }
978
979 static int
980 arc_cksum_equal(arc_buf_t *buf)
981 {
982 zio_cksum_t zc;
983 int equal;
984
985 mutex_enter(&buf->b_hdr->b_freeze_lock);
986 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
987 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
988 mutex_exit(&buf->b_hdr->b_freeze_lock);
989
990 return (equal);
991 }
992
993 static void
994 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
995 {
996 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
997 return;
998
999 mutex_enter(&buf->b_hdr->b_freeze_lock);
1000 if (buf->b_hdr->b_freeze_cksum != NULL) {
1001 mutex_exit(&buf->b_hdr->b_freeze_lock);
1002 return;
1003 }
1004 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1005 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1006 buf->b_hdr->b_freeze_cksum);
1007 mutex_exit(&buf->b_hdr->b_freeze_lock);
1008 arc_buf_watch(buf);
1009 }
1010
1011 #ifndef _KERNEL
1012 typedef struct procctl {
1013 long cmd;
1014 prwatch_t prwatch;
1015 } procctl_t;
1016 #endif
1017
1018 /* ARGSUSED */
1019 static void
1020 arc_buf_unwatch(arc_buf_t *buf)
1021 {
1022 #ifndef _KERNEL
1023 if (arc_watch) {
1024 int result;
1025 procctl_t ctl;
1026 ctl.cmd = PCWATCH;
1027 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1028 ctl.prwatch.pr_size = 0;
1029 ctl.prwatch.pr_wflags = 0;
1030 result = write(arc_procfd, &ctl, sizeof (ctl));
1031 ASSERT3U(result, ==, sizeof (ctl));
1032 }
1033 #endif
1034 }
1035
1036 /* ARGSUSED */
1037 static void
1038 arc_buf_watch(arc_buf_t *buf)
1039 {
1040 #ifndef _KERNEL
1041 if (arc_watch) {
1042 int result;
1043 procctl_t ctl;
1044 ctl.cmd = PCWATCH;
1045 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1046 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1047 ctl.prwatch.pr_wflags = WA_WRITE;
1048 result = write(arc_procfd, &ctl, sizeof (ctl));
1049 ASSERT3U(result, ==, sizeof (ctl));
1050 }
1051 #endif
1052 }
1053
1054 void
1055 arc_buf_thaw(arc_buf_t *buf)
1056 {
1057 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1058 if (buf->b_hdr->b_state != arc_anon)
1059 panic("modifying non-anon buffer!");
1060 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1061 panic("modifying buffer while i/o in progress!");
1062 arc_cksum_verify(buf);
1063 }
1064
1065 mutex_enter(&buf->b_hdr->b_freeze_lock);
1066 if (buf->b_hdr->b_freeze_cksum != NULL) {
1067 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1068 buf->b_hdr->b_freeze_cksum = NULL;
1069 }
1070
1071 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1072 if (buf->b_hdr->b_thawed)
1073 kmem_free(buf->b_hdr->b_thawed, 1);
1074 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
1075 }
1076
1077 mutex_exit(&buf->b_hdr->b_freeze_lock);
1078
1079 arc_buf_unwatch(buf);
1080 }
1081
1082 void
1083 arc_buf_freeze(arc_buf_t *buf)
1084 {
1085 kmutex_t *hash_lock;
1086
1087 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1088 return;
1089
1090 hash_lock = HDR_LOCK(buf->b_hdr);
1091 mutex_enter(hash_lock);
1092
1093 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1094 buf->b_hdr->b_state == arc_anon);
1095 arc_cksum_compute(buf, B_FALSE);
1096 mutex_exit(hash_lock);
1097
1098 }
1099
1100 static void
1101 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1102 {
1103 ASSERT(MUTEX_HELD(hash_lock));
1104
1105 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1106 (ab->b_state != arc_anon)) {
1107 uint64_t delta = ab->b_size * ab->b_datacnt;
1108 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1109 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1110
1111 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1112 mutex_enter(&ab->b_state->arcs_mtx);
1113 ASSERT(list_link_active(&ab->b_arc_node));
1114 list_remove(list, ab);
1115 if (GHOST_STATE(ab->b_state)) {
1116 ASSERT0(ab->b_datacnt);
1117 ASSERT3P(ab->b_buf, ==, NULL);
1118 delta = ab->b_size;
1119 }
1120 ASSERT(delta > 0);
1121 ASSERT3U(*size, >=, delta);
1122 atomic_add_64(size, -delta);
1123 mutex_exit(&ab->b_state->arcs_mtx);
1124 /* remove the prefetch flag if we get a reference */
1125 if (ab->b_flags & ARC_PREFETCH)
1126 ab->b_flags &= ~ARC_PREFETCH;
1127 }
1128 }
1129
1130 static int
1131 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1132 {
1133 int cnt;
1134 arc_state_t *state = ab->b_state;
1135
1136 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1137 ASSERT(!GHOST_STATE(state));
1138
1139 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1140 (state != arc_anon)) {
1141 uint64_t *size = &state->arcs_lsize[ab->b_type];
1142
1143 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1144 mutex_enter(&state->arcs_mtx);
1145 ASSERT(!list_link_active(&ab->b_arc_node));
1146 list_insert_head(&state->arcs_list[ab->b_type], ab);
1147 ASSERT(ab->b_datacnt > 0);
1148 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1149 mutex_exit(&state->arcs_mtx);
1150 }
1151 return (cnt);
1152 }
1153
1154 /*
1155 * Move the supplied buffer to the indicated state. The mutex
1156 * for the buffer must be held by the caller.
1157 */
1158 static void
1159 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1160 {
1161 arc_state_t *old_state = ab->b_state;
1162 int64_t refcnt = refcount_count(&ab->b_refcnt);
1163 uint64_t from_delta, to_delta;
1164
1165 ASSERT(MUTEX_HELD(hash_lock));
1166 ASSERT(new_state != old_state);
1167 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1168 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1169 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1170
1171 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1172
1173 /*
1174 * If this buffer is evictable, transfer it from the
1175 * old state list to the new state list.
1176 */
1177 if (refcnt == 0) {
1178 if (old_state != arc_anon) {
1179 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1180 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1181
1182 if (use_mutex)
1183 mutex_enter(&old_state->arcs_mtx);
1184
1185 ASSERT(list_link_active(&ab->b_arc_node));
1186 list_remove(&old_state->arcs_list[ab->b_type], ab);
1187
1188 /*
1189 * If prefetching out of the ghost cache,
1190 * we will have a non-zero datacnt.
1191 */
1192 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1193 /* ghost elements have a ghost size */
1194 ASSERT(ab->b_buf == NULL);
1195 from_delta = ab->b_size;
1196 }
1197 ASSERT3U(*size, >=, from_delta);
1198 atomic_add_64(size, -from_delta);
1199
1200 if (use_mutex)
1201 mutex_exit(&old_state->arcs_mtx);
1202 }
1203 if (new_state != arc_anon) {
1204 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1205 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1206
1207 if (use_mutex)
1208 mutex_enter(&new_state->arcs_mtx);
1209
1210 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1211
1212 /* ghost elements have a ghost size */
1213 if (GHOST_STATE(new_state)) {
1214 ASSERT(ab->b_datacnt == 0);
1215 ASSERT(ab->b_buf == NULL);
1216 to_delta = ab->b_size;
1217 }
1218 atomic_add_64(size, to_delta);
1219
1220 if (use_mutex)
1221 mutex_exit(&new_state->arcs_mtx);
1222 }
1223 }
1224
1225 ASSERT(!BUF_EMPTY(ab));
1226 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1227 buf_hash_remove(ab);
1228
1229 /* adjust state sizes */
1230 if (to_delta)
1231 atomic_add_64(&new_state->arcs_size, to_delta);
1232 if (from_delta) {
1233 ASSERT3U(old_state->arcs_size, >=, from_delta);
1234 atomic_add_64(&old_state->arcs_size, -from_delta);
1235 }
1236 ab->b_state = new_state;
1237
1238 /* adjust l2arc hdr stats */
1239 if (new_state == arc_l2c_only)
1240 l2arc_hdr_stat_add();
1241 else if (old_state == arc_l2c_only)
1242 l2arc_hdr_stat_remove();
1243 }
1244
1245 void
1246 arc_space_consume(uint64_t space, arc_space_type_t type)
1247 {
1248 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1249
1250 switch (type) {
1251 case ARC_SPACE_DATA:
1252 ARCSTAT_INCR(arcstat_data_size, space);
1253 break;
1254 case ARC_SPACE_OTHER:
1255 ARCSTAT_INCR(arcstat_other_size, space);
1256 break;
1257 case ARC_SPACE_HDRS:
1258 ARCSTAT_INCR(arcstat_hdr_size, space);
1259 break;
1260 case ARC_SPACE_L2HDRS:
1261 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1262 break;
1263 }
1264
1265 ARCSTAT_INCR(arcstat_meta_used, space);
1266 atomic_add_64(&arc_size, space);
1267 }
1268
1269 void
1270 arc_space_return(uint64_t space, arc_space_type_t type)
1271 {
1272 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1273
1274 switch (type) {
1275 case ARC_SPACE_DATA:
1276 ARCSTAT_INCR(arcstat_data_size, -space);
1277 break;
1278 case ARC_SPACE_OTHER:
1279 ARCSTAT_INCR(arcstat_other_size, -space);
1280 break;
1281 case ARC_SPACE_HDRS:
1282 ARCSTAT_INCR(arcstat_hdr_size, -space);
1283 break;
1284 case ARC_SPACE_L2HDRS:
1285 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1286 break;
1287 }
1288
1289 ASSERT(arc_meta_used >= space);
1290 if (arc_meta_max < arc_meta_used)
1291 arc_meta_max = arc_meta_used;
1292 ARCSTAT_INCR(arcstat_meta_used, -space);
1293 ASSERT(arc_size >= space);
1294 atomic_add_64(&arc_size, -space);
1295 }
1296
1297 void *
1298 arc_data_buf_alloc(uint64_t size)
1299 {
1300 if (arc_evict_needed(ARC_BUFC_DATA))
1301 cv_signal(&arc_reclaim_thr_cv);
1302 atomic_add_64(&arc_size, size);
1303 return (zio_data_buf_alloc(size));
1304 }
1305
1306 void
1307 arc_data_buf_free(void *buf, uint64_t size)
1308 {
1309 zio_data_buf_free(buf, size);
1310 ASSERT(arc_size >= size);
1311 atomic_add_64(&arc_size, -size);
1312 }
1313
1314 arc_buf_t *
1315 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1316 {
1317 arc_buf_hdr_t *hdr;
1318 arc_buf_t *buf;
1319
1320 ASSERT3U(size, >, 0);
1321 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1322 ASSERT(BUF_EMPTY(hdr));
1323 hdr->b_size = size;
1324 hdr->b_type = type;
1325 hdr->b_spa = spa_load_guid(spa);
1326 hdr->b_state = arc_anon;
1327 hdr->b_arc_access = 0;
1328 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1329 buf->b_hdr = hdr;
1330 buf->b_data = NULL;
1331 buf->b_efunc = NULL;
1332 buf->b_private = NULL;
1333 buf->b_next = NULL;
1334 hdr->b_buf = buf;
1335 arc_get_data_buf(buf);
1336 hdr->b_datacnt = 1;
1337 hdr->b_flags = 0;
1338 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1339 (void) refcount_add(&hdr->b_refcnt, tag);
1340
1341 return (buf);
1342 }
1343
1344 static char *arc_onloan_tag = "onloan";
1345
1346 /*
1347 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1348 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1349 * buffers must be returned to the arc before they can be used by the DMU or
1350 * freed.
1351 */
1352 arc_buf_t *
1353 arc_loan_buf(spa_t *spa, int size)
1354 {
1355 arc_buf_t *buf;
1356
1357 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1358
1359 atomic_add_64(&arc_loaned_bytes, size);
1360 return (buf);
1361 }
1362
1363 /*
1364 * Return a loaned arc buffer to the arc.
1365 */
1366 void
1367 arc_return_buf(arc_buf_t *buf, void *tag)
1368 {
1369 arc_buf_hdr_t *hdr = buf->b_hdr;
1370
1371 ASSERT(buf->b_data != NULL);
1372 (void) refcount_add(&hdr->b_refcnt, tag);
1373 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1374
1375 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1376 }
1377
1378 /* Detach an arc_buf from a dbuf (tag) */
1379 void
1380 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1381 {
1382 arc_buf_hdr_t *hdr;
1383
1384 ASSERT(buf->b_data != NULL);
1385 hdr = buf->b_hdr;
1386 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1387 (void) refcount_remove(&hdr->b_refcnt, tag);
1388 buf->b_efunc = NULL;
1389 buf->b_private = NULL;
1390
1391 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1392 }
1393
1394 static arc_buf_t *
1395 arc_buf_clone(arc_buf_t *from)
1396 {
1397 arc_buf_t *buf;
1398 arc_buf_hdr_t *hdr = from->b_hdr;
1399 uint64_t size = hdr->b_size;
1400
1401 ASSERT(hdr->b_state != arc_anon);
1402
1403 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1404 buf->b_hdr = hdr;
1405 buf->b_data = NULL;
1406 buf->b_efunc = NULL;
1407 buf->b_private = NULL;
1408 buf->b_next = hdr->b_buf;
1409 hdr->b_buf = buf;
1410 arc_get_data_buf(buf);
1411 bcopy(from->b_data, buf->b_data, size);
1412
1413 /*
1414 * This buffer already exists in the arc so create a duplicate
1415 * copy for the caller. If the buffer is associated with user data
1416 * then track the size and number of duplicates. These stats will be
1417 * updated as duplicate buffers are created and destroyed.
1418 */
1419 if (hdr->b_type == ARC_BUFC_DATA) {
1420 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1421 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1422 }
1423 hdr->b_datacnt += 1;
1424 return (buf);
1425 }
1426
1427 void
1428 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1429 {
1430 arc_buf_hdr_t *hdr;
1431 kmutex_t *hash_lock;
1432
1433 /*
1434 * Check to see if this buffer is evicted. Callers
1435 * must verify b_data != NULL to know if the add_ref
1436 * was successful.
1437 */
1438 mutex_enter(&buf->b_evict_lock);
1439 if (buf->b_data == NULL) {
1440 mutex_exit(&buf->b_evict_lock);
1441 return;
1442 }
1443 hash_lock = HDR_LOCK(buf->b_hdr);
1444 mutex_enter(hash_lock);
1445 hdr = buf->b_hdr;
1446 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1447 mutex_exit(&buf->b_evict_lock);
1448
1449 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1450 add_reference(hdr, hash_lock, tag);
1451 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1452 arc_access(hdr, hash_lock);
1453 mutex_exit(hash_lock);
1454 ARCSTAT_BUMP(arcstat_hits);
1455 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1456 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1457 data, metadata, hits);
1458 }
1459
1460 /*
1461 * Free the arc data buffer. If it is an l2arc write in progress,
1462 * the buffer is placed on l2arc_free_on_write to be freed later.
1463 */
1464 static void
1465 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1466 {
1467 arc_buf_hdr_t *hdr = buf->b_hdr;
1468
1469 if (HDR_L2_WRITING(hdr)) {
1470 l2arc_data_free_t *df;
1471 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1472 df->l2df_data = buf->b_data;
1473 df->l2df_size = hdr->b_size;
1474 df->l2df_func = free_func;
1475 mutex_enter(&l2arc_free_on_write_mtx);
1476 list_insert_head(l2arc_free_on_write, df);
1477 mutex_exit(&l2arc_free_on_write_mtx);
1478 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1479 } else {
1480 free_func(buf->b_data, hdr->b_size);
1481 }
1482 }
1483
1484 static void
1485 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1486 {
1487 arc_buf_t **bufp;
1488
1489 /* free up data associated with the buf */
1490 if (buf->b_data) {
1491 arc_state_t *state = buf->b_hdr->b_state;
1492 uint64_t size = buf->b_hdr->b_size;
1493 arc_buf_contents_t type = buf->b_hdr->b_type;
1494
1495 arc_cksum_verify(buf);
1496 arc_buf_unwatch(buf);
1497
1498 if (!recycle) {
1499 if (type == ARC_BUFC_METADATA) {
1500 arc_buf_data_free(buf, zio_buf_free);
1501 arc_space_return(size, ARC_SPACE_DATA);
1502 } else {
1503 ASSERT(type == ARC_BUFC_DATA);
1504 arc_buf_data_free(buf, zio_data_buf_free);
1505 ARCSTAT_INCR(arcstat_data_size, -size);
1506 atomic_add_64(&arc_size, -size);
1507 }
1508 }
1509 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1510 uint64_t *cnt = &state->arcs_lsize[type];
1511
1512 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1513 ASSERT(state != arc_anon);
1514
1515 ASSERT3U(*cnt, >=, size);
1516 atomic_add_64(cnt, -size);
1517 }
1518 ASSERT3U(state->arcs_size, >=, size);
1519 atomic_add_64(&state->arcs_size, -size);
1520 buf->b_data = NULL;
1521
1522 /*
1523 * If we're destroying a duplicate buffer make sure
1524 * that the appropriate statistics are updated.
