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) 2012 by Delphix. All rights reserved.
25 */
26
27 /*
28 * DVA-based Adjustable Replacement Cache
29 *
30 * While much of the theory of operation used here is
31 * based on the self-tuning, low overhead replacement cache
32 * presented by Megiddo and Modha at FAST 2003, there are some
33 * significant differences:
34 *
35 * 1. The Megiddo and Modha model assumes any page is evictable.
36 * Pages in its cache cannot be "locked" into memory. This makes
37 * the eviction algorithm simple: evict the last page in the list.
38 * This also make the performance characteristics easy to reason
39 * about. Our cache is not so simple. At any given moment, some
40 * subset of the blocks in the cache are un-evictable because we
41 * have handed out a reference to them. Blocks are only evictable
42 * when there are no external references active. This makes
43 * eviction far more problematic: we choose to evict the evictable
44 * blocks that are the "lowest" in the list.
45 *
46 * There are times when it is not possible to evict the requested
47 * space. In these circumstances we are unable to adjust the cache
48 * size. To prevent the cache growing unbounded at these times we
49 * implement a "cache throttle" that slows the flow of new data
50 * into the cache until we can make space available.
51 *
52 * 2. The Megiddo and Modha model assumes a fixed cache size.
53 * Pages are evicted when the cache is full and there is a cache
54 * miss. Our model has a variable sized cache. It grows with
55 * high use, but also tries to react to memory pressure from the
56 * operating system: decreasing its size when system memory is
57 * tight.
58 *
59 * 3. The Megiddo and Modha model assumes a fixed page size. All
60 * elements of the cache are therefor exactly the same size. So
61 * when adjusting the cache size following a cache miss, its simply
62 * a matter of choosing a single page to evict. In our model, we
63 * have variable sized cache blocks (rangeing from 512 bytes to
64 * 128K bytes). We therefor choose a set of blocks to evict to make
65 * space for a cache miss that approximates as closely as possible
66 * the space used by the new block.
67 *
68 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
69 * by N. Megiddo & D. Modha, FAST 2003
70 */
71
72 /*
73 * The locking model:
74 *
75 * A new reference to a cache buffer can be obtained in two
76 * ways: 1) via a hash table lookup using the DVA as a key,
77 * or 2) via one of the ARC lists. The arc_read() interface
78 * uses method 1, while the internal arc algorithms for
79 * adjusting the cache use method 2. We therefor provide two
80 * types of locks: 1) the hash table lock array, and 2) the
81 * arc list locks.
82 *
83 * Buffers do not have their own mutexs, rather they rely on the
84 * hash table mutexs for the bulk of their protection (i.e. most
85 * fields in the arc_buf_hdr_t are protected by these mutexs).
86 *
87 * buf_hash_find() returns the appropriate mutex (held) when it
88 * locates the requested buffer in the hash table. It returns
89 * NULL for the mutex if the buffer was not in the table.
90 *
91 * buf_hash_remove() expects the appropriate hash mutex to be
92 * already held before it is invoked.
93 *
94 * Each arc state also has a mutex which is used to protect the
95 * buffer list associated with the state. When attempting to
96 * obtain a hash table lock while holding an arc list lock you
97 * must use: mutex_tryenter() to avoid deadlock. Also note that
98 * the active state mutex must be held before the ghost state mutex.
99 *
100 * Arc buffers may have an associated eviction callback function.
101 * This function will be invoked prior to removing the buffer (e.g.
102 * in arc_do_user_evicts()). Note however that the data associated
103 * with the buffer may be evicted prior to the callback. The callback
104 * must be made with *no locks held* (to prevent deadlock). Additionally,
105 * the users of callbacks must ensure that their private data is
106 * protected from simultaneous callbacks from arc_buf_evict()
107 * and arc_do_user_evicts().
108 *
109 * Note that the majority of the performance stats are manipulated
110 * with atomic operations.
111 *
112 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
113 *
114 * - L2ARC buflist creation
115 * - L2ARC buflist eviction
116 * - L2ARC write completion, which walks L2ARC buflists
117 * - ARC header destruction, as it removes from L2ARC buflists
118 * - ARC header release, as it removes from L2ARC buflists
119 */
120
121 #include <sys/spa.h>
122 #include <sys/zio.h>
123 #include <sys/zfs_context.h>
124 #include <sys/arc.h>
125 #include <sys/refcount.h>
126 #include <sys/vdev.h>
127 #include <sys/vdev_impl.h>
128 #ifdef _KERNEL
129 #include <sys/vmsystm.h>
130 #include <vm/anon.h>
131 #include <sys/fs/swapnode.h>
132 #include <sys/dnlc.h>
133 #endif
134 #include <sys/callb.h>
135 #include <sys/kstat.h>
136 #include <zfs_fletcher.h>
137
138 static kmutex_t arc_reclaim_thr_lock;
139 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
140 static uint8_t arc_thread_exit;
141
142 extern int zfs_write_limit_shift;
143 extern uint64_t zfs_write_limit_max;
144 extern kmutex_t zfs_write_limit_lock;
145
146 #define ARC_REDUCE_DNLC_PERCENT 3
147 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
148
149 typedef enum arc_reclaim_strategy {
150 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
151 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
152 } arc_reclaim_strategy_t;
153
154 /* number of seconds before growing cache again */
155 static int arc_grow_retry = 60;
156
157 /* shift of arc_c for calculating both min and max arc_p */
158 static int arc_p_min_shift = 4;
159
160 /* log2(fraction of arc to reclaim) */
161 static int arc_shrink_shift = 5;
162
163 /*
164 * minimum lifespan of a prefetch block in clock ticks
165 * (initialized in arc_init())
166 */
167 static int arc_min_prefetch_lifespan;
168
169 static int arc_dead;
170
171 /*
172 * The arc has filled available memory and has now warmed up.
173 */
174 static boolean_t arc_warm;
175
176 /*
177 * These tunables are for performance analysis.
178 */
179 uint64_t zfs_arc_max;
180 uint64_t zfs_arc_min;
181 uint64_t zfs_arc_meta_limit = 0;
182 int zfs_arc_grow_retry = 0;
183 int zfs_arc_shrink_shift = 0;
184 int zfs_arc_p_min_shift = 0;
185
186 /*
187 * Note that buffers can be in one of 6 states:
188 * ARC_anon - anonymous (discussed below)
189 * ARC_mru - recently used, currently cached
190 * ARC_mru_ghost - recentely used, no longer in cache
191 * ARC_mfu - frequently used, currently cached
192 * ARC_mfu_ghost - frequently used, no longer in cache
193 * ARC_l2c_only - exists in L2ARC but not other states
194 * When there are no active references to the buffer, they are
195 * are linked onto a list in one of these arc states. These are
196 * the only buffers that can be evicted or deleted. Within each
197 * state there are multiple lists, one for meta-data and one for
198 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
199 * etc.) is tracked separately so that it can be managed more
200 * explicitly: favored over data, limited explicitly.
201 *
202 * Anonymous buffers are buffers that are not associated with
203 * a DVA. These are buffers that hold dirty block copies
204 * before they are written to stable storage. By definition,
205 * they are "ref'd" and are considered part of arc_mru
206 * that cannot be freed. Generally, they will aquire a DVA
207 * as they are written and migrate onto the arc_mru list.
208 *
209 * The ARC_l2c_only state is for buffers that are in the second
210 * level ARC but no longer in any of the ARC_m* lists. The second
211 * level ARC itself may also contain buffers that are in any of
212 * the ARC_m* states - meaning that a buffer can exist in two
213 * places. The reason for the ARC_l2c_only state is to keep the
214 * buffer header in the hash table, so that reads that hit the
215 * second level ARC benefit from these fast lookups.
216 */
217
218 typedef struct arc_state {
219 list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */
220 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
221 uint64_t arcs_size; /* total amount of data in this state */
222 kmutex_t arcs_mtx;
223 } arc_state_t;
224
225 /* The 6 states: */
226 static arc_state_t ARC_anon;
227 static arc_state_t ARC_mru;
228 static arc_state_t ARC_mru_ghost;
229 static arc_state_t ARC_mfu;
230 static arc_state_t ARC_mfu_ghost;
231 static arc_state_t ARC_l2c_only;
232
233 typedef struct arc_stats {
234 kstat_named_t arcstat_hits;
235 kstat_named_t arcstat_misses;
236 kstat_named_t arcstat_demand_data_hits;
237 kstat_named_t arcstat_demand_data_misses;
238 kstat_named_t arcstat_demand_metadata_hits;
239 kstat_named_t arcstat_demand_metadata_misses;
240 kstat_named_t arcstat_prefetch_data_hits;
241 kstat_named_t arcstat_prefetch_data_misses;
242 kstat_named_t arcstat_prefetch_metadata_hits;
243 kstat_named_t arcstat_prefetch_metadata_misses;
244 kstat_named_t arcstat_mru_hits;
245 kstat_named_t arcstat_mru_ghost_hits;
246 kstat_named_t arcstat_mfu_hits;
247 kstat_named_t arcstat_mfu_ghost_hits;
248 kstat_named_t arcstat_deleted;
249 kstat_named_t arcstat_recycle_miss;
250 kstat_named_t arcstat_mutex_miss;
251 kstat_named_t arcstat_evict_skip;
252 kstat_named_t arcstat_evict_l2_cached;
253 kstat_named_t arcstat_evict_l2_eligible;
254 kstat_named_t arcstat_evict_l2_ineligible;
255 kstat_named_t arcstat_hash_elements;
256 kstat_named_t arcstat_hash_elements_max;
257 kstat_named_t arcstat_hash_collisions;
258 kstat_named_t arcstat_hash_chains;
259 kstat_named_t arcstat_hash_chain_max;
260 kstat_named_t arcstat_p;
261 kstat_named_t arcstat_c;
262 kstat_named_t arcstat_c_min;
263 kstat_named_t arcstat_c_max;
264 kstat_named_t arcstat_size;
265 kstat_named_t arcstat_hdr_size;
266 kstat_named_t arcstat_data_size;
267 kstat_named_t arcstat_other_size;
268 kstat_named_t arcstat_l2_hits;
269 kstat_named_t arcstat_l2_misses;
270 kstat_named_t arcstat_l2_feeds;
271 kstat_named_t arcstat_l2_rw_clash;
272 kstat_named_t arcstat_l2_read_bytes;
273 kstat_named_t arcstat_l2_write_bytes;
274 kstat_named_t arcstat_l2_writes_sent;
275 kstat_named_t arcstat_l2_writes_done;
276 kstat_named_t arcstat_l2_writes_error;
277 kstat_named_t arcstat_l2_writes_hdr_miss;
278 kstat_named_t arcstat_l2_evict_lock_retry;
279 kstat_named_t arcstat_l2_evict_reading;
280 kstat_named_t arcstat_l2_free_on_write;
281 kstat_named_t arcstat_l2_abort_lowmem;
282 kstat_named_t arcstat_l2_cksum_bad;
283 kstat_named_t arcstat_l2_io_error;
284 kstat_named_t arcstat_l2_size;
285 kstat_named_t arcstat_l2_hdr_size;
286 kstat_named_t arcstat_memory_throttle_count;
287 } arc_stats_t;
288
289 static arc_stats_t arc_stats = {
290 { "hits", KSTAT_DATA_UINT64 },
291 { "misses", KSTAT_DATA_UINT64 },
292 { "demand_data_hits", KSTAT_DATA_UINT64 },
293 { "demand_data_misses", KSTAT_DATA_UINT64 },
294 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
295 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
296 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
297 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
298 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
299 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
300 { "mru_hits", KSTAT_DATA_UINT64 },
301 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
302 { "mfu_hits", KSTAT_DATA_UINT64 },
303 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
304 { "deleted", KSTAT_DATA_UINT64 },
305 { "recycle_miss", KSTAT_DATA_UINT64 },
306 { "mutex_miss", KSTAT_DATA_UINT64 },
307 { "evict_skip", KSTAT_DATA_UINT64 },
308 { "evict_l2_cached", KSTAT_DATA_UINT64 },
309 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
310 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
311 { "hash_elements", KSTAT_DATA_UINT64 },
312 { "hash_elements_max", KSTAT_DATA_UINT64 },
313 { "hash_collisions", KSTAT_DATA_UINT64 },
314 { "hash_chains", KSTAT_DATA_UINT64 },
315 { "hash_chain_max", KSTAT_DATA_UINT64 },
316 { "p", KSTAT_DATA_UINT64 },
317 { "c", KSTAT_DATA_UINT64 },
318 { "c_min", KSTAT_DATA_UINT64 },
319 { "c_max", KSTAT_DATA_UINT64 },
320 { "size", KSTAT_DATA_UINT64 },
321 { "hdr_size", KSTAT_DATA_UINT64 },
322 { "data_size", KSTAT_DATA_UINT64 },
323 { "other_size", KSTAT_DATA_UINT64 },
324 { "l2_hits", KSTAT_DATA_UINT64 },
325 { "l2_misses", KSTAT_DATA_UINT64 },
326 { "l2_feeds", KSTAT_DATA_UINT64 },
327 { "l2_rw_clash", KSTAT_DATA_UINT64 },
328 { "l2_read_bytes", KSTAT_DATA_UINT64 },
329 { "l2_write_bytes", KSTAT_DATA_UINT64 },
330 { "l2_writes_sent", KSTAT_DATA_UINT64 },
331 { "l2_writes_done", KSTAT_DATA_UINT64 },
332 { "l2_writes_error", KSTAT_DATA_UINT64 },
333 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
334 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
335 { "l2_evict_reading", KSTAT_DATA_UINT64 },
336 { "l2_free_on_write", KSTAT_DATA_UINT64 },
337 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
338 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
339 { "l2_io_error", KSTAT_DATA_UINT64 },
340 { "l2_size", KSTAT_DATA_UINT64 },
341 { "l2_hdr_size", KSTAT_DATA_UINT64 },
342 { "memory_throttle_count", KSTAT_DATA_UINT64 }
343 };
344
345 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
346
347 #define ARCSTAT_INCR(stat, val) \
348 atomic_add_64(&arc_stats.stat.value.ui64, (val));
349
350 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
351 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
352
353 #define ARCSTAT_MAX(stat, val) { \
354 uint64_t m; \
355 while ((val) > (m = arc_stats.stat.value.ui64) && \
356 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
357 continue; \
358 }
359
360 #define ARCSTAT_MAXSTAT(stat) \
361 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
362
363 /*
364 * We define a macro to allow ARC hits/misses to be easily broken down by
365 * two separate conditions, giving a total of four different subtypes for
366 * each of hits and misses (so eight statistics total).
367 */
368 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
369 if (cond1) { \
370 if (cond2) { \
371 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
372 } else { \
373 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
374 } \
375 } else { \
376 if (cond2) { \
377 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
378 } else { \
379 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
380 } \
381 }
382
383 kstat_t *arc_ksp;
384 static arc_state_t *arc_anon;
385 static arc_state_t *arc_mru;
386 static arc_state_t *arc_mru_ghost;
387 static arc_state_t *arc_mfu;
388 static arc_state_t *arc_mfu_ghost;
389 static arc_state_t *arc_l2c_only;
390
391 /*
392 * There are several ARC variables that are critical to export as kstats --
393 * but we don't want to have to grovel around in the kstat whenever we wish to
394 * manipulate them. For these variables, we therefore define them to be in
395 * terms of the statistic variable. This assures that we are not introducing
396 * the possibility of inconsistency by having shadow copies of the variables,
397 * while still allowing the code to be readable.
398 */
399 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
400 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
401 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
402 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
403 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
404
405 static int arc_no_grow; /* Don't try to grow cache size */
406 static uint64_t arc_tempreserve;
407 static uint64_t arc_loaned_bytes;
408 static uint64_t arc_meta_used;
409 static uint64_t arc_meta_limit;
410 static uint64_t arc_meta_max = 0;
411
412 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
413
414 typedef struct arc_callback arc_callback_t;
415
416 struct arc_callback {
417 void *acb_private;
418 arc_done_func_t *acb_done;
419 arc_buf_t *acb_buf;
420 zio_t *acb_zio_dummy;
421 arc_callback_t *acb_next;
422 };
423
424 typedef struct arc_write_callback arc_write_callback_t;
425
426 struct arc_write_callback {
427 void *awcb_private;
428 arc_done_func_t *awcb_ready;
429 arc_done_func_t *awcb_done;
430 arc_buf_t *awcb_buf;
431 };
432
433 struct arc_buf_hdr {
434 /* protected by hash lock */
435 dva_t b_dva;
436 uint64_t b_birth;
437 uint64_t b_cksum0;
438
439 kmutex_t b_freeze_lock;
440 zio_cksum_t *b_freeze_cksum;
441 void *b_thawed;
442
443 arc_buf_hdr_t *b_hash_next;
444 arc_buf_t *b_buf;
445 uint32_t b_flags;
446 uint32_t b_datacnt;
447
448 arc_callback_t *b_acb;
449 kcondvar_t b_cv;
450
451 /* immutable */
452 arc_buf_contents_t b_type;
453 uint64_t b_size;
454 uint64_t b_spa;
455
456 /* protected by arc state mutex */
457 arc_state_t *b_state;
458 list_node_t b_arc_node;
459
460 /* updated atomically */
461 clock_t b_arc_access;
462
463 /* self protecting */
464 refcount_t b_refcnt;
465
466 l2arc_buf_hdr_t *b_l2hdr;
467 list_node_t b_l2node;
468 };
469
470 static arc_buf_t *arc_eviction_list;
471 static kmutex_t arc_eviction_mtx;
472 static arc_buf_hdr_t arc_eviction_hdr;
473 static void arc_get_data_buf(arc_buf_t *buf);
474 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
475 static int arc_evict_needed(arc_buf_contents_t type);
476 static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes);
477
478 static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab);
479
480 #define GHOST_STATE(state) \
481 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
482 (state) == arc_l2c_only)
483
484 /*
485 * Private ARC flags. These flags are private ARC only flags that will show up
486 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
487 * be passed in as arc_flags in things like arc_read. However, these flags
488 * should never be passed and should only be set by ARC code. When adding new
489 * public flags, make sure not to smash the private ones.
