1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
27 */
28
29 /*
30 * DVA-based Adjustable Replacement Cache
31 *
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
36 *
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
47 *
48 * There are times when it is not possible to evict the requested
49 * space. In these circumstances we are unable to adjust the cache
50 * size. To prevent the cache growing unbounded at these times we
51 * implement a "cache throttle" that slows the flow of new data
52 * into the cache until we can make space available.
53 *
54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 * Pages are evicted when the cache is full and there is a cache
56 * miss. Our model has a variable sized cache. It grows with
57 * high use, but also tries to react to memory pressure from the
58 * operating system: decreasing its size when system memory is
59 * tight.
60 *
61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 * elements of the cache are therefore exactly the same size. So
63 * when adjusting the cache size following a cache miss, its simply
64 * a matter of choosing a single page to evict. In our model, we
65 * have variable sized cache blocks (rangeing from 512 bytes to
66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 * space for a cache miss that approximates as closely as possible
68 * the space used by the new block.
69 *
70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 * by N. Megiddo & D. Modha, FAST 2003
72 */
73
74 /*
75 * The locking model:
76 *
77 * A new reference to a cache buffer can be obtained in two
78 * ways: 1) via a hash table lookup using the DVA as a key,
79 * or 2) via one of the ARC lists. The arc_read() interface
80 * uses method 1, while the internal arc algorithms for
81 * adjusting the cache use method 2. We therefore provide two
82 * types of locks: 1) the hash table lock array, and 2) the
83 * arc list locks.
84 *
85 * Buffers do not have their own mutexes, rather they rely on the
86 * hash table mutexes for the bulk of their protection (i.e. most
87 * fields in the arc_buf_hdr_t are protected by these mutexes).
88 *
89 * buf_hash_find() returns the appropriate mutex (held) when it
90 * locates the requested buffer in the hash table. It returns
91 * NULL for the mutex if the buffer was not in the table.
92 *
93 * buf_hash_remove() expects the appropriate hash mutex to be
94 * already held before it is invoked.
95 *
96 * Each arc state also has a mutex which is used to protect the
97 * buffer list associated with the state. When attempting to
98 * obtain a hash table lock while holding an arc list lock you
99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 * the active state mutex must be held before the ghost state mutex.
101 *
102 * Arc buffers may have an associated eviction callback function.
103 * This function will be invoked prior to removing the buffer (e.g.
104 * in arc_do_user_evicts()). Note however that the data associated
105 * with the buffer may be evicted prior to the callback. The callback
106 * must be made with *no locks held* (to prevent deadlock). Additionally,
107 * the users of callbacks must ensure that their private data is
108 * protected from simultaneous callbacks from arc_clear_callback()
109 * and arc_do_user_evicts().
110 *
111 * Note that the majority of the performance stats are manipulated
112 * with atomic operations.
113 *
114 * The L2ARC uses the l2ad_mtx on each vdev for the following:
115 *
116 * - L2ARC buflist creation
117 * - L2ARC buflist eviction
118 * - L2ARC write completion, which walks L2ARC buflists
119 * - ARC header destruction, as it removes from L2ARC buflists
120 * - ARC header release, as it removes from L2ARC buflists
121 */
122
123 #include <sys/spa.h>
124 #include <sys/zio.h>
125 #include <sys/zio_compress.h>
126 #include <sys/zfs_context.h>
127 #include <sys/arc.h>
128 #include <sys/refcount.h>
129 #include <sys/vdev.h>
130 #include <sys/vdev_impl.h>
131 #include <sys/dsl_pool.h>
132 #include <sys/multilist.h>
133 #ifdef _KERNEL
134 #include <sys/vmsystm.h>
135 #include <vm/anon.h>
136 #include <sys/fs/swapnode.h>
137 #include <sys/dnlc.h>
138 #endif
139 #include <sys/callb.h>
140 #include <sys/kstat.h>
141 #include <zfs_fletcher.h>
142
143 #ifndef _KERNEL
144 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
145 boolean_t arc_watch = B_FALSE;
146 int arc_procfd;
147 #endif
148
149 static kmutex_t arc_reclaim_lock;
150 static kcondvar_t arc_reclaim_thread_cv;
151 static boolean_t arc_reclaim_thread_exit;
152 static kcondvar_t arc_reclaim_waiters_cv;
153
154 static kmutex_t arc_user_evicts_lock;
155 static kcondvar_t arc_user_evicts_cv;
156 static boolean_t arc_user_evicts_thread_exit;
157
158 uint_t arc_reduce_dnlc_percent = 3;
159
160 /*
161 * The number of headers to evict in arc_evict_state_impl() before
162 * dropping the sublist lock and evicting from another sublist. A lower
163 * value means we're more likely to evict the "correct" header (i.e. the
164 * oldest header in the arc state), but comes with higher overhead
165 * (i.e. more invocations of arc_evict_state_impl()).