1525 */
1526 if (buf->b_hdr->b_datacnt > 1 &&
1527 buf->b_hdr->b_type == ARC_BUFC_DATA) {
1528 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
1529 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
1530 }
1531 ASSERT(buf->b_hdr->b_datacnt > 0);
1532 buf->b_hdr->b_datacnt -= 1;
1533 }
1534
1535 /* only remove the buf if requested */
1536 if (!all)
1537 return;
1538
1539 /* remove the buf from the hdr list */
1540 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1541 continue;
1542 *bufp = buf->b_next;
1543 buf->b_next = NULL;
1544
1545 ASSERT(buf->b_efunc == NULL);
1546
1547 /* clean up the buf */
1548 buf->b_hdr = NULL;
1549 kmem_cache_free(buf_cache, buf);
1550 }
1551
1552 static void
1553 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1554 {
1555 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1556 ASSERT3P(hdr->b_state, ==, arc_anon);
1557 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1558 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1559
1560 if (l2hdr != NULL) {
1561 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1562 /*
1563 * To prevent arc_free() and l2arc_evict() from
1564 * attempting to free the same buffer at the same time,
1565 * a FREE_IN_PROGRESS flag is given to arc_free() to
1566 * give it priority. l2arc_evict() can't destroy this
1567 * header while we are waiting on l2arc_buflist_mtx.
1568 *
1569 * The hdr may be removed from l2ad_buflist before we
1570 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1571 */
1572 if (!buflist_held) {
1573 mutex_enter(&l2arc_buflist_mtx);
1574 l2hdr = hdr->b_l2hdr;
1575 }
1576
1577 if (l2hdr != NULL) {
1578 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1579 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1580 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
1581 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1582 if (hdr->b_state == arc_l2c_only)
1583 l2arc_hdr_stat_remove();
1584 hdr->b_l2hdr = NULL;
1585 }
1586
1587 if (!buflist_held)
1588 mutex_exit(&l2arc_buflist_mtx);
1589 }
1590
1591 if (!BUF_EMPTY(hdr)) {
1592 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1593 buf_discard_identity(hdr);
1594 }
1595 while (hdr->b_buf) {
1596 arc_buf_t *buf = hdr->b_buf;
1597
1598 if (buf->b_efunc) {
1599 mutex_enter(&arc_eviction_mtx);
1600 mutex_enter(&buf->b_evict_lock);
1601 ASSERT(buf->b_hdr != NULL);
1602 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1603 hdr->b_buf = buf->b_next;
1604 buf->b_hdr = &arc_eviction_hdr;
1605 buf->b_next = arc_eviction_list;
1606 arc_eviction_list = buf;
1607 mutex_exit(&buf->b_evict_lock);
1608 mutex_exit(&arc_eviction_mtx);
1609 } else {
1610 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1611 }
1612 }
1613 if (hdr->b_freeze_cksum != NULL) {
1614 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1615 hdr->b_freeze_cksum = NULL;
1616 }
1617 if (hdr->b_thawed) {
1618 kmem_free(hdr->b_thawed, 1);
1619 hdr->b_thawed = NULL;
1620 }
1621
1622 ASSERT(!list_link_active(&hdr->b_arc_node));
1623 ASSERT3P(hdr->b_hash_next, ==, NULL);
1624 ASSERT3P(hdr->b_acb, ==, NULL);
1625 kmem_cache_free(hdr_cache, hdr);
1626 }
1627
1628 void
1629 arc_buf_free(arc_buf_t *buf, void *tag)
1630 {
1631 arc_buf_hdr_t *hdr = buf->b_hdr;
1632 int hashed = hdr->b_state != arc_anon;
1633
1634 ASSERT(buf->b_efunc == NULL);
1635 ASSERT(buf->b_data != NULL);
1636
1637 if (hashed) {
1638 kmutex_t *hash_lock = HDR_LOCK(hdr);
1639
1640 mutex_enter(hash_lock);
1641 hdr = buf->b_hdr;
1642 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1643
1644 (void) remove_reference(hdr, hash_lock, tag);
1645 if (hdr->b_datacnt > 1) {
1646 arc_buf_destroy(buf, FALSE, TRUE);
1647 } else {
1648 ASSERT(buf == hdr->b_buf);
1649 ASSERT(buf->b_efunc == NULL);
1650 hdr->b_flags |= ARC_BUF_AVAILABLE;
1651 }
1652 mutex_exit(hash_lock);
1653 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1654 int destroy_hdr;
1655 /*
1656 * We are in the middle of an async write. Don't destroy
1657 * this buffer unless the write completes before we finish
1658 * decrementing the reference count.
1659 */
1660 mutex_enter(&arc_eviction_mtx);
1661 (void) remove_reference(hdr, NULL, tag);
1662 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1663 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1664 mutex_exit(&arc_eviction_mtx);
1665 if (destroy_hdr)
1666 arc_hdr_destroy(hdr);
1667 } else {
1668 if (remove_reference(hdr, NULL, tag) > 0)
1669 arc_buf_destroy(buf, FALSE, TRUE);
1670 else
1671 arc_hdr_destroy(hdr);
1672 }
1673 }
1674
1675 boolean_t
1676 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1677 {
1678 arc_buf_hdr_t *hdr = buf->b_hdr;
1679 kmutex_t *hash_lock = HDR_LOCK(hdr);
1680 boolean_t no_callback = (buf->b_efunc == NULL);
1681
1682 if (hdr->b_state == arc_anon) {
1683 ASSERT(hdr->b_datacnt == 1);
1684 arc_buf_free(buf, tag);
1685 return (no_callback);
1686 }
1687
1688 mutex_enter(hash_lock);
1689 hdr = buf->b_hdr;
1690 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1691 ASSERT(hdr->b_state != arc_anon);
1692 ASSERT(buf->b_data != NULL);
1693
1694 (void) remove_reference(hdr, hash_lock, tag);
1695 if (hdr->b_datacnt > 1) {
1696 if (no_callback)
1697 arc_buf_destroy(buf, FALSE, TRUE);
1698 } else if (no_callback) {
1699 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1700 ASSERT(buf->b_efunc == NULL);
1701 hdr->b_flags |= ARC_BUF_AVAILABLE;
1702 }
1703 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1704 refcount_is_zero(&hdr->b_refcnt));
1705 mutex_exit(hash_lock);
1706 return (no_callback);
1707 }
1708
1709 int
1710 arc_buf_size(arc_buf_t *buf)
1711 {
1712 return (buf->b_hdr->b_size);
1713 }
1714
1715 /*
1716 * Called from the DMU to determine if the current buffer should be
1717 * evicted. In order to ensure proper locking, the eviction must be initiated
1718 * from the DMU. Return true if the buffer is associated with user data and
1719 * duplicate buffers still exist.
1720 */
1721 boolean_t
1722 arc_buf_eviction_needed(arc_buf_t *buf)
1723 {
1724 arc_buf_hdr_t *hdr;
1725 boolean_t evict_needed = B_FALSE;
1726
1727 if (zfs_disable_dup_eviction)
1728 return (B_FALSE);
1729
1730 mutex_enter(&buf->b_evict_lock);
1731 hdr = buf->b_hdr;
1732 if (hdr == NULL) {
1733 /*
1734 * We are in arc_do_user_evicts(); let that function
1735 * perform the eviction.
1736 */
1737 ASSERT(buf->b_data == NULL);
1738 mutex_exit(&buf->b_evict_lock);
1739 return (B_FALSE);
1740 } else if (buf->b_data == NULL) {
1741 /*
1742 * We have already been added to the arc eviction list;
1743 * recommend eviction.
1744 */
1745 ASSERT3P(hdr, ==, &arc_eviction_hdr);
1746 mutex_exit(&buf->b_evict_lock);
1747 return (B_TRUE);
1748 }
1749
1750 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA)
1751 evict_needed = B_TRUE;
1752
1753 mutex_exit(&buf->b_evict_lock);
1754 return (evict_needed);
1755 }
1756
1757 /*
1758 * Evict buffers from list until we've removed the specified number of
1759 * bytes. Move the removed buffers to the appropriate evict state.
1760 * If the recycle flag is set, then attempt to "recycle" a buffer:
1761 * - look for a buffer to evict that is `bytes' long.
1762 * - return the data block from this buffer rather than freeing it.
1763 * This flag is used by callers that are trying to make space for a
1764 * new buffer in a full arc cache.
1765 *
1766 * This function makes a "best effort". It skips over any buffers
1767 * it can't get a hash_lock on, and so may not catch all candidates.
1768 * It may also return without evicting as much space as requested.
1769 */
1770 static void *
1771 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1772 arc_buf_contents_t type)
1773 {
1774 arc_state_t *evicted_state;
1775 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1776 arc_buf_hdr_t *ab, *ab_prev = NULL;
1777 list_t *list = &state->arcs_list[type];
1778 kmutex_t *hash_lock;
1779 boolean_t have_lock;
1780 void *stolen = NULL;
1781
1782 ASSERT(state == arc_mru || state == arc_mfu);
1783
1784 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1785
1786 mutex_enter(&state->arcs_mtx);
1787 mutex_enter(&evicted_state->arcs_mtx);
1788
1789 for (ab = list_tail(list); ab; ab = ab_prev) {
1790 ab_prev = list_prev(list, ab);
1791 /* prefetch buffers have a minimum lifespan */
1792 if (HDR_IO_IN_PROGRESS(ab) ||
1793 (spa && ab->b_spa != spa) ||
1794 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1795 ddi_get_lbolt() - ab->b_arc_access <
1796 arc_min_prefetch_lifespan)) {
1797 skipped++;
1798 continue;
1799 }
1800 /* "lookahead" for better eviction candidate */
1801 if (recycle && ab->b_size != bytes &&
1802 ab_prev && ab_prev->b_size == bytes)
1803 continue;
1804 hash_lock = HDR_LOCK(ab);
1805 have_lock = MUTEX_HELD(hash_lock);
1806 if (have_lock || mutex_tryenter(hash_lock)) {
1807 ASSERT0(refcount_count(&ab->b_refcnt));
1808 ASSERT(ab->b_datacnt > 0);
1809 while (ab->b_buf) {
1810 arc_buf_t *buf = ab->b_buf;
1811 if (!mutex_tryenter(&buf->b_evict_lock)) {
1812 missed += 1;
1813 break;
1814 }
1815 if (buf->b_data) {
1816 bytes_evicted += ab->b_size;
1817 if (recycle && ab->b_type == type &&
1818 ab->b_size == bytes &&
1819 !HDR_L2_WRITING(ab)) {
1820 stolen = buf->b_data;
1821 recycle = FALSE;
1822 }
1823 }
1824 if (buf->b_efunc) {
1825 mutex_enter(&arc_eviction_mtx);
1826 arc_buf_destroy(buf,
1827 buf->b_data == stolen, FALSE);
1828 ab->b_buf = buf->b_next;
1829 buf->b_hdr = &arc_eviction_hdr;
1830 buf->b_next = arc_eviction_list;
1831 arc_eviction_list = buf;
1832 mutex_exit(&arc_eviction_mtx);
1833 mutex_exit(&buf->b_evict_lock);
1834 } else {
1835 mutex_exit(&buf->b_evict_lock);
1836 arc_buf_destroy(buf,
1837 buf->b_data == stolen, TRUE);
1838 }
1839 }
1840
1841 if (ab->b_l2hdr) {
1842 ARCSTAT_INCR(arcstat_evict_l2_cached,
1843 ab->b_size);
1844 } else {
1845 if (l2arc_write_eligible(ab->b_spa, ab)) {
1846 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1847 ab->b_size);
1848 } else {
1849 ARCSTAT_INCR(
1850 arcstat_evict_l2_ineligible,
1851 ab->b_size);
1852 }
1853 }
1854
1855 if (ab->b_datacnt == 0) {
1856 arc_change_state(evicted_state, ab, hash_lock);
1857 ASSERT(HDR_IN_HASH_TABLE(ab));
1858 ab->b_flags |= ARC_IN_HASH_TABLE;
1859 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1860 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1861 }
1862 if (!have_lock)
1863 mutex_exit(hash_lock);
1864 if (bytes >= 0 && bytes_evicted >= bytes)
1865 break;
1866 } else {
1867 missed += 1;
1868 }
1869 }
1870
1871 mutex_exit(&evicted_state->arcs_mtx);
1872 mutex_exit(&state->arcs_mtx);
1873
1874 if (bytes_evicted < bytes)
1875 dprintf("only evicted %lld bytes from %x",
1876 (longlong_t)bytes_evicted, state);
1877
1878 if (skipped)
1879 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1880
1881 if (missed)
1882 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1883
1884 /*
1885 * We have just evicted some data into the ghost state, make
1886 * sure we also adjust the ghost state size if necessary.
1887 */
1888 if (arc_no_grow &&
1889 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1890 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1891 arc_mru_ghost->arcs_size - arc_c;
1892
1893 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1894 int64_t todelete =
1895 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1896 arc_evict_ghost(arc_mru_ghost, NULL, todelete);
1897 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1898 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1899 arc_mru_ghost->arcs_size +
1900 arc_mfu_ghost->arcs_size - arc_c);
1901 arc_evict_ghost(arc_mfu_ghost, NULL, todelete);
1902 }
1903 }
1904
1905 return (stolen);
1906 }
1907
1908 /*
1909 * Remove buffers from list until we've removed the specified number of
1910 * bytes. Destroy the buffers that are removed.
1911 */
1912 static void
1913 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1914 {
1915 arc_buf_hdr_t *ab, *ab_prev;
1916 arc_buf_hdr_t marker = { 0 };
1917 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1918 kmutex_t *hash_lock;
1919 uint64_t bytes_deleted = 0;
1920 uint64_t bufs_skipped = 0;
1921
1922 ASSERT(GHOST_STATE(state));
1923 top:
1924 mutex_enter(&state->arcs_mtx);
1925 for (ab = list_tail(list); ab; ab = ab_prev) {
1926 ab_prev = list_prev(list, ab);
1927 if (spa && ab->b_spa != spa)
1928 continue;
1929
1930 /* ignore markers */
1931 if (ab->b_spa == 0)
1932 continue;
1933
1934 hash_lock = HDR_LOCK(ab);
1935 /* caller may be trying to modify this buffer, skip it */
1936 if (MUTEX_HELD(hash_lock))
1937 continue;
1938 if (mutex_tryenter(hash_lock)) {
1939 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1940 ASSERT(ab->b_buf == NULL);
1941 ARCSTAT_BUMP(arcstat_deleted);
1942 bytes_deleted += ab->b_size;
1943
1944 if (ab->b_l2hdr != NULL) {
1945 /*
1946 * This buffer is cached on the 2nd Level ARC;
1947 * don't destroy the header.
1948 */
1949 arc_change_state(arc_l2c_only, ab, hash_lock);
1950 mutex_exit(hash_lock);
1951 } else {
1952 arc_change_state(arc_anon, ab, hash_lock);
1953 mutex_exit(hash_lock);
1954 arc_hdr_destroy(ab);
1955 }
1956
1957 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1958 if (bytes >= 0 && bytes_deleted >= bytes)
1959 break;
1960 } else if (bytes < 0) {
1961 /*
1962 * Insert a list marker and then wait for the
1963 * hash lock to become available. Once its
1964 * available, restart from where we left off.