490 */
491
492 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
493 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
494 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
495 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
496 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
497 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
498 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
499 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
500 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
501 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
502
503 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
504 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
505 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
506 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
507 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
508 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
509 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
510 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
511 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
512 (hdr)->b_l2hdr != NULL)
513 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
514 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
515 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
516
517 /*
518 * Other sizes
519 */
520
521 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
522 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
523
524 /*
525 * Hash table routines
526 */
527
528 #define HT_LOCK_PAD 64
529
530 struct ht_lock {
531 kmutex_t ht_lock;
532 #ifdef _KERNEL
533 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
534 #endif
535 };
536
537 #define BUF_LOCKS 256
538 typedef struct buf_hash_table {
539 uint64_t ht_mask;
540 arc_buf_hdr_t **ht_table;
541 struct ht_lock ht_locks[BUF_LOCKS];
542 } buf_hash_table_t;
543
544 static buf_hash_table_t buf_hash_table;
545
546 #define BUF_HASH_INDEX(spa, dva, birth) \
547 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
548 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
549 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
550 #define HDR_LOCK(hdr) \
551 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
552
553 uint64_t zfs_crc64_table[256];
554
555 /*
556 * Level 2 ARC
557 */
558
559 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
560 #define L2ARC_HEADROOM 2 /* num of writes */
561 #define L2ARC_FEED_SECS 1 /* caching interval secs */
562 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
563
564 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
565 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
566
567 /*
568 * L2ARC Performance Tunables
569 */
570 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
571 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
572 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
573 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
574 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
575 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
576 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
577 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
578
579 /*
580 * L2ARC Internals
581 */
582 typedef struct l2arc_dev {
583 vdev_t *l2ad_vdev; /* vdev */
584 spa_t *l2ad_spa; /* spa */
585 uint64_t l2ad_hand; /* next write location */
586 uint64_t l2ad_write; /* desired write size, bytes */
587 uint64_t l2ad_boost; /* warmup write boost, bytes */
588 uint64_t l2ad_start; /* first addr on device */
589 uint64_t l2ad_end; /* last addr on device */
590 uint64_t l2ad_evict; /* last addr eviction reached */
591 boolean_t l2ad_first; /* first sweep through */
592 boolean_t l2ad_writing; /* currently writing */
593 list_t *l2ad_buflist; /* buffer list */
594 list_node_t l2ad_node; /* device list node */
595 } l2arc_dev_t;
596
597 static list_t L2ARC_dev_list; /* device list */
598 static list_t *l2arc_dev_list; /* device list pointer */
599 static kmutex_t l2arc_dev_mtx; /* device list mutex */
600 static l2arc_dev_t *l2arc_dev_last; /* last device used */
601 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
602 static list_t L2ARC_free_on_write; /* free after write buf list */
603 static list_t *l2arc_free_on_write; /* free after write list ptr */
604 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
605 static uint64_t l2arc_ndev; /* number of devices */
606
607 typedef struct l2arc_read_callback {
608 arc_buf_t *l2rcb_buf; /* read buffer */
609 spa_t *l2rcb_spa; /* spa */
610 blkptr_t l2rcb_bp; /* original blkptr */
611 zbookmark_t l2rcb_zb; /* original bookmark */
612 int l2rcb_flags; /* original flags */
613 } l2arc_read_callback_t;
614
615 typedef struct l2arc_write_callback {
616 l2arc_dev_t *l2wcb_dev; /* device info */
617 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
618 } l2arc_write_callback_t;
619
620 struct l2arc_buf_hdr {
621 /* protected by arc_buf_hdr mutex */
622 l2arc_dev_t *b_dev; /* L2ARC device */
623 uint64_t b_daddr; /* disk address, offset byte */
624 };
625
626 typedef struct l2arc_data_free {
627 /* protected by l2arc_free_on_write_mtx */
628 void *l2df_data;
629 size_t l2df_size;
630 void (*l2df_func)(void *, size_t);
631 list_node_t l2df_list_node;
632 } l2arc_data_free_t;
633
634 static kmutex_t l2arc_feed_thr_lock;
635 static kcondvar_t l2arc_feed_thr_cv;
636 static uint8_t l2arc_thread_exit;
637
638 static void l2arc_read_done(zio_t *zio);
639 static void l2arc_hdr_stat_add(void);
640 static void l2arc_hdr_stat_remove(void);
641
642 static uint64_t
643 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
644 {
645 uint8_t *vdva = (uint8_t *)dva;
646 uint64_t crc = -1ULL;
647 int i;
648
649 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
650
651 for (i = 0; i < sizeof (dva_t); i++)
652 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
653
654 crc ^= (spa>>8) ^ birth;
655
656 return (crc);
657 }
658
659 #define BUF_EMPTY(buf) \
660 ((buf)->b_dva.dva_word[0] == 0 && \
661 (buf)->b_dva.dva_word[1] == 0 && \
662 (buf)->b_birth == 0)
663
664 #define BUF_EQUAL(spa, dva, birth, buf) \
665 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
666 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
667 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
668
669 static void
670 buf_discard_identity(arc_buf_hdr_t *hdr)
671 {
672 hdr->b_dva.dva_word[0] = 0;
673 hdr->b_dva.dva_word[1] = 0;
674 hdr->b_birth = 0;
675 hdr->b_cksum0 = 0;
676 }
677
678 static arc_buf_hdr_t *
679 buf_hash_find(uint64_t spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
680 {
681 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
682 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
683 arc_buf_hdr_t *buf;
684
685 mutex_enter(hash_lock);
686 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
687 buf = buf->b_hash_next) {
688 if (BUF_EQUAL(spa, dva, birth, buf)) {
689 *lockp = hash_lock;
690 return (buf);
691 }
692 }
693 mutex_exit(hash_lock);
694 *lockp = NULL;
695 return (NULL);
696 }
697
698 /*
699 * Insert an entry into the hash table. If there is already an element
700 * equal to elem in the hash table, then the already existing element
701 * will be returned and the new element will not be inserted.
702 * Otherwise returns NULL.
703 */
704 static arc_buf_hdr_t *
705 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
706 {
707 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
708 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
709 arc_buf_hdr_t *fbuf;
710 uint32_t i;
711
712 ASSERT(!HDR_IN_HASH_TABLE(buf));
713 *lockp = hash_lock;
714 mutex_enter(hash_lock);
715 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
716 fbuf = fbuf->b_hash_next, i++) {
717 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
718 return (fbuf);
719 }
720
721 buf->b_hash_next = buf_hash_table.ht_table[idx];
722 buf_hash_table.ht_table[idx] = buf;
723 buf->b_flags |= ARC_IN_HASH_TABLE;
724
725 /* collect some hash table performance data */
726 if (i > 0) {
727 ARCSTAT_BUMP(arcstat_hash_collisions);
728 if (i == 1)
729 ARCSTAT_BUMP(arcstat_hash_chains);
730
731 ARCSTAT_MAX(arcstat_hash_chain_max, i);
732 }
733
734 ARCSTAT_BUMP(arcstat_hash_elements);
735 ARCSTAT_MAXSTAT(arcstat_hash_elements);
736
737 return (NULL);
738 }
739
740 static void
741 buf_hash_remove(arc_buf_hdr_t *buf)
742 {
743 arc_buf_hdr_t *fbuf, **bufp;
744 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
745
746 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
747 ASSERT(HDR_IN_HASH_TABLE(buf));
748
749 bufp = &buf_hash_table.ht_table[idx];
750 while ((fbuf = *bufp) != buf) {
751 ASSERT(fbuf != NULL);
752 bufp = &fbuf->b_hash_next;
753 }
754 *bufp = buf->b_hash_next;
755 buf->b_hash_next = NULL;
756 buf->b_flags &= ~ARC_IN_HASH_TABLE;
757
758 /* collect some hash table performance data */
759 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
760
761 if (buf_hash_table.ht_table[idx] &&
762 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
763 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
764 }
765
766 /*
767 * Global data structures and functions for the buf kmem cache.
768 */
769 static kmem_cache_t *hdr_cache;
770 static kmem_cache_t *buf_cache;
771
772 static void
773 buf_fini(void)
774 {
775 int i;
776
777 kmem_free(buf_hash_table.ht_table,
778 (buf_hash_table.ht_mask + 1) * sizeof (void *));
779 for (i = 0; i < BUF_LOCKS; i++)
780 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
781 kmem_cache_destroy(hdr_cache);
782 kmem_cache_destroy(buf_cache);
783 }
784
785 /*
786 * Constructor callback - called when the cache is empty
787 * and a new buf is requested.
788 */
789 /* ARGSUSED */
790 static int
791 hdr_cons(void *vbuf, void *unused, int kmflag)
792 {
793 arc_buf_hdr_t *buf = vbuf;
794
795 bzero(buf, sizeof (arc_buf_hdr_t));
796 refcount_create(&buf->b_refcnt);
797 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
798 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
799 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
800
801 return (0);
802 }
803
804 /* ARGSUSED */
805 static int
806 buf_cons(void *vbuf, void *unused, int kmflag)
807 {
808 arc_buf_t *buf = vbuf;
809
810 bzero(buf, sizeof (arc_buf_t));
811 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
812 rw_init(&buf->b_data_lock, NULL, RW_DEFAULT, NULL);
813 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
814
815 return (0);
816 }
817
818 /*
819 * Destructor callback - called when a cached buf is
820 * no longer required.
821 */
822 /* ARGSUSED */
823 static void
824 hdr_dest(void *vbuf, void *unused)
825 {
826 arc_buf_hdr_t *buf = vbuf;
827
828 ASSERT(BUF_EMPTY(buf));
829 refcount_destroy(&buf->b_refcnt);
830 cv_destroy(&buf->b_cv);
831 mutex_destroy(&buf->b_freeze_lock);
832 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
833 }
834
835 /* ARGSUSED */
836 static void
837 buf_dest(void *vbuf, void *unused)
838 {
839 arc_buf_t *buf = vbuf;
840
841 mutex_destroy(&buf->b_evict_lock);
842 rw_destroy(&buf->b_data_lock);
843 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
844 }
845
846 /*
847 * Reclaim callback -- invoked when memory is low.
848 */
849 /* ARGSUSED */
850 static void
851 hdr_recl(void *unused)
852 {
853 dprintf("hdr_recl called\n");
854 /*
855 * umem calls the reclaim func when we destroy the buf cache,
856 * which is after we do arc_fini().
857 */
858 if (!arc_dead)
859 cv_signal(&arc_reclaim_thr_cv);
860 }
861
862 static void
863 buf_init(void)
864 {
865 uint64_t *ct;
866 uint64_t hsize = 1ULL << 12;
867 int i, j;
868
869 /*
870 * The hash table is big enough to fill all of physical memory
871 * with an average 64K block size. The table will take up
872 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
873 */
874 while (hsize * 65536 < physmem * PAGESIZE)
875 hsize <<= 1;
876 retry:
877 buf_hash_table.ht_mask = hsize - 1;
878 buf_hash_table.ht_table =
879 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
880 if (buf_hash_table.ht_table == NULL) {
881 ASSERT(hsize > (1ULL << 8));
882 hsize >>= 1;
883 goto retry;
884 }
885
886 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
887 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
888 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
889 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
890
891 for (i = 0; i < 256; i++)
892 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
893 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
894
895 for (i = 0; i < BUF_LOCKS; i++) {
896 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
897 NULL, MUTEX_DEFAULT, NULL);
898 }
899 }
900
901 #define ARC_MINTIME (hz>>4) /* 62 ms */
902
903 static void
904 arc_cksum_verify(arc_buf_t *buf)
905 {
906 zio_cksum_t zc;
907
908 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
909 return;
910
911 mutex_enter(&buf->b_hdr->b_freeze_lock);
912 if (buf->b_hdr->b_freeze_cksum == NULL ||
913 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
914 mutex_exit(&buf->b_hdr->b_freeze_lock);
915 return;
916 }
917 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
918 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
919 panic("buffer modified while frozen!");
920 mutex_exit(&buf->b_hdr->b_freeze_lock);
921 }
922
923 static int
924 arc_cksum_equal(arc_buf_t *buf)
925 {
926 zio_cksum_t zc;
927 int equal;
928
929 mutex_enter(&buf->b_hdr->b_freeze_lock);
930 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
931 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
932 mutex_exit(&buf->b_hdr->b_freeze_lock);
933
934 return (equal);
935 }
936
937 static void
938 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
939 {
940 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
941 return;
942
943 mutex_enter(&buf->b_hdr->b_freeze_lock);
944 if (buf->b_hdr->b_freeze_cksum != NULL) {
945 mutex_exit(&buf->b_hdr->b_freeze_lock);
946 return;
947 }
948 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
949 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
950 buf->b_hdr->b_freeze_cksum);
951 mutex_exit(&buf->b_hdr->b_freeze_lock);
952 }
953
954 void
955 arc_buf_thaw(arc_buf_t *buf)
956 {
957 if (zfs_flags & ZFS_DEBUG_MODIFY) {
958 if (buf->b_hdr->b_state != arc_anon)
959 panic("modifying non-anon buffer!");
960 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
961 panic("modifying buffer while i/o in progress!");
962 arc_cksum_verify(buf);
963 }
964
965 mutex_enter(&buf->b_hdr->b_freeze_lock);
966 if (buf->b_hdr->b_freeze_cksum != NULL) {
967 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
968 buf->b_hdr->b_freeze_cksum = NULL;
969 }
970
971 if (zfs_flags & ZFS_DEBUG_MODIFY) {
972 if (buf->b_hdr->b_thawed)
973 kmem_free(buf->b_hdr->b_thawed, 1);
974 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP);
975 }
976
977 mutex_exit(&buf->b_hdr->b_freeze_lock);
978 }
979
980 void
981 arc_buf_freeze(arc_buf_t *buf)
982 {
983 kmutex_t *hash_lock;
984
985 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
986 return;
987
988 hash_lock = HDR_LOCK(buf->b_hdr);
989 mutex_enter(hash_lock);
990
991 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
992 buf->b_hdr->b_state == arc_anon);
993 arc_cksum_compute(buf, B_FALSE);
994 mutex_exit(hash_lock);
995 }
996
997 static void
998 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
999 {
1000 ASSERT(MUTEX_HELD(hash_lock));
1001
1002 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1003 (ab->b_state != arc_anon)) {
1004 uint64_t delta = ab->b_size * ab->b_datacnt;
1005 list_t *list = &ab->b_state->arcs_list[ab->b_type];
1006 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1007
1008 ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx));
1009 mutex_enter(&ab->b_state->arcs_mtx);
1010 ASSERT(list_link_active(&ab->b_arc_node));
1011 list_remove(list, ab);
1012 if (GHOST_STATE(ab->b_state)) {
1013 ASSERT3U(ab->b_datacnt, ==, 0);
1014 ASSERT3P(ab->b_buf, ==, NULL);
1015 delta = ab->b_size;
1016 }
1017 ASSERT(delta > 0);
1018 ASSERT3U(*size, >=, delta);
1019 atomic_add_64(size, -delta);
1020 mutex_exit(&ab->b_state->arcs_mtx);
1021 /* remove the prefetch flag if we get a reference */
1022 if (ab->b_flags & ARC_PREFETCH)
1023 ab->b_flags &= ~ARC_PREFETCH;
1024 }
1025 }
1026
1027 static int
1028 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1029 {
1030 int cnt;
1031 arc_state_t *state = ab->b_state;
1032
1033 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1034 ASSERT(!GHOST_STATE(state));
1035
1036 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1037 (state != arc_anon)) {
1038 uint64_t *size = &state->arcs_lsize[ab->b_type];
1039
1040 ASSERT(!MUTEX_HELD(&state->arcs_mtx));
1041 mutex_enter(&state->arcs_mtx);
1042 ASSERT(!list_link_active(&ab->b_arc_node));
1043 list_insert_head(&state->arcs_list[ab->b_type], ab);
1044 ASSERT(ab->b_datacnt > 0);
1045 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1046 mutex_exit(&state->arcs_mtx);
1047 }
1048 return (cnt);
1049 }
1050
1051 /*
1052 * Move the supplied buffer to the indicated state. The mutex
1053 * for the buffer must be held by the caller.
1054 */
1055 static void
1056 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1057 {
1058 arc_state_t *old_state = ab->b_state;
1059 int64_t refcnt = refcount_count(&ab->b_refcnt);
1060 uint64_t from_delta, to_delta;
1061
1062 ASSERT(MUTEX_HELD(hash_lock));
1063 ASSERT(new_state != old_state);
1064 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1065 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1066 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon);
1067
1068 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1069
1070 /*
1071 * If this buffer is evictable, transfer it from the
1072 * old state list to the new state list.
1073 */
1074 if (refcnt == 0) {
1075 if (old_state != arc_anon) {
1076 int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx);
1077 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1078
1079 if (use_mutex)
1080 mutex_enter(&old_state->arcs_mtx);
1081
1082 ASSERT(list_link_active(&ab->b_arc_node));
1083 list_remove(&old_state->arcs_list[ab->b_type], ab);
1084
1085 /*
1086 * If prefetching out of the ghost cache,
1087 * we will have a non-zero datacnt.
1088 */
1089 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1090 /* ghost elements have a ghost size */
1091 ASSERT(ab->b_buf == NULL);
1092 from_delta = ab->b_size;
1093 }
1094 ASSERT3U(*size, >=, from_delta);
1095 atomic_add_64(size, -from_delta);
1096
1097 if (use_mutex)
1098 mutex_exit(&old_state->arcs_mtx);
1099 }
1100 if (new_state != arc_anon) {
1101 int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx);
1102 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1103
1104 if (use_mutex)
1105 mutex_enter(&new_state->arcs_mtx);
1106
1107 list_insert_head(&new_state->arcs_list[ab->b_type], ab);
1108
1109 /* ghost elements have a ghost size */
1110 if (GHOST_STATE(new_state)) {
1111 ASSERT(ab->b_datacnt == 0);
1112 ASSERT(ab->b_buf == NULL);
1113 to_delta = ab->b_size;
1114 }
1115 atomic_add_64(size, to_delta);
1116
1117 if (use_mutex)
1118 mutex_exit(&new_state->arcs_mtx);
1119 }
1120 }
1121
1122 ASSERT(!BUF_EMPTY(ab));
1123 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab))
1124 buf_hash_remove(ab);
1125
1126 /* adjust state sizes */
1127 if (to_delta)
1128 atomic_add_64(&new_state->arcs_size, to_delta);
1129 if (from_delta) {
1130 ASSERT3U(old_state->arcs_size, >=, from_delta);
1131 atomic_add_64(&old_state->arcs_size, -from_delta);
1132 }
1133 ab->b_state = new_state;
1134
1135 /* adjust l2arc hdr stats */
1136 if (new_state == arc_l2c_only)
1137 l2arc_hdr_stat_add();
1138 else if (old_state == arc_l2c_only)
1139 l2arc_hdr_stat_remove();
1140 }
1141
1142 void
1143 arc_space_consume(uint64_t space, arc_space_type_t type)
1144 {
1145 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1146
1147 switch (type) {
1148 case ARC_SPACE_DATA:
1149 ARCSTAT_INCR(arcstat_data_size, space);
1150 break;
1151 case ARC_SPACE_OTHER:
1152 ARCSTAT_INCR(arcstat_other_size, space);
1153 break;
1154 case ARC_SPACE_HDRS:
1155 ARCSTAT_INCR(arcstat_hdr_size, space);
1156 break;
1157 case ARC_SPACE_L2HDRS:
1158 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1159 break;
1160 }
1161
1162 atomic_add_64(&arc_meta_used, space);
1163 atomic_add_64(&arc_size, space);
1164 }
1165
1166 void
1167 arc_space_return(uint64_t space, arc_space_type_t type)
1168 {
1169 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1170
1171 switch (type) {
1172 case ARC_SPACE_DATA:
1173 ARCSTAT_INCR(arcstat_data_size, -space);
1174 break;
1175 case ARC_SPACE_OTHER:
1176 ARCSTAT_INCR(arcstat_other_size, -space);
1177 break;
1178 case ARC_SPACE_HDRS:
1179 ARCSTAT_INCR(arcstat_hdr_size, -space);
1180 break;
1181 case ARC_SPACE_L2HDRS:
1182 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1183 break;
1184 }
1185
1186 ASSERT(arc_meta_used >= space);
1187 if (arc_meta_max < arc_meta_used)
1188 arc_meta_max = arc_meta_used;
1189 atomic_add_64(&arc_meta_used, -space);
1190 ASSERT(arc_size >= space);
1191 atomic_add_64(&arc_size, -space);
1192 }
1193
1194 void *
1195 arc_data_buf_alloc(uint64_t size)
1196 {
1197 if (arc_evict_needed(ARC_BUFC_DATA))
1198 cv_signal(&arc_reclaim_thr_cv);
1199 atomic_add_64(&arc_size, size);
1200 return (zio_data_buf_alloc(size));
1201 }
1202
1203 void
1204 arc_data_buf_free(void *buf, uint64_t size)
1205 {
1206 zio_data_buf_free(buf, size);
1207 ASSERT(arc_size >= size);
1208 atomic_add_64(&arc_size, -size);
1209 }
1210
1211 arc_buf_t *
1212 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1213 {
1214 arc_buf_hdr_t *hdr;
1215 arc_buf_t *buf;
1216
1217 ASSERT3U(size, >, 0);
1218 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1219 ASSERT(BUF_EMPTY(hdr));
1220 hdr->b_size = size;
1221 hdr->b_type = type;
1222 hdr->b_spa = spa_load_guid(spa);
1223 hdr->b_state = arc_anon;
1224 hdr->b_arc_access = 0;
1225 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1226 buf->b_hdr = hdr;
1227 buf->b_data = NULL;
1228 buf->b_efunc = NULL;
1229 buf->b_private = NULL;
1230 buf->b_next = NULL;
1231 hdr->b_buf = buf;
1232 arc_get_data_buf(buf);
1233 hdr->b_datacnt = 1;
1234 hdr->b_flags = 0;
1235 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1236 (void) refcount_add(&hdr->b_refcnt, tag);
1237
1238 return (buf);
1239 }
1240
1241 static char *arc_onloan_tag = "onloan";
1242
1243 /*
1244 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1245 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1246 * buffers must be returned to the arc before they can be used by the DMU or
1247 * freed.
