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