1965 */
1966 list_insert_after(list, ab, &marker);
1967 mutex_exit(&state->arcs_mtx);
1968 mutex_enter(hash_lock);
1969 mutex_exit(hash_lock);
1970 mutex_enter(&state->arcs_mtx);
1971 ab_prev = list_prev(list, &marker);
1972 list_remove(list, &marker);
1973 } else
1974 bufs_skipped += 1;
1975 }
1976 mutex_exit(&state->arcs_mtx);
1977
1978 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1979 (bytes < 0 || bytes_deleted < bytes)) {
1980 list = &state->arcs_list[ARC_BUFC_METADATA];
1981 goto top;
1982 }
1983
1984 if (bufs_skipped) {
1985 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1986 ASSERT(bytes >= 0);
1987 }
1988
1989 if (bytes_deleted < bytes)
1990 dprintf("only deleted %lld bytes from %p",
1991 (longlong_t)bytes_deleted, state);
1992 }
1993
1994 static void
1995 arc_adjust(void)
1996 {
1997 int64_t adjustment, delta;
1998
1999 /*
2000 * Adjust MRU size
2001 */
2002
2003 adjustment = MIN((int64_t)(arc_size - arc_c),
2004 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
2005 arc_p));
2006
2007 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
2008 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
2009 (void) arc_evict(arc_mru, NULL, delta, FALSE, ARC_BUFC_DATA);
2010 adjustment -= delta;
2011 }
2012
2013 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2014 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
2015 (void) arc_evict(arc_mru, NULL, delta, FALSE,
2016 ARC_BUFC_METADATA);
2017 }
2018
2019 /*
2020 * Adjust MFU size
2021 */
2022
2023 adjustment = arc_size - arc_c;
2024
2025 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
2026 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
2027 (void) arc_evict(arc_mfu, NULL, delta, FALSE, ARC_BUFC_DATA);
2028 adjustment -= delta;
2029 }
2030
2031 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
2032 int64_t delta = MIN(adjustment,
2033 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
2034 (void) arc_evict(arc_mfu, NULL, delta, FALSE,
2035 ARC_BUFC_METADATA);
2036 }
2037
2038 /*
2039 * Adjust ghost lists
2040 */
2041
2042 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
2043
2044 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
2045 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
2046 arc_evict_ghost(arc_mru_ghost, NULL, delta);
2047 }
2048
2049 adjustment =
2050 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2051
2052 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2053 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2054 arc_evict_ghost(arc_mfu_ghost, NULL, delta);
2055 }
2056 }
2057
2058 static void
2059 arc_do_user_evicts(void)
2060 {
2061 mutex_enter(&arc_eviction_mtx);
2062 while (arc_eviction_list != NULL) {
2063 arc_buf_t *buf = arc_eviction_list;
2064 arc_eviction_list = buf->b_next;
2065 mutex_enter(&buf->b_evict_lock);
2066 buf->b_hdr = NULL;
2067 mutex_exit(&buf->b_evict_lock);
2068 mutex_exit(&arc_eviction_mtx);
2069
2070 if (buf->b_efunc != NULL)
2071 VERIFY(buf->b_efunc(buf) == 0);
2072
2073 buf->b_efunc = NULL;
2074 buf->b_private = NULL;
2075 kmem_cache_free(buf_cache, buf);
2076 mutex_enter(&arc_eviction_mtx);
2077 }
2078 mutex_exit(&arc_eviction_mtx);
2079 }
2080
2081 /*
2082 * Flush all *evictable* data from the cache for the given spa.
2083 * NOTE: this will not touch "active" (i.e. referenced) data.
2084 */
2085 void
2086 arc_flush(spa_t *spa)
2087 {
2088 uint64_t guid = 0;
2089
2090 if (spa)
2091 guid = spa_load_guid(spa);
2092
2093 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
2094 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
2095 if (spa)
2096 break;
2097 }
2098 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
2099 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
2100 if (spa)
2101 break;
2102 }
2103 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
2104 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
2105 if (spa)
2106 break;
2107 }
2108 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
2109 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
2110 if (spa)
2111 break;
2112 }
2113
2114 arc_evict_ghost(arc_mru_ghost, guid, -1);
2115 arc_evict_ghost(arc_mfu_ghost, guid, -1);
2116
2117 mutex_enter(&arc_reclaim_thr_lock);
2118 arc_do_user_evicts();
2119 mutex_exit(&arc_reclaim_thr_lock);
2120 ASSERT(spa || arc_eviction_list == NULL);
2121 }
2122
2123 void
2124 arc_shrink(void)
2125 {
2126 if (arc_c > arc_c_min) {
2127 uint64_t to_free;
2128
2129 #ifdef _KERNEL
2130 to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
2131 #else
2132 to_free = arc_c >> arc_shrink_shift;
2133 #endif
2134 if (arc_c > arc_c_min + to_free)
2135 atomic_add_64(&arc_c, -to_free);
2136 else
2137 arc_c = arc_c_min;
2138
2139 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2140 if (arc_c > arc_size)
2141 arc_c = MAX(arc_size, arc_c_min);
2142 if (arc_p > arc_c)
2143 arc_p = (arc_c >> 1);
2144 ASSERT(arc_c >= arc_c_min);
2145 ASSERT((int64_t)arc_p >= 0);
2146 }
2147
2148 if (arc_size > arc_c)
2149 arc_adjust();
2150 }
2151
2152 /*
2153 * Determine if the system is under memory pressure and is asking
2154 * to reclaim memory. A return value of 1 indicates that the system
2155 * is under memory pressure and that the arc should adjust accordingly.
2156 */
2157 static int
2158 arc_reclaim_needed(void)
2159 {
2160 uint64_t extra;
2161
2162 #ifdef _KERNEL
2163
2164 if (needfree)
2165 return (1);
2166
2167 /*
2168 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2169 */
2170 extra = desfree;
2171
2172 /*
2173 * check that we're out of range of the pageout scanner. It starts to
2174 * schedule paging if freemem is less than lotsfree and needfree.
2175 * lotsfree is the high-water mark for pageout, and needfree is the
2176 * number of needed free pages. We add extra pages here to make sure
2177 * the scanner doesn't start up while we're freeing memory.
2178 */
2179 if (freemem < lotsfree + needfree + extra)
2180 return (1);
2181
2182 /*
2183 * check to make sure that swapfs has enough space so that anon
2184 * reservations can still succeed. anon_resvmem() checks that the
2185 * availrmem is greater than swapfs_minfree, and the number of reserved
2186 * swap pages. We also add a bit of extra here just to prevent
2187 * circumstances from getting really dire.
2188 */
2189 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2190 return (1);
2191
2192 #if defined(__i386)
2193 /*
2194 * If we're on an i386 platform, it's possible that we'll exhaust the
2195 * kernel heap space before we ever run out of available physical
2196 * memory. Most checks of the size of the heap_area compare against
2197 * tune.t_minarmem, which is the minimum available real memory that we
2198 * can have in the system. However, this is generally fixed at 25 pages
2199 * which is so low that it's useless. In this comparison, we seek to
2200 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2201 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2202 * free)
2203 */
2204 if (vmem_size(heap_arena, VMEM_FREE) <
2205 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2))
2206 return (1);
2207 #endif
2208
2209 /*
2210 * If zio data pages are being allocated out of a separate heap segment,
2211 * then enforce that the size of available vmem for this arena remains
2212 * above about 1/16th free.
2213 *
2214 * Note: The 1/16th arena free requirement was put in place
2215 * to aggressively evict memory from the arc in order to avoid
2216 * memory fragmentation issues.
2217 */
2218 if (zio_arena != NULL &&
2219 vmem_size(zio_arena, VMEM_FREE) <
2220 (vmem_size(zio_arena, VMEM_ALLOC) >> 4))
2221 return (1);
2222 #else
2223 if (spa_get_random(100) == 0)
2224 return (1);
2225 #endif
2226 return (0);
2227 }
2228
2229 static void
2230 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2231 {
2232 size_t i;
2233 kmem_cache_t *prev_cache = NULL;
2234 kmem_cache_t *prev_data_cache = NULL;
2235 extern kmem_cache_t *zio_buf_cache[];
2236 extern kmem_cache_t *zio_data_buf_cache[];
2237
2238 #ifdef _KERNEL
2239 if (arc_meta_used >= arc_meta_limit) {
2240 /*
2241 * We are exceeding our meta-data cache limit.
2242 * Purge some DNLC entries to release holds on meta-data.
2243 */
2244 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2245 }
2246 #if defined(__i386)
2247 /*
2248 * Reclaim unused memory from all kmem caches.
2249 */
2250 kmem_reap();
2251 #endif
2252 #endif
2253
2254 /*
2255 * An aggressive reclamation will shrink the cache size as well as
2256 * reap free buffers from the arc kmem caches.
2257 */
2258 if (strat == ARC_RECLAIM_AGGR)
2259 arc_shrink();
2260
2261 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2262 if (zio_buf_cache[i] != prev_cache) {
2263 prev_cache = zio_buf_cache[i];
2264 kmem_cache_reap_now(zio_buf_cache[i]);
2265 }
2266 if (zio_data_buf_cache[i] != prev_data_cache) {
2267 prev_data_cache = zio_data_buf_cache[i];
2268 kmem_cache_reap_now(zio_data_buf_cache[i]);
2269 }
2270 }
2271 kmem_cache_reap_now(buf_cache);
2272 kmem_cache_reap_now(hdr_cache);
2273
2274 /*
2275 * Ask the vmem areana to reclaim unused memory from its
2276 * quantum caches.
2277 */
2278 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR)
2279 vmem_qcache_reap(zio_arena);
2280 }
2281
2282 static void
2283 arc_reclaim_thread(void)
2284 {
2285 clock_t growtime = 0;
2286 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2287 callb_cpr_t cpr;
2288
2289 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2290
2291 mutex_enter(&arc_reclaim_thr_lock);
2292 while (arc_thread_exit == 0) {
2293 if (arc_reclaim_needed()) {
2294
2295 if (arc_no_grow) {
2296 if (last_reclaim == ARC_RECLAIM_CONS) {
2297 last_reclaim = ARC_RECLAIM_AGGR;
2298 } else {
2299 last_reclaim = ARC_RECLAIM_CONS;
2300 }
2301 } else {
2302 arc_no_grow = TRUE;
2303 last_reclaim = ARC_RECLAIM_AGGR;
2304 membar_producer();
2305 }
2306
2307 /* reset the growth delay for every reclaim */
2308 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2309
2310 arc_kmem_reap_now(last_reclaim);
2311 arc_warm = B_TRUE;
2312
2313 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2314 arc_no_grow = FALSE;
2315 }
2316
2317 arc_adjust();
2318
2319 if (arc_eviction_list != NULL)
2320 arc_do_user_evicts();
2321
2322 /* block until needed, or one second, whichever is shorter */
2323 CALLB_CPR_SAFE_BEGIN(&cpr);
2324 (void) cv_timedwait(&arc_reclaim_thr_cv,
2325 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2326 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2327 }
2328
2329 arc_thread_exit = 0;
2330 cv_broadcast(&arc_reclaim_thr_cv);
2331 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2332 thread_exit();
2333 }
2334
2335 /*
2336 * Adapt arc info given the number of bytes we are trying to add and
2337 * the state that we are comming from. This function is only called
2338 * when we are adding new content to the cache.
2339 */
2340 static void
2341 arc_adapt(int bytes, arc_state_t *state)
2342 {
2343 int mult;
2344 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2345
2346 if (state == arc_l2c_only)
2347 return;
2348
2349 ASSERT(bytes > 0);
2350 /*
2351 * Adapt the target size of the MRU list:
2352 * - if we just hit in the MRU ghost list, then increase
2353 * the target size of the MRU list.
2354 * - if we just hit in the MFU ghost list, then increase
2355 * the target size of the MFU list by decreasing the
2356 * target size of the MRU list.
2357 */
2358 if (state == arc_mru_ghost) {
2359 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2360 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2361 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2362
2363 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2364 } else if (state == arc_mfu_ghost) {
2365 uint64_t delta;
2366
2367 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2368 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2369 mult = MIN(mult, 10);
2370
2371 delta = MIN(bytes * mult, arc_p);
2372 arc_p = MAX(arc_p_min, arc_p - delta);
2373 }
2374 ASSERT((int64_t)arc_p >= 0);
2375
2376 if (arc_reclaim_needed()) {
2377 cv_signal(&arc_reclaim_thr_cv);
2378 return;
2379 }
2380
2381 if (arc_no_grow)
2382 return;
2383
2384 if (arc_c >= arc_c_max)
2385 return;
2386
2387 /*
2388 * If we're within (2 * maxblocksize) bytes of the target
2389 * cache size, increment the target cache size
2390 */
2391 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2392 atomic_add_64(&arc_c, (int64_t)bytes);
2393 if (arc_c > arc_c_max)
2394 arc_c = arc_c_max;
2395 else if (state == arc_anon)
2396 atomic_add_64(&arc_p, (int64_t)bytes);
2397 if (arc_p > arc_c)
2398 arc_p = arc_c;
2399 }
2400 ASSERT((int64_t)arc_p >= 0);
2401 }
2402
2403 /*
2404 * Check if the cache has reached its limits and eviction is required
2405 * prior to insert.
2406 */
2407 static int
2408 arc_evict_needed(arc_buf_contents_t type)
2409 {
2410 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2411 return (1);
2412
2413 if (arc_reclaim_needed())
2414 return (1);
2415
2416 return (arc_size > arc_c);
2417 }
2418
2419 /*
2420 * The buffer, supplied as the first argument, needs a data block.
2421 * So, if we are at cache max, determine which cache should be victimized.
2422 * We have the following cases:
2423 *
2424 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2425 * In this situation if we're out of space, but the resident size of the MFU is
2426 * under the limit, victimize the MFU cache to satisfy this insertion request.
2427 *
2428 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2429 * Here, we've used up all of the available space for the MRU, so we need to
2430 * evict from our own cache instead. Evict from the set of resident MRU
2431 * entries.
2432 *
2433 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2434 * c minus p represents the MFU space in the cache, since p is the size of the
2435 * cache that is dedicated to the MRU. In this situation there's still space on
2436 * the MFU side, so the MRU side needs to be victimized.
2437 *
2438 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2439 * MFU's resident set is consuming more space than it has been allotted. In
2440 * this situation, we must victimize our own cache, the MFU, for this insertion.
2441 */
2442 static void
2443 arc_get_data_buf(arc_buf_t *buf)
2444 {
2445 arc_state_t *state = buf->b_hdr->b_state;
2446 uint64_t size = buf->b_hdr->b_size;
2447 arc_buf_contents_t type = buf->b_hdr->b_type;
2448
2449 arc_adapt(size, state);
2450
2451 /*
2452 * We have not yet reached cache maximum size,
2453 * just allocate a new buffer.
2454 */
2455 if (!arc_evict_needed(type)) {
2456 if (type == ARC_BUFC_METADATA) {
2457 buf->b_data = zio_buf_alloc(size);
2458 arc_space_consume(size, ARC_SPACE_DATA);
2459 } else {
2460 ASSERT(type == ARC_BUFC_DATA);
2461 buf->b_data = zio_data_buf_alloc(size);
2462 ARCSTAT_INCR(arcstat_data_size, size);
2463 atomic_add_64(&arc_size, size);
2464 }
2465 goto out;
2466 }
2467
2468 /*
2469 * If we are prefetching from the mfu ghost list, this buffer
2470 * will end up on the mru list; so steal space from there.
2471 */
2472 if (state == arc_mfu_ghost)
2473 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2474 else if (state == arc_mru_ghost)
2475 state = arc_mru;
2476
2477 if (state == arc_mru || state == arc_anon) {
2478 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2479 state = (arc_mfu->arcs_lsize[type] >= size &&
2480 arc_p > mru_used) ? arc_mfu : arc_mru;
2481 } else {
2482 /* MFU cases */
2483 uint64_t mfu_space = arc_c - arc_p;
2484 state = (arc_mru->arcs_lsize[type] >= size &&
2485 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2486 }
2487 if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) {
2488 if (type == ARC_BUFC_METADATA) {
2489 buf->b_data = zio_buf_alloc(size);
2490 arc_space_consume(size, ARC_SPACE_DATA);
2491 } else {
2492 ASSERT(type == ARC_BUFC_DATA);
2493 buf->b_data = zio_data_buf_alloc(size);
2494 ARCSTAT_INCR(arcstat_data_size, size);
2495 atomic_add_64(&arc_size, size);
2496 }
2497 ARCSTAT_BUMP(arcstat_recycle_miss);
2498 }
2499 ASSERT(buf->b_data != NULL);
2500 out:
2501 /*
2502 * Update the state size. Note that ghost states have a
2503 * "ghost size" and so don't need to be updated.
2504 */
2505 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2506 arc_buf_hdr_t *hdr = buf->b_hdr;
2507
2508 atomic_add_64(&hdr->b_state->arcs_size, size);
2509 if (list_link_active(&hdr->b_arc_node)) {
2510 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2511 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2512 }
2513 /*
2514 * If we are growing the cache, and we are adding anonymous
2515 * data, and we have outgrown arc_p, update arc_p
2516 */
2517 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2518 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2519 arc_p = MIN(arc_c, arc_p + size);
2520 }
2521 }
2522
2523 /*
2524 * This routine is called whenever a buffer is accessed.
2525 * NOTE: the hash lock is dropped in this function.
2526 */
2527 static void
2528 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2529 {
2530 clock_t now;
2531
2532 ASSERT(MUTEX_HELD(hash_lock));
2533
2534 if (buf->b_state == arc_anon) {
2535 /*
2536 * This buffer is not in the cache, and does not
2537 * appear in our "ghost" list. Add the new buffer
2538 * to the MRU state.
2539 */
2540
2541 ASSERT(buf->b_arc_access == 0);
2542 buf->b_arc_access = ddi_get_lbolt();
2543 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2544 arc_change_state(arc_mru, buf, hash_lock);
2545
2546 } else if (buf->b_state == arc_mru) {
2547 now = ddi_get_lbolt();
2548
2549 /*
2550 * If this buffer is here because of a prefetch, then either:
2551 * - clear the flag if this is a "referencing" read
2552 * (any subsequent access will bump this into the MFU state).