1248 */
1249 arc_buf_t *
1250 arc_loan_buf(spa_t *spa, int size)
1251 {
1252 arc_buf_t *buf;
1253
1254 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1255
1256 atomic_add_64(&arc_loaned_bytes, size);
1257 return (buf);
1258 }
1259
1260 /*
1261 * Return a loaned arc buffer to the arc.
1262 */
1263 void
1264 arc_return_buf(arc_buf_t *buf, void *tag)
1265 {
1266 arc_buf_hdr_t *hdr = buf->b_hdr;
1267
1268 ASSERT(buf->b_data != NULL);
1269 (void) refcount_add(&hdr->b_refcnt, tag);
1270 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag);
1271
1272 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1273 }
1274
1275 /* Detach an arc_buf from a dbuf (tag) */
1276 void
1277 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1278 {
1279 arc_buf_hdr_t *hdr;
1280
1281 ASSERT(buf->b_data != NULL);
1282 hdr = buf->b_hdr;
1283 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag);
1284 (void) refcount_remove(&hdr->b_refcnt, tag);
1285 buf->b_efunc = NULL;
1286 buf->b_private = NULL;
1287
1288 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1289 }
1290
1291 static arc_buf_t *
1292 arc_buf_clone(arc_buf_t *from)
1293 {
1294 arc_buf_t *buf;
1295 arc_buf_hdr_t *hdr = from->b_hdr;
1296 uint64_t size = hdr->b_size;
1297
1298 ASSERT(hdr->b_state != arc_anon);
1299
1300 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1301 buf->b_hdr = hdr;
1302 buf->b_data = NULL;
1303 buf->b_efunc = NULL;
1304 buf->b_private = NULL;
1305 buf->b_next = hdr->b_buf;
1306 hdr->b_buf = buf;
1307 arc_get_data_buf(buf);
1308 bcopy(from->b_data, buf->b_data, size);
1309 hdr->b_datacnt += 1;
1310 return (buf);
1311 }
1312
1313 void
1314 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1315 {
1316 arc_buf_hdr_t *hdr;
1317 kmutex_t *hash_lock;
1318
1319 /*
1320 * Check to see if this buffer is evicted. Callers
1321 * must verify b_data != NULL to know if the add_ref
1322 * was successful.
1323 */
1324 mutex_enter(&buf->b_evict_lock);
1325 if (buf->b_data == NULL) {
1326 mutex_exit(&buf->b_evict_lock);
1327 return;
1328 }
1329 hash_lock = HDR_LOCK(buf->b_hdr);
1330 mutex_enter(hash_lock);
1331 hdr = buf->b_hdr;
1332 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1333 mutex_exit(&buf->b_evict_lock);
1334
1335 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1336 add_reference(hdr, hash_lock, tag);
1337 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1338 arc_access(hdr, hash_lock);
1339 mutex_exit(hash_lock);
1340 ARCSTAT_BUMP(arcstat_hits);
1341 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1342 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1343 data, metadata, hits);
1344 }
1345
1346 /*
1347 * Free the arc data buffer. If it is an l2arc write in progress,
1348 * the buffer is placed on l2arc_free_on_write to be freed later.
1349 */
1350 static void
1351 arc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t),
1352 void *data, size_t size)
1353 {
1354 if (HDR_L2_WRITING(hdr)) {
1355 l2arc_data_free_t *df;
1356 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1357 df->l2df_data = data;
1358 df->l2df_size = size;
1359 df->l2df_func = free_func;
1360 mutex_enter(&l2arc_free_on_write_mtx);
1361 list_insert_head(l2arc_free_on_write, df);
1362 mutex_exit(&l2arc_free_on_write_mtx);
1363 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1364 } else {
1365 free_func(data, size);
1366 }
1367 }
1368
1369 static void
1370 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1371 {
1372 arc_buf_t **bufp;
1373
1374 /* free up data associated with the buf */
1375 if (buf->b_data) {
1376 arc_state_t *state = buf->b_hdr->b_state;
1377 uint64_t size = buf->b_hdr->b_size;
1378 arc_buf_contents_t type = buf->b_hdr->b_type;
1379
1380 arc_cksum_verify(buf);
1381
1382 if (!recycle) {
1383 if (type == ARC_BUFC_METADATA) {
1384 arc_buf_data_free(buf->b_hdr, zio_buf_free,
1385 buf->b_data, size);
1386 arc_space_return(size, ARC_SPACE_DATA);
1387 } else {
1388 ASSERT(type == ARC_BUFC_DATA);
1389 arc_buf_data_free(buf->b_hdr,
1390 zio_data_buf_free, buf->b_data, size);
1391 ARCSTAT_INCR(arcstat_data_size, -size);
1392 atomic_add_64(&arc_size, -size);
1393 }
1394 }
1395 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1396 uint64_t *cnt = &state->arcs_lsize[type];
1397
1398 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1399 ASSERT(state != arc_anon);
1400
1401 ASSERT3U(*cnt, >=, size);
1402 atomic_add_64(cnt, -size);
1403 }
1404 ASSERT3U(state->arcs_size, >=, size);
1405 atomic_add_64(&state->arcs_size, -size);
1406 buf->b_data = NULL;
1407 ASSERT(buf->b_hdr->b_datacnt > 0);
1408 buf->b_hdr->b_datacnt -= 1;
1409 }
1410
1411 /* only remove the buf if requested */
1412 if (!all)
1413 return;
1414
1415 /* remove the buf from the hdr list */
1416 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1417 continue;
1418 *bufp = buf->b_next;
1419 buf->b_next = NULL;
1420
1421 ASSERT(buf->b_efunc == NULL);
1422
1423 /* clean up the buf */
1424 buf->b_hdr = NULL;
1425 kmem_cache_free(buf_cache, buf);
1426 }
1427
1428 static void
1429 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1430 {
1431 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1432 ASSERT3P(hdr->b_state, ==, arc_anon);
1433 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1434 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr;
1435
1436 if (l2hdr != NULL) {
1437 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx);
1438 /*
1439 * To prevent arc_free() and l2arc_evict() from
1440 * attempting to free the same buffer at the same time,
1441 * a FREE_IN_PROGRESS flag is given to arc_free() to
1442 * give it priority. l2arc_evict() can't destroy this
1443 * header while we are waiting on l2arc_buflist_mtx.
1444 *
1445 * The hdr may be removed from l2ad_buflist before we
1446 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1447 */
1448 if (!buflist_held) {
1449 mutex_enter(&l2arc_buflist_mtx);
1450 l2hdr = hdr->b_l2hdr;
1451 }
1452
1453 if (l2hdr != NULL) {
1454 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
1455 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1456 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
1457 if (hdr->b_state == arc_l2c_only)
1458 l2arc_hdr_stat_remove();
1459 hdr->b_l2hdr = NULL;
1460 }
1461
1462 if (!buflist_held)
1463 mutex_exit(&l2arc_buflist_mtx);
1464 }
1465
1466 if (!BUF_EMPTY(hdr)) {
1467 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1468 buf_discard_identity(hdr);
1469 }
1470 while (hdr->b_buf) {
1471 arc_buf_t *buf = hdr->b_buf;
1472
1473 if (buf->b_efunc) {
1474 mutex_enter(&arc_eviction_mtx);
1475 mutex_enter(&buf->b_evict_lock);
1476 ASSERT(buf->b_hdr != NULL);
1477 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1478 hdr->b_buf = buf->b_next;
1479 buf->b_hdr = &arc_eviction_hdr;
1480 buf->b_next = arc_eviction_list;
1481 arc_eviction_list = buf;
1482 mutex_exit(&buf->b_evict_lock);
1483 mutex_exit(&arc_eviction_mtx);
1484 } else {
1485 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1486 }
1487 }
1488 if (hdr->b_freeze_cksum != NULL) {
1489 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1490 hdr->b_freeze_cksum = NULL;
1491 }
1492 if (hdr->b_thawed) {
1493 kmem_free(hdr->b_thawed, 1);
1494 hdr->b_thawed = NULL;
1495 }
1496
1497 ASSERT(!list_link_active(&hdr->b_arc_node));
1498 ASSERT3P(hdr->b_hash_next, ==, NULL);
1499 ASSERT3P(hdr->b_acb, ==, NULL);
1500 kmem_cache_free(hdr_cache, hdr);
1501 }
1502
1503 void
1504 arc_buf_free(arc_buf_t *buf, void *tag)
1505 {
1506 arc_buf_hdr_t *hdr = buf->b_hdr;
1507 int hashed = hdr->b_state != arc_anon;
1508
1509 ASSERT(buf->b_efunc == NULL);
1510 ASSERT(buf->b_data != NULL);
1511
1512 if (hashed) {
1513 kmutex_t *hash_lock = HDR_LOCK(hdr);
1514
1515 mutex_enter(hash_lock);
1516 hdr = buf->b_hdr;
1517 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1518
1519 (void) remove_reference(hdr, hash_lock, tag);
1520 if (hdr->b_datacnt > 1) {
1521 arc_buf_destroy(buf, FALSE, TRUE);
1522 } else {
1523 ASSERT(buf == hdr->b_buf);
1524 ASSERT(buf->b_efunc == NULL);
1525 hdr->b_flags |= ARC_BUF_AVAILABLE;
1526 }
1527 mutex_exit(hash_lock);
1528 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1529 int destroy_hdr;
1530 /*
1531 * We are in the middle of an async write. Don't destroy
1532 * this buffer unless the write completes before we finish
1533 * decrementing the reference count.
1534 */
1535 mutex_enter(&arc_eviction_mtx);
1536 (void) remove_reference(hdr, NULL, tag);
1537 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1538 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1539 mutex_exit(&arc_eviction_mtx);
1540 if (destroy_hdr)
1541 arc_hdr_destroy(hdr);
1542 } else {
1543 if (remove_reference(hdr, NULL, tag) > 0)
1544 arc_buf_destroy(buf, FALSE, TRUE);
1545 else
1546 arc_hdr_destroy(hdr);
1547 }
1548 }
1549
1550 int
1551 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1552 {
1553 arc_buf_hdr_t *hdr = buf->b_hdr;
1554 kmutex_t *hash_lock = HDR_LOCK(hdr);
1555 int no_callback = (buf->b_efunc == NULL);
1556
1557 if (hdr->b_state == arc_anon) {
1558 ASSERT(hdr->b_datacnt == 1);
1559 arc_buf_free(buf, tag);
1560 return (no_callback);
1561 }
1562
1563 mutex_enter(hash_lock);
1564 hdr = buf->b_hdr;
1565 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1566 ASSERT(hdr->b_state != arc_anon);
1567 ASSERT(buf->b_data != NULL);
1568
1569 (void) remove_reference(hdr, hash_lock, tag);
1570 if (hdr->b_datacnt > 1) {
1571 if (no_callback)
1572 arc_buf_destroy(buf, FALSE, TRUE);
1573 } else if (no_callback) {
1574 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1575 ASSERT(buf->b_efunc == NULL);
1576 hdr->b_flags |= ARC_BUF_AVAILABLE;
1577 }
1578 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1579 refcount_is_zero(&hdr->b_refcnt));
1580 mutex_exit(hash_lock);
1581 return (no_callback);
1582 }
1583
1584 int
1585 arc_buf_size(arc_buf_t *buf)
1586 {
1587 return (buf->b_hdr->b_size);
1588 }
1589
1590 /*
1591 * Evict buffers from list until we've removed the specified number of
1592 * bytes. Move the removed buffers to the appropriate evict state.
1593 * If the recycle flag is set, then attempt to "recycle" a buffer:
1594 * - look for a buffer to evict that is `bytes' long.
1595 * - return the data block from this buffer rather than freeing it.
1596 * This flag is used by callers that are trying to make space for a
1597 * new buffer in a full arc cache.
1598 *
1599 * This function makes a "best effort". It skips over any buffers
1600 * it can't get a hash_lock on, and so may not catch all candidates.
1601 * It may also return without evicting as much space as requested.
1602 */
1603 static void *
1604 arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle,
1605 arc_buf_contents_t type)
1606 {
1607 arc_state_t *evicted_state;
1608 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1609 arc_buf_hdr_t *ab, *ab_prev = NULL;
1610 list_t *list = &state->arcs_list[type];
1611 kmutex_t *hash_lock;
1612 boolean_t have_lock;
1613 void *stolen = NULL;
1614
1615 ASSERT(state == arc_mru || state == arc_mfu);
1616
1617 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1618
1619 mutex_enter(&state->arcs_mtx);
1620 mutex_enter(&evicted_state->arcs_mtx);
1621
1622 for (ab = list_tail(list); ab; ab = ab_prev) {
1623 ab_prev = list_prev(list, ab);
1624 /* prefetch buffers have a minimum lifespan */
1625 if (HDR_IO_IN_PROGRESS(ab) ||
1626 (spa && ab->b_spa != spa) ||
1627 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1628 ddi_get_lbolt() - ab->b_arc_access <
1629 arc_min_prefetch_lifespan)) {
1630 skipped++;
1631 continue;
1632 }
1633 /* "lookahead" for better eviction candidate */
1634 if (recycle && ab->b_size != bytes &&
1635 ab_prev && ab_prev->b_size == bytes)
1636 continue;
1637 hash_lock = HDR_LOCK(ab);
1638 have_lock = MUTEX_HELD(hash_lock);
1639 if (have_lock || mutex_tryenter(hash_lock)) {
1640 ASSERT3U(refcount_count(&ab->b_refcnt), ==, 0);
1641 ASSERT(ab->b_datacnt > 0);
1642 while (ab->b_buf) {
1643 arc_buf_t *buf = ab->b_buf;
1644 if (!mutex_tryenter(&buf->b_evict_lock)) {
1645 missed += 1;
1646 break;
1647 }
1648 if (buf->b_data) {
1649 bytes_evicted += ab->b_size;
1650 if (recycle && ab->b_type == type &&
1651 ab->b_size == bytes &&
1652 !HDR_L2_WRITING(ab)) {
1653 stolen = buf->b_data;
1654 recycle = FALSE;
1655 }
1656 }
1657 if (buf->b_efunc) {
1658 mutex_enter(&arc_eviction_mtx);
1659 arc_buf_destroy(buf,
1660 buf->b_data == stolen, FALSE);
1661 ab->b_buf = buf->b_next;
1662 buf->b_hdr = &arc_eviction_hdr;
1663 buf->b_next = arc_eviction_list;
1664 arc_eviction_list = buf;
1665 mutex_exit(&arc_eviction_mtx);
1666 mutex_exit(&buf->b_evict_lock);
1667 } else {
1668 mutex_exit(&buf->b_evict_lock);
1669 arc_buf_destroy(buf,
1670 buf->b_data == stolen, TRUE);
1671 }
1672 }
1673
1674 if (ab->b_l2hdr) {
1675 ARCSTAT_INCR(arcstat_evict_l2_cached,
1676 ab->b_size);
1677 } else {
1678 if (l2arc_write_eligible(ab->b_spa, ab)) {
1679 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1680 ab->b_size);
1681 } else {
1682 ARCSTAT_INCR(
1683 arcstat_evict_l2_ineligible,
1684 ab->b_size);
1685 }
1686 }
1687
1688 if (ab->b_datacnt == 0) {
1689 arc_change_state(evicted_state, ab, hash_lock);
1690 ASSERT(HDR_IN_HASH_TABLE(ab));
1691 ab->b_flags |= ARC_IN_HASH_TABLE;
1692 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1693 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1694 }
1695 if (!have_lock)
1696 mutex_exit(hash_lock);
1697 if (bytes >= 0 && bytes_evicted >= bytes)
1698 break;
1699 } else {
1700 missed += 1;
1701 }
1702 }
1703
1704 mutex_exit(&evicted_state->arcs_mtx);
1705 mutex_exit(&state->arcs_mtx);
1706
1707 if (bytes_evicted < bytes)
1708 dprintf("only evicted %lld bytes from %x",
1709 (longlong_t)bytes_evicted, state);
1710
1711 if (skipped)
1712 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1713
1714 if (missed)
1715 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1716
1717 /*
1718 * We have just evicted some date into the ghost state, make
1719 * sure we also adjust the ghost state size if necessary.
1720 */
1721 if (arc_no_grow &&
1722 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1723 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1724 arc_mru_ghost->arcs_size - arc_c;
1725
1726 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1727 int64_t todelete =
1728 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1729 arc_evict_ghost(arc_mru_ghost, NULL, todelete);
1730 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1731 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1732 arc_mru_ghost->arcs_size +
1733 arc_mfu_ghost->arcs_size - arc_c);
1734 arc_evict_ghost(arc_mfu_ghost, NULL, todelete);
1735 }
1736 }
1737
1738 return (stolen);
1739 }
1740
1741 /*
1742 * Remove buffers from list until we've removed the specified number of
1743 * bytes. Destroy the buffers that are removed.