166 */
167 int zfs_arc_evict_batch_limit = 10;
168
169 /*
170 * The number of sublists used for each of the arc state lists. If this
171 * is not set to a suitable value by the user, it will be configured to
172 * the number of CPUs on the system in arc_init().
173 */
174 int zfs_arc_num_sublists_per_state = 0;
175
176 /* number of seconds before growing cache again */
177 static int arc_grow_retry = 60;
178
179 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
180 int zfs_arc_overflow_shift = 8;
181
182 /* shift of arc_c for calculating both min and max arc_p */
183 static int arc_p_min_shift = 4;
184
185 /* log2(fraction of arc to reclaim) */
186 static int arc_shrink_shift = 7;
187
188 /*
189 * log2(fraction of ARC which must be free to allow growing).
190 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
191 * when reading a new block into the ARC, we will evict an equal-sized block
192 * from the ARC.
193 *
194 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
195 * we will still not allow it to grow.
196 */
197 int arc_no_grow_shift = 5;
198
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 uint64_t zfs_arc_meta_min = 0;
225 int zfs_arc_grow_retry = 0;
226 int zfs_arc_shrink_shift = 0;
227 int zfs_arc_p_min_shift = 0;
228 int zfs_disable_dup_eviction = 0;
229 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
230
231 /*
232 * Note that buffers can be in one of 6 states:
233 * ARC_anon - anonymous (discussed below)
234 * ARC_mru - recently used, currently cached
235 * ARC_mru_ghost - recentely used, no longer in cache
236 * ARC_mfu - frequently used, currently cached
237 * ARC_mfu_ghost - frequently used, no longer in cache
238 * ARC_l2c_only - exists in L2ARC but not other states
239 * When there are no active references to the buffer, they are
240 * are linked onto a list in one of these arc states. These are
241 * the only buffers that can be evicted or deleted. Within each
242 * state there are multiple lists, one for meta-data and one for
243 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
244 * etc.) is tracked separately so that it can be managed more
245 * explicitly: favored over data, limited explicitly.
246 *
247 * Anonymous buffers are buffers that are not associated with
248 * a DVA. These are buffers that hold dirty block copies
249 * before they are written to stable storage. By definition,
250 * they are "ref'd" and are considered part of arc_mru
251 * that cannot be freed. Generally, they will aquire a DVA
252 * as they are written and migrate onto the arc_mru list.
253 *
254 * The ARC_l2c_only state is for buffers that are in the second
255 * level ARC but no longer in any of the ARC_m* lists. The second
256 * level ARC itself may also contain buffers that are in any of
257 * the ARC_m* states - meaning that a buffer can exist in two
258 * places. The reason for the ARC_l2c_only state is to keep the
259 * buffer header in the hash table, so that reads that hit the
260 * second level ARC benefit from these fast lookups.
261 */
262
263 typedef struct arc_state {
264 /*
265 * list of evictable buffers
266 */
267 multilist_t arcs_list[ARC_BUFC_NUMTYPES];
268 /*
269 * total amount of evictable data in this state
270 */
271 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES];
272 /*
273 * total amount of data in this state; this includes: evictable,
274 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
275 */
276 refcount_t arcs_size;
277 } arc_state_t;
278
279 /* The 6 states: */
280 static arc_state_t ARC_anon;
281 static arc_state_t ARC_mru;
282 static arc_state_t ARC_mru_ghost;
283 static arc_state_t ARC_mfu;
284 static arc_state_t ARC_mfu_ghost;
285 static arc_state_t ARC_l2c_only;
286
287 typedef struct arc_stats {
288 kstat_named_t arcstat_hits;
289 kstat_named_t arcstat_misses;
290 kstat_named_t arcstat_demand_data_hits;
291 kstat_named_t arcstat_demand_data_misses;
292 kstat_named_t arcstat_demand_metadata_hits;
293 kstat_named_t arcstat_demand_metadata_misses;
294 kstat_named_t arcstat_prefetch_data_hits;
295 kstat_named_t arcstat_prefetch_data_misses;
296 kstat_named_t arcstat_prefetch_metadata_hits;
297 kstat_named_t arcstat_prefetch_metadata_misses;
298 kstat_named_t arcstat_mru_hits;
299 kstat_named_t arcstat_mru_ghost_hits;
300 kstat_named_t arcstat_mfu_hits;
301 kstat_named_t arcstat_mfu_ghost_hits;
302 kstat_named_t arcstat_deleted;
303 /*
304 * Number of buffers that could not be evicted because the hash lock
305 * was held by another thread. The lock may not necessarily be held
306 * by something using the same buffer, since hash locks are shared
307 * by multiple buffers.