2553 * or
2554 * - move the buffer to the head of the list if this is
2555 * another prefetch (to make it less likely to be evicted).
2556 */
2557 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2558 if (refcount_count(&buf->b_refcnt) == 0) {
2559 ASSERT(list_link_active(&buf->b_arc_node));
2560 } else {
2561 buf->b_flags &= ~ARC_PREFETCH;
2562 ARCSTAT_BUMP(arcstat_mru_hits);
2563 }
2564 buf->b_arc_access = now;
2565 return;
2566 }
2567
2568 /*
2569 * This buffer has been "accessed" only once so far,
2570 * but it is still in the cache. Move it to the MFU
2571 * state.
2572 */
2573 if (now > buf->b_arc_access + ARC_MINTIME) {
2574 /*
2575 * More than 125ms have passed since we
2576 * instantiated this buffer. Move it to the
2577 * most frequently used state.
2578 */
2579 buf->b_arc_access = now;
2580 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2581 arc_change_state(arc_mfu, buf, hash_lock);
2582 }
2583 ARCSTAT_BUMP(arcstat_mru_hits);
2584 } else if (buf->b_state == arc_mru_ghost) {
2585 arc_state_t *new_state;
2586 /*
2587 * This buffer has been "accessed" recently, but
2588 * was evicted from the cache. Move it to the
2589 * MFU state.
2590 */
2591
2592 if (buf->b_flags & ARC_PREFETCH) {
2593 new_state = arc_mru;
2594 if (refcount_count(&buf->b_refcnt) > 0)
2595 buf->b_flags &= ~ARC_PREFETCH;
2596 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2597 } else {
2598 new_state = arc_mfu;
2599 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2600 }
2601
2602 buf->b_arc_access = ddi_get_lbolt();
2603 arc_change_state(new_state, buf, hash_lock);
2604
2605 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2606 } else if (buf->b_state == arc_mfu) {
2607 /*
2608 * This buffer has been accessed more than once and is
2609 * still in the cache. Keep it in the MFU state.
2610 *
2611 * NOTE: an add_reference() that occurred when we did
2612 * the arc_read() will have kicked this off the list.
2613 * If it was a prefetch, we will explicitly move it to
2614 * the head of the list now.
2615 */
2616 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2617 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2618 ASSERT(list_link_active(&buf->b_arc_node));
2619 }
2620 ARCSTAT_BUMP(arcstat_mfu_hits);
2621 buf->b_arc_access = ddi_get_lbolt();
2622 } else if (buf->b_state == arc_mfu_ghost) {
2623 arc_state_t *new_state = arc_mfu;
2624 /*
2625 * This buffer has been accessed more than once but has
2626 * been evicted from the cache. Move it back to the
2627 * MFU state.
2628 */
2629
2630 if (buf->b_flags & ARC_PREFETCH) {
2631 /*
2632 * This is a prefetch access...
2633 * move this block back to the MRU state.
2634 */
2635 ASSERT0(refcount_count(&buf->b_refcnt));
2636 new_state = arc_mru;
2637 }
2638
2639 buf->b_arc_access = ddi_get_lbolt();
2640 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2641 arc_change_state(new_state, buf, hash_lock);
2642
2643 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2644 } else if (buf->b_state == arc_l2c_only) {
2645 /*
2646 * This buffer is on the 2nd Level ARC.
2647 */
2648
2649 buf->b_arc_access = ddi_get_lbolt();
2650 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2651 arc_change_state(arc_mfu, buf, hash_lock);
2652 } else {
2653 ASSERT(!"invalid arc state");
2654 }
2655 }
2656
2657 /* a generic arc_done_func_t which you can use */
2658 /* ARGSUSED */
2659 void
2660 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2661 {
2662 if (zio == NULL || zio->io_error == 0)
2663 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2664 VERIFY(arc_buf_remove_ref(buf, arg));
2665 }
2666
2667 /* a generic arc_done_func_t */
2668 void
2669 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2670 {
2671 arc_buf_t **bufp = arg;
2672 if (zio && zio->io_error) {
2673 VERIFY(arc_buf_remove_ref(buf, arg));
2674 *bufp = NULL;
2675 } else {
2676 *bufp = buf;
2677 ASSERT(buf->b_data);
2678 }
2679 }
2680
2681 static void
2682 arc_read_done(zio_t *zio)
2683 {
2684 arc_buf_hdr_t *hdr, *found;
2685 arc_buf_t *buf;
2686 arc_buf_t *abuf; /* buffer we're assigning to callback */
2687 kmutex_t *hash_lock;
2688 arc_callback_t *callback_list, *acb;
2689 int freeable = FALSE;
2690
2691 buf = zio->io_private;
2692 hdr = buf->b_hdr;
2693
2694 /*
2695 * The hdr was inserted into hash-table and removed from lists
2696 * prior to starting I/O. We should find this header, since
2697 * it's in the hash table, and it should be legit since it's
2698 * not possible to evict it during the I/O. The only possible
2699 * reason for it not to be found is if we were freed during the
2700 * read.
2701 */
2702 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2703 &hash_lock);
2704
2705 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2706 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2707 (found == hdr && HDR_L2_READING(hdr)));
2708
2709 hdr->b_flags &= ~ARC_L2_EVICTED;
2710 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2711 hdr->b_flags &= ~ARC_L2CACHE;
2712
2713 /* byteswap if necessary */
2714 callback_list = hdr->b_acb;
2715 ASSERT(callback_list != NULL);
2716 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2717 dmu_object_byteswap_t bswap =
2718 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
2719 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2720 byteswap_uint64_array :
2721 dmu_ot_byteswap[bswap].ob_func;
2722 func(buf->b_data, hdr->b_size);
2723 }
2724
2725 arc_cksum_compute(buf, B_FALSE);
2726 arc_buf_watch(buf);
2727
2728 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2729 /*
2730 * Only call arc_access on anonymous buffers. This is because
2731 * if we've issued an I/O for an evicted buffer, we've already
2732 * called arc_access (to prevent any simultaneous readers from
2733 * getting confused).
2734 */
2735 arc_access(hdr, hash_lock);
2736 }
2737
2738 /* create copies of the data buffer for the callers */
2739 abuf = buf;
2740 for (acb = callback_list; acb; acb = acb->acb_next) {
2741 if (acb->acb_done) {
2742 if (abuf == NULL) {
2743 ARCSTAT_BUMP(arcstat_duplicate_reads);
2744 abuf = arc_buf_clone(buf);
2745 }
2746 acb->acb_buf = abuf;
2747 abuf = NULL;
2748 }
2749 }
2750 hdr->b_acb = NULL;
2751 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2752 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2753 if (abuf == buf) {
2754 ASSERT(buf->b_efunc == NULL);
2755 ASSERT(hdr->b_datacnt == 1);
2756 hdr->b_flags |= ARC_BUF_AVAILABLE;
2757 }
2758
2759 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2760
2761 if (zio->io_error != 0) {
2762 hdr->b_flags |= ARC_IO_ERROR;
2763 if (hdr->b_state != arc_anon)
2764 arc_change_state(arc_anon, hdr, hash_lock);
2765 if (HDR_IN_HASH_TABLE(hdr))
2766 buf_hash_remove(hdr);
2767 freeable = refcount_is_zero(&hdr->b_refcnt);
2768 }
2769
2770 /*
2771 * Broadcast before we drop the hash_lock to avoid the possibility
2772 * that the hdr (and hence the cv) might be freed before we get to
2773 * the cv_broadcast().
2774 */
2775 cv_broadcast(&hdr->b_cv);
2776
2777 if (hash_lock) {
2778 mutex_exit(hash_lock);
2779 } else {
2780 /*
2781 * This block was freed while we waited for the read to
2782 * complete. It has been removed from the hash table and
2783 * moved to the anonymous state (so that it won't show up
2784 * in the cache).
2785 */
2786 ASSERT3P(hdr->b_state, ==, arc_anon);
2787 freeable = refcount_is_zero(&hdr->b_refcnt);
2788 }
2789
2790 /* execute each callback and free its structure */
2791 while ((acb = callback_list) != NULL) {
2792 if (acb->acb_done)
2793 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2794
2795 if (acb->acb_zio_dummy != NULL) {
2796 acb->acb_zio_dummy->io_error = zio->io_error;
2797 zio_nowait(acb->acb_zio_dummy);
2798 }
2799
2800 callback_list = acb->acb_next;
2801 kmem_free(acb, sizeof (arc_callback_t));
2802 }
2803
2804 if (freeable)
2805 arc_hdr_destroy(hdr);
2806 }
2807
2808 /*
2809 * "Read" the block at the specified DVA (in bp) via the
2810 * cache. If the block is found in the cache, invoke the provided
2811 * callback immediately and return. Note that the `zio' parameter
2812 * in the callback will be NULL in this case, since no IO was
2813 * required. If the block is not in the cache pass the read request
2814 * on to the spa with a substitute callback function, so that the
2815 * requested block will be added to the cache.
2816 *
2817 * If a read request arrives for a block that has a read in-progress,
2818 * either wait for the in-progress read to complete (and return the
2819 * results); or, if this is a read with a "done" func, add a record
2820 * to the read to invoke the "done" func when the read completes,
2821 * and return; or just return.
2822 *
2823 * arc_read_done() will invoke all the requested "done" functions
2824 * for readers of this block.
2825 */
2826 int
2827 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
2828 void *private, int priority, int zio_flags, uint32_t *arc_flags,
2829 const zbookmark_t *zb)
2830 {
2831 arc_buf_hdr_t *hdr;
2832 arc_buf_t *buf = NULL;
2833 kmutex_t *hash_lock;
2834 zio_t *rzio;
2835 uint64_t guid = spa_load_guid(spa);
2836
2837 top:
2838 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2839 &hash_lock);
2840 if (hdr && hdr->b_datacnt > 0) {
2841
2842 *arc_flags |= ARC_CACHED;
2843
2844 if (HDR_IO_IN_PROGRESS(hdr)) {
2845
2846 if (*arc_flags & ARC_WAIT) {
2847 cv_wait(&hdr->b_cv, hash_lock);
2848 mutex_exit(hash_lock);
2849 goto top;
2850 }
2851 ASSERT(*arc_flags & ARC_NOWAIT);
2852
2853 if (done) {
2854 arc_callback_t *acb = NULL;
2855
2856 acb = kmem_zalloc(sizeof (arc_callback_t),
2857 KM_SLEEP);
2858 acb->acb_done = done;
2859 acb->acb_private = private;
2860 if (pio != NULL)
2861 acb->acb_zio_dummy = zio_null(pio,
2862 spa, NULL, NULL, NULL, zio_flags);
2863
2864 ASSERT(acb->acb_done != NULL);
2865 acb->acb_next = hdr->b_acb;
2866 hdr->b_acb = acb;
2867 add_reference(hdr, hash_lock, private);
2868 mutex_exit(hash_lock);
2869 return (0);
2870 }
2871 mutex_exit(hash_lock);
2872 return (0);
2873 }
2874
2875 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2876
2877 if (done) {
2878 add_reference(hdr, hash_lock, private);
2879 /*
2880 * If this block is already in use, create a new
2881 * copy of the data so that we will be guaranteed
2882 * that arc_release() will always succeed.
2883 */
2884 buf = hdr->b_buf;
2885 ASSERT(buf);
2886 ASSERT(buf->b_data);
2887 if (HDR_BUF_AVAILABLE(hdr)) {
2888 ASSERT(buf->b_efunc == NULL);
2889 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2890 } else {
2891 buf = arc_buf_clone(buf);
2892 }
2893
2894 } else if (*arc_flags & ARC_PREFETCH &&
2895 refcount_count(&hdr->b_refcnt) == 0) {
2896 hdr->b_flags |= ARC_PREFETCH;
2897 }
2898 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2899 arc_access(hdr, hash_lock);
2900 if (*arc_flags & ARC_L2CACHE)
2901 hdr->b_flags |= ARC_L2CACHE;
2902 if (*arc_flags & ARC_L2COMPRESS)
2903 hdr->b_flags |= ARC_L2COMPRESS;
2904 mutex_exit(hash_lock);
2905 ARCSTAT_BUMP(arcstat_hits);
2906 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2907 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2908 data, metadata, hits);
2909
2910 if (done)
2911 done(NULL, buf, private);
2912 } else {
2913 uint64_t size = BP_GET_LSIZE(bp);
2914 arc_callback_t *acb;
2915 vdev_t *vd = NULL;
2916 uint64_t addr = 0;
2917 boolean_t devw = B_FALSE;
2918
2919 if (hdr == NULL) {
2920 /* this block is not in the cache */
2921 arc_buf_hdr_t *exists;
2922 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2923 buf = arc_buf_alloc(spa, size, private, type);
2924 hdr = buf->b_hdr;
2925 hdr->b_dva = *BP_IDENTITY(bp);
2926 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
2927 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2928 exists = buf_hash_insert(hdr, &hash_lock);
2929 if (exists) {
2930 /* somebody beat us to the hash insert */
2931 mutex_exit(hash_lock);
2932 buf_discard_identity(hdr);
2933 (void) arc_buf_remove_ref(buf, private);
2934 goto top; /* restart the IO request */
2935 }
2936 /* if this is a prefetch, we don't have a reference */
2937 if (*arc_flags & ARC_PREFETCH) {
2938 (void) remove_reference(hdr, hash_lock,
2939 private);
2940 hdr->b_flags |= ARC_PREFETCH;
2941 }
2942 if (*arc_flags & ARC_L2CACHE)
2943 hdr->b_flags |= ARC_L2CACHE;
2944 if (*arc_flags & ARC_L2COMPRESS)
2945 hdr->b_flags |= ARC_L2COMPRESS;
2946 if (BP_GET_LEVEL(bp) > 0)
2947 hdr->b_flags |= ARC_INDIRECT;
2948 } else {
2949 /* this block is in the ghost cache */
2950 ASSERT(GHOST_STATE(hdr->b_state));
2951 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2952 ASSERT0(refcount_count(&hdr->b_refcnt));
2953 ASSERT(hdr->b_buf == NULL);
2954
2955 /* if this is a prefetch, we don't have a reference */
2956 if (*arc_flags & ARC_PREFETCH)
2957 hdr->b_flags |= ARC_PREFETCH;
2958 else
2959 add_reference(hdr, hash_lock, private);
2960 if (*arc_flags & ARC_L2CACHE)
2961 hdr->b_flags |= ARC_L2CACHE;
2962 if (*arc_flags & ARC_L2COMPRESS)
2963 hdr->b_flags |= ARC_L2COMPRESS;
2964 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2965 buf->b_hdr = hdr;
2966 buf->b_data = NULL;
2967 buf->b_efunc = NULL;
2968 buf->b_private = NULL;
2969 buf->b_next = NULL;
2970 hdr->b_buf = buf;
2971 ASSERT(hdr->b_datacnt == 0);
2972 hdr->b_datacnt = 1;
2973 arc_get_data_buf(buf);
2974 arc_access(hdr, hash_lock);
2975 }
2976
2977 ASSERT(!GHOST_STATE(hdr->b_state));
2978
2979 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
2980 acb->acb_done = done;
2981 acb->acb_private = private;
2982
2983 ASSERT(hdr->b_acb == NULL);
2984 hdr->b_acb = acb;
2985 hdr->b_flags |= ARC_IO_IN_PROGRESS;
2986
2987 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
2988 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
2989 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
2990 addr = hdr->b_l2hdr->b_daddr;
2991 /*
2992 * Lock out device removal.
2993 */
2994 if (vdev_is_dead(vd) ||
2995 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
2996 vd = NULL;
2997 }
2998
2999 mutex_exit(hash_lock);
3000
3001 /*
3002 * At this point, we have a level 1 cache miss. Try again in
3003 * L2ARC if possible.
3004 */
3005 ASSERT3U(hdr->b_size, ==, size);
3006 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
3007 uint64_t, size, zbookmark_t *, zb);
3008 ARCSTAT_BUMP(arcstat_misses);
3009 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
3010 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
3011 data, metadata, misses);
3012
3013 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3014 /*
3015 * Read from the L2ARC if the following are true:
3016 * 1. The L2ARC vdev was previously cached.
3017 * 2. This buffer still has L2ARC metadata.
3018 * 3. This buffer isn't currently writing to the L2ARC.
3019 * 4. The L2ARC entry wasn't evicted, which may
3020 * also have invalidated the vdev.
3021 * 5. This isn't prefetch and l2arc_noprefetch is set.