1744 */
1745 static void
1746 arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes)
1747 {
1748 arc_buf_hdr_t *ab, *ab_prev;
1749 arc_buf_hdr_t marker = { 0 };
1750 list_t *list = &state->arcs_list[ARC_BUFC_DATA];
1751 kmutex_t *hash_lock;
1752 uint64_t bytes_deleted = 0;
1753 uint64_t bufs_skipped = 0;
1754
1755 ASSERT(GHOST_STATE(state));
1756 top:
1757 mutex_enter(&state->arcs_mtx);
1758 for (ab = list_tail(list); ab; ab = ab_prev) {
1759 ab_prev = list_prev(list, ab);
1760 if (spa && ab->b_spa != spa)
1761 continue;
1762
1763 /* ignore markers */
1764 if (ab->b_spa == 0)
1765 continue;
1766
1767 hash_lock = HDR_LOCK(ab);
1768 /* caller may be trying to modify this buffer, skip it */
1769 if (MUTEX_HELD(hash_lock))
1770 continue;
1771 if (mutex_tryenter(hash_lock)) {
1772 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1773 ASSERT(ab->b_buf == NULL);
1774 ARCSTAT_BUMP(arcstat_deleted);
1775 bytes_deleted += ab->b_size;
1776
1777 if (ab->b_l2hdr != NULL) {
1778 /*
1779 * This buffer is cached on the 2nd Level ARC;
1780 * don't destroy the header.
1781 */
1782 arc_change_state(arc_l2c_only, ab, hash_lock);
1783 mutex_exit(hash_lock);
1784 } else {
1785 arc_change_state(arc_anon, ab, hash_lock);
1786 mutex_exit(hash_lock);
1787 arc_hdr_destroy(ab);
1788 }
1789
1790 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1791 if (bytes >= 0 && bytes_deleted >= bytes)
1792 break;
1793 } else if (bytes < 0) {
1794 /*
1795 * Insert a list marker and then wait for the
1796 * hash lock to become available. Once its
1797 * available, restart from where we left off.
1798 */
1799 list_insert_after(list, ab, &marker);
1800 mutex_exit(&state->arcs_mtx);
1801 mutex_enter(hash_lock);
1802 mutex_exit(hash_lock);
1803 mutex_enter(&state->arcs_mtx);
1804 ab_prev = list_prev(list, &marker);
1805 list_remove(list, &marker);
1806 } else
1807 bufs_skipped += 1;
1808 }
1809 mutex_exit(&state->arcs_mtx);
1810
1811 if (list == &state->arcs_list[ARC_BUFC_DATA] &&
1812 (bytes < 0 || bytes_deleted < bytes)) {
1813 list = &state->arcs_list[ARC_BUFC_METADATA];
1814 goto top;
1815 }
1816
1817 if (bufs_skipped) {
1818 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1819 ASSERT(bytes >= 0);
1820 }
1821
1822 if (bytes_deleted < bytes)
1823 dprintf("only deleted %lld bytes from %p",
1824 (longlong_t)bytes_deleted, state);
1825 }
1826
1827 static void
1828 arc_adjust(void)
1829 {
1830 int64_t adjustment, delta;
1831
1832 /*
1833 * Adjust MRU size
1834 */
1835
1836 adjustment = MIN((int64_t)(arc_size - arc_c),
1837 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used -
1838 arc_p));
1839
1840 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1841 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
1842 (void) arc_evict(arc_mru, NULL, delta, FALSE, ARC_BUFC_DATA);
1843 adjustment -= delta;
1844 }
1845
1846 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1847 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
1848 (void) arc_evict(arc_mru, NULL, delta, FALSE,
1849 ARC_BUFC_METADATA);
1850 }
1851
1852 /*
1853 * Adjust MFU size
1854 */
1855
1856 adjustment = arc_size - arc_c;
1857
1858 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1859 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
1860 (void) arc_evict(arc_mfu, NULL, delta, FALSE, ARC_BUFC_DATA);
1861 adjustment -= delta;
1862 }
1863
1864 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1865 int64_t delta = MIN(adjustment,
1866 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
1867 (void) arc_evict(arc_mfu, NULL, delta, FALSE,
1868 ARC_BUFC_METADATA);
1869 }
1870
1871 /*
1872 * Adjust ghost lists
1873 */
1874
1875 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
1876
1877 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
1878 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
1879 arc_evict_ghost(arc_mru_ghost, NULL, delta);
1880 }
1881
1882 adjustment =
1883 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
1884
1885 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
1886 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
1887 arc_evict_ghost(arc_mfu_ghost, NULL, delta);
1888 }
1889 }
1890
1891 static void
1892 arc_do_user_evicts(void)
1893 {
1894 mutex_enter(&arc_eviction_mtx);
1895 while (arc_eviction_list != NULL) {
1896 arc_buf_t *buf = arc_eviction_list;
1897 arc_eviction_list = buf->b_next;
1898 mutex_enter(&buf->b_evict_lock);
1899 buf->b_hdr = NULL;
1900 mutex_exit(&buf->b_evict_lock);
1901 mutex_exit(&arc_eviction_mtx);
1902
1903 if (buf->b_efunc != NULL)
1904 VERIFY(buf->b_efunc(buf) == 0);
1905
1906 buf->b_efunc = NULL;
1907 buf->b_private = NULL;
1908 kmem_cache_free(buf_cache, buf);
1909 mutex_enter(&arc_eviction_mtx);
1910 }
1911 mutex_exit(&arc_eviction_mtx);
1912 }
1913
1914 /*
1915 * Flush all *evictable* data from the cache for the given spa.
1916 * NOTE: this will not touch "active" (i.e. referenced) data.
1917 */
1918 void
1919 arc_flush(spa_t *spa)
1920 {
1921 uint64_t guid = 0;
1922
1923 if (spa)
1924 guid = spa_load_guid(spa);
1925
1926 while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) {
1927 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA);
1928 if (spa)
1929 break;
1930 }
1931 while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) {
1932 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA);
1933 if (spa)
1934 break;
1935 }
1936 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) {
1937 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA);
1938 if (spa)
1939 break;
1940 }
1941 while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) {
1942 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA);
1943 if (spa)
1944 break;
1945 }
1946
1947 arc_evict_ghost(arc_mru_ghost, guid, -1);
1948 arc_evict_ghost(arc_mfu_ghost, guid, -1);
1949
1950 mutex_enter(&arc_reclaim_thr_lock);
1951 arc_do_user_evicts();
1952 mutex_exit(&arc_reclaim_thr_lock);
1953 ASSERT(spa || arc_eviction_list == NULL);
1954 }
1955
1956 void
1957 arc_shrink(void)
1958 {
1959 if (arc_c > arc_c_min) {
1960 uint64_t to_free;
1961
1962 #ifdef _KERNEL
1963 to_free = MAX(arc_c >> arc_shrink_shift, ptob(needfree));
1964 #else
1965 to_free = arc_c >> arc_shrink_shift;
1966 #endif
1967 if (arc_c > arc_c_min + to_free)
1968 atomic_add_64(&arc_c, -to_free);
1969 else
1970 arc_c = arc_c_min;
1971
1972 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
1973 if (arc_c > arc_size)
1974 arc_c = MAX(arc_size, arc_c_min);
1975 if (arc_p > arc_c)
1976 arc_p = (arc_c >> 1);
1977 ASSERT(arc_c >= arc_c_min);
1978 ASSERT((int64_t)arc_p >= 0);
1979 }
1980
1981 if (arc_size > arc_c)
1982 arc_adjust();
1983 }
1984
1985 /*
1986 * Determine if the system is under memory pressure and is asking
1987 * to reclaim memory. A return value of 1 indicates that the system
1988 * is under memory pressure and that the arc should adjust accordingly.
1989 */
1990 static int
1991 arc_reclaim_needed(void)
1992 {
1993 uint64_t extra;
1994
1995 #ifdef _KERNEL
1996
1997 if (needfree)
1998 return (1);
1999
2000 /*
2001 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2002 */
2003 extra = desfree;
2004
2005 /*
2006 * check that we're out of range of the pageout scanner. It starts to
2007 * schedule paging if freemem is less than lotsfree and needfree.
2008 * lotsfree is the high-water mark for pageout, and needfree is the
2009 * number of needed free pages. We add extra pages here to make sure
2010 * the scanner doesn't start up while we're freeing memory.
2011 */
2012 if (freemem < lotsfree + needfree + extra)
2013 return (1);
2014
2015 /*
2016 * check to make sure that swapfs has enough space so that anon
2017 * reservations can still succeed. anon_resvmem() checks that the
2018 * availrmem is greater than swapfs_minfree, and the number of reserved
2019 * swap pages. We also add a bit of extra here just to prevent
2020 * circumstances from getting really dire.
2021 */
2022 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2023 return (1);
2024
2025 #if defined(__i386)
2026 /*
2027 * If we're on an i386 platform, it's possible that we'll exhaust the
2028 * kernel heap space before we ever run out of available physical
2029 * memory. Most checks of the size of the heap_area compare against
2030 * tune.t_minarmem, which is the minimum available real memory that we
2031 * can have in the system. However, this is generally fixed at 25 pages
2032 * which is so low that it's useless. In this comparison, we seek to
2033 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2034 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2035 * free)
2036 */
2037 if (vmem_size(heap_arena, VMEM_FREE) <
2038 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2))
2039 return (1);
2040 #endif
2041
2042 /*
2043 * If zio data pages are being allocated out of a separate heap segment,
2044 * then enforce that the size of available vmem for this arena remains
2045 * above about 1/16th free.
2046 *
2047 * Note: The 1/16th arena free requirement was put in place
2048 * to aggressively evict memory from the arc in order to avoid
2049 * memory fragmentation issues.
2050 */
2051 if (zio_arena != NULL &&
2052 vmem_size(zio_arena, VMEM_FREE) <
2053 (vmem_size(zio_arena, VMEM_ALLOC) >> 4))
2054 return (1);
2055 #else
2056 if (spa_get_random(100) == 0)
2057 return (1);
2058 #endif
2059 return (0);
2060 }
2061
2062 static void
2063 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2064 {
2065 size_t i;
2066 kmem_cache_t *prev_cache = NULL;
2067 kmem_cache_t *prev_data_cache = NULL;
2068 extern kmem_cache_t *zio_buf_cache[];
2069 extern kmem_cache_t *zio_data_buf_cache[];
2070
2071 #ifdef _KERNEL
2072 if (arc_meta_used >= arc_meta_limit) {
2073 /*
2074 * We are exceeding our meta-data cache limit.
2075 * Purge some DNLC entries to release holds on meta-data.
2076 */
2077 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2078 }
2079 #if defined(__i386)
2080 /*
2081 * Reclaim unused memory from all kmem caches.
2082 */
2083 kmem_reap();
2084 #endif
2085 #endif
2086
2087 /*
2088 * An aggressive reclamation will shrink the cache size as well as
2089 * reap free buffers from the arc kmem caches.
2090 */
2091 if (strat == ARC_RECLAIM_AGGR)
2092 arc_shrink();
2093
2094 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2095 if (zio_buf_cache[i] != prev_cache) {
2096 prev_cache = zio_buf_cache[i];
2097 kmem_cache_reap_now(zio_buf_cache[i]);
2098 }
2099 if (zio_data_buf_cache[i] != prev_data_cache) {
2100 prev_data_cache = zio_data_buf_cache[i];
2101 kmem_cache_reap_now(zio_data_buf_cache[i]);
2102 }
2103 }
2104 kmem_cache_reap_now(buf_cache);
2105 kmem_cache_reap_now(hdr_cache);
2106
2107 /*
2108 * Ask the vmem areana to reclaim unused memory from its
2109 * quantum caches.
2110 */
2111 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR)
2112 vmem_qcache_reap(zio_arena);
2113 }
2114
2115 static void
2116 arc_reclaim_thread(void)
2117 {
2118 clock_t growtime = 0;
2119 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2120 callb_cpr_t cpr;
2121
2122 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2123
2124 mutex_enter(&arc_reclaim_thr_lock);
2125 while (arc_thread_exit == 0) {
2126 if (arc_reclaim_needed()) {
2127
2128 if (arc_no_grow) {
2129 if (last_reclaim == ARC_RECLAIM_CONS) {
2130 last_reclaim = ARC_RECLAIM_AGGR;
2131 } else {
2132 last_reclaim = ARC_RECLAIM_CONS;
2133 }
2134 } else {
2135 arc_no_grow = TRUE;
2136 last_reclaim = ARC_RECLAIM_AGGR;
2137 membar_producer();
2138 }
2139
2140 /* reset the growth delay for every reclaim */
2141 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
2142
2143 arc_kmem_reap_now(last_reclaim);
2144 arc_warm = B_TRUE;
2145
2146 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) {
2147 arc_no_grow = FALSE;
2148 }
2149
2150 arc_adjust();
2151
2152 if (arc_eviction_list != NULL)
2153 arc_do_user_evicts();
2154
2155 /* block until needed, or one second, whichever is shorter */
2156 CALLB_CPR_SAFE_BEGIN(&cpr);
2157 (void) cv_timedwait(&arc_reclaim_thr_cv,
2158 &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz));
2159 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2160 }
2161
2162 arc_thread_exit = 0;
2163 cv_broadcast(&arc_reclaim_thr_cv);
2164 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2165 thread_exit();
2166 }
2167
2168 /*
2169 * Adapt arc info given the number of bytes we are trying to add and
2170 * the state that we are comming from. This function is only called
2171 * when we are adding new content to the cache.
2172 */
2173 static void
2174 arc_adapt(int bytes, arc_state_t *state)
2175 {
2176 int mult;
2177 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2178
2179 if (state == arc_l2c_only)
2180 return;
2181
2182 ASSERT(bytes > 0);
2183 /*
2184 * Adapt the target size of the MRU list:
2185 * - if we just hit in the MRU ghost list, then increase
2186 * the target size of the MRU list.
2187 * - if we just hit in the MFU ghost list, then increase
2188 * the target size of the MFU list by decreasing the
2189 * target size of the MRU list.
2190 */
2191 if (state == arc_mru_ghost) {
2192 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2193 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2194 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
2195
2196 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2197 } else if (state == arc_mfu_ghost) {
2198 uint64_t delta;
2199
2200 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2201 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2202 mult = MIN(mult, 10);
2203
2204 delta = MIN(bytes * mult, arc_p);
2205 arc_p = MAX(arc_p_min, arc_p - delta);
2206 }
2207 ASSERT((int64_t)arc_p >= 0);
2208
2209 if (arc_reclaim_needed()) {
2210 cv_signal(&arc_reclaim_thr_cv);
2211 return;
2212 }
2213
2214 if (arc_no_grow)
2215 return;
2216
2217 if (arc_c >= arc_c_max)
2218 return;
2219
2220 /*
2221 * If we're within (2 * maxblocksize) bytes of the target
2222 * cache size, increment the target cache size
2223 */
2224 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2225 atomic_add_64(&arc_c, (int64_t)bytes);
2226 if (arc_c > arc_c_max)
2227 arc_c = arc_c_max;
2228 else if (state == arc_anon)
2229 atomic_add_64(&arc_p, (int64_t)bytes);
2230 if (arc_p > arc_c)
2231 arc_p = arc_c;
2232 }
2233 ASSERT((int64_t)arc_p >= 0);
2234 }
2235
2236 /*
2237 * Check if the cache has reached its limits and eviction is required
2238 * prior to insert.
2239 */
2240 static int
2241 arc_evict_needed(arc_buf_contents_t type)
2242 {
2243 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2244 return (1);
2245
2246 if (arc_reclaim_needed())
2247 return (1);
2248
2249 return (arc_size > arc_c);
2250 }
2251
2252 /*
2253 * The buffer, supplied as the first argument, needs a data block.
2254 * So, if we are at cache max, determine which cache should be victimized.
2255 * We have the following cases:
2256 *
2257 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2258 * In this situation if we're out of space, but the resident size of the MFU is
2259 * under the limit, victimize the MFU cache to satisfy this insertion request.
2260 *
2261 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2262 * Here, we've used up all of the available space for the MRU, so we need to
2263 * evict from our own cache instead. Evict from the set of resident MRU
2264 * entries.
2265 *
2266 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2267 * c minus p represents the MFU space in the cache, since p is the size of the
2268 * cache that is dedicated to the MRU. In this situation there's still space on
2269 * the MFU side, so the MRU side needs to be victimized.
2270 *
2271 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2272 * MFU's resident set is consuming more space than it has been allotted. In
2273 * this situation, we must victimize our own cache, the MFU, for this insertion.
2274 */
2275 static void
2276 arc_get_data_buf(arc_buf_t *buf)
2277 {
2278 arc_state_t *state = buf->b_hdr->b_state;
2279 uint64_t size = buf->b_hdr->b_size;
2280 arc_buf_contents_t type = buf->b_hdr->b_type;
2281
2282 arc_adapt(size, state);
2283
2284 /*
2285 * We have not yet reached cache maximum size,
2286 * just allocate a new buffer.
2287 */
2288 if (!arc_evict_needed(type)) {
2289 if (type == ARC_BUFC_METADATA) {
2290 buf->b_data = zio_buf_alloc(size);
2291 arc_space_consume(size, ARC_SPACE_DATA);
2292 } else {
2293 ASSERT(type == ARC_BUFC_DATA);
2294 buf->b_data = zio_data_buf_alloc(size);
2295 ARCSTAT_INCR(arcstat_data_size, size);
2296 atomic_add_64(&arc_size, size);
2297 }
2298 goto out;
2299 }
2300
2301 /*
2302 * If we are prefetching from the mfu ghost list, this buffer
2303 * will end up on the mru list; so steal space from there.
2304 */
2305 if (state == arc_mfu_ghost)
2306 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2307 else if (state == arc_mru_ghost)
2308 state = arc_mru;
2309
2310 if (state == arc_mru || state == arc_anon) {
2311 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2312 state = (arc_mfu->arcs_lsize[type] >= size &&
2313 arc_p > mru_used) ? arc_mfu : arc_mru;
2314 } else {
2315 /* MFU cases */
2316 uint64_t mfu_space = arc_c - arc_p;
2317 state = (arc_mru->arcs_lsize[type] >= size &&
2318 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2319 }
2320 if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) {
2321 if (type == ARC_BUFC_METADATA) {
2322 buf->b_data = zio_buf_alloc(size);
2323 arc_space_consume(size, ARC_SPACE_DATA);
2324 } else {
2325 ASSERT(type == ARC_BUFC_DATA);
2326 buf->b_data = zio_data_buf_alloc(size);
2327 ARCSTAT_INCR(arcstat_data_size, size);
2328 atomic_add_64(&arc_size, size);
2329 }
2330 ARCSTAT_BUMP(arcstat_recycle_miss);
2331 }
2332 ASSERT(buf->b_data != NULL);
2333 out:
2334 /*
2335 * Update the state size. Note that ghost states have a
2336 * "ghost size" and so don't need to be updated.