308 */
309 kstat_named_t arcstat_mutex_miss;
310 /*
311 * Number of buffers skipped because they have I/O in progress, are
312 * indrect prefetch buffers that have not lived long enough, or are
313 * not from the spa we're trying to evict from.
314 */
315 kstat_named_t arcstat_evict_skip;
316 /*
317 * Number of times arc_evict_state() was unable to evict enough
318 * buffers to reach it's target amount.
319 */
320 kstat_named_t arcstat_evict_not_enough;
321 kstat_named_t arcstat_evict_l2_cached;
322 kstat_named_t arcstat_evict_l2_eligible;
323 kstat_named_t arcstat_evict_l2_ineligible;
324 kstat_named_t arcstat_evict_l2_skip;
325 kstat_named_t arcstat_hash_elements;
326 kstat_named_t arcstat_hash_elements_max;
327 kstat_named_t arcstat_hash_collisions;
328 kstat_named_t arcstat_hash_chains;
329 kstat_named_t arcstat_hash_chain_max;
330 kstat_named_t arcstat_p;
331 kstat_named_t arcstat_c;
332 kstat_named_t arcstat_c_min;
333 kstat_named_t arcstat_c_max;
334 kstat_named_t arcstat_size;
335 /*
336 * Number of bytes consumed by internal ARC structures necessary
337 * for tracking purposes; these structures are not actually
338 * backed by ARC buffers. This includes arc_buf_hdr_t structures
339 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
340 * caches), and arc_buf_t structures (allocated via arc_buf_t
341 * cache).
342 */
343 kstat_named_t arcstat_hdr_size;
344 /*
345 * Number of bytes consumed by ARC buffers of type equal to
346 * ARC_BUFC_DATA. This is generally consumed by buffers backing
347 * on disk user data (e.g. plain file contents).
348 */
349 kstat_named_t arcstat_data_size;
350 /*
351 * Number of bytes consumed by ARC buffers of type equal to
352 * ARC_BUFC_METADATA. This is generally consumed by buffers
353 * backing on disk data that is used for internal ZFS
354 * structures (e.g. ZAP, dnode, indirect blocks, etc).
355 */
356 kstat_named_t arcstat_metadata_size;
357 /*
358 * Number of bytes consumed by various buffers and structures
359 * not actually backed with ARC buffers. This includes bonus
360 * buffers (allocated directly via zio_buf_* functions),
361 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
362 * cache), and dnode_t structures (allocated via dnode_t cache).
363 */
364 kstat_named_t arcstat_other_size;
365 /*
366 * Total number of bytes consumed by ARC buffers residing in the
367 * arc_anon state. This includes *all* buffers in the arc_anon
368 * state; e.g. data, metadata, evictable, and unevictable buffers
369 * are all included in this value.
370 */
371 kstat_named_t arcstat_anon_size;
372 /*
373 * Number of bytes consumed by ARC buffers that meet the
374 * following criteria: backing buffers of type ARC_BUFC_DATA,
375 * residing in the arc_anon state, and are eligible for eviction
376 * (e.g. have no outstanding holds on the buffer).
377 */
378 kstat_named_t arcstat_anon_evictable_data;
379 /*
380 * Number of bytes consumed by ARC buffers that meet the
381 * following criteria: backing buffers of type ARC_BUFC_METADATA,
382 * residing in the arc_anon state, and are eligible for eviction
383 * (e.g. have no outstanding holds on the buffer).