3022 */
3023 if (hdr->b_l2hdr != NULL &&
3024 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3025 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3026 l2arc_read_callback_t *cb;
3027
3028 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3029 ARCSTAT_BUMP(arcstat_l2_hits);
3030
3031 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3032 KM_SLEEP);
3033 cb->l2rcb_buf = buf;
3034 cb->l2rcb_spa = spa;
3035 cb->l2rcb_bp = *bp;
3036 cb->l2rcb_zb = *zb;
3037 cb->l2rcb_flags = zio_flags;
3038 cb->l2rcb_compress = hdr->b_l2hdr->b_compress;
3039
3040 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
3041 addr + size < vd->vdev_psize -
3042 VDEV_LABEL_END_SIZE);
3043
3044 /*
3045 * l2arc read. The SCL_L2ARC lock will be
3046 * released by l2arc_read_done().
3047 * Issue a null zio if the underlying buffer
3048 * was squashed to zero size by compression.
3049 */
3050 if (hdr->b_l2hdr->b_compress ==
3051 ZIO_COMPRESS_EMPTY) {
3052 rzio = zio_null(pio, spa, vd,
3053 l2arc_read_done, cb,
3054 zio_flags | ZIO_FLAG_DONT_CACHE |
3055 ZIO_FLAG_CANFAIL |
3056 ZIO_FLAG_DONT_PROPAGATE |
3057 ZIO_FLAG_DONT_RETRY);
3058 } else {
3059 rzio = zio_read_phys(pio, vd, addr,
3060 hdr->b_l2hdr->b_asize,
3061 buf->b_data, ZIO_CHECKSUM_OFF,
3062 l2arc_read_done, cb, priority,
3063 zio_flags | ZIO_FLAG_DONT_CACHE |
3064 ZIO_FLAG_CANFAIL |
3065 ZIO_FLAG_DONT_PROPAGATE |
3066 ZIO_FLAG_DONT_RETRY, B_FALSE);
3067 }
3068 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3069 zio_t *, rzio);
3070 ARCSTAT_INCR(arcstat_l2_read_bytes,
3071 hdr->b_l2hdr->b_asize);
3072
3073 if (*arc_flags & ARC_NOWAIT) {
3074 zio_nowait(rzio);
3075 return (0);
3076 }
3077
3078 ASSERT(*arc_flags & ARC_WAIT);
3079 if (zio_wait(rzio) == 0)
3080 return (0);
3081
3082 /* l2arc read error; goto zio_read() */
3083 } else {
3084 DTRACE_PROBE1(l2arc__miss,
3085 arc_buf_hdr_t *, hdr);
3086 ARCSTAT_BUMP(arcstat_l2_misses);
3087 if (HDR_L2_WRITING(hdr))
3088 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3089 spa_config_exit(spa, SCL_L2ARC, vd);
3090 }
3091 } else {
3092 if (vd != NULL)
3093 spa_config_exit(spa, SCL_L2ARC, vd);
3094 if (l2arc_ndev != 0) {
3095 DTRACE_PROBE1(l2arc__miss,
3096 arc_buf_hdr_t *, hdr);
3097 ARCSTAT_BUMP(arcstat_l2_misses);
3098 }
3099 }
3100
3101 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3102 arc_read_done, buf, priority, zio_flags, zb);
3103
3104 if (*arc_flags & ARC_WAIT)
3105 return (zio_wait(rzio));
3106
3107 ASSERT(*arc_flags & ARC_NOWAIT);
3108 zio_nowait(rzio);
3109 }
3110 return (0);
3111 }
3112
3113 void
3114 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3115 {
3116 ASSERT(buf->b_hdr != NULL);
3117 ASSERT(buf->b_hdr->b_state != arc_anon);
3118 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3119 ASSERT(buf->b_efunc == NULL);
3120 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
3121
3122 buf->b_efunc = func;
3123 buf->b_private = private;
3124 }
3125
3126 /*
3127 * Notify the arc that a block was freed, and thus will never be used again.
3128 */
3129 void
3130 arc_freed(spa_t *spa, const blkptr_t *bp)
3131 {
3132 arc_buf_hdr_t *hdr;
3133 kmutex_t *hash_lock;
3134 uint64_t guid = spa_load_guid(spa);
3135
3136 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
3137 &hash_lock);
3138 if (hdr == NULL)
3139 return;
3140 if (HDR_BUF_AVAILABLE(hdr)) {
3141 arc_buf_t *buf = hdr->b_buf;
3142 add_reference(hdr, hash_lock, FTAG);
3143 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3144 mutex_exit(hash_lock);
3145
3146 arc_release(buf, FTAG);
3147 (void) arc_buf_remove_ref(buf, FTAG);
3148 } else {
3149 mutex_exit(hash_lock);
3150 }
3151
3152 }
3153
3154 /*
3155 * This is used by the DMU to let the ARC know that a buffer is
3156 * being evicted, so the ARC should clean up. If this arc buf
3157 * is not yet in the evicted state, it will be put there.
3158 */
3159 int
3160 arc_buf_evict(arc_buf_t *buf)
3161 {
3162 arc_buf_hdr_t *hdr;
3163 kmutex_t *hash_lock;
3164 arc_buf_t **bufp;
3165
3166 mutex_enter(&buf->b_evict_lock);
3167 hdr = buf->b_hdr;
3168 if (hdr == NULL) {
3169 /*
3170 * We are in arc_do_user_evicts().
3171 */
3172 ASSERT(buf->b_data == NULL);
3173 mutex_exit(&buf->b_evict_lock);
3174 return (0);
3175 } else if (buf->b_data == NULL) {
3176 arc_buf_t copy = *buf; /* structure assignment */
3177 /*
3178 * We are on the eviction list; process this buffer now
3179 * but let arc_do_user_evicts() do the reaping.
3180 */
3181 buf->b_efunc = NULL;
3182 mutex_exit(&buf->b_evict_lock);
3183 VERIFY(copy.b_efunc(©) == 0);
3184 return (1);
3185 }
3186 hash_lock = HDR_LOCK(hdr);
3187 mutex_enter(hash_lock);
3188 hdr = buf->b_hdr;
3189 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3190
3191 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3192 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3193
3194 /*
3195 * Pull this buffer off of the hdr
3196 */
3197 bufp = &hdr->b_buf;
3198 while (*bufp != buf)
3199 bufp = &(*bufp)->b_next;
3200 *bufp = buf->b_next;
3201
3202 ASSERT(buf->b_data != NULL);
3203 arc_buf_destroy(buf, FALSE, FALSE);
3204
3205 if (hdr->b_datacnt == 0) {
3206 arc_state_t *old_state = hdr->b_state;
3207 arc_state_t *evicted_state;
3208
3209 ASSERT(hdr->b_buf == NULL);
3210 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3211
3212 evicted_state =
3213 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3214
3215 mutex_enter(&old_state->arcs_mtx);
3216 mutex_enter(&evicted_state->arcs_mtx);
3217
3218 arc_change_state(evicted_state, hdr, hash_lock);
3219 ASSERT(HDR_IN_HASH_TABLE(hdr));
3220 hdr->b_flags |= ARC_IN_HASH_TABLE;
3221 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3222
3223 mutex_exit(&evicted_state->arcs_mtx);
3224 mutex_exit(&old_state->arcs_mtx);
3225 }
3226 mutex_exit(hash_lock);
3227 mutex_exit(&buf->b_evict_lock);
3228
3229 VERIFY(buf->b_efunc(buf) == 0);
3230 buf->b_efunc = NULL;
3231 buf->b_private = NULL;
3232 buf->b_hdr = NULL;
3233 buf->b_next = NULL;
3234 kmem_cache_free(buf_cache, buf);
3235 return (1);
3236 }
3237
3238 /*
3239 * Release this buffer from the cache, making it an anonymous buffer. This
3240 * must be done after a read and prior to modifying the buffer contents.
3241 * If the buffer has more than one reference, we must make
3242 * a new hdr for the buffer.
3243 */
3244 void
3245 arc_release(arc_buf_t *buf, void *tag)
3246 {
3247 arc_buf_hdr_t *hdr;
3248 kmutex_t *hash_lock = NULL;
3249 l2arc_buf_hdr_t *l2hdr;
3250 uint64_t buf_size;
3251
3252 /*
3253 * It would be nice to assert that if it's DMU metadata (level >
3254 * 0 || it's the dnode file), then it must be syncing context.
3255 * But we don't know that information at this level.
3256 */
3257
3258 mutex_enter(&buf->b_evict_lock);
3259 hdr = buf->b_hdr;
3260
3261 /* this buffer is not on any list */
3262 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3263
3264 if (hdr->b_state == arc_anon) {
3265 /* this buffer is already released */
3266 ASSERT(buf->b_efunc == NULL);
3267 } else {
3268 hash_lock = HDR_LOCK(hdr);
3269 mutex_enter(hash_lock);
3270 hdr = buf->b_hdr;
3271 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3272 }
3273
3274 l2hdr = hdr->b_l2hdr;
3275 if (l2hdr) {
3276 mutex_enter(&l2arc_buflist_mtx);
3277 hdr->b_l2hdr = NULL;
3278 }
3279 buf_size = hdr->b_size;
3280
3281 /*
3282 * Do we have more than one buf?
3283 */
3284 if (hdr->b_datacnt > 1) {
3285 arc_buf_hdr_t *nhdr;
3286 arc_buf_t **bufp;
3287 uint64_t blksz = hdr->b_size;
3288 uint64_t spa = hdr->b_spa;
3289 arc_buf_contents_t type = hdr->b_type;
3290 uint32_t flags = hdr->b_flags;
3291
3292 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3293 /*
3294 * Pull the data off of this hdr and attach it to
3295 * a new anonymous hdr.
3296 */
3297 (void) remove_reference(hdr, hash_lock, tag);
3298 bufp = &hdr->b_buf;
3299 while (*bufp != buf)
3300 bufp = &(*bufp)->b_next;
3301 *bufp = buf->b_next;
3302 buf->b_next = NULL;
3303
3304 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3305 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3306 if (refcount_is_zero(&hdr->b_refcnt)) {
3307 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3308 ASSERT3U(*size, >=, hdr->b_size);
3309 atomic_add_64(size, -hdr->b_size);
3310 }
3311
3312 /*
3313 * We're releasing a duplicate user data buffer, update
3314 * our statistics accordingly.
3315 */
3316 if (hdr->b_type == ARC_BUFC_DATA) {
3317 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
3318 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
3319 -hdr->b_size);
3320 }
3321 hdr->b_datacnt -= 1;
3322 arc_cksum_verify(buf);
3323 arc_buf_unwatch(buf);
3324
3325 mutex_exit(hash_lock);
3326
3327 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3328 nhdr->b_size = blksz;
3329 nhdr->b_spa = spa;
3330 nhdr->b_type = type;
3331 nhdr->b_buf = buf;
3332 nhdr->b_state = arc_anon;
3333 nhdr->b_arc_access = 0;
3334 nhdr->b_flags = flags & ARC_L2_WRITING;
3335 nhdr->b_l2hdr = NULL;
3336 nhdr->b_datacnt = 1;
3337 nhdr->b_freeze_cksum = NULL;
3338 (void) refcount_add(&nhdr->b_refcnt, tag);
3339 buf->b_hdr = nhdr;
3340 mutex_exit(&buf->b_evict_lock);
3341 atomic_add_64(&arc_anon->arcs_size, blksz);
3342 } else {
3343 mutex_exit(&buf->b_evict_lock);
3344 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3345 ASSERT(!list_link_active(&hdr->b_arc_node));
3346 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3347 if (hdr->b_state != arc_anon)
3348 arc_change_state(arc_anon, hdr, hash_lock);
3349 hdr->b_arc_access = 0;
3350 if (hash_lock)
3351 mutex_exit(hash_lock);
3352
3353 buf_discard_identity(hdr);
3354 arc_buf_thaw(buf);
3355 }
3356 buf->b_efunc = NULL;
3357 buf->b_private = NULL;
3358
3359 if (l2hdr) {
3360 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
3361 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3362 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3363 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3364 mutex_exit(&l2arc_buflist_mtx);
3365 }
3366 }
3367
3368 int
3369 arc_released(arc_buf_t *buf)
3370 {
3371 int released;
3372
3373 mutex_enter(&buf->b_evict_lock);
3374 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3375 mutex_exit(&buf->b_evict_lock);
3376 return (released);
3377 }
3378
3379 int
3380 arc_has_callback(arc_buf_t *buf)
3381 {
3382 int callback;
3383
3384 mutex_enter(&buf->b_evict_lock);
3385 callback = (buf->b_efunc != NULL);
3386 mutex_exit(&buf->b_evict_lock);
3387 return (callback);
3388 }
3389
3390 #ifdef ZFS_DEBUG
3391 int
3392 arc_referenced(arc_buf_t *buf)
3393 {
3394 int referenced;
3395
3396 mutex_enter(&buf->b_evict_lock);
3397 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3398 mutex_exit(&buf->b_evict_lock);
3399 return (referenced);
3400 }
3401 #endif
3402
3403 static void
3404 arc_write_ready(zio_t *zio)
3405 {
3406 arc_write_callback_t *callback = zio->io_private;
3407 arc_buf_t *buf = callback->awcb_buf;
3408 arc_buf_hdr_t *hdr = buf->b_hdr;
3409
3410 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3411 callback->awcb_ready(zio, buf, callback->awcb_private);
3412
3413 /*
3414 * If the IO is already in progress, then this is a re-write
3415 * attempt, so we need to thaw and re-compute the cksum.
3416 * It is the responsibility of the callback to handle the
3417 * accounting for any re-write attempt.
3418 */
3419 if (HDR_IO_IN_PROGRESS(hdr)) {
3420 mutex_enter(&hdr->b_freeze_lock);
3421 if (hdr->b_freeze_cksum != NULL) {
3422 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3423 hdr->b_freeze_cksum = NULL;
3424 }
3425 mutex_exit(&hdr->b_freeze_lock);
3426 }
3427 arc_cksum_compute(buf, B_FALSE);
3428 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3429 }
3430
3431 static void
3432 arc_write_done(zio_t *zio)
3433 {
3434 arc_write_callback_t *callback = zio->io_private;
3435 arc_buf_t *buf = callback->awcb_buf;
3436 arc_buf_hdr_t *hdr = buf->b_hdr;
3437
3438 ASSERT(hdr->b_acb == NULL);
3439
3440 if (zio->io_error == 0) {
3441 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3442 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3443 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3444 } else {
3445 ASSERT(BUF_EMPTY(hdr));
3446 }
3447
3448 /*
3449 * If the block to be written was all-zero, we may have
3450 * compressed it away. In this case no write was performed
3451 * so there will be no dva/birth/checksum. The buffer must
3452 * therefore remain anonymous (and uncached).
3453 */
3454 if (!BUF_EMPTY(hdr)) {
3455 arc_buf_hdr_t *exists;
3456 kmutex_t *hash_lock;
3457
3458 ASSERT(zio->io_error == 0);
3459
3460 arc_cksum_verify(buf);
3461
3462 exists = buf_hash_insert(hdr, &hash_lock);
3463 if (exists) {
3464 /*
3465 * This can only happen if we overwrite for
3466 * sync-to-convergence, because we remove
3467 * buffers from the hash table when we arc_free().
3468 */
3469 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3470 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3471 panic("bad overwrite, hdr=%p exists=%p",
3472 (void *)hdr, (void *)exists);
3473 ASSERT(refcount_is_zero(&exists->b_refcnt));
3474 arc_change_state(arc_anon, exists, hash_lock);
3475 mutex_exit(hash_lock);
3476 arc_hdr_destroy(exists);
3477 exists = buf_hash_insert(hdr, &hash_lock);
3478 ASSERT3P(exists, ==, NULL);
3479 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
3480 /* nopwrite */
3481 ASSERT(zio->io_prop.zp_nopwrite);
3482 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3483 panic("bad nopwrite, hdr=%p exists=%p",
3484 (void *)hdr, (void *)exists);
3485 } else {
3486 /* Dedup */
3487 ASSERT(hdr->b_datacnt == 1);
3488 ASSERT(hdr->b_state == arc_anon);
3489 ASSERT(BP_GET_DEDUP(zio->io_bp));
3490 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3491 }
3492 }
3493 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3494 /* if it's not anon, we are doing a scrub */
3495 if (!exists && hdr->b_state == arc_anon)
3496 arc_access(hdr, hash_lock);
3497 mutex_exit(hash_lock);
3498 } else {
3499 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3500 }
3501
3502 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3503 callback->awcb_done(zio, buf, callback->awcb_private);
3504
3505 kmem_free(callback, sizeof (arc_write_callback_t));
3506 }
3507
3508 zio_t *
3509 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3510 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
3511 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *done,
3512 void *private, int priority, int zio_flags, const zbookmark_t *zb)
3513 {
3514 arc_buf_hdr_t *hdr = buf->b_hdr;
3515 arc_write_callback_t *callback;
3516 zio_t *zio;
3517
3518 ASSERT(ready != NULL);
3519 ASSERT(done != NULL);
3520 ASSERT(!HDR_IO_ERROR(hdr));
3521 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3522 ASSERT(hdr->b_acb == NULL);
3523 if (l2arc)
3524 hdr->b_flags |= ARC_L2CACHE;
3525 if (l2arc_compress)
3526 hdr->b_flags |= ARC_L2COMPRESS;
3527 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3528 callback->awcb_ready = ready;
3529 callback->awcb_done = done;
3530 callback->awcb_private = private;
3531 callback->awcb_buf = buf;
3532
3533 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3534 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3535
3536 return (zio);
3537 }
3538
3539 static int
3540 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3541 {
3542 #ifdef _KERNEL
3543 uint64_t available_memory = ptob(freemem);
3544 static uint64_t page_load = 0;
3545 static uint64_t last_txg = 0;
3546
3547 #if defined(__i386)
3548 available_memory =
3549 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3550 #endif
3551 if (available_memory >= zfs_write_limit_max)
3552 return (0);
3553
3554 if (txg > last_txg) {
3555 last_txg = txg;
3556 page_load = 0;
3557 }
3558 /*
3559 * If we are in pageout, we know that memory is already tight,
3560 * the arc is already going to be evicting, so we just want to
3561 * continue to let page writes occur as quickly as possible.