2337 */
2338 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2339 arc_buf_hdr_t *hdr = buf->b_hdr;
2340
2341 atomic_add_64(&hdr->b_state->arcs_size, size);
2342 if (list_link_active(&hdr->b_arc_node)) {
2343 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2344 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2345 }
2346 /*
2347 * If we are growing the cache, and we are adding anonymous
2348 * data, and we have outgrown arc_p, update arc_p
2349 */
2350 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2351 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2352 arc_p = MIN(arc_c, arc_p + size);
2353 }
2354 }
2355
2356 /*
2357 * This routine is called whenever a buffer is accessed.
2358 * NOTE: the hash lock is dropped in this function.
2359 */
2360 static void
2361 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2362 {
2363 clock_t now;
2364
2365 ASSERT(MUTEX_HELD(hash_lock));
2366
2367 if (buf->b_state == arc_anon) {
2368 /*
2369 * This buffer is not in the cache, and does not
2370 * appear in our "ghost" list. Add the new buffer
2371 * to the MRU state.
2372 */
2373
2374 ASSERT(buf->b_arc_access == 0);
2375 buf->b_arc_access = ddi_get_lbolt();
2376 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2377 arc_change_state(arc_mru, buf, hash_lock);
2378
2379 } else if (buf->b_state == arc_mru) {
2380 now = ddi_get_lbolt();
2381
2382 /*
2383 * If this buffer is here because of a prefetch, then either:
2384 * - clear the flag if this is a "referencing" read
2385 * (any subsequent access will bump this into the MFU state).
2386 * or
2387 * - move the buffer to the head of the list if this is
2388 * another prefetch (to make it less likely to be evicted).
2389 */
2390 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2391 if (refcount_count(&buf->b_refcnt) == 0) {
2392 ASSERT(list_link_active(&buf->b_arc_node));
2393 } else {
2394 buf->b_flags &= ~ARC_PREFETCH;
2395 ARCSTAT_BUMP(arcstat_mru_hits);
2396 }
2397 buf->b_arc_access = now;
2398 return;
2399 }
2400
2401 /*
2402 * This buffer has been "accessed" only once so far,
2403 * but it is still in the cache. Move it to the MFU
2404 * state.
2405 */
2406 if (now > buf->b_arc_access + ARC_MINTIME) {
2407 /*
2408 * More than 125ms have passed since we
2409 * instantiated this buffer. Move it to the
2410 * most frequently used state.
2411 */
2412 buf->b_arc_access = now;
2413 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2414 arc_change_state(arc_mfu, buf, hash_lock);
2415 }
2416 ARCSTAT_BUMP(arcstat_mru_hits);
2417 } else if (buf->b_state == arc_mru_ghost) {
2418 arc_state_t *new_state;
2419 /*
2420 * This buffer has been "accessed" recently, but
2421 * was evicted from the cache. Move it to the
2422 * MFU state.
2423 */
2424
2425 if (buf->b_flags & ARC_PREFETCH) {
2426 new_state = arc_mru;
2427 if (refcount_count(&buf->b_refcnt) > 0)
2428 buf->b_flags &= ~ARC_PREFETCH;
2429 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2430 } else {
2431 new_state = arc_mfu;
2432 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2433 }
2434
2435 buf->b_arc_access = ddi_get_lbolt();
2436 arc_change_state(new_state, buf, hash_lock);
2437
2438 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2439 } else if (buf->b_state == arc_mfu) {
2440 /*
2441 * This buffer has been accessed more than once and is
2442 * still in the cache. Keep it in the MFU state.
2443 *
2444 * NOTE: an add_reference() that occurred when we did
2445 * the arc_read() will have kicked this off the list.
2446 * If it was a prefetch, we will explicitly move it to
2447 * the head of the list now.
2448 */
2449 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2450 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2451 ASSERT(list_link_active(&buf->b_arc_node));
2452 }
2453 ARCSTAT_BUMP(arcstat_mfu_hits);
2454 buf->b_arc_access = ddi_get_lbolt();
2455 } else if (buf->b_state == arc_mfu_ghost) {
2456 arc_state_t *new_state = arc_mfu;
2457 /*
2458 * This buffer has been accessed more than once but has
2459 * been evicted from the cache. Move it back to the
2460 * MFU state.
2461 */
2462
2463 if (buf->b_flags & ARC_PREFETCH) {
2464 /*
2465 * This is a prefetch access...
2466 * move this block back to the MRU state.
2467 */
2468 ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0);
2469 new_state = arc_mru;
2470 }
2471
2472 buf->b_arc_access = ddi_get_lbolt();
2473 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2474 arc_change_state(new_state, buf, hash_lock);
2475
2476 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2477 } else if (buf->b_state == arc_l2c_only) {
2478 /*
2479 * This buffer is on the 2nd Level ARC.
2480 */
2481
2482 buf->b_arc_access = ddi_get_lbolt();
2483 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2484 arc_change_state(arc_mfu, buf, hash_lock);
2485 } else {
2486 ASSERT(!"invalid arc state");
2487 }
2488 }
2489
2490 /* a generic arc_done_func_t which you can use */
2491 /* ARGSUSED */
2492 void
2493 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2494 {
2495 if (zio == NULL || zio->io_error == 0)
2496 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2497 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2498 }
2499
2500 /* a generic arc_done_func_t */
2501 void
2502 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2503 {
2504 arc_buf_t **bufp = arg;
2505 if (zio && zio->io_error) {
2506 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2507 *bufp = NULL;
2508 } else {
2509 *bufp = buf;
2510 ASSERT(buf->b_data);
2511 }
2512 }
2513
2514 static void
2515 arc_read_done(zio_t *zio)
2516 {
2517 arc_buf_hdr_t *hdr, *found;
2518 arc_buf_t *buf;
2519 arc_buf_t *abuf; /* buffer we're assigning to callback */
2520 kmutex_t *hash_lock;
2521 arc_callback_t *callback_list, *acb;
2522 int freeable = FALSE;
2523
2524 buf = zio->io_private;
2525 hdr = buf->b_hdr;
2526
2527 /*
2528 * The hdr was inserted into hash-table and removed from lists
2529 * prior to starting I/O. We should find this header, since
2530 * it's in the hash table, and it should be legit since it's
2531 * not possible to evict it during the I/O. The only possible
2532 * reason for it not to be found is if we were freed during the
2533 * read.
2534 */
2535 found = buf_hash_find(hdr->b_spa, &hdr->b_dva, hdr->b_birth,
2536 &hash_lock);
2537
2538 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2539 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2540 (found == hdr && HDR_L2_READING(hdr)));
2541
2542 hdr->b_flags &= ~ARC_L2_EVICTED;
2543 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2544 hdr->b_flags &= ~ARC_L2CACHE;
2545
2546 /* byteswap if necessary */
2547 callback_list = hdr->b_acb;
2548 ASSERT(callback_list != NULL);
2549 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
2550 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2551 byteswap_uint64_array :
2552 dmu_ot[BP_GET_TYPE(zio->io_bp)].ot_byteswap;
2553 func(buf->b_data, hdr->b_size);
2554 }
2555
2556 arc_cksum_compute(buf, B_FALSE);
2557
2558 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) {
2559 /*
2560 * Only call arc_access on anonymous buffers. This is because
2561 * if we've issued an I/O for an evicted buffer, we've already
2562 * called arc_access (to prevent any simultaneous readers from
2563 * getting confused).
2564 */
2565 arc_access(hdr, hash_lock);
2566 }
2567
2568 /* create copies of the data buffer for the callers */
2569 abuf = buf;
2570 for (acb = callback_list; acb; acb = acb->acb_next) {
2571 if (acb->acb_done) {
2572 if (abuf == NULL)
2573 abuf = arc_buf_clone(buf);
2574 acb->acb_buf = abuf;
2575 abuf = NULL;
2576 }
2577 }
2578 hdr->b_acb = NULL;
2579 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2580 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2581 if (abuf == buf) {
2582 ASSERT(buf->b_efunc == NULL);
2583 ASSERT(hdr->b_datacnt == 1);
2584 hdr->b_flags |= ARC_BUF_AVAILABLE;
2585 }
2586
2587 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2588
2589 if (zio->io_error != 0) {
2590 hdr->b_flags |= ARC_IO_ERROR;
2591 if (hdr->b_state != arc_anon)
2592 arc_change_state(arc_anon, hdr, hash_lock);
2593 if (HDR_IN_HASH_TABLE(hdr))
2594 buf_hash_remove(hdr);
2595 freeable = refcount_is_zero(&hdr->b_refcnt);
2596 }
2597
2598 /*
2599 * Broadcast before we drop the hash_lock to avoid the possibility
2600 * that the hdr (and hence the cv) might be freed before we get to
2601 * the cv_broadcast().
2602 */
2603 cv_broadcast(&hdr->b_cv);
2604
2605 if (hash_lock) {
2606 mutex_exit(hash_lock);
2607 } else {
2608 /*
2609 * This block was freed while we waited for the read to
2610 * complete. It has been removed from the hash table and
2611 * moved to the anonymous state (so that it won't show up
2612 * in the cache).
2613 */
2614 ASSERT3P(hdr->b_state, ==, arc_anon);
2615 freeable = refcount_is_zero(&hdr->b_refcnt);
2616 }
2617
2618 /* execute each callback and free its structure */
2619 while ((acb = callback_list) != NULL) {
2620 if (acb->acb_done)
2621 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2622
2623 if (acb->acb_zio_dummy != NULL) {
2624 acb->acb_zio_dummy->io_error = zio->io_error;
2625 zio_nowait(acb->acb_zio_dummy);
2626 }
2627
2628 callback_list = acb->acb_next;
2629 kmem_free(acb, sizeof (arc_callback_t));
2630 }
2631
2632 if (freeable)
2633 arc_hdr_destroy(hdr);
2634 }
2635
2636 /*
2637 * "Read" the block block at the specified DVA (in bp) via the
2638 * cache. If the block is found in the cache, invoke the provided
2639 * callback immediately and return. Note that the `zio' parameter
2640 * in the callback will be NULL in this case, since no IO was
2641 * required. If the block is not in the cache pass the read request
2642 * on to the spa with a substitute callback function, so that the
2643 * requested block will be added to the cache.
2644 *
2645 * If a read request arrives for a block that has a read in-progress,
2646 * either wait for the in-progress read to complete (and return the
2647 * results); or, if this is a read with a "done" func, add a record
2648 * to the read to invoke the "done" func when the read completes,
2649 * and return; or just return.
2650 *
2651 * arc_read_done() will invoke all the requested "done" functions
2652 * for readers of this block.
2653 *
2654 * Normal callers should use arc_read and pass the arc buffer and offset
2655 * for the bp. But if you know you don't need locking, you can use
2656 * arc_read_bp.
2657 */
2658 int
2659 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_buf_t *pbuf,
2660 arc_done_func_t *done, void *private, int priority, int zio_flags,
2661 uint32_t *arc_flags, const zbookmark_t *zb)
2662 {
2663 int err;
2664
2665 if (pbuf == NULL) {
2666 /*
2667 * XXX This happens from traverse callback funcs, for
2668 * the objset_phys_t block.
2669 */
2670 return (arc_read_nolock(pio, spa, bp, done, private, priority,
2671 zio_flags, arc_flags, zb));
2672 }
2673
2674 ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt));
2675 ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size);
2676 rw_enter(&pbuf->b_data_lock, RW_READER);
2677
2678 err = arc_read_nolock(pio, spa, bp, done, private, priority,
2679 zio_flags, arc_flags, zb);
2680 rw_exit(&pbuf->b_data_lock);
2681
2682 return (err);
2683 }
2684
2685 int
2686 arc_read_nolock(zio_t *pio, spa_t *spa, const blkptr_t *bp,
2687 arc_done_func_t *done, void *private, int priority, int zio_flags,
2688 uint32_t *arc_flags, const zbookmark_t *zb)
2689 {
2690 arc_buf_hdr_t *hdr;
2691 arc_buf_t *buf;
2692 kmutex_t *hash_lock;
2693 zio_t *rzio;
2694 uint64_t guid = spa_load_guid(spa);
2695
2696 top:
2697 hdr = buf_hash_find(guid, BP_IDENTITY(bp), BP_PHYSICAL_BIRTH(bp),
2698 &hash_lock);
2699 if (hdr && hdr->b_datacnt > 0) {
2700
2701 *arc_flags |= ARC_CACHED;
2702
2703 if (HDR_IO_IN_PROGRESS(hdr)) {
2704
2705 if (*arc_flags & ARC_WAIT) {
2706 cv_wait(&hdr->b_cv, hash_lock);
2707 mutex_exit(hash_lock);
2708 goto top;
2709 }
2710 ASSERT(*arc_flags & ARC_NOWAIT);
2711
2712 if (done) {
2713 arc_callback_t *acb = NULL;
2714
2715 acb = kmem_zalloc(sizeof (arc_callback_t),
2716 KM_SLEEP);
2717 acb->acb_done = done;
2718 acb->acb_private = private;
2719 if (pio != NULL)
2720 acb->acb_zio_dummy = zio_null(pio,
2721 spa, NULL, NULL, NULL, zio_flags);
2722
2723 ASSERT(acb->acb_done != NULL);
2724 acb->acb_next = hdr->b_acb;
2725 hdr->b_acb = acb;
2726 add_reference(hdr, hash_lock, private);
2727 mutex_exit(hash_lock);
2728 return (0);
2729 }
2730 mutex_exit(hash_lock);
2731 return (0);
2732 }
2733
2734 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2735
2736 if (done) {
2737 add_reference(hdr, hash_lock, private);
2738 /*
2739 * If this block is already in use, create a new
2740 * copy of the data so that we will be guaranteed
2741 * that arc_release() will always succeed.
2742 */
2743 buf = hdr->b_buf;
2744 ASSERT(buf);
2745 ASSERT(buf->b_data);
2746 if (HDR_BUF_AVAILABLE(hdr)) {
2747 ASSERT(buf->b_efunc == NULL);
2748 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2749 } else {
2750 buf = arc_buf_clone(buf);
2751 }
2752
2753 } else if (*arc_flags & ARC_PREFETCH &&
2754 refcount_count(&hdr->b_refcnt) == 0) {
2755 hdr->b_flags |= ARC_PREFETCH;
2756 }
2757 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2758 arc_access(hdr, hash_lock);
2759 if (*arc_flags & ARC_L2CACHE)
2760 hdr->b_flags |= ARC_L2CACHE;
2761 mutex_exit(hash_lock);
2762 ARCSTAT_BUMP(arcstat_hits);
2763 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2764 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2765 data, metadata, hits);
2766
2767 if (done)
2768 done(NULL, buf, private);
2769 } else {
2770 uint64_t size = BP_GET_LSIZE(bp);
2771 arc_callback_t *acb;
2772 vdev_t *vd = NULL;
2773 uint64_t addr;
2774 boolean_t devw = B_FALSE;
2775
2776 if (hdr == NULL) {
2777 /* this block is not in the cache */
2778 arc_buf_hdr_t *exists;
2779 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2780 buf = arc_buf_alloc(spa, size, private, type);
2781 hdr = buf->b_hdr;
2782 hdr->b_dva = *BP_IDENTITY(bp);
2783 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
2784 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2785 exists = buf_hash_insert(hdr, &hash_lock);
2786 if (exists) {
2787 /* somebody beat us to the hash insert */
2788 mutex_exit(hash_lock);
2789 buf_discard_identity(hdr);
2790 (void) arc_buf_remove_ref(buf, private);
2791 goto top; /* restart the IO request */
2792 }
2793 /* if this is a prefetch, we don't have a reference */
2794 if (*arc_flags & ARC_PREFETCH) {
2795 (void) remove_reference(hdr, hash_lock,
2796 private);
2797 hdr->b_flags |= ARC_PREFETCH;
2798 }
2799 if (*arc_flags & ARC_L2CACHE)
2800 hdr->b_flags |= ARC_L2CACHE;
2801 if (BP_GET_LEVEL(bp) > 0)
2802 hdr->b_flags |= ARC_INDIRECT;
2803 } else {
2804 /* this block is in the ghost cache */
2805 ASSERT(GHOST_STATE(hdr->b_state));
2806 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2807 ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0);
2808 ASSERT(hdr->b_buf == NULL);
2809
2810 /* if this is a prefetch, we don't have a reference */
2811 if (*arc_flags & ARC_PREFETCH)
2812 hdr->b_flags |= ARC_PREFETCH;
2813 else
2814 add_reference(hdr, hash_lock, private);
2815 if (*arc_flags & ARC_L2CACHE)
2816 hdr->b_flags |= ARC_L2CACHE;
2817 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2818 buf->b_hdr = hdr;
2819 buf->b_data = NULL;
2820 buf->b_efunc = NULL;
2821 buf->b_private = NULL;
2822 buf->b_next = NULL;
2823 hdr->b_buf = buf;
2824 ASSERT(hdr->b_datacnt == 0);
2825 hdr->b_datacnt = 1;
2826 arc_get_data_buf(buf);
2827 arc_access(hdr, hash_lock);
2828 }
2829
2830 ASSERT(!GHOST_STATE(hdr->b_state));
2831
2832 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
2833 acb->acb_done = done;
2834 acb->acb_private = private;
2835
2836 ASSERT(hdr->b_acb == NULL);
2837 hdr->b_acb = acb;
2838 hdr->b_flags |= ARC_IO_IN_PROGRESS;
2839
2840 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
2841 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
2842 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
2843 addr = hdr->b_l2hdr->b_daddr;
2844 /*
2845 * Lock out device removal.
2846 */
2847 if (vdev_is_dead(vd) ||
2848 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
2849 vd = NULL;
2850 }
2851
2852 mutex_exit(hash_lock);
2853
2854 ASSERT3U(hdr->b_size, ==, size);
2855 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
2856 uint64_t, size, zbookmark_t *, zb);
2857 ARCSTAT_BUMP(arcstat_misses);
2858 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2859 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2860 data, metadata, misses);
2861
2862 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
2863 /*
2864 * Read from the L2ARC if the following are true:
2865 * 1. The L2ARC vdev was previously cached.
2866 * 2. This buffer still has L2ARC metadata.
2867 * 3. This buffer isn't currently writing to the L2ARC.
2868 * 4. The L2ARC entry wasn't evicted, which may
2869 * also have invalidated the vdev.
2870 * 5. This isn't prefetch and l2arc_noprefetch is set.
2871 */
2872 if (hdr->b_l2hdr != NULL &&
2873 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
2874 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
2875 l2arc_read_callback_t *cb;
2876
2877 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
2878 ARCSTAT_BUMP(arcstat_l2_hits);
2879
2880 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
2881 KM_SLEEP);
2882 cb->l2rcb_buf = buf;
2883 cb->l2rcb_spa = spa;
2884 cb->l2rcb_bp = *bp;
2885 cb->l2rcb_zb = *zb;
2886 cb->l2rcb_flags = zio_flags;
2887
2888 /*
2889 * l2arc read. The SCL_L2ARC lock will be
2890 * released by l2arc_read_done().