384 */
385 kstat_named_t arcstat_anon_evictable_metadata;
386 /*
387 * Total number of bytes consumed by ARC buffers residing in the
388 * arc_mru state. This includes *all* buffers in the arc_mru
389 * state; e.g. data, metadata, evictable, and unevictable buffers
390 * are all included in this value.
391 */
392 kstat_named_t arcstat_mru_size;
393 /*
394 * Number of bytes consumed by ARC buffers that meet the
395 * following criteria: backing buffers of type ARC_BUFC_DATA,
396 * residing in the arc_mru state, and are eligible for eviction
397 * (e.g. have no outstanding holds on the buffer).
398 */
399 kstat_named_t arcstat_mru_evictable_data;
400 /*
401 * Number of bytes consumed by ARC buffers that meet the
402 * following criteria: backing buffers of type ARC_BUFC_METADATA,
403 * residing in the arc_mru state, and are eligible for eviction
404 * (e.g. have no outstanding holds on the buffer).
405 */
406 kstat_named_t arcstat_mru_evictable_metadata;
407 /*
408 * Total number of bytes that *would have been* consumed by ARC
409 * buffers in the arc_mru_ghost state. The key thing to note
410 * here, is the fact that this size doesn't actually indicate
411 * RAM consumption. The ghost lists only consist of headers and
412 * don't actually have ARC buffers linked off of these headers.
413 * Thus, *if* the headers had associated ARC buffers, these
414 * buffers *would have* consumed this number of bytes.
415 */
416 kstat_named_t arcstat_mru_ghost_size;
417 /*
418 * Number of bytes that *would have been* consumed by ARC
419 * buffers that are eligible for eviction, of type
420 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
421 */
422 kstat_named_t arcstat_mru_ghost_evictable_data;
423 /*
424 * Number of bytes that *would have been* consumed by ARC
425 * buffers that are eligible for eviction, of type
426 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
427 */
428 kstat_named_t arcstat_mru_ghost_evictable_metadata;
429 /*
430 * Total number of bytes consumed by ARC buffers residing in the
431 * arc_mfu state. This includes *all* buffers in the arc_mfu
432 * state; e.g. data, metadata, evictable, and unevictable buffers
433 * are all included in this value.
434 */
435 kstat_named_t arcstat_mfu_size;
436 /*
437 * Number of bytes consumed by ARC buffers that are eligible for
438 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
439 * state.
440 */
441 kstat_named_t arcstat_mfu_evictable_data;
442 /*
443 * Number of bytes consumed by ARC buffers that are eligible for
444 * eviction, of type ARC_BUFC_METADATA, and reside in the
445 * arc_mfu state.
446 */
447 kstat_named_t arcstat_mfu_evictable_metadata;
448 /*
449 * Total number of bytes that *would have been* consumed by ARC
450 * buffers in the arc_mfu_ghost state. See the comment above
451 * arcstat_mru_ghost_size for more details.
452 */
453 kstat_named_t arcstat_mfu_ghost_size;
454 /*
455 * Number of bytes that *would have been* consumed by ARC
456 * buffers that are eligible for eviction, of type
457 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
458 */
459 kstat_named_t arcstat_mfu_ghost_evictable_data;
460 /*
461 * Number of bytes that *would have been* consumed by ARC
462 * buffers that are eligible for eviction, of type
463 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
464 */
465 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
466 kstat_named_t arcstat_l2_hits;
467 kstat_named_t arcstat_l2_misses;
468 kstat_named_t arcstat_l2_feeds;
469 kstat_named_t arcstat_l2_rw_clash;
470 kstat_named_t arcstat_l2_read_bytes;
471 kstat_named_t arcstat_l2_write_bytes;
472 kstat_named_t arcstat_l2_writes_sent;
473 kstat_named_t arcstat_l2_writes_done;
474 kstat_named_t arcstat_l2_writes_error;
475 kstat_named_t arcstat_l2_writes_lock_retry;
476 kstat_named_t arcstat_l2_evict_lock_retry;
477 kstat_named_t arcstat_l2_evict_reading;
478 kstat_named_t arcstat_l2_evict_l1cached;
479 kstat_named_t arcstat_l2_free_on_write;
480 kstat_named_t arcstat_l2_cdata_free_on_write;
481 kstat_named_t arcstat_l2_abort_lowmem;
482 kstat_named_t arcstat_l2_cksum_bad;
483 kstat_named_t arcstat_l2_io_error;
484 kstat_named_t arcstat_l2_size;
485 kstat_named_t arcstat_l2_asize;
486 kstat_named_t arcstat_l2_hdr_size;
487 kstat_named_t arcstat_l2_compress_successes;
488 kstat_named_t arcstat_l2_compress_zeros;
489 kstat_named_t arcstat_l2_compress_failures;