3562 */
3563 if (curproc == proc_pageout) {
3564 if (page_load > MAX(ptob(minfree), available_memory) / 4)
3565 return (SET_ERROR(ERESTART));
3566 /* Note: reserve is inflated, so we deflate */
3567 page_load += reserve / 8;
3568 return (0);
3569 } else if (page_load > 0 && arc_reclaim_needed()) {
3570 /* memory is low, delay before restarting */
3571 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3572 return (SET_ERROR(EAGAIN));
3573 }
3574 page_load = 0;
3575
3576 if (arc_size > arc_c_min) {
3577 uint64_t evictable_memory =
3578 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3579 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3580 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3581 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3582 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3583 }
3584
3585 if (inflight_data > available_memory / 4) {
3586 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3587 return (SET_ERROR(ERESTART));
3588 }
3589 #endif
3590 return (0);
3591 }
3592
3593 void
3594 arc_tempreserve_clear(uint64_t reserve)
3595 {
3596 atomic_add_64(&arc_tempreserve, -reserve);
3597 ASSERT((int64_t)arc_tempreserve >= 0);
3598 }
3599
3600 int
3601 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3602 {
3603 int error;
3604 uint64_t anon_size;
3605
3606 #ifdef ZFS_DEBUG
3607 /*
3608 * Once in a while, fail for no reason. Everything should cope.
3609 */
3610 if (spa_get_random(10000) == 0) {
3611 dprintf("forcing random failure\n");
3612 return (SET_ERROR(ERESTART));
3613 }
3614 #endif
3615 if (reserve > arc_c/4 && !arc_no_grow)
3616 arc_c = MIN(arc_c_max, reserve * 4);
3617 if (reserve > arc_c)
3618 return (SET_ERROR(ENOMEM));
3619
3620 /*
3621 * Don't count loaned bufs as in flight dirty data to prevent long
3622 * network delays from blocking transactions that are ready to be
3623 * assigned to a txg.
3624 */
3625 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3626
3627 /*
3628 * Writes will, almost always, require additional memory allocations
3629 * in order to compress/encrypt/etc the data. We therefore need to
3630 * make sure that there is sufficient available memory for this.
3631 */
3632 if (error = arc_memory_throttle(reserve, anon_size, txg))
3633 return (error);
3634
3635 /*
3636 * Throttle writes when the amount of dirty data in the cache
3637 * gets too large. We try to keep the cache less than half full
3638 * of dirty blocks so that our sync times don't grow too large.
3639 * Note: if two requests come in concurrently, we might let them
3640 * both succeed, when one of them should fail. Not a huge deal.
3641 */
3642
3643 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3644 anon_size > arc_c / 4) {
3645 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3646 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3647 arc_tempreserve>>10,
3648 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3649 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3650 reserve>>10, arc_c>>10);
3651 return (SET_ERROR(ERESTART));
3652 }
3653 atomic_add_64(&arc_tempreserve, reserve);
3654 return (0);
3655 }
3656
3657 void
3658 arc_init(void)
3659 {
3660 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3661 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3662
3663 /* Convert seconds to clock ticks */
3664 arc_min_prefetch_lifespan = 1 * hz;
3665
3666 /* Start out with 1/8 of all memory */
3667 arc_c = physmem * PAGESIZE / 8;
3668
3669 #ifdef _KERNEL
3670 /*
3671 * On architectures where the physical memory can be larger
3672 * than the addressable space (intel in 32-bit mode), we may
3673 * need to limit the cache to 1/8 of VM size.
3674 */
3675 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3676 #endif
3677
3678 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3679 arc_c_min = MAX(arc_c / 4, 64<<20);
3680 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
3681 if (arc_c * 8 >= 1<<30)
3682 arc_c_max = (arc_c * 8) - (1<<30);
3683 else
3684 arc_c_max = arc_c_min;
3685 arc_c_max = MAX(arc_c * 6, arc_c_max);
3686
3687 /*
3688 * Allow the tunables to override our calculations if they are
3689 * reasonable (ie. over 64MB)
3690 */
3691 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3692 arc_c_max = zfs_arc_max;
3693 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3694 arc_c_min = zfs_arc_min;
3695
3696 arc_c = arc_c_max;
3697 arc_p = (arc_c >> 1);
3698
3699 /* limit meta-data to 1/4 of the arc capacity */
3700 arc_meta_limit = arc_c_max / 4;
3701
3702 /* Allow the tunable to override if it is reasonable */
3703 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3704 arc_meta_limit = zfs_arc_meta_limit;
3705
3706 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3707 arc_c_min = arc_meta_limit / 2;
3708
3709 if (zfs_arc_grow_retry > 0)
3710 arc_grow_retry = zfs_arc_grow_retry;
3711
3712 if (zfs_arc_shrink_shift > 0)
3713 arc_shrink_shift = zfs_arc_shrink_shift;
3714
3715 if (zfs_arc_p_min_shift > 0)
3716 arc_p_min_shift = zfs_arc_p_min_shift;
3717
3718 /* if kmem_flags are set, lets try to use less memory */
3719 if (kmem_debugging())
3720 arc_c = arc_c / 2;
3721 if (arc_c < arc_c_min)
3722 arc_c = arc_c_min;
3723
3724 arc_anon = &ARC_anon;
3725 arc_mru = &ARC_mru;
3726 arc_mru_ghost = &ARC_mru_ghost;
3727 arc_mfu = &ARC_mfu;
3728 arc_mfu_ghost = &ARC_mfu_ghost;
3729 arc_l2c_only = &ARC_l2c_only;
3730 arc_size = 0;
3731
3732 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3733 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3734 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3735 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3736 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3737 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3738
3739 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3740 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3741 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3742 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3743 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3744 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3745 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3746 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3747 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3748 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3749 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3750 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3751 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3752 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3753 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3754 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3755 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3756 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3757 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3758 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3759
3760 buf_init();
3761
3762 arc_thread_exit = 0;
3763 arc_eviction_list = NULL;
3764 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3765 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3766
3767 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3768 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3769
3770 if (arc_ksp != NULL) {
3771 arc_ksp->ks_data = &arc_stats;
3772 kstat_install(arc_ksp);
3773 }
3774
3775 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3776 TS_RUN, minclsyspri);
3777
3778 arc_dead = FALSE;
3779 arc_warm = B_FALSE;
3780
3781 if (zfs_write_limit_max == 0)
3782 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3783 else
3784 zfs_write_limit_shift = 0;
3785 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3786 }
3787
3788 void
3789 arc_fini(void)
3790 {
3791 mutex_enter(&arc_reclaim_thr_lock);
3792 arc_thread_exit = 1;
3793 while (arc_thread_exit != 0)
3794 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3795 mutex_exit(&arc_reclaim_thr_lock);
3796
3797 arc_flush(NULL);
3798
3799 arc_dead = TRUE;
3800
3801 if (arc_ksp != NULL) {
3802 kstat_delete(arc_ksp);
3803 arc_ksp = NULL;
3804 }
3805
3806 mutex_destroy(&arc_eviction_mtx);
3807 mutex_destroy(&arc_reclaim_thr_lock);
3808 cv_destroy(&arc_reclaim_thr_cv);
3809
3810 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3811 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3812 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3813 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3814 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3815 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3816 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3817 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3818
3819 mutex_destroy(&arc_anon->arcs_mtx);
3820 mutex_destroy(&arc_mru->arcs_mtx);
3821 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3822 mutex_destroy(&arc_mfu->arcs_mtx);
3823 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3824 mutex_destroy(&arc_l2c_only->arcs_mtx);
3825
3826 mutex_destroy(&zfs_write_limit_lock);
3827
3828 buf_fini();
3829
3830 ASSERT(arc_loaned_bytes == 0);
3831 }
3832
3833 /*
3834 * Level 2 ARC
3835 *
3836 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3837 * It uses dedicated storage devices to hold cached data, which are populated
3838 * using large infrequent writes. The main role of this cache is to boost
3839 * the performance of random read workloads. The intended L2ARC devices
3840 * include short-stroked disks, solid state disks, and other media with
3841 * substantially faster read latency than disk.
3842 *
3843 * +-----------------------+
3844 * | ARC |
3845 * +-----------------------+
3846 * | ^ ^
3847 * | | |
3848 * l2arc_feed_thread() arc_read()
3849 * | | |
3850 * | l2arc read |
3851 * V | |
3852 * +---------------+ |
3853 * | L2ARC | |
3854 * +---------------+ |
3855 * | ^ |
3856 * l2arc_write() | |
3857 * | | |
3858 * V | |
3859 * +-------+ +-------+
3860 * | vdev | | vdev |
3861 * | cache | | cache |
3862 * +-------+ +-------+
3863 * +=========+ .-----.
3864 * : L2ARC : |-_____-|
3865 * : devices : | Disks |
3866 * +=========+ `-_____-'
3867 *
3868 * Read requests are satisfied from the following sources, in order:
3869 *
3870 * 1) ARC
3871 * 2) vdev cache of L2ARC devices
3872 * 3) L2ARC devices
3873 * 4) vdev cache of disks
3874 * 5) disks
3875 *
3876 * Some L2ARC device types exhibit extremely slow write performance.
3877 * To accommodate for this there are some significant differences between
3878 * the L2ARC and traditional cache design:
3879 *
3880 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3881 * the ARC behave as usual, freeing buffers and placing headers on ghost
3882 * lists. The ARC does not send buffers to the L2ARC during eviction as
3883 * this would add inflated write latencies for all ARC memory pressure.
3884 *
3885 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3886 * It does this by periodically scanning buffers from the eviction-end of
3887 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3888 * not already there. It scans until a headroom of buffers is satisfied,
3889 * which itself is a buffer for ARC eviction. If a compressible buffer is
3890 * found during scanning and selected for writing to an L2ARC device, we
3891 * temporarily boost scanning headroom during the next scan cycle to make
3892 * sure we adapt to compression effects (which might significantly reduce
3893 * the data volume we write to L2ARC). The thread that does this is
3894 * l2arc_feed_thread(), illustrated below; example sizes are included to
3895 * provide a better sense of ratio than this diagram:
3896 *
3897 * head --> tail
3898 * +---------------------+----------+
3899 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
3900 * +---------------------+----------+ | o L2ARC eligible
3901 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
3902 * +---------------------+----------+ |
3903 * 15.9 Gbytes ^ 32 Mbytes |
3904 * headroom |
3905 * l2arc_feed_thread()
3906 * |
3907 * l2arc write hand <--[oooo]--'
3908 * | 8 Mbyte
3909 * | write max
3910 * V
3911 * +==============================+
3912 * L2ARC dev |####|#|###|###| |####| ... |
3913 * +==============================+
3914 * 32 Gbytes
3915 *
3916 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
3917 * evicted, then the L2ARC has cached a buffer much sooner than it probably
3918 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
3919 * safe to say that this is an uncommon case, since buffers at the end of
3920 * the ARC lists have moved there due to inactivity.
3921 *
3922 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
3923 * then the L2ARC simply misses copying some buffers. This serves as a
3924 * pressure valve to prevent heavy read workloads from both stalling the ARC
3925 * with waits and clogging the L2ARC with writes. This also helps prevent
3926 * the potential for the L2ARC to churn if it attempts to cache content too
3927 * quickly, such as during backups of the entire pool.
3928 *
3929 * 5. After system boot and before the ARC has filled main memory, there are
3930 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
3931 * lists can remain mostly static. Instead of searching from tail of these
3932 * lists as pictured, the l2arc_feed_thread() will search from the list heads
3933 * for eligible buffers, greatly increasing its chance of finding them.
3934 *
3935 * The L2ARC device write speed is also boosted during this time so that
3936 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
3937 * there are no L2ARC reads, and no fear of degrading read performance
3938 * through increased writes.
3939 *
3940 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
3941 * the vdev queue can aggregate them into larger and fewer writes. Each
3942 * device is written to in a rotor fashion, sweeping writes through
3943 * available space then repeating.
3944 *
3945 * 7. The L2ARC does not store dirty content. It never needs to flush
3946 * write buffers back to disk based storage.
3947 *
3948 * 8. If an ARC buffer is written (and dirtied) which also exists in the
3949 * L2ARC, the now stale L2ARC buffer is immediately dropped.
3950 *
3951 * The performance of the L2ARC can be tweaked by a number of tunables, which
3952 * may be necessary for different workloads:
3953 *
3954 * l2arc_write_max max write bytes per interval
3955 * l2arc_write_boost extra write bytes during device warmup
3956 * l2arc_noprefetch skip caching prefetched buffers
3957 * l2arc_headroom number of max device writes to precache
3958 * l2arc_headroom_boost when we find compressed buffers during ARC
3959 * scanning, we multiply headroom by this
3960 * percentage factor for the next scan cycle,
3961 * since more compressed buffers are likely to
3962 * be present
3963 * l2arc_feed_secs seconds between L2ARC writing
3964 *
3965 * Tunables may be removed or added as future performance improvements are
3966 * integrated, and also may become zpool properties.
3967 *
3968 * There are three key functions that control how the L2ARC warms up:
3969 *
3970 * l2arc_write_eligible() check if a buffer is eligible to cache
3971 * l2arc_write_size() calculate how much to write
3972 * l2arc_write_interval() calculate sleep delay between writes
3973 *
3974 * These three functions determine what to write, how much, and how quickly
3975 * to send writes.
3976 */
3977
3978 static boolean_t
3979 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
3980 {
3981 /*
3982 * A buffer is *not* eligible for the L2ARC if it:
3983 * 1. belongs to a different spa.
3984 * 2. is already cached on the L2ARC.
3985 * 3. has an I/O in progress (it may be an incomplete read).
3986 * 4. is flagged not eligible (zfs property).
3987 */
3988 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
3989 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
3990 return (B_FALSE);
3991
3992 return (B_TRUE);
3993 }
3994
3995 static uint64_t
3996 l2arc_write_size(void)
3997 {
3998 uint64_t size;
3999
4000 /*
4001 * Make sure our globals have meaningful values in case the user
4002 * altered them.
4003 */
4004 size = l2arc_write_max;
4005 if (size == 0) {
4006 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
4007 "be greater than zero, resetting it to the default (%d)",
4008 L2ARC_WRITE_SIZE);
4009 size = l2arc_write_max = L2ARC_WRITE_SIZE;
4010 }
4011
4012 if (arc_warm == B_FALSE)
4013 size += l2arc_write_boost;
4014
4015 return (size);
4016
4017 }
4018
4019 static clock_t
4020 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4021 {
4022 clock_t interval, next, now;
4023
4024 /*
4025 * If the ARC lists are busy, increase our write rate; if the
4026 * lists are stale, idle back. This is achieved by checking
4027 * how much we previously wrote - if it was more than half of
4028 * what we wanted, schedule the next write much sooner.
4029 */
4030 if (l2arc_feed_again && wrote > (wanted / 2))
4031 interval = (hz * l2arc_feed_min_ms) / 1000;
4032 else
4033 interval = hz * l2arc_feed_secs;
4034
4035 now = ddi_get_lbolt();
4036 next = MAX(now, MIN(now + interval, began + interval));
4037
4038 return (next);
4039 }
4040
4041 static void
4042 l2arc_hdr_stat_add(void)
4043 {
4044 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4045 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4046 }
4047
4048 static void
4049 l2arc_hdr_stat_remove(void)
4050 {
4051 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4052 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4053 }
4054
4055 /*
4056 * Cycle through L2ARC devices. This is how L2ARC load balances.
4057 * If a device is returned, this also returns holding the spa config lock.
4058 */
4059 static l2arc_dev_t *
4060 l2arc_dev_get_next(void)
4061 {
4062 l2arc_dev_t *first, *next = NULL;
4063
4064 /*
4065 * Lock out the removal of spas (spa_namespace_lock), then removal
4066 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4067 * both locks will be dropped and a spa config lock held instead.