2891 */
2892 rzio = zio_read_phys(pio, vd, addr, size,
2893 buf->b_data, ZIO_CHECKSUM_OFF,
2894 l2arc_read_done, cb, priority, zio_flags |
2895 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
2896 ZIO_FLAG_DONT_PROPAGATE |
2897 ZIO_FLAG_DONT_RETRY, B_FALSE);
2898 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
2899 zio_t *, rzio);
2900 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
2901
2902 if (*arc_flags & ARC_NOWAIT) {
2903 zio_nowait(rzio);
2904 return (0);
2905 }
2906
2907 ASSERT(*arc_flags & ARC_WAIT);
2908 if (zio_wait(rzio) == 0)
2909 return (0);
2910
2911 /* l2arc read error; goto zio_read() */
2912 } else {
2913 DTRACE_PROBE1(l2arc__miss,
2914 arc_buf_hdr_t *, hdr);
2915 ARCSTAT_BUMP(arcstat_l2_misses);
2916 if (HDR_L2_WRITING(hdr))
2917 ARCSTAT_BUMP(arcstat_l2_rw_clash);
2918 spa_config_exit(spa, SCL_L2ARC, vd);
2919 }
2920 } else {
2921 if (vd != NULL)
2922 spa_config_exit(spa, SCL_L2ARC, vd);
2923 if (l2arc_ndev != 0) {
2924 DTRACE_PROBE1(l2arc__miss,
2925 arc_buf_hdr_t *, hdr);
2926 ARCSTAT_BUMP(arcstat_l2_misses);
2927 }
2928 }
2929
2930 rzio = zio_read(pio, spa, bp, buf->b_data, size,
2931 arc_read_done, buf, priority, zio_flags, zb);
2932
2933 if (*arc_flags & ARC_WAIT)
2934 return (zio_wait(rzio));
2935
2936 ASSERT(*arc_flags & ARC_NOWAIT);
2937 zio_nowait(rzio);
2938 }
2939 return (0);
2940 }
2941
2942 void
2943 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
2944 {
2945 ASSERT(buf->b_hdr != NULL);
2946 ASSERT(buf->b_hdr->b_state != arc_anon);
2947 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
2948 ASSERT(buf->b_efunc == NULL);
2949 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
2950
2951 buf->b_efunc = func;
2952 buf->b_private = private;
2953 }
2954
2955 /*
2956 * This is used by the DMU to let the ARC know that a buffer is
2957 * being evicted, so the ARC should clean up. If this arc buf
2958 * is not yet in the evicted state, it will be put there.
2959 */
2960 int
2961 arc_buf_evict(arc_buf_t *buf)
2962 {
2963 arc_buf_hdr_t *hdr;
2964 kmutex_t *hash_lock;
2965 arc_buf_t **bufp;
2966
2967 mutex_enter(&buf->b_evict_lock);
2968 hdr = buf->b_hdr;
2969 if (hdr == NULL) {
2970 /*
2971 * We are in arc_do_user_evicts().
2972 */
2973 ASSERT(buf->b_data == NULL);
2974 mutex_exit(&buf->b_evict_lock);
2975 return (0);
2976 } else if (buf->b_data == NULL) {
2977 arc_buf_t copy = *buf; /* structure assignment */
2978 /*
2979 * We are on the eviction list; process this buffer now
2980 * but let arc_do_user_evicts() do the reaping.
2981 */
2982 buf->b_efunc = NULL;
2983 mutex_exit(&buf->b_evict_lock);
2984 VERIFY(copy.b_efunc(©) == 0);
2985 return (1);
2986 }
2987 hash_lock = HDR_LOCK(hdr);
2988 mutex_enter(hash_lock);
2989 hdr = buf->b_hdr;
2990 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2991
2992 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
2993 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2994
2995 /*
2996 * Pull this buffer off of the hdr
2997 */
2998 bufp = &hdr->b_buf;
2999 while (*bufp != buf)
3000 bufp = &(*bufp)->b_next;
3001 *bufp = buf->b_next;
3002
3003 ASSERT(buf->b_data != NULL);
3004 arc_buf_destroy(buf, FALSE, FALSE);
3005
3006 if (hdr->b_datacnt == 0) {
3007 arc_state_t *old_state = hdr->b_state;
3008 arc_state_t *evicted_state;
3009
3010 ASSERT(hdr->b_buf == NULL);
3011 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3012
3013 evicted_state =
3014 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3015
3016 mutex_enter(&old_state->arcs_mtx);
3017 mutex_enter(&evicted_state->arcs_mtx);
3018
3019 arc_change_state(evicted_state, hdr, hash_lock);
3020 ASSERT(HDR_IN_HASH_TABLE(hdr));
3021 hdr->b_flags |= ARC_IN_HASH_TABLE;
3022 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3023
3024 mutex_exit(&evicted_state->arcs_mtx);
3025 mutex_exit(&old_state->arcs_mtx);
3026 }
3027 mutex_exit(hash_lock);
3028 mutex_exit(&buf->b_evict_lock);
3029
3030 VERIFY(buf->b_efunc(buf) == 0);
3031 buf->b_efunc = NULL;
3032 buf->b_private = NULL;
3033 buf->b_hdr = NULL;
3034 buf->b_next = NULL;
3035 kmem_cache_free(buf_cache, buf);
3036 return (1);
3037 }
3038
3039 /*
3040 * Release this buffer from the cache. This must be done
3041 * after a read and prior to modifying the buffer contents.
3042 * If the buffer has more than one reference, we must make
3043 * a new hdr for the buffer.
3044 */
3045 void
3046 arc_release(arc_buf_t *buf, void *tag)
3047 {
3048 arc_buf_hdr_t *hdr;
3049 kmutex_t *hash_lock = NULL;
3050 l2arc_buf_hdr_t *l2hdr;
3051 uint64_t buf_size;
3052
3053 /*
3054 * It would be nice to assert that if it's DMU metadata (level >
3055 * 0 || it's the dnode file), then it must be syncing context.
3056 * But we don't know that information at this level.
3057 */
3058
3059 mutex_enter(&buf->b_evict_lock);
3060 hdr = buf->b_hdr;
3061
3062 /* this buffer is not on any list */
3063 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3064
3065 if (hdr->b_state == arc_anon) {
3066 /* this buffer is already released */
3067 ASSERT(buf->b_efunc == NULL);
3068 } else {
3069 hash_lock = HDR_LOCK(hdr);
3070 mutex_enter(hash_lock);
3071 hdr = buf->b_hdr;
3072 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3073 }
3074
3075 l2hdr = hdr->b_l2hdr;
3076 if (l2hdr) {
3077 mutex_enter(&l2arc_buflist_mtx);
3078 hdr->b_l2hdr = NULL;
3079 buf_size = hdr->b_size;
3080 }
3081
3082 /*
3083 * Do we have more than one buf?
3084 */
3085 if (hdr->b_datacnt > 1) {
3086 arc_buf_hdr_t *nhdr;
3087 arc_buf_t **bufp;
3088 uint64_t blksz = hdr->b_size;
3089 uint64_t spa = hdr->b_spa;
3090 arc_buf_contents_t type = hdr->b_type;
3091 uint32_t flags = hdr->b_flags;
3092
3093 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3094 /*
3095 * Pull the data off of this hdr and attach it to
3096 * a new anonymous hdr.
3097 */
3098 (void) remove_reference(hdr, hash_lock, tag);
3099 bufp = &hdr->b_buf;
3100 while (*bufp != buf)
3101 bufp = &(*bufp)->b_next;
3102 *bufp = buf->b_next;
3103 buf->b_next = NULL;
3104
3105 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3106 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3107 if (refcount_is_zero(&hdr->b_refcnt)) {
3108 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3109 ASSERT3U(*size, >=, hdr->b_size);
3110 atomic_add_64(size, -hdr->b_size);
3111 }
3112 hdr->b_datacnt -= 1;
3113 arc_cksum_verify(buf);
3114
3115 mutex_exit(hash_lock);
3116
3117 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3118 nhdr->b_size = blksz;
3119 nhdr->b_spa = spa;
3120 nhdr->b_type = type;
3121 nhdr->b_buf = buf;
3122 nhdr->b_state = arc_anon;
3123 nhdr->b_arc_access = 0;
3124 nhdr->b_flags = flags & ARC_L2_WRITING;
3125 nhdr->b_l2hdr = NULL;
3126 nhdr->b_datacnt = 1;
3127 nhdr->b_freeze_cksum = NULL;
3128 (void) refcount_add(&nhdr->b_refcnt, tag);
3129 buf->b_hdr = nhdr;
3130 mutex_exit(&buf->b_evict_lock);
3131 atomic_add_64(&arc_anon->arcs_size, blksz);
3132 } else {
3133 mutex_exit(&buf->b_evict_lock);
3134 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3135 ASSERT(!list_link_active(&hdr->b_arc_node));
3136 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3137 if (hdr->b_state != arc_anon)
3138 arc_change_state(arc_anon, hdr, hash_lock);
3139 hdr->b_arc_access = 0;
3140 if (hash_lock)
3141 mutex_exit(hash_lock);
3142
3143 buf_discard_identity(hdr);
3144 arc_buf_thaw(buf);
3145 }
3146 buf->b_efunc = NULL;
3147 buf->b_private = NULL;
3148
3149 if (l2hdr) {
3150 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3151 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3152 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3153 mutex_exit(&l2arc_buflist_mtx);
3154 }
3155 }
3156
3157 /*
3158 * Release this buffer. If it does not match the provided BP, fill it
3159 * with that block's contents.
3160 */
3161 /* ARGSUSED */
3162 int
3163 arc_release_bp(arc_buf_t *buf, void *tag, blkptr_t *bp, spa_t *spa,
3164 zbookmark_t *zb)
3165 {
3166 arc_release(buf, tag);
3167 return (0);
3168 }
3169
3170 int
3171 arc_released(arc_buf_t *buf)
3172 {
3173 int released;
3174
3175 mutex_enter(&buf->b_evict_lock);
3176 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3177 mutex_exit(&buf->b_evict_lock);
3178 return (released);
3179 }
3180
3181 int
3182 arc_has_callback(arc_buf_t *buf)
3183 {
3184 int callback;
3185
3186 mutex_enter(&buf->b_evict_lock);
3187 callback = (buf->b_efunc != NULL);
3188 mutex_exit(&buf->b_evict_lock);
3189 return (callback);
3190 }
3191
3192 #ifdef ZFS_DEBUG
3193 int
3194 arc_referenced(arc_buf_t *buf)
3195 {
3196 int referenced;
3197
3198 mutex_enter(&buf->b_evict_lock);
3199 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3200 mutex_exit(&buf->b_evict_lock);
3201 return (referenced);
3202 }
3203 #endif
3204
3205 static void
3206 arc_write_ready(zio_t *zio)
3207 {
3208 arc_write_callback_t *callback = zio->io_private;
3209 arc_buf_t *buf = callback->awcb_buf;
3210 arc_buf_hdr_t *hdr = buf->b_hdr;
3211
3212 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3213 callback->awcb_ready(zio, buf, callback->awcb_private);
3214
3215 /*
3216 * If the IO is already in progress, then this is a re-write
3217 * attempt, so we need to thaw and re-compute the cksum.
3218 * It is the responsibility of the callback to handle the
3219 * accounting for any re-write attempt.
3220 */
3221 if (HDR_IO_IN_PROGRESS(hdr)) {
3222 mutex_enter(&hdr->b_freeze_lock);
3223 if (hdr->b_freeze_cksum != NULL) {
3224 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3225 hdr->b_freeze_cksum = NULL;
3226 }
3227 mutex_exit(&hdr->b_freeze_lock);
3228 }
3229 arc_cksum_compute(buf, B_FALSE);
3230 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3231 }
3232
3233 static void
3234 arc_write_done(zio_t *zio)
3235 {
3236 arc_write_callback_t *callback = zio->io_private;
3237 arc_buf_t *buf = callback->awcb_buf;
3238 arc_buf_hdr_t *hdr = buf->b_hdr;
3239
3240 ASSERT(hdr->b_acb == NULL);
3241
3242 if (zio->io_error == 0) {
3243 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3244 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
3245 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3246 } else {
3247 ASSERT(BUF_EMPTY(hdr));
3248 }
3249
3250 /*
3251 * If the block to be written was all-zero, we may have
3252 * compressed it away. In this case no write was performed
3253 * so there will be no dva/birth/checksum. The buffer must
3254 * therefore remain anonymous (and uncached).
3255 */
3256 if (!BUF_EMPTY(hdr)) {
3257 arc_buf_hdr_t *exists;
3258 kmutex_t *hash_lock;
3259
3260 ASSERT(zio->io_error == 0);
3261
3262 arc_cksum_verify(buf);
3263
3264 exists = buf_hash_insert(hdr, &hash_lock);
3265 if (exists) {
3266 /*
3267 * This can only happen if we overwrite for
3268 * sync-to-convergence, because we remove
3269 * buffers from the hash table when we arc_free().
3270 */
3271 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
3272 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
3273 panic("bad overwrite, hdr=%p exists=%p",
3274 (void *)hdr, (void *)exists);
3275 ASSERT(refcount_is_zero(&exists->b_refcnt));
3276 arc_change_state(arc_anon, exists, hash_lock);
3277 mutex_exit(hash_lock);
3278 arc_hdr_destroy(exists);
3279 exists = buf_hash_insert(hdr, &hash_lock);
3280 ASSERT3P(exists, ==, NULL);
3281 } else {
3282 /* Dedup */
3283 ASSERT(hdr->b_datacnt == 1);
3284 ASSERT(hdr->b_state == arc_anon);
3285 ASSERT(BP_GET_DEDUP(zio->io_bp));
3286 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
3287 }
3288 }
3289 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3290 /* if it's not anon, we are doing a scrub */
3291 if (!exists && hdr->b_state == arc_anon)
3292 arc_access(hdr, hash_lock);
3293 mutex_exit(hash_lock);
3294 } else {
3295 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3296 }
3297
3298 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3299 callback->awcb_done(zio, buf, callback->awcb_private);
3300
3301 kmem_free(callback, sizeof (arc_write_callback_t));
3302 }
3303
3304 zio_t *
3305 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
3306 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, const zio_prop_t *zp,
3307 arc_done_func_t *ready, arc_done_func_t *done, void *private,
3308 int priority, int zio_flags, const zbookmark_t *zb)
3309 {
3310 arc_buf_hdr_t *hdr = buf->b_hdr;
3311 arc_write_callback_t *callback;
3312 zio_t *zio;
3313
3314 ASSERT(ready != NULL);
3315 ASSERT(done != NULL);
3316 ASSERT(!HDR_IO_ERROR(hdr));
3317 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3318 ASSERT(hdr->b_acb == NULL);
3319 if (l2arc)
3320 hdr->b_flags |= ARC_L2CACHE;
3321 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3322 callback->awcb_ready = ready;
3323 callback->awcb_done = done;
3324 callback->awcb_private = private;
3325 callback->awcb_buf = buf;
3326
3327 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
3328 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3329
3330 return (zio);
3331 }
3332
3333 static int
3334 arc_memory_throttle(uint64_t reserve, uint64_t inflight_data, uint64_t txg)
3335 {
3336 #ifdef _KERNEL
3337 uint64_t available_memory = ptob(freemem);
3338 static uint64_t page_load = 0;
3339 static uint64_t last_txg = 0;
3340
3341 #if defined(__i386)
3342 available_memory =
3343 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3344 #endif
3345 if (available_memory >= zfs_write_limit_max)
3346 return (0);
3347
3348 if (txg > last_txg) {
3349 last_txg = txg;
3350 page_load = 0;
3351 }
3352 /*
3353 * If we are in pageout, we know that memory is already tight,
3354 * the arc is already going to be evicting, so we just want to
3355 * continue to let page writes occur as quickly as possible.
3356 */
3357 if (curproc == proc_pageout) {
3358 if (page_load > MAX(ptob(minfree), available_memory) / 4)
3359 return (ERESTART);
3360 /* Note: reserve is inflated, so we deflate */
3361 page_load += reserve / 8;
3362 return (0);
3363 } else if (page_load > 0 && arc_reclaim_needed()) {
3364 /* memory is low, delay before restarting */
3365 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3366 return (EAGAIN);
3367 }
3368 page_load = 0;
3369
3370 if (arc_size > arc_c_min) {
3371 uint64_t evictable_memory =
3372 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3373 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3374 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3375 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3376 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3377 }
3378
3379 if (inflight_data > available_memory / 4) {
3380 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3381 return (ERESTART);
3382 }
3383 #endif
3384 return (0);
3385 }
3386
3387 void
3388 arc_tempreserve_clear(uint64_t reserve)
3389 {
3390 atomic_add_64(&arc_tempreserve, -reserve);
3391 ASSERT((int64_t)arc_tempreserve >= 0);
3392 }
3393
3394 int
3395 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3396 {
3397 int error;
3398 uint64_t anon_size;
3399
3400 #ifdef ZFS_DEBUG
3401 /*
3402 * Once in a while, fail for no reason. Everything should cope.
3403 */
3404 if (spa_get_random(10000) == 0) {
3405 dprintf("forcing random failure\n");
3406 return (ERESTART);
3407 }
3408 #endif
3409 if (reserve > arc_c/4 && !arc_no_grow)
3410 arc_c = MIN(arc_c_max, reserve * 4);
3411 if (reserve > arc_c)
3412 return (ENOMEM);
3413
3414 /*
3415 * Don't count loaned bufs as in flight dirty data to prevent long
3416 * network delays from blocking transactions that are ready to be
3417 * assigned to a txg.
3418 */
3419 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0);
3420
3421 /*
3422 * Writes will, almost always, require additional memory allocations
3423 * in order to compress/encrypt/etc the data. We therefor need to
3424 * make sure that there is sufficient available memory for this.
3425 */
3426 if (error = arc_memory_throttle(reserve, anon_size, txg))
3427 return (error);
3428
3429 /*
3430 * Throttle writes when the amount of dirty data in the cache
3431 * gets too large. We try to keep the cache less than half full
3432 * of dirty blocks so that our sync times don't grow too large.
3433 * Note: if two requests come in concurrently, we might let them
3434 * both succeed, when one of them should fail. Not a huge deal.
3435 */
3436
3437 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
3438 anon_size > arc_c / 4) {
3439 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3440 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3441 arc_tempreserve>>10,
3442 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3443 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3444 reserve>>10, arc_c>>10);
3445 return (ERESTART);
3446 }
3447 atomic_add_64(&arc_tempreserve, reserve);
3448 return (0);
3449 }
3450
3451 void
3452 arc_init(void)
3453 {
3454 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3455 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3456
3457 /* Convert seconds to clock ticks */
3458 arc_min_prefetch_lifespan = 1 * hz;
3459
3460 /* Start out with 1/8 of all memory */
3461 arc_c = physmem * PAGESIZE / 8;
3462
3463 #ifdef _KERNEL
3464 /*
3465 * On architectures where the physical memory can be larger
3466 * than the addressable space (intel in 32-bit mode), we may
3467 * need to limit the cache to 1/8 of VM size.