490 kstat_named_t arcstat_memory_throttle_count;
491 kstat_named_t arcstat_duplicate_buffers;
492 kstat_named_t arcstat_duplicate_buffers_size;
493 kstat_named_t arcstat_duplicate_reads;
494 kstat_named_t arcstat_meta_used;
495 kstat_named_t arcstat_meta_limit;
496 kstat_named_t arcstat_meta_max;
497 kstat_named_t arcstat_meta_min;
498 } arc_stats_t;
499
500 static arc_stats_t arc_stats = {
501 { "hits", KSTAT_DATA_UINT64 },
502 { "misses", KSTAT_DATA_UINT64 },
503 { "demand_data_hits", KSTAT_DATA_UINT64 },
504 { "demand_data_misses", KSTAT_DATA_UINT64 },
505 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
506 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
507 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
508 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
509 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
510 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
511 { "mru_hits", KSTAT_DATA_UINT64 },
512 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
513 { "mfu_hits", KSTAT_DATA_UINT64 },
514 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
515 { "deleted", KSTAT_DATA_UINT64 },
516 { "mutex_miss", KSTAT_DATA_UINT64 },
517 { "evict_skip", KSTAT_DATA_UINT64 },
518 { "evict_not_enough", KSTAT_DATA_UINT64 },
519 { "evict_l2_cached", KSTAT_DATA_UINT64 },
520 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
521 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
522 { "evict_l2_skip", KSTAT_DATA_UINT64 },
523 { "hash_elements", KSTAT_DATA_UINT64 },
524 { "hash_elements_max", KSTAT_DATA_UINT64 },
525 { "hash_collisions", KSTAT_DATA_UINT64 },
526 { "hash_chains", KSTAT_DATA_UINT64 },
527 { "hash_chain_max", KSTAT_DATA_UINT64 },
528 { "p", KSTAT_DATA_UINT64 },
529 { "c", KSTAT_DATA_UINT64 },
530 { "c_min", KSTAT_DATA_UINT64 },
531 { "c_max", KSTAT_DATA_UINT64 },
532 { "size", KSTAT_DATA_UINT64 },
533 { "hdr_size", KSTAT_DATA_UINT64 },
534 { "data_size", KSTAT_DATA_UINT64 },
535 { "metadata_size", KSTAT_DATA_UINT64 },
536 { "other_size", KSTAT_DATA_UINT64 },
537 { "anon_size", KSTAT_DATA_UINT64 },
538 { "anon_evictable_data", KSTAT_DATA_UINT64 },
539 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
540 { "mru_size", KSTAT_DATA_UINT64 },
541 { "mru_evictable_data", KSTAT_DATA_UINT64 },
542 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
543 { "mru_ghost_size", KSTAT_DATA_UINT64 },
544 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
545 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
546 { "mfu_size", KSTAT_DATA_UINT64 },
547 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
548 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
549 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
550 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
551 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
552 { "l2_hits", KSTAT_DATA_UINT64 },
553 { "l2_misses", KSTAT_DATA_UINT64 },
554 { "l2_feeds", KSTAT_DATA_UINT64 },
555 { "l2_rw_clash", KSTAT_DATA_UINT64 },
556 { "l2_read_bytes", KSTAT_DATA_UINT64 },
557 { "l2_write_bytes", KSTAT_DATA_UINT64 },
558 { "l2_writes_sent", KSTAT_DATA_UINT64 },
559 { "l2_writes_done", KSTAT_DATA_UINT64 },
560 { "l2_writes_error", KSTAT_DATA_UINT64 },
561 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
562 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
563 { "l2_evict_reading", KSTAT_DATA_UINT64 },
564 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
565 { "l2_free_on_write", KSTAT_DATA_UINT64 },
566 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 },
567 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
568 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
569 { "l2_io_error", KSTAT_DATA_UINT64 },
570 { "l2_size", KSTAT_DATA_UINT64 },
571 { "l2_asize", KSTAT_DATA_UINT64 },
572 { "l2_hdr_size", KSTAT_DATA_UINT64 },
573 { "l2_compress_successes", KSTAT_DATA_UINT64 },
574 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
575 { "l2_compress_failures", KSTAT_DATA_UINT64 },
576 { "memory_throttle_count", KSTAT_DATA_UINT64 },
577 { "duplicate_buffers", KSTAT_DATA_UINT64 },
578 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
579 { "duplicate_reads", KSTAT_DATA_UINT64 },
580 { "arc_meta_used", KSTAT_DATA_UINT64 },
581 { "arc_meta_limit", KSTAT_DATA_UINT64 },
582 { "arc_meta_max", KSTAT_DATA_UINT64 },
583 { "arc_meta_min", KSTAT_DATA_UINT64 }
584 };
585
586 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