4068 */
4069 mutex_enter(&spa_namespace_lock);
4070 mutex_enter(&l2arc_dev_mtx);
4071
4072 /* if there are no vdevs, there is nothing to do */
4073 if (l2arc_ndev == 0)
4074 goto out;
4075
4076 first = NULL;
4077 next = l2arc_dev_last;
4078 do {
4079 /* loop around the list looking for a non-faulted vdev */
4080 if (next == NULL) {
4081 next = list_head(l2arc_dev_list);
4082 } else {
4083 next = list_next(l2arc_dev_list, next);
4084 if (next == NULL)
4085 next = list_head(l2arc_dev_list);
4086 }
4087
4088 /* if we have come back to the start, bail out */
4089 if (first == NULL)
4090 first = next;
4091 else if (next == first)
4092 break;
4093
4094 } while (vdev_is_dead(next->l2ad_vdev));
4095
4096 /* if we were unable to find any usable vdevs, return NULL */
4097 if (vdev_is_dead(next->l2ad_vdev))
4098 next = NULL;
4099
4100 l2arc_dev_last = next;
4101
4102 out:
4103 mutex_exit(&l2arc_dev_mtx);
4104
4105 /*
4106 * Grab the config lock to prevent the 'next' device from being
4107 * removed while we are writing to it.
4108 */
4109 if (next != NULL)
4110 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4111 mutex_exit(&spa_namespace_lock);
4112
4113 return (next);
4114 }
4115
4116 /*
4117 * Free buffers that were tagged for destruction.
4118 */
4119 static void
4120 l2arc_do_free_on_write()
4121 {
4122 list_t *buflist;
4123 l2arc_data_free_t *df, *df_prev;
4124
4125 mutex_enter(&l2arc_free_on_write_mtx);
4126 buflist = l2arc_free_on_write;
4127
4128 for (df = list_tail(buflist); df; df = df_prev) {
4129 df_prev = list_prev(buflist, df);
4130 ASSERT(df->l2df_data != NULL);
4131 ASSERT(df->l2df_func != NULL);
4132 df->l2df_func(df->l2df_data, df->l2df_size);
4133 list_remove(buflist, df);
4134 kmem_free(df, sizeof (l2arc_data_free_t));
4135 }
4136
4137 mutex_exit(&l2arc_free_on_write_mtx);
4138 }
4139
4140 /*
4141 * A write to a cache device has completed. Update all headers to allow
4142 * reads from these buffers to begin.
4143 */
4144 static void
4145 l2arc_write_done(zio_t *zio)
4146 {
4147 l2arc_write_callback_t *cb;
4148 l2arc_dev_t *dev;
4149 list_t *buflist;
4150 arc_buf_hdr_t *head, *ab, *ab_prev;
4151 l2arc_buf_hdr_t *abl2;
4152 kmutex_t *hash_lock;
4153
4154 cb = zio->io_private;
4155 ASSERT(cb != NULL);
4156 dev = cb->l2wcb_dev;
4157 ASSERT(dev != NULL);
4158 head = cb->l2wcb_head;
4159 ASSERT(head != NULL);
4160 buflist = dev->l2ad_buflist;
4161 ASSERT(buflist != NULL);
4162 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4163 l2arc_write_callback_t *, cb);
4164
4165 if (zio->io_error != 0)
4166 ARCSTAT_BUMP(arcstat_l2_writes_error);
4167
4168 mutex_enter(&l2arc_buflist_mtx);
4169
4170 /*
4171 * All writes completed, or an error was hit.
4172 */
4173 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4174 ab_prev = list_prev(buflist, ab);
4175
4176 hash_lock = HDR_LOCK(ab);
4177 if (!mutex_tryenter(hash_lock)) {
4178 /*
4179 * This buffer misses out. It may be in a stage
4180 * of eviction. Its ARC_L2_WRITING flag will be
4181 * left set, denying reads to this buffer.
4182 */
4183 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4184 continue;
4185 }
4186
4187 abl2 = ab->b_l2hdr;
4188
4189 /*
4190 * Release the temporary compressed buffer as soon as possible.
4191 */
4192 if (abl2->b_compress != ZIO_COMPRESS_OFF)
4193 l2arc_release_cdata_buf(ab);
4194
4195 if (zio->io_error != 0) {
4196 /*
4197 * Error - drop L2ARC entry.
4198 */
4199 list_remove(buflist, ab);
4200 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4201 ab->b_l2hdr = NULL;
4202 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4203 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4204 }
4205
4206 /*
4207 * Allow ARC to begin reads to this L2ARC entry.
4208 */
4209 ab->b_flags &= ~ARC_L2_WRITING;
4210
4211 mutex_exit(hash_lock);
4212 }
4213
4214 atomic_inc_64(&l2arc_writes_done);
4215 list_remove(buflist, head);
4216 kmem_cache_free(hdr_cache, head);
4217 mutex_exit(&l2arc_buflist_mtx);
4218
4219 l2arc_do_free_on_write();
4220
4221 kmem_free(cb, sizeof (l2arc_write_callback_t));
4222 }
4223
4224 /*
4225 * A read to a cache device completed. Validate buffer contents before
4226 * handing over to the regular ARC routines.
4227 */
4228 static void
4229 l2arc_read_done(zio_t *zio)
4230 {
4231 l2arc_read_callback_t *cb;
4232 arc_buf_hdr_t *hdr;
4233 arc_buf_t *buf;
4234 kmutex_t *hash_lock;
4235 int equal;
4236
4237 ASSERT(zio->io_vd != NULL);
4238 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4239
4240 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4241
4242 cb = zio->io_private;
4243 ASSERT(cb != NULL);
4244 buf = cb->l2rcb_buf;
4245 ASSERT(buf != NULL);
4246
4247 hash_lock = HDR_LOCK(buf->b_hdr);
4248 mutex_enter(hash_lock);
4249 hdr = buf->b_hdr;
4250 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4251
4252 /*
4253 * If the buffer was compressed, decompress it first.
4254 */
4255 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
4256 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
4257 ASSERT(zio->io_data != NULL);
4258
4259 /*
4260 * Check this survived the L2ARC journey.
4261 */
4262 equal = arc_cksum_equal(buf);
4263 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4264 mutex_exit(hash_lock);
4265 zio->io_private = buf;
4266 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4267 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4268 arc_read_done(zio);
4269 } else {
4270 mutex_exit(hash_lock);
4271 /*
4272 * Buffer didn't survive caching. Increment stats and
4273 * reissue to the original storage device.
4274 */
4275 if (zio->io_error != 0) {
4276 ARCSTAT_BUMP(arcstat_l2_io_error);
4277 } else {
4278 zio->io_error = SET_ERROR(EIO);
4279 }
4280 if (!equal)
4281 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4282
4283 /*
4284 * If there's no waiter, issue an async i/o to the primary
4285 * storage now. If there *is* a waiter, the caller must
4286 * issue the i/o in a context where it's OK to block.
4287 */
4288 if (zio->io_waiter == NULL) {
4289 zio_t *pio = zio_unique_parent(zio);
4290
4291 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4292
4293 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4294 buf->b_data, zio->io_size, arc_read_done, buf,
4295 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4296 }
4297 }
4298
4299 kmem_free(cb, sizeof (l2arc_read_callback_t));
4300 }
4301
4302 /*
4303 * This is the list priority from which the L2ARC will search for pages to
4304 * cache. This is used within loops (0..3) to cycle through lists in the
4305 * desired order. This order can have a significant effect on cache
4306 * performance.
4307 *
4308 * Currently the metadata lists are hit first, MFU then MRU, followed by
4309 * the data lists. This function returns a locked list, and also returns
4310 * the lock pointer.
4311 */
4312 static list_t *
4313 l2arc_list_locked(int list_num, kmutex_t **lock)
4314 {
4315 list_t *list = NULL;
4316
4317 ASSERT(list_num >= 0 && list_num <= 3);
4318
4319 switch (list_num) {
4320 case 0:
4321 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4322 *lock = &arc_mfu->arcs_mtx;
4323 break;
4324 case 1:
4325 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4326 *lock = &arc_mru->arcs_mtx;
4327 break;
4328 case 2:
4329 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4330 *lock = &arc_mfu->arcs_mtx;
4331 break;
4332 case 3:
4333 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4334 *lock = &arc_mru->arcs_mtx;
4335 break;
4336 }
4337
4338 ASSERT(!(MUTEX_HELD(*lock)));
4339 mutex_enter(*lock);
4340 return (list);
4341 }
4342
4343 /*
4344 * Evict buffers from the device write hand to the distance specified in
4345 * bytes. This distance may span populated buffers, it may span nothing.
4346 * This is clearing a region on the L2ARC device ready for writing.
4347 * If the 'all' boolean is set, every buffer is evicted.
4348 */
4349 static void
4350 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4351 {
4352 list_t *buflist;
4353 l2arc_buf_hdr_t *abl2;
4354 arc_buf_hdr_t *ab, *ab_prev;
4355 kmutex_t *hash_lock;
4356 uint64_t taddr;
4357
4358 buflist = dev->l2ad_buflist;
4359
4360 if (buflist == NULL)
4361 return;
4362
4363 if (!all && dev->l2ad_first) {
4364 /*
4365 * This is the first sweep through the device. There is
4366 * nothing to evict.
4367 */
4368 return;
4369 }
4370
4371 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4372 /*
4373 * When nearing the end of the device, evict to the end
4374 * before the device write hand jumps to the start.
4375 */
4376 taddr = dev->l2ad_end;
4377 } else {
4378 taddr = dev->l2ad_hand + distance;
4379 }
4380 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4381 uint64_t, taddr, boolean_t, all);
4382
4383 top:
4384 mutex_enter(&l2arc_buflist_mtx);
4385 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4386 ab_prev = list_prev(buflist, ab);
4387
4388 hash_lock = HDR_LOCK(ab);
4389 if (!mutex_tryenter(hash_lock)) {
4390 /*
4391 * Missed the hash lock. Retry.
4392 */
4393 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4394 mutex_exit(&l2arc_buflist_mtx);
4395 mutex_enter(hash_lock);
4396 mutex_exit(hash_lock);
4397 goto top;
4398 }
4399
4400 if (HDR_L2_WRITE_HEAD(ab)) {
4401 /*
4402 * We hit a write head node. Leave it for
4403 * l2arc_write_done().
4404 */
4405 list_remove(buflist, ab);
4406 mutex_exit(hash_lock);
4407 continue;
4408 }
4409
4410 if (!all && ab->b_l2hdr != NULL &&
4411 (ab->b_l2hdr->b_daddr > taddr ||
4412 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4413 /*
4414 * We've evicted to the target address,
4415 * or the end of the device.
4416 */
4417 mutex_exit(hash_lock);
4418 break;
4419 }
4420
4421 if (HDR_FREE_IN_PROGRESS(ab)) {
4422 /*
4423 * Already on the path to destruction.
4424 */
4425 mutex_exit(hash_lock);
4426 continue;
4427 }
4428
4429 if (ab->b_state == arc_l2c_only) {
4430 ASSERT(!HDR_L2_READING(ab));
4431 /*
4432 * This doesn't exist in the ARC. Destroy.
4433 * arc_hdr_destroy() will call list_remove()
4434 * and decrement arcstat_l2_size.
4435 */
4436 arc_change_state(arc_anon, ab, hash_lock);
4437 arc_hdr_destroy(ab);
4438 } else {
4439 /*
4440 * Invalidate issued or about to be issued
4441 * reads, since we may be about to write
4442 * over this location.
4443 */
4444 if (HDR_L2_READING(ab)) {
4445 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4446 ab->b_flags |= ARC_L2_EVICTED;
4447 }
4448
4449 /*
4450 * Tell ARC this no longer exists in L2ARC.
4451 */
4452 if (ab->b_l2hdr != NULL) {
4453 abl2 = ab->b_l2hdr;
4454 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize);
4455 ab->b_l2hdr = NULL;
4456 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4457 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4458 }
4459 list_remove(buflist, ab);
4460
4461 /*
4462 * This may have been leftover after a
4463 * failed write.
4464 */
4465 ab->b_flags &= ~ARC_L2_WRITING;
4466 }
4467 mutex_exit(hash_lock);
4468 }
4469 mutex_exit(&l2arc_buflist_mtx);
4470
4471 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4472 dev->l2ad_evict = taddr;
4473 }
4474
4475 /*
4476 * Find and write ARC buffers to the L2ARC device.
4477 *
4478 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4479 * for reading until they have completed writing.
4480 * The headroom_boost is an in-out parameter used to maintain headroom boost
4481 * state between calls to this function.
4482 *
4483 * Returns the number of bytes actually written (which may be smaller than
4484 * the delta by which the device hand has changed due to alignment).
4485 */
4486 static uint64_t
4487 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
4488 boolean_t *headroom_boost)
4489 {
4490 arc_buf_hdr_t *ab, *ab_prev, *head;
4491 list_t *list;
4492 uint64_t write_asize, write_psize, write_sz, headroom,
4493 buf_compress_minsz;
4494 void *buf_data;
4495 kmutex_t *list_lock;
4496 boolean_t full;
4497 l2arc_write_callback_t *cb;
4498 zio_t *pio, *wzio;
4499 uint64_t guid = spa_load_guid(spa);
4500 const boolean_t do_headroom_boost = *headroom_boost;
4501
4502 ASSERT(dev->l2ad_vdev != NULL);
4503
4504 /* Lower the flag now, we might want to raise it again later. */
4505 *headroom_boost = B_FALSE;
4506
4507 pio = NULL;
4508 write_sz = write_asize = write_psize = 0;
4509 full = B_FALSE;
4510 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4511 head->b_flags |= ARC_L2_WRITE_HEAD;
4512
4513 /*
4514 * We will want to try to compress buffers that are at least 2x the
4515 * device sector size.
4516 */
4517 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
4518
4519 /*
4520 * Copy buffers for L2ARC writing.
4521 */
4522 mutex_enter(&l2arc_buflist_mtx);
4523 for (int try = 0; try <= 3; try++) {
4524 uint64_t passed_sz = 0;
4525
4526 list = l2arc_list_locked(try, &list_lock);
4527
4528 /*
4529 * L2ARC fast warmup.
4530 *
4531 * Until the ARC is warm and starts to evict, read from the
4532 * head of the ARC lists rather than the tail.
4533 */
4534 if (arc_warm == B_FALSE)
4535 ab = list_head(list);
4536 else
4537 ab = list_tail(list);
4538
4539 headroom = target_sz * l2arc_headroom;
4540 if (do_headroom_boost)
4541 headroom = (headroom * l2arc_headroom_boost) / 100;
4542
4543 for (; ab; ab = ab_prev) {
4544 l2arc_buf_hdr_t *l2hdr;
4545 kmutex_t *hash_lock;
4546 uint64_t buf_sz;
4547
4548 if (arc_warm == B_FALSE)
4549 ab_prev = list_next(list, ab);
4550 else
4551 ab_prev = list_prev(list, ab);
4552
4553 hash_lock = HDR_LOCK(ab);
4554 if (!mutex_tryenter(hash_lock)) {
4555 /*
4556 * Skip this buffer rather than waiting.
4557 */
4558 continue;
4559 }
4560
4561 passed_sz += ab->b_size;
4562 if (passed_sz > headroom) {
4563 /*
4564 * Searched too far.
4565 */
4566 mutex_exit(hash_lock);
4567 break;
4568 }
4569
4570 if (!l2arc_write_eligible(guid, ab)) {
4571 mutex_exit(hash_lock);
4572 continue;
4573 }
4574
4575 if ((write_sz + ab->b_size) > target_sz) {
4576 full = B_TRUE;
4577 mutex_exit(hash_lock);
4578 break;
4579 }
4580
4581 if (pio == NULL) {
4582 /*
4583 * Insert a dummy header on the buflist so
4584 * l2arc_write_done() can find where the
4585 * write buffers begin without searching.
4586 */
4587 list_insert_head(dev->l2ad_buflist, head);
4588
4589 cb = kmem_alloc(
4590 sizeof (l2arc_write_callback_t), KM_SLEEP);
4591 cb->l2wcb_dev = dev;
4592 cb->l2wcb_head = head;
4593 pio = zio_root(spa, l2arc_write_done, cb,
4594 ZIO_FLAG_CANFAIL);
4595 }
4596
4597 /*
4598 * Create and add a new L2ARC header.
4599 */
4600 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4601 l2hdr->b_dev = dev;
4602 ab->b_flags |= ARC_L2_WRITING;
4603
4604 /*
4605 * Temporarily stash the data buffer in b_tmp_cdata.
4606 * The subsequent write step will pick it up from
4607 * there. This is because can't access ab->b_buf
4608 * without holding the hash_lock, which we in turn
4609 * can't access without holding the ARC list locks
4610 * (which we want to avoid during compression/writing).