3468 */
3469 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3470 #endif
3471
3472 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
3473 arc_c_min = MAX(arc_c / 4, 64<<20);
3474 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
3475 if (arc_c * 8 >= 1<<30)
3476 arc_c_max = (arc_c * 8) - (1<<30);
3477 else
3478 arc_c_max = arc_c_min;
3479 arc_c_max = MAX(arc_c * 6, arc_c_max);
3480
3481 /*
3482 * Allow the tunables to override our calculations if they are
3483 * reasonable (ie. over 64MB)
3484 */
3485 if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE)
3486 arc_c_max = zfs_arc_max;
3487 if (zfs_arc_min > 64<<20 && zfs_arc_min <= arc_c_max)
3488 arc_c_min = zfs_arc_min;
3489
3490 arc_c = arc_c_max;
3491 arc_p = (arc_c >> 1);
3492
3493 /* limit meta-data to 1/4 of the arc capacity */
3494 arc_meta_limit = arc_c_max / 4;
3495
3496 /* Allow the tunable to override if it is reasonable */
3497 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3498 arc_meta_limit = zfs_arc_meta_limit;
3499
3500 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3501 arc_c_min = arc_meta_limit / 2;
3502
3503 if (zfs_arc_grow_retry > 0)
3504 arc_grow_retry = zfs_arc_grow_retry;
3505
3506 if (zfs_arc_shrink_shift > 0)
3507 arc_shrink_shift = zfs_arc_shrink_shift;
3508
3509 if (zfs_arc_p_min_shift > 0)
3510 arc_p_min_shift = zfs_arc_p_min_shift;
3511
3512 /* if kmem_flags are set, lets try to use less memory */
3513 if (kmem_debugging())
3514 arc_c = arc_c / 2;
3515 if (arc_c < arc_c_min)
3516 arc_c = arc_c_min;
3517
3518 arc_anon = &ARC_anon;
3519 arc_mru = &ARC_mru;
3520 arc_mru_ghost = &ARC_mru_ghost;
3521 arc_mfu = &ARC_mfu;
3522 arc_mfu_ghost = &ARC_mfu_ghost;
3523 arc_l2c_only = &ARC_l2c_only;
3524 arc_size = 0;
3525
3526 mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3527 mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3528 mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3529 mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3530 mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3531 mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL);
3532
3533 list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
3534 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3535 list_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
3536 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3537 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
3538 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3539 list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
3540 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3541 list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
3542 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3543 list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
3544 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3545 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
3546 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3547 list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
3548 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3549 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
3550 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3551 list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
3552 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3553
3554 buf_init();
3555
3556 arc_thread_exit = 0;
3557 arc_eviction_list = NULL;
3558 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3559 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3560
3561 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3562 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3563
3564 if (arc_ksp != NULL) {
3565 arc_ksp->ks_data = &arc_stats;
3566 kstat_install(arc_ksp);
3567 }
3568
3569 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3570 TS_RUN, minclsyspri);
3571
3572 arc_dead = FALSE;
3573 arc_warm = B_FALSE;
3574
3575 if (zfs_write_limit_max == 0)
3576 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3577 else
3578 zfs_write_limit_shift = 0;
3579 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3580 }
3581
3582 void
3583 arc_fini(void)
3584 {
3585 mutex_enter(&arc_reclaim_thr_lock);
3586 arc_thread_exit = 1;
3587 while (arc_thread_exit != 0)
3588 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3589 mutex_exit(&arc_reclaim_thr_lock);
3590
3591 arc_flush(NULL);
3592
3593 arc_dead = TRUE;
3594
3595 if (arc_ksp != NULL) {
3596 kstat_delete(arc_ksp);
3597 arc_ksp = NULL;
3598 }
3599
3600 mutex_destroy(&arc_eviction_mtx);
3601 mutex_destroy(&arc_reclaim_thr_lock);
3602 cv_destroy(&arc_reclaim_thr_cv);
3603
3604 list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
3605 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
3606 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
3607 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
3608 list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
3609 list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
3610 list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
3611 list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
3612
3613 mutex_destroy(&arc_anon->arcs_mtx);
3614 mutex_destroy(&arc_mru->arcs_mtx);
3615 mutex_destroy(&arc_mru_ghost->arcs_mtx);
3616 mutex_destroy(&arc_mfu->arcs_mtx);
3617 mutex_destroy(&arc_mfu_ghost->arcs_mtx);
3618 mutex_destroy(&arc_l2c_only->arcs_mtx);
3619
3620 mutex_destroy(&zfs_write_limit_lock);
3621
3622 buf_fini();
3623
3624 ASSERT(arc_loaned_bytes == 0);
3625 }
3626
3627 /*
3628 * Level 2 ARC
3629 *
3630 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3631 * It uses dedicated storage devices to hold cached data, which are populated
3632 * using large infrequent writes. The main role of this cache is to boost
3633 * the performance of random read workloads. The intended L2ARC devices
3634 * include short-stroked disks, solid state disks, and other media with
3635 * substantially faster read latency than disk.
3636 *
3637 * +-----------------------+
3638 * | ARC |
3639 * +-----------------------+
3640 * | ^ ^
3641 * | | |
3642 * l2arc_feed_thread() arc_read()
3643 * | | |
3644 * | l2arc read |
3645 * V | |
3646 * +---------------+ |
3647 * | L2ARC | |
3648 * +---------------+ |
3649 * | ^ |
3650 * l2arc_write() | |
3651 * | | |
3652 * V | |
3653 * +-------+ +-------+
3654 * | vdev | | vdev |
3655 * | cache | | cache |
3656 * +-------+ +-------+
3657 * +=========+ .-----.
3658 * : L2ARC : |-_____-|
3659 * : devices : | Disks |
3660 * +=========+ `-_____-'
3661 *
3662 * Read requests are satisfied from the following sources, in order:
3663 *
3664 * 1) ARC
3665 * 2) vdev cache of L2ARC devices
3666 * 3) L2ARC devices
3667 * 4) vdev cache of disks
3668 * 5) disks
3669 *
3670 * Some L2ARC device types exhibit extremely slow write performance.
3671 * To accommodate for this there are some significant differences between
3672 * the L2ARC and traditional cache design:
3673 *
3674 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
3675 * the ARC behave as usual, freeing buffers and placing headers on ghost
3676 * lists. The ARC does not send buffers to the L2ARC during eviction as
3677 * this would add inflated write latencies for all ARC memory pressure.
3678 *
3679 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
3680 * It does this by periodically scanning buffers from the eviction-end of
3681 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
3682 * not already there. It scans until a headroom of buffers is satisfied,
3683 * which itself is a buffer for ARC eviction. The thread that does this is
3684 * l2arc_feed_thread(), illustrated below; example sizes are included to
3685 * provide a better sense of ratio than this diagram:
3686 *
3687 * head --> tail
3688 * +---------------------+----------+
3689 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
3690 * +---------------------+----------+ | o L2ARC eligible
3691 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
3692 * +---------------------+----------+ |
3693 * 15.9 Gbytes ^ 32 Mbytes |
3694 * headroom |
3695 * l2arc_feed_thread()
3696 * |
3697 * l2arc write hand <--[oooo]--'
3698 * | 8 Mbyte
3699 * | write max
3700 * V
3701 * +==============================+
3702 * L2ARC dev |####|#|###|###| |####| ... |
3703 * +==============================+
3704 * 32 Gbytes
3705 *
3706 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
3707 * evicted, then the L2ARC has cached a buffer much sooner than it probably
3708 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
3709 * safe to say that this is an uncommon case, since buffers at the end of
3710 * the ARC lists have moved there due to inactivity.
3711 *
3712 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
3713 * then the L2ARC simply misses copying some buffers. This serves as a
3714 * pressure valve to prevent heavy read workloads from both stalling the ARC
3715 * with waits and clogging the L2ARC with writes. This also helps prevent
3716 * the potential for the L2ARC to churn if it attempts to cache content too
3717 * quickly, such as during backups of the entire pool.
3718 *
3719 * 5. After system boot and before the ARC has filled main memory, there are
3720 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
3721 * lists can remain mostly static. Instead of searching from tail of these
3722 * lists as pictured, the l2arc_feed_thread() will search from the list heads
3723 * for eligible buffers, greatly increasing its chance of finding them.
3724 *
3725 * The L2ARC device write speed is also boosted during this time so that
3726 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
3727 * there are no L2ARC reads, and no fear of degrading read performance
3728 * through increased writes.
3729 *
3730 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
3731 * the vdev queue can aggregate them into larger and fewer writes. Each
3732 * device is written to in a rotor fashion, sweeping writes through
3733 * available space then repeating.
3734 *
3735 * 7. The L2ARC does not store dirty content. It never needs to flush
3736 * write buffers back to disk based storage.
3737 *
3738 * 8. If an ARC buffer is written (and dirtied) which also exists in the
3739 * L2ARC, the now stale L2ARC buffer is immediately dropped.
3740 *
3741 * The performance of the L2ARC can be tweaked by a number of tunables, which
3742 * may be necessary for different workloads:
3743 *
3744 * l2arc_write_max max write bytes per interval
3745 * l2arc_write_boost extra write bytes during device warmup
3746 * l2arc_noprefetch skip caching prefetched buffers
3747 * l2arc_headroom number of max device writes to precache
3748 * l2arc_feed_secs seconds between L2ARC writing
3749 *
3750 * Tunables may be removed or added as future performance improvements are
3751 * integrated, and also may become zpool properties.
3752 *
3753 * There are three key functions that control how the L2ARC warms up:
3754 *
3755 * l2arc_write_eligible() check if a buffer is eligible to cache
3756 * l2arc_write_size() calculate how much to write
3757 * l2arc_write_interval() calculate sleep delay between writes
3758 *
3759 * These three functions determine what to write, how much, and how quickly
3760 * to send writes.
3761 */
3762
3763 static boolean_t
3764 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab)
3765 {
3766 /*
3767 * A buffer is *not* eligible for the L2ARC if it:
3768 * 1. belongs to a different spa.
3769 * 2. is already cached on the L2ARC.
3770 * 3. has an I/O in progress (it may be an incomplete read).
3771 * 4. is flagged not eligible (zfs property).
3772 */
3773 if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL ||
3774 HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab))
3775 return (B_FALSE);
3776
3777 return (B_TRUE);
3778 }
3779
3780 static uint64_t
3781 l2arc_write_size(l2arc_dev_t *dev)
3782 {
3783 uint64_t size;
3784
3785 size = dev->l2ad_write;
3786
3787 if (arc_warm == B_FALSE)
3788 size += dev->l2ad_boost;
3789
3790 return (size);
3791
3792 }
3793
3794 static clock_t
3795 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
3796 {
3797 clock_t interval, next, now;
3798
3799 /*
3800 * If the ARC lists are busy, increase our write rate; if the
3801 * lists are stale, idle back. This is achieved by checking
3802 * how much we previously wrote - if it was more than half of
3803 * what we wanted, schedule the next write much sooner.
3804 */
3805 if (l2arc_feed_again && wrote > (wanted / 2))
3806 interval = (hz * l2arc_feed_min_ms) / 1000;
3807 else
3808 interval = hz * l2arc_feed_secs;
3809
3810 now = ddi_get_lbolt();
3811 next = MAX(now, MIN(now + interval, began + interval));
3812
3813 return (next);
3814 }
3815
3816 static void
3817 l2arc_hdr_stat_add(void)
3818 {
3819 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
3820 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
3821 }
3822
3823 static void
3824 l2arc_hdr_stat_remove(void)
3825 {
3826 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
3827 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
3828 }
3829
3830 /*
3831 * Cycle through L2ARC devices. This is how L2ARC load balances.
3832 * If a device is returned, this also returns holding the spa config lock.
3833 */
3834 static l2arc_dev_t *
3835 l2arc_dev_get_next(void)
3836 {
3837 l2arc_dev_t *first, *next = NULL;
3838
3839 /*
3840 * Lock out the removal of spas (spa_namespace_lock), then removal
3841 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
3842 * both locks will be dropped and a spa config lock held instead.
3843 */
3844 mutex_enter(&spa_namespace_lock);
3845 mutex_enter(&l2arc_dev_mtx);
3846
3847 /* if there are no vdevs, there is nothing to do */
3848 if (l2arc_ndev == 0)
3849 goto out;
3850
3851 first = NULL;
3852 next = l2arc_dev_last;
3853 do {
3854 /* loop around the list looking for a non-faulted vdev */
3855 if (next == NULL) {
3856 next = list_head(l2arc_dev_list);
3857 } else {
3858 next = list_next(l2arc_dev_list, next);
3859 if (next == NULL)
3860 next = list_head(l2arc_dev_list);
3861 }
3862
3863 /* if we have come back to the start, bail out */
3864 if (first == NULL)
3865 first = next;
3866 else if (next == first)
3867 break;
3868
3869 } while (vdev_is_dead(next->l2ad_vdev));
3870
3871 /* if we were unable to find any usable vdevs, return NULL */
3872 if (vdev_is_dead(next->l2ad_vdev))
3873 next = NULL;
3874
3875 l2arc_dev_last = next;
3876
3877 out:
3878 mutex_exit(&l2arc_dev_mtx);
3879
3880 /*
3881 * Grab the config lock to prevent the 'next' device from being
3882 * removed while we are writing to it.
3883 */
3884 if (next != NULL)
3885 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
3886 mutex_exit(&spa_namespace_lock);
3887
3888 return (next);
3889 }
3890
3891 /*
3892 * Free buffers that were tagged for destruction.
3893 */
3894 static void
3895 l2arc_do_free_on_write()
3896 {
3897 list_t *buflist;
3898 l2arc_data_free_t *df, *df_prev;
3899
3900 mutex_enter(&l2arc_free_on_write_mtx);
3901 buflist = l2arc_free_on_write;
3902
3903 for (df = list_tail(buflist); df; df = df_prev) {
3904 df_prev = list_prev(buflist, df);
3905 ASSERT(df->l2df_data != NULL);
3906 ASSERT(df->l2df_func != NULL);
3907 df->l2df_func(df->l2df_data, df->l2df_size);
3908 list_remove(buflist, df);
3909 kmem_free(df, sizeof (l2arc_data_free_t));
3910 }
3911
3912 mutex_exit(&l2arc_free_on_write_mtx);
3913 }
3914
3915 /*
3916 * A write to a cache device has completed. Update all headers to allow
3917 * reads from these buffers to begin.
3918 */
3919 static void
3920 l2arc_write_done(zio_t *zio)
3921 {
3922 l2arc_write_callback_t *cb;
3923 l2arc_dev_t *dev;
3924 list_t *buflist;
3925 arc_buf_hdr_t *head, *ab, *ab_prev;
3926 l2arc_buf_hdr_t *abl2;
3927 kmutex_t *hash_lock;
3928
3929 cb = zio->io_private;
3930 ASSERT(cb != NULL);
3931 dev = cb->l2wcb_dev;
3932 ASSERT(dev != NULL);
3933 head = cb->l2wcb_head;
3934 ASSERT(head != NULL);
3935 buflist = dev->l2ad_buflist;
3936 ASSERT(buflist != NULL);
3937 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
3938 l2arc_write_callback_t *, cb);
3939
3940 if (zio->io_error != 0)
3941 ARCSTAT_BUMP(arcstat_l2_writes_error);
3942
3943 mutex_enter(&l2arc_buflist_mtx);
3944
3945 /*
3946 * All writes completed, or an error was hit.
3947 */
3948 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
3949 ab_prev = list_prev(buflist, ab);
3950
3951 hash_lock = HDR_LOCK(ab);
3952 if (!mutex_tryenter(hash_lock)) {
3953 /*
3954 * This buffer misses out. It may be in a stage
3955 * of eviction. Its ARC_L2_WRITING flag will be
3956 * left set, denying reads to this buffer.
3957 */
3958 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
3959 continue;
3960 }
3961
3962 if (zio->io_error != 0) {
3963 /*
3964 * Error - drop L2ARC entry.
3965 */
3966 list_remove(buflist, ab);
3967 abl2 = ab->b_l2hdr;
3968 ab->b_l2hdr = NULL;
3969 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
3970 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
3971 }
3972
3973 /*
3974 * Allow ARC to begin reads to this L2ARC entry.
3975 */
3976 ab->b_flags &= ~ARC_L2_WRITING;
3977
3978 mutex_exit(hash_lock);
3979 }
3980
3981 atomic_inc_64(&l2arc_writes_done);
3982 list_remove(buflist, head);
3983 kmem_cache_free(hdr_cache, head);
3984 mutex_exit(&l2arc_buflist_mtx);
3985
3986 l2arc_do_free_on_write();
3987
3988 kmem_free(cb, sizeof (l2arc_write_callback_t));
3989 }
3990
3991 /*
3992 * A read to a cache device completed. Validate buffer contents before
3993 * handing over to the regular ARC routines.
3994 */
3995 static void
3996 l2arc_read_done(zio_t *zio)
3997 {
3998 l2arc_read_callback_t *cb;
3999 arc_buf_hdr_t *hdr;
4000 arc_buf_t *buf;
4001 kmutex_t *hash_lock;
4002 int equal;
4003
4004 ASSERT(zio->io_vd != NULL);
4005 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4006
4007 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4008
4009 cb = zio->io_private;
4010 ASSERT(cb != NULL);
4011 buf = cb->l2rcb_buf;
4012 ASSERT(buf != NULL);
4013
4014 hash_lock = HDR_LOCK(buf->b_hdr);
4015 mutex_enter(hash_lock);
4016 hdr = buf->b_hdr;
4017 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4018
4019 /*
4020 * Check this survived the L2ARC journey.
4021 */
4022 equal = arc_cksum_equal(buf);
4023 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4024 mutex_exit(hash_lock);
4025 zio->io_private = buf;
4026 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4027 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4028 arc_read_done(zio);
4029 } else {
4030 mutex_exit(hash_lock);
4031 /*
4032 * Buffer didn't survive caching. Increment stats and
4033 * reissue to the original storage device.
4034 */
4035 if (zio->io_error != 0) {
4036 ARCSTAT_BUMP(arcstat_l2_io_error);
4037 } else {
4038 zio->io_error = EIO;
4039 }
4040 if (!equal)
4041 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4042
4043 /*
4044 * If there's no waiter, issue an async i/o to the primary
4045 * storage now. If there *is* a waiter, the caller must
4046 * issue the i/o in a context where it's OK to block.
4047 */
4048 if (zio->io_waiter == NULL) {
4049 zio_t *pio = zio_unique_parent(zio);
4050
4051 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
4052
4053 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
4054 buf->b_data, zio->io_size, arc_read_done, buf,
4055 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4056 }
4057 }
4058
4059 kmem_free(cb, sizeof (l2arc_read_callback_t));
4060 }
4061
4062 /*
4063 * This is the list priority from which the L2ARC will search for pages to
4064 * cache. This is used within loops (0..3) to cycle through lists in the
4065 * desired order. This order can have a significant effect on cache
4066 * performance.