587
588 #define ARCSTAT_INCR(stat, val) \
589 atomic_add_64(&arc_stats.stat.value.ui64, (val))
590
591 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
592 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
593
594 #define ARCSTAT_MAX(stat, val) { \
595 uint64_t m; \
596 while ((val) > (m = arc_stats.stat.value.ui64) && \
597 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
598 continue; \
599 }
600
601 #define ARCSTAT_MAXSTAT(stat) \
602 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
603
604 /*
605 * We define a macro to allow ARC hits/misses to be easily broken down by
606 * two separate conditions, giving a total of four different subtypes for
607 * each of hits and misses (so eight statistics total).
608 */
609 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
610 if (cond1) { \
611 if (cond2) { \
612 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
613 } else { \
614 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
615 } \
616 } else { \
617 if (cond2) { \
618 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
619 } else { \
620 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
621 } \
622 }
623
624 kstat_t *arc_ksp;
625 static arc_state_t *arc_anon;
626 static arc_state_t *arc_mru;
627 static arc_state_t *arc_mru_ghost;
628 static arc_state_t *arc_mfu;
629 static arc_state_t *arc_mfu_ghost;
630 static arc_state_t *arc_l2c_only;
631
632 /*
633 * There are several ARC variables that are critical to export as kstats --
634 * but we don't want to have to grovel around in the kstat whenever we wish to
635 * manipulate them. For these variables, we therefore define them to be in
636 * terms of the statistic variable. This assures that we are not introducing
637 * the possibility of inconsistency by having shadow copies of the variables,
638 * while still allowing the code to be readable.
639 */
640 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
641 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
642 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
643 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
644 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
645 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
646 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
647 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
648 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
649
650 #define L2ARC_IS_VALID_COMPRESS(_c_) \
651 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
652
653 static int arc_no_grow; /* Don't try to grow cache size */
654 static uint64_t arc_tempreserve;
655 static uint64_t arc_loaned_bytes;
656
657 typedef struct arc_callback arc_callback_t;
658
659 struct arc_callback {
660 void *acb_private;
661 arc_done_func_t *acb_done;
662 arc_buf_t *acb_buf;
663 zio_t *acb_zio_dummy;
664 arc_callback_t *acb_next;
665 };
666
667 typedef struct arc_write_callback arc_write_callback_t;
668
669 struct arc_write_callback {
670 void *awcb_private;
671 arc_done_func_t *awcb_ready;
672 arc_done_func_t *awcb_physdone;
673 arc_done_func_t *awcb_done;
674 arc_buf_t *awcb_buf;
675 };
676
677 /*
678 * ARC buffers are separated into multiple structs as a memory saving measure:
679 * - Common fields struct, always defined, and embedded within it:
680 * - L2-only fields, always allocated but undefined when not in L2ARC
681 * - L1-only fields, only allocated when in L1ARC
682 *
683 * Buffer in L1 Buffer only in L2
684 * +------------------------+ +------------------------+
685 * | arc_buf_hdr_t | | arc_buf_hdr_t |
686 * | | | |
687 * | | | |
688 * | | | |
689 * +------------------------+ +------------------------+
690 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
691 * | (undefined if L1-only) | | |
692 * +------------------------+ +------------------------+
693 * | l1arc_buf_hdr_t |
694 * | |
695 * | |
696 * | |
697 * | |
698 * +------------------------+
699 *
700 * Because it's possible for the L2ARC to become extremely large, we can wind
701 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
702 * is minimized by only allocating the fields necessary for an L1-cached buffer
703 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
704 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
705 * words in pointers. arc_hdr_realloc() is used to switch a header between
706 * these two allocation states.