4611 */
4612 l2hdr->b_compress = ZIO_COMPRESS_OFF;
4613 l2hdr->b_asize = ab->b_size;
4614 l2hdr->b_tmp_cdata = ab->b_buf->b_data;
4615
4616 buf_sz = ab->b_size;
4617 ab->b_l2hdr = l2hdr;
4618
4619 list_insert_head(dev->l2ad_buflist, ab);
4620
4621 /*
4622 * Compute and store the buffer cksum before
4623 * writing. On debug the cksum is verified first.
4624 */
4625 arc_cksum_verify(ab->b_buf);
4626 arc_cksum_compute(ab->b_buf, B_TRUE);
4627
4628 mutex_exit(hash_lock);
4629
4630 write_sz += buf_sz;
4631 }
4632
4633 mutex_exit(list_lock);
4634
4635 if (full == B_TRUE)
4636 break;
4637 }
4638
4639 /* No buffers selected for writing? */
4640 if (pio == NULL) {
4641 ASSERT0(write_sz);
4642 mutex_exit(&l2arc_buflist_mtx);
4643 kmem_cache_free(hdr_cache, head);
4644 return (0);
4645 }
4646
4647 /*
4648 * Now start writing the buffers. We're starting at the write head
4649 * and work backwards, retracing the course of the buffer selector
4650 * loop above.
4651 */
4652 for (ab = list_prev(dev->l2ad_buflist, head); ab;
4653 ab = list_prev(dev->l2ad_buflist, ab)) {
4654 l2arc_buf_hdr_t *l2hdr;
4655 uint64_t buf_sz;
4656
4657 /*
4658 * We shouldn't need to lock the buffer here, since we flagged
4659 * it as ARC_L2_WRITING in the previous step, but we must take
4660 * care to only access its L2 cache parameters. In particular,
4661 * ab->b_buf may be invalid by now due to ARC eviction.
4662 */
4663 l2hdr = ab->b_l2hdr;
4664 l2hdr->b_daddr = dev->l2ad_hand;
4665
4666 if ((ab->b_flags & ARC_L2COMPRESS) &&
4667 l2hdr->b_asize >= buf_compress_minsz) {
4668 if (l2arc_compress_buf(l2hdr)) {
4669 /*
4670 * If compression succeeded, enable headroom
4671 * boost on the next scan cycle.
4672 */
4673 *headroom_boost = B_TRUE;
4674 }
4675 }
4676
4677 /*
4678 * Pick up the buffer data we had previously stashed away
4679 * (and now potentially also compressed).
4680 */
4681 buf_data = l2hdr->b_tmp_cdata;
4682 buf_sz = l2hdr->b_asize;
4683
4684 /* Compression may have squashed the buffer to zero length. */
4685 if (buf_sz != 0) {
4686 uint64_t buf_p_sz;
4687
4688 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4689 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4690 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4691 ZIO_FLAG_CANFAIL, B_FALSE);
4692
4693 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4694 zio_t *, wzio);
4695 (void) zio_nowait(wzio);
4696
4697 write_asize += buf_sz;
4698 /*
4699 * Keep the clock hand suitably device-aligned.
4700 */
4701 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4702 write_psize += buf_p_sz;
4703 dev->l2ad_hand += buf_p_sz;
4704 }
4705 }
4706
4707 mutex_exit(&l2arc_buflist_mtx);
4708
4709 ASSERT3U(write_asize, <=, target_sz);
4710 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4711 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
4712 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4713 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
4714 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
4715
4716 /*
4717 * Bump device hand to the device start if it is approaching the end.
4718 * l2arc_evict() will already have evicted ahead for this case.
4719 */
4720 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4721 vdev_space_update(dev->l2ad_vdev,
4722 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4723 dev->l2ad_hand = dev->l2ad_start;
4724 dev->l2ad_evict = dev->l2ad_start;
4725 dev->l2ad_first = B_FALSE;
4726 }
4727
4728 dev->l2ad_writing = B_TRUE;
4729 (void) zio_wait(pio);
4730 dev->l2ad_writing = B_FALSE;
4731
4732 return (write_asize);
4733 }
4734
4735 /*
4736 * Compresses an L2ARC buffer.
4737 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its
4738 * size in l2hdr->b_asize. This routine tries to compress the data and
4739 * depending on the compression result there are three possible outcomes:
4740 * *) The buffer was incompressible. The original l2hdr contents were left
4741 * untouched and are ready for writing to an L2 device.
4742 * *) The buffer was all-zeros, so there is no need to write it to an L2
4743 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
4744 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
4745 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
4746 * data buffer which holds the compressed data to be written, and b_asize
4747 * tells us how much data there is. b_compress is set to the appropriate
4748 * compression algorithm. Once writing is done, invoke
4749 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
4750 *
4751 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
4752 * buffer was incompressible).
4753 */
4754 static boolean_t
4755 l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr)
4756 {
4757 void *cdata;
4758 size_t csize, len;
4759
4760 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF);
4761 ASSERT(l2hdr->b_tmp_cdata != NULL);
4762
4763 len = l2hdr->b_asize;
4764 cdata = zio_data_buf_alloc(len);
4765 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata,
4766 cdata, l2hdr->b_asize);
4767
4768 if (csize == 0) {
4769 /* zero block, indicate that there's nothing to write */
4770 zio_data_buf_free(cdata, len);
4771 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
4772 l2hdr->b_asize = 0;
4773 l2hdr->b_tmp_cdata = NULL;
4774 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
4775 return (B_TRUE);
4776 } else if (csize > 0 && csize < len) {
4777 /*
4778 * Compression succeeded, we'll keep the cdata around for
4779 * writing and release it afterwards.
4780 */
4781 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
4782 l2hdr->b_asize = csize;
4783 l2hdr->b_tmp_cdata = cdata;
4784 ARCSTAT_BUMP(arcstat_l2_compress_successes);
4785 return (B_TRUE);
4786 } else {
4787 /*
4788 * Compression failed, release the compressed buffer.
4789 * l2hdr will be left unmodified.
4790 */
4791 zio_data_buf_free(cdata, len);
4792 ARCSTAT_BUMP(arcstat_l2_compress_failures);
4793 return (B_FALSE);
4794 }
4795 }
4796
4797 /*
4798 * Decompresses a zio read back from an l2arc device. On success, the
4799 * underlying zio's io_data buffer is overwritten by the uncompressed
4800 * version. On decompression error (corrupt compressed stream), the
4801 * zio->io_error value is set to signal an I/O error.
4802 *
4803 * Please note that the compressed data stream is not checksummed, so
4804 * if the underlying device is experiencing data corruption, we may feed
4805 * corrupt data to the decompressor, so the decompressor needs to be
4806 * able to handle this situation (LZ4 does).
4807 */
4808 static void
4809 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
4810 {
4811 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
4812
4813 if (zio->io_error != 0) {
4814 /*
4815 * An io error has occured, just restore the original io
4816 * size in preparation for a main pool read.
4817 */
4818 zio->io_orig_size = zio->io_size = hdr->b_size;
4819 return;
4820 }
4821
4822 if (c == ZIO_COMPRESS_EMPTY) {
4823 /*
4824 * An empty buffer results in a null zio, which means we
4825 * need to fill its io_data after we're done restoring the
4826 * buffer's contents.
4827 */
4828 ASSERT(hdr->b_buf != NULL);
4829 bzero(hdr->b_buf->b_data, hdr->b_size);
4830 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data;
4831 } else {
4832 ASSERT(zio->io_data != NULL);
4833 /*
4834 * We copy the compressed data from the start of the arc buffer
4835 * (the zio_read will have pulled in only what we need, the
4836 * rest is garbage which we will overwrite at decompression)
4837 * and then decompress back to the ARC data buffer. This way we
4838 * can minimize copying by simply decompressing back over the
4839 * original compressed data (rather than decompressing to an
4840 * aux buffer and then copying back the uncompressed buffer,
4841 * which is likely to be much larger).
4842 */
4843 uint64_t csize;
4844 void *cdata;
4845
4846 csize = zio->io_size;
4847 cdata = zio_data_buf_alloc(csize);
4848 bcopy(zio->io_data, cdata, csize);
4849 if (zio_decompress_data(c, cdata, zio->io_data, csize,
4850 hdr->b_size) != 0)
4851 zio->io_error = EIO;
4852 zio_data_buf_free(cdata, csize);
4853 }
4854
4855 /* Restore the expected uncompressed IO size. */
4856 zio->io_orig_size = zio->io_size = hdr->b_size;
4857 }
4858
4859 /*
4860 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
4861 * This buffer serves as a temporary holder of compressed data while
4862 * the buffer entry is being written to an l2arc device. Once that is
4863 * done, we can dispose of it.
4864 */
4865 static void
4866 l2arc_release_cdata_buf(arc_buf_hdr_t *ab)
4867 {
4868 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr;
4869
4870 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) {
4871 /*
4872 * If the data was compressed, then we've allocated a
4873 * temporary buffer for it, so now we need to release it.
4874 */
4875 ASSERT(l2hdr->b_tmp_cdata != NULL);
4876 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size);
4877 }
4878 l2hdr->b_tmp_cdata = NULL;
4879 }
4880
4881 /*
4882 * This thread feeds the L2ARC at regular intervals. This is the beating
4883 * heart of the L2ARC.
4884 */
4885 static void
4886 l2arc_feed_thread(void)
4887 {
4888 callb_cpr_t cpr;
4889 l2arc_dev_t *dev;
4890 spa_t *spa;
4891 uint64_t size, wrote;
4892 clock_t begin, next = ddi_get_lbolt();
4893 boolean_t headroom_boost = B_FALSE;
4894
4895 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4896
4897 mutex_enter(&l2arc_feed_thr_lock);
4898
4899 while (l2arc_thread_exit == 0) {
4900 CALLB_CPR_SAFE_BEGIN(&cpr);
4901 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4902 next);
4903 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4904 next = ddi_get_lbolt() + hz;
4905
4906 /*
4907 * Quick check for L2ARC devices.
4908 */
4909 mutex_enter(&l2arc_dev_mtx);
4910 if (l2arc_ndev == 0) {
4911 mutex_exit(&l2arc_dev_mtx);
4912 continue;
4913 }
4914 mutex_exit(&l2arc_dev_mtx);
4915 begin = ddi_get_lbolt();
4916
4917 /*
4918 * This selects the next l2arc device to write to, and in
4919 * doing so the next spa to feed from: dev->l2ad_spa. This
4920 * will return NULL if there are now no l2arc devices or if
4921 * they are all faulted.
4922 *
4923 * If a device is returned, its spa's config lock is also
4924 * held to prevent device removal. l2arc_dev_get_next()
4925 * will grab and release l2arc_dev_mtx.
4926 */
4927 if ((dev = l2arc_dev_get_next()) == NULL)
4928 continue;
4929
4930 spa = dev->l2ad_spa;
4931 ASSERT(spa != NULL);
4932
4933 /*
4934 * If the pool is read-only then force the feed thread to
4935 * sleep a little longer.
4936 */
4937 if (!spa_writeable(spa)) {
4938 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4939 spa_config_exit(spa, SCL_L2ARC, dev);
4940 continue;
4941 }
4942
4943 /*
4944 * Avoid contributing to memory pressure.
4945 */
4946 if (arc_reclaim_needed()) {
4947 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4948 spa_config_exit(spa, SCL_L2ARC, dev);
4949 continue;
4950 }
4951
4952 ARCSTAT_BUMP(arcstat_l2_feeds);
4953
4954 size = l2arc_write_size();
4955
4956 /*
4957 * Evict L2ARC buffers that will be overwritten.
4958 */
4959 l2arc_evict(dev, size, B_FALSE);
4960
4961 /*
4962 * Write ARC buffers.
4963 */
4964 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
4965
4966 /*
4967 * Calculate interval between writes.
4968 */
4969 next = l2arc_write_interval(begin, size, wrote);
4970 spa_config_exit(spa, SCL_L2ARC, dev);
4971 }
4972
4973 l2arc_thread_exit = 0;
4974 cv_broadcast(&l2arc_feed_thr_cv);
4975 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4976 thread_exit();
4977 }
4978
4979 boolean_t
4980 l2arc_vdev_present(vdev_t *vd)
4981 {
4982 l2arc_dev_t *dev;
4983
4984 mutex_enter(&l2arc_dev_mtx);
4985 for (dev = list_head(l2arc_dev_list); dev != NULL;
4986 dev = list_next(l2arc_dev_list, dev)) {
4987 if (dev->l2ad_vdev == vd)
4988 break;
4989 }
4990 mutex_exit(&l2arc_dev_mtx);
4991
4992 return (dev != NULL);
4993 }
4994
4995 /*
4996 * Add a vdev for use by the L2ARC. By this point the spa has already
4997 * validated the vdev and opened it.
4998 */
4999 void
5000 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
5001 {
5002 l2arc_dev_t *adddev;
5003
5004 ASSERT(!l2arc_vdev_present(vd));
5005
5006 /*
5007 * Create a new l2arc device entry.
5008 */
5009 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
5010 adddev->l2ad_spa = spa;
5011 adddev->l2ad_vdev = vd;
5012 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
5013 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
5014 adddev->l2ad_hand = adddev->l2ad_start;
5015 adddev->l2ad_evict = adddev->l2ad_start;
5016 adddev->l2ad_first = B_TRUE;
5017 adddev->l2ad_writing = B_FALSE;
5018
5019 /*
5020 * This is a list of all ARC buffers that are still valid on the
5021 * device.
5022 */
5023 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
5024 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
5025 offsetof(arc_buf_hdr_t, b_l2node));
5026
5027 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
5028
5029 /*
5030 * Add device to global list
5031 */
5032 mutex_enter(&l2arc_dev_mtx);
5033 list_insert_head(l2arc_dev_list, adddev);
5034 atomic_inc_64(&l2arc_ndev);
5035 mutex_exit(&l2arc_dev_mtx);
5036 }
5037
5038 /*
5039 * Remove a vdev from the L2ARC.
5040 */
5041 void
5042 l2arc_remove_vdev(vdev_t *vd)
5043 {
5044 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
5045
5046 /*
5047 * Find the device by vdev
5048 */
5049 mutex_enter(&l2arc_dev_mtx);
5050 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
5051 nextdev = list_next(l2arc_dev_list, dev);
5052 if (vd == dev->l2ad_vdev) {
5053 remdev = dev;
5054 break;
5055 }
5056 }
5057 ASSERT(remdev != NULL);
5058
5059 /*
5060 * Remove device from global list
5061 */
5062 list_remove(l2arc_dev_list, remdev);
5063 l2arc_dev_last = NULL; /* may have been invalidated */
5064 atomic_dec_64(&l2arc_ndev);
5065 mutex_exit(&l2arc_dev_mtx);
5066
5067 /*
5068 * Clear all buflists and ARC references. L2ARC device flush.
5069 */
5070 l2arc_evict(remdev, 0, B_TRUE);
5071 list_destroy(remdev->l2ad_buflist);
5072 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
5073 kmem_free(remdev, sizeof (l2arc_dev_t));
5074 }
5075
5076 void
5077 l2arc_init(void)
5078 {
5079 l2arc_thread_exit = 0;
5080 l2arc_ndev = 0;
5081 l2arc_writes_sent = 0;
5082 l2arc_writes_done = 0;
5083
5084 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
5085 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
5086 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
5087 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
5088 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
5089
5090 l2arc_dev_list = &L2ARC_dev_list;
5091 l2arc_free_on_write = &L2ARC_free_on_write;
5092 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
5093 offsetof(l2arc_dev_t, l2ad_node));
5094 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
5095 offsetof(l2arc_data_free_t, l2df_list_node));
5096 }
5097
5098 void
5099 l2arc_fini(void)
5100 {
5101 /*
5102 * This is called from dmu_fini(), which is called from spa_fini();
5103 * Because of this, we can assume that all l2arc devices have
5104 * already been removed when the pools themselves were removed.
5105 */
5106
5107 l2arc_do_free_on_write();
5108
5109 mutex_destroy(&l2arc_feed_thr_lock);
5110 cv_destroy(&l2arc_feed_thr_cv);
5111 mutex_destroy(&l2arc_dev_mtx);
5112 mutex_destroy(&l2arc_buflist_mtx);
5113 mutex_destroy(&l2arc_free_on_write_mtx);
5114
5115 list_destroy(l2arc_dev_list);
5116 list_destroy(l2arc_free_on_write);
5117 }
5118
5119 void
5120 l2arc_start(void)
5121 {
5122 if (!(spa_mode_global & FWRITE))
5123 return;
5124
5125 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5126 TS_RUN, minclsyspri);
5127 }
5128
5129 void
5130 l2arc_stop(void)
5131 {
5132 if (!(spa_mode_global & FWRITE))
5133 return;
5134
5135 mutex_enter(&l2arc_feed_thr_lock);
5136 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5137 l2arc_thread_exit = 1;
5138 while (l2arc_thread_exit != 0)
5139 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5140 mutex_exit(&l2arc_feed_thr_lock);
5141 }