4067 *
4068 * Currently the metadata lists are hit first, MFU then MRU, followed by
4069 * the data lists. This function returns a locked list, and also returns
4070 * the lock pointer.
4071 */
4072 static list_t *
4073 l2arc_list_locked(int list_num, kmutex_t **lock)
4074 {
4075 list_t *list;
4076
4077 ASSERT(list_num >= 0 && list_num <= 3);
4078
4079 switch (list_num) {
4080 case 0:
4081 list = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
4082 *lock = &arc_mfu->arcs_mtx;
4083 break;
4084 case 1:
4085 list = &arc_mru->arcs_list[ARC_BUFC_METADATA];
4086 *lock = &arc_mru->arcs_mtx;
4087 break;
4088 case 2:
4089 list = &arc_mfu->arcs_list[ARC_BUFC_DATA];
4090 *lock = &arc_mfu->arcs_mtx;
4091 break;
4092 case 3:
4093 list = &arc_mru->arcs_list[ARC_BUFC_DATA];
4094 *lock = &arc_mru->arcs_mtx;
4095 break;
4096 }
4097
4098 ASSERT(!(MUTEX_HELD(*lock)));
4099 mutex_enter(*lock);
4100 return (list);
4101 }
4102
4103 /*
4104 * Evict buffers from the device write hand to the distance specified in
4105 * bytes. This distance may span populated buffers, it may span nothing.
4106 * This is clearing a region on the L2ARC device ready for writing.
4107 * If the 'all' boolean is set, every buffer is evicted.
4108 */
4109 static void
4110 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4111 {
4112 list_t *buflist;
4113 l2arc_buf_hdr_t *abl2;
4114 arc_buf_hdr_t *ab, *ab_prev;
4115 kmutex_t *hash_lock;
4116 uint64_t taddr;
4117
4118 buflist = dev->l2ad_buflist;
4119
4120 if (buflist == NULL)
4121 return;
4122
4123 if (!all && dev->l2ad_first) {
4124 /*
4125 * This is the first sweep through the device. There is
4126 * nothing to evict.
4127 */
4128 return;
4129 }
4130
4131 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4132 /*
4133 * When nearing the end of the device, evict to the end
4134 * before the device write hand jumps to the start.
4135 */
4136 taddr = dev->l2ad_end;
4137 } else {
4138 taddr = dev->l2ad_hand + distance;
4139 }
4140 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4141 uint64_t, taddr, boolean_t, all);
4142
4143 top:
4144 mutex_enter(&l2arc_buflist_mtx);
4145 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4146 ab_prev = list_prev(buflist, ab);
4147
4148 hash_lock = HDR_LOCK(ab);
4149 if (!mutex_tryenter(hash_lock)) {
4150 /*
4151 * Missed the hash lock. Retry.
4152 */
4153 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4154 mutex_exit(&l2arc_buflist_mtx);
4155 mutex_enter(hash_lock);
4156 mutex_exit(hash_lock);
4157 goto top;
4158 }
4159
4160 if (HDR_L2_WRITE_HEAD(ab)) {
4161 /*
4162 * We hit a write head node. Leave it for
4163 * l2arc_write_done().
4164 */
4165 list_remove(buflist, ab);
4166 mutex_exit(hash_lock);
4167 continue;
4168 }
4169
4170 if (!all && ab->b_l2hdr != NULL &&
4171 (ab->b_l2hdr->b_daddr > taddr ||
4172 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4173 /*
4174 * We've evicted to the target address,
4175 * or the end of the device.
4176 */
4177 mutex_exit(hash_lock);
4178 break;
4179 }
4180
4181 if (HDR_FREE_IN_PROGRESS(ab)) {
4182 /*
4183 * Already on the path to destruction.
4184 */
4185 mutex_exit(hash_lock);
4186 continue;
4187 }
4188
4189 if (ab->b_state == arc_l2c_only) {
4190 ASSERT(!HDR_L2_READING(ab));
4191 /*
4192 * This doesn't exist in the ARC. Destroy.
4193 * arc_hdr_destroy() will call list_remove()
4194 * and decrement arcstat_l2_size.
4195 */
4196 arc_change_state(arc_anon, ab, hash_lock);
4197 arc_hdr_destroy(ab);
4198 } else {
4199 /*
4200 * Invalidate issued or about to be issued
4201 * reads, since we may be about to write
4202 * over this location.
4203 */
4204 if (HDR_L2_READING(ab)) {
4205 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4206 ab->b_flags |= ARC_L2_EVICTED;
4207 }
4208
4209 /*
4210 * Tell ARC this no longer exists in L2ARC.
4211 */
4212 if (ab->b_l2hdr != NULL) {
4213 abl2 = ab->b_l2hdr;
4214 ab->b_l2hdr = NULL;
4215 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4216 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4217 }
4218 list_remove(buflist, ab);
4219
4220 /*
4221 * This may have been leftover after a
4222 * failed write.
4223 */
4224 ab->b_flags &= ~ARC_L2_WRITING;
4225 }
4226 mutex_exit(hash_lock);
4227 }
4228 mutex_exit(&l2arc_buflist_mtx);
4229
4230 vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0);
4231 dev->l2ad_evict = taddr;
4232 }
4233
4234 /*
4235 * Find and write ARC buffers to the L2ARC device.
4236 *
4237 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4238 * for reading until they have completed writing.
4239 */
4240 static uint64_t
4241 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4242 {
4243 arc_buf_hdr_t *ab, *ab_prev, *head;
4244 l2arc_buf_hdr_t *hdrl2;
4245 list_t *list;
4246 uint64_t passed_sz, write_sz, buf_sz, headroom;
4247 void *buf_data;
4248 kmutex_t *hash_lock, *list_lock;
4249 boolean_t have_lock, full;
4250 l2arc_write_callback_t *cb;
4251 zio_t *pio, *wzio;
4252 uint64_t guid = spa_load_guid(spa);
4253
4254 ASSERT(dev->l2ad_vdev != NULL);
4255
4256 pio = NULL;
4257 write_sz = 0;
4258 full = B_FALSE;
4259 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4260 head->b_flags |= ARC_L2_WRITE_HEAD;
4261
4262 /*
4263 * Copy buffers for L2ARC writing.
4264 */
4265 mutex_enter(&l2arc_buflist_mtx);
4266 for (int try = 0; try <= 3; try++) {
4267 list = l2arc_list_locked(try, &list_lock);
4268 passed_sz = 0;
4269
4270 /*
4271 * L2ARC fast warmup.
4272 *
4273 * Until the ARC is warm and starts to evict, read from the
4274 * head of the ARC lists rather than the tail.
4275 */
4276 headroom = target_sz * l2arc_headroom;
4277 if (arc_warm == B_FALSE)
4278 ab = list_head(list);
4279 else
4280 ab = list_tail(list);
4281
4282 for (; ab; ab = ab_prev) {
4283 if (arc_warm == B_FALSE)
4284 ab_prev = list_next(list, ab);
4285 else
4286 ab_prev = list_prev(list, ab);
4287
4288 hash_lock = HDR_LOCK(ab);
4289 have_lock = MUTEX_HELD(hash_lock);
4290 if (!have_lock && !mutex_tryenter(hash_lock)) {
4291 /*
4292 * Skip this buffer rather than waiting.
4293 */
4294 continue;
4295 }
4296
4297 passed_sz += ab->b_size;
4298 if (passed_sz > headroom) {
4299 /*
4300 * Searched too far.
4301 */
4302 mutex_exit(hash_lock);
4303 break;
4304 }
4305
4306 if (!l2arc_write_eligible(guid, ab)) {
4307 mutex_exit(hash_lock);
4308 continue;
4309 }
4310
4311 if ((write_sz + ab->b_size) > target_sz) {
4312 full = B_TRUE;
4313 mutex_exit(hash_lock);
4314 break;
4315 }
4316
4317 if (pio == NULL) {
4318 /*
4319 * Insert a dummy header on the buflist so
4320 * l2arc_write_done() can find where the
4321 * write buffers begin without searching.
4322 */
4323 list_insert_head(dev->l2ad_buflist, head);
4324
4325 cb = kmem_alloc(
4326 sizeof (l2arc_write_callback_t), KM_SLEEP);
4327 cb->l2wcb_dev = dev;
4328 cb->l2wcb_head = head;
4329 pio = zio_root(spa, l2arc_write_done, cb,
4330 ZIO_FLAG_CANFAIL);
4331 }
4332
4333 /*
4334 * Create and add a new L2ARC header.
4335 */
4336 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4337 hdrl2->b_dev = dev;
4338 hdrl2->b_daddr = dev->l2ad_hand;
4339
4340 ab->b_flags |= ARC_L2_WRITING;
4341 ab->b_l2hdr = hdrl2;
4342 list_insert_head(dev->l2ad_buflist, ab);
4343 buf_data = ab->b_buf->b_data;
4344 buf_sz = ab->b_size;
4345
4346 /*
4347 * Compute and store the buffer cksum before
4348 * writing. On debug the cksum is verified first.
4349 */
4350 arc_cksum_verify(ab->b_buf);
4351 arc_cksum_compute(ab->b_buf, B_TRUE);
4352
4353 mutex_exit(hash_lock);
4354
4355 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4356 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4357 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4358 ZIO_FLAG_CANFAIL, B_FALSE);
4359
4360 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4361 zio_t *, wzio);
4362 (void) zio_nowait(wzio);
4363
4364 /*
4365 * Keep the clock hand suitably device-aligned.
4366 */
4367 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4368
4369 write_sz += buf_sz;
4370 dev->l2ad_hand += buf_sz;
4371 }
4372
4373 mutex_exit(list_lock);
4374
4375 if (full == B_TRUE)
4376 break;
4377 }
4378 mutex_exit(&l2arc_buflist_mtx);
4379
4380 if (pio == NULL) {
4381 ASSERT3U(write_sz, ==, 0);
4382 kmem_cache_free(hdr_cache, head);
4383 return (0);
4384 }
4385
4386 ASSERT3U(write_sz, <=, target_sz);
4387 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4388 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4389 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4390 vdev_space_update(dev->l2ad_vdev, write_sz, 0, 0);
4391
4392 /*
4393 * Bump device hand to the device start if it is approaching the end.
4394 * l2arc_evict() will already have evicted ahead for this case.
4395 */
4396 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4397 vdev_space_update(dev->l2ad_vdev,
4398 dev->l2ad_end - dev->l2ad_hand, 0, 0);
4399 dev->l2ad_hand = dev->l2ad_start;
4400 dev->l2ad_evict = dev->l2ad_start;
4401 dev->l2ad_first = B_FALSE;
4402 }
4403
4404 dev->l2ad_writing = B_TRUE;
4405 (void) zio_wait(pio);
4406 dev->l2ad_writing = B_FALSE;
4407
4408 return (write_sz);
4409 }
4410
4411 /*
4412 * This thread feeds the L2ARC at regular intervals. This is the beating
4413 * heart of the L2ARC.
4414 */
4415 static void
4416 l2arc_feed_thread(void)
4417 {
4418 callb_cpr_t cpr;
4419 l2arc_dev_t *dev;
4420 spa_t *spa;
4421 uint64_t size, wrote;
4422 clock_t begin, next = ddi_get_lbolt();
4423
4424 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4425
4426 mutex_enter(&l2arc_feed_thr_lock);
4427
4428 while (l2arc_thread_exit == 0) {
4429 CALLB_CPR_SAFE_BEGIN(&cpr);
4430 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4431 next);
4432 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4433 next = ddi_get_lbolt() + hz;
4434
4435 /*
4436 * Quick check for L2ARC devices.
4437 */
4438 mutex_enter(&l2arc_dev_mtx);
4439 if (l2arc_ndev == 0) {
4440 mutex_exit(&l2arc_dev_mtx);
4441 continue;
4442 }
4443 mutex_exit(&l2arc_dev_mtx);
4444 begin = ddi_get_lbolt();
4445
4446 /*
4447 * This selects the next l2arc device to write to, and in
4448 * doing so the next spa to feed from: dev->l2ad_spa. This
4449 * will return NULL if there are now no l2arc devices or if
4450 * they are all faulted.
4451 *
4452 * If a device is returned, its spa's config lock is also
4453 * held to prevent device removal. l2arc_dev_get_next()
4454 * will grab and release l2arc_dev_mtx.
4455 */
4456 if ((dev = l2arc_dev_get_next()) == NULL)
4457 continue;
4458
4459 spa = dev->l2ad_spa;
4460 ASSERT(spa != NULL);
4461
4462 /*
4463 * If the pool is read-only then force the feed thread to
4464 * sleep a little longer.
4465 */
4466 if (!spa_writeable(spa)) {
4467 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
4468 spa_config_exit(spa, SCL_L2ARC, dev);
4469 continue;
4470 }
4471
4472 /*
4473 * Avoid contributing to memory pressure.
4474 */
4475 if (arc_reclaim_needed()) {
4476 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4477 spa_config_exit(spa, SCL_L2ARC, dev);
4478 continue;
4479 }
4480
4481 ARCSTAT_BUMP(arcstat_l2_feeds);
4482
4483 size = l2arc_write_size(dev);
4484
4485 /*
4486 * Evict L2ARC buffers that will be overwritten.
4487 */
4488 l2arc_evict(dev, size, B_FALSE);
4489
4490 /*
4491 * Write ARC buffers.
4492 */
4493 wrote = l2arc_write_buffers(spa, dev, size);
4494
4495 /*
4496 * Calculate interval between writes.
4497 */
4498 next = l2arc_write_interval(begin, size, wrote);
4499 spa_config_exit(spa, SCL_L2ARC, dev);
4500 }
4501
4502 l2arc_thread_exit = 0;
4503 cv_broadcast(&l2arc_feed_thr_cv);
4504 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4505 thread_exit();
4506 }
4507
4508 boolean_t
4509 l2arc_vdev_present(vdev_t *vd)
4510 {
4511 l2arc_dev_t *dev;
4512
4513 mutex_enter(&l2arc_dev_mtx);
4514 for (dev = list_head(l2arc_dev_list); dev != NULL;
4515 dev = list_next(l2arc_dev_list, dev)) {
4516 if (dev->l2ad_vdev == vd)
4517 break;
4518 }
4519 mutex_exit(&l2arc_dev_mtx);
4520
4521 return (dev != NULL);
4522 }
4523
4524 /*
4525 * Add a vdev for use by the L2ARC. By this point the spa has already
4526 * validated the vdev and opened it.
4527 */
4528 void
4529 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
4530 {
4531 l2arc_dev_t *adddev;
4532
4533 ASSERT(!l2arc_vdev_present(vd));
4534
4535 /*
4536 * Create a new l2arc device entry.
4537 */
4538 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
4539 adddev->l2ad_spa = spa;
4540 adddev->l2ad_vdev = vd;
4541 adddev->l2ad_write = l2arc_write_max;
4542 adddev->l2ad_boost = l2arc_write_boost;
4543 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
4544 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
4545 adddev->l2ad_hand = adddev->l2ad_start;
4546 adddev->l2ad_evict = adddev->l2ad_start;
4547 adddev->l2ad_first = B_TRUE;
4548 adddev->l2ad_writing = B_FALSE;
4549 ASSERT3U(adddev->l2ad_write, >, 0);
4550
4551 /*
4552 * This is a list of all ARC buffers that are still valid on the
4553 * device.
4554 */
4555 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
4556 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
4557 offsetof(arc_buf_hdr_t, b_l2node));
4558
4559 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
4560
4561 /*
4562 * Add device to global list
4563 */
4564 mutex_enter(&l2arc_dev_mtx);
4565 list_insert_head(l2arc_dev_list, adddev);
4566 atomic_inc_64(&l2arc_ndev);
4567 mutex_exit(&l2arc_dev_mtx);
4568 }
4569
4570 /*
4571 * Remove a vdev from the L2ARC.
4572 */
4573 void
4574 l2arc_remove_vdev(vdev_t *vd)
4575 {
4576 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
4577
4578 /*
4579 * Find the device by vdev
4580 */
4581 mutex_enter(&l2arc_dev_mtx);
4582 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
4583 nextdev = list_next(l2arc_dev_list, dev);
4584 if (vd == dev->l2ad_vdev) {
4585 remdev = dev;
4586 break;
4587 }
4588 }
4589 ASSERT(remdev != NULL);
4590
4591 /*
4592 * Remove device from global list
4593 */
4594 list_remove(l2arc_dev_list, remdev);
4595 l2arc_dev_last = NULL; /* may have been invalidated */
4596 atomic_dec_64(&l2arc_ndev);
4597 mutex_exit(&l2arc_dev_mtx);
4598
4599 /*
4600 * Clear all buflists and ARC references. L2ARC device flush.
4601 */
4602 l2arc_evict(remdev, 0, B_TRUE);
4603 list_destroy(remdev->l2ad_buflist);
4604 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
4605 kmem_free(remdev, sizeof (l2arc_dev_t));
4606 }
4607
4608 void
4609 l2arc_init(void)
4610 {
4611 l2arc_thread_exit = 0;
4612 l2arc_ndev = 0;
4613 l2arc_writes_sent = 0;
4614 l2arc_writes_done = 0;
4615
4616 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4617 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
4618 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
4619 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
4620 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
4621
4622 l2arc_dev_list = &L2ARC_dev_list;
4623 l2arc_free_on_write = &L2ARC_free_on_write;
4624 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
4625 offsetof(l2arc_dev_t, l2ad_node));
4626 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
4627 offsetof(l2arc_data_free_t, l2df_list_node));
4628 }
4629
4630 void
4631 l2arc_fini(void)
4632 {
4633 /*
4634 * This is called from dmu_fini(), which is called from spa_fini();
4635 * Because of this, we can assume that all l2arc devices have
4636 * already been removed when the pools themselves were removed.
4637 */
4638
4639 l2arc_do_free_on_write();
4640
4641 mutex_destroy(&l2arc_feed_thr_lock);
4642 cv_destroy(&l2arc_feed_thr_cv);
4643 mutex_destroy(&l2arc_dev_mtx);
4644 mutex_destroy(&l2arc_buflist_mtx);
4645 mutex_destroy(&l2arc_free_on_write_mtx);
4646
4647 list_destroy(l2arc_dev_list);
4648 list_destroy(l2arc_free_on_write);
4649 }
4650
4651 void
4652 l2arc_start(void)
4653 {
4654 if (!(spa_mode_global & FWRITE))
4655 return;
4656
4657 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
4658 TS_RUN, minclsyspri);
4659 }
4660
4661 void
4662 l2arc_stop(void)
4663 {
4664 if (!(spa_mode_global & FWRITE))
4665 return;
4666
4667 mutex_enter(&l2arc_feed_thr_lock);
4668 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
4669 l2arc_thread_exit = 1;
4670 while (l2arc_thread_exit != 0)
4671 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
4672 mutex_exit(&l2arc_feed_thr_lock);
4673 }