707 */
708 typedef struct l1arc_buf_hdr {
709 kmutex_t b_freeze_lock;
710 #ifdef ZFS_DEBUG
711 /*
712 * used for debugging wtih kmem_flags - by allocating and freeing
713 * b_thawed when the buffer is thawed, we get a record of the stack
714 * trace that thawed it.
715 */
716 void *b_thawed;
717 #endif
718
719 arc_buf_t *b_buf;
720 uint32_t b_datacnt;
721 /* for waiting on writes to complete */
722 kcondvar_t b_cv;
723
724 /* protected by arc state mutex */
725 arc_state_t *b_state;
726 multilist_node_t b_arc_node;
727
728 /* updated atomically */
729 clock_t b_arc_access;
730
731 /* self protecting */
732 refcount_t b_refcnt;
733
734 arc_callback_t *b_acb;
735 /* temporary buffer holder for in-flight compressed data */
736 void *b_tmp_cdata;
737 } l1arc_buf_hdr_t;
738
739 typedef struct l2arc_dev l2arc_dev_t;
740
741 typedef struct l2arc_buf_hdr {
742 /* protected by arc_buf_hdr mutex */
743 l2arc_dev_t *b_dev; /* L2ARC device */
744 uint64_t b_daddr; /* disk address, offset byte */
745 /* real alloc'd buffer size depending on b_compress applied */
746 int32_t b_asize;
747 uint8_t b_compress;
748
749 list_node_t b_l2node;
750 } l2arc_buf_hdr_t;
751
752 struct arc_buf_hdr {
753 /* protected by hash lock */
754 dva_t b_dva;
755 uint64_t b_birth;
756 /*
757 * Even though this checksum is only set/verified when a buffer is in
758 * the L1 cache, it needs to be in the set of common fields because it
759 * must be preserved from the time before a buffer is written out to
760 * L2ARC until after it is read back in.
761 */
762 zio_cksum_t *b_freeze_cksum;
763
764 arc_buf_hdr_t *b_hash_next;
765 arc_flags_t b_flags;
766
767 /* immutable */
768 int32_t b_size;
769 uint64_t b_spa;
770
771 /* L2ARC fields. Undefined when not in L2ARC. */
772 l2arc_buf_hdr_t b_l2hdr;
773 /* L1ARC fields. Undefined when in l2arc_only state */
774 l1arc_buf_hdr_t b_l1hdr;
775 };
776
777 static arc_buf_t *arc_eviction_list;
778 static arc_buf_hdr_t arc_eviction_hdr;
779
780 #define GHOST_STATE(state) \
781 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
782 (state) == arc_l2c_only)
783
784 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
785 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
786 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
787 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
788 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
789 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
790
791 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
792 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
793 #define HDR_L2_READING(hdr) \
794 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
795 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
796 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
797 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
798 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
799
800 #define HDR_ISTYPE_METADATA(hdr) \
801 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
802 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
803
804 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
805 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
806
807 /*
808 * Other sizes
809 */
810
811 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
812 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
813
814 /*
815 * Hash table routines
816 */
817
818 #define HT_LOCK_PAD 64
819
820 struct ht_lock {
821 kmutex_t ht_lock;
822 #ifdef _KERNEL
823 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
824 #endif
825 };
826
827 #define BUF_LOCKS 256
828 typedef struct buf_hash_table {
829 uint64_t ht_mask;
830 arc_buf_hdr_t **ht_table;
831 struct ht_lock ht_locks[BUF_LOCKS];
832 } buf_hash_table_t;
833
834 static buf_hash_table_t buf_hash_table;
835
836 #define BUF_HASH_INDEX(spa, dva, birth) \
837 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
838 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
839 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
840 #define HDR_LOCK(hdr) \
841 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
842
843 uint64_t zfs_crc64_table[256];
844
845 /*
846 * Level 2 ARC
847 */
848
849 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
850 #define L2ARC_HEADROOM 2 /* num of writes */
851 /*
852 * If we discover during ARC scan any buffers to be compressed, we boost
853 * our headroom for the next scanning cycle by this percentage multiple.
854 */
855 #define L2ARC_HEADROOM_BOOST 200
856 #define L2ARC_FEED_SECS 1 /* caching interval secs */
857 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
858
859 /*
860 * Used to distinguish headers that are being process by
861 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
862 * address. This can happen when the header is added to the l2arc's list
863 * of buffers to write in the first stage of l2arc_write_buffers(), but
864 * has not yet been written out which happens in the second stage of
865 * l2arc_write_buffers().
866 */
867 #define L2ARC_ADDR_UNSET ((uint64_t)(-1))
868
869 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
870 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
871
872 /* L2ARC Performance Tunables */
873 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
874 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
875 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
876 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
877 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
878 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
879 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
880 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
881 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
882
883 /*
884 * L2ARC Internals
885 */
886 struct l2arc_dev {
887 vdev_t *l2ad_vdev; /* vdev */
888 spa_t *l2ad_spa; /* spa */
889 uint64_t l2ad_hand; /* next write location */
890 uint64_t l2ad_start; /* first addr on device */
891 uint64_t l2ad_end; /* last addr on device */
892 boolean_t l2ad_first; /* first sweep through */
893 boolean_t l2ad_writing; /* currently writing */
894 kmutex_t l2ad_mtx; /* lock for buffer list */
895 list_t l2ad_buflist; /* buffer list */
896 list_node_t l2ad_node; /* device list node */
897 refcount_t l2ad_alloc; /* allocated bytes */
898 };
899
900 static list_t L2ARC_dev_list; /* device list */
901 static list_t *l2arc_dev_list; /* device list pointer */
902 static kmutex_t l2arc_dev_mtx; /* device list mutex */
903 static l2arc_dev_t *l2arc_dev_last; /* last device used */
904 static list_t L2ARC_free_on_write; /* free after write buf list */
905 static list_t *l2arc_free_on_write; /* free after write list ptr */
906 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
907 static uint64_t l2arc_ndev; /* number of devices */
908
909 typedef struct l2arc_read_callback {
910 arc_buf_t *l2rcb_buf; /* read buffer */
911 spa_t *l2rcb_spa; /* spa */
912 blkptr_t l2rcb_bp; /* original blkptr */
913 zbookmark_phys_t l2rcb_zb; /* original bookmark */
914 int l2rcb_flags; /* original flags */
915 enum zio_compress l2rcb_compress; /* applied compress */
916 } l2arc_read_callback_t;
917
918 typedef struct l2arc_write_callback {
919 l2arc_dev_t *l2wcb_dev; /* device info */
920 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
921 } l2arc_write_callback_t;
922
923 typedef struct l2arc_data_free {
924 /* protected by l2arc_free_on_write_mtx */
925 void *l2df_data;
926 size_t l2df_size;
927 void (*l2df_func)(void *, size_t);
928 list_node_t l2df_list_node;
929 } l2arc_data_free_t;
930
931 static kmutex_t l2arc_feed_thr_lock;
932 static kcondvar_t l2arc_feed_thr_cv;
933 static uint8_t l2arc_thread_exit;
934
935 static void arc_get_data_buf(arc_buf_t *);
936 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
937 static boolean_t arc_is_overflowing();
938 static void arc_buf_watch(arc_buf_t *);
939
940 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
941 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
942
943 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
944 static void l2arc_read_done(zio_t *);
945
946 static boolean_t l2arc_compress_buf(arc_buf_hdr_t *);
947 static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress);
948 static void l2arc_release_cdata_buf(arc_buf_hdr_t *);
949
950 static uint64_t
951 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
952 {
953 uint8_t *vdva = (uint8_t *)dva;
954 uint64_t crc = -1ULL;
955 int i;
956
957 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
958
959 for (i = 0; i < sizeof (dva_t); i++)
960 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
961
962 crc ^= (spa>>8) ^ birth;
963
964 return (crc);
965 }
966
967 #define BUF_EMPTY(buf) \
968 ((buf)->b_dva.dva_word[0] == 0 && \
969 (buf)->b_dva.dva_word[1] == 0)
970
971 #define BUF_EQUAL(spa, dva, birth, buf) \
972 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
973 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
974 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
975
976 static void
977 buf_discard_identity(arc_buf_hdr_t *hdr)
978 {
979 hdr->b_dva.dva_word[0] = 0;
980 hdr->b_dva.dva_word[1] = 0;
981 hdr->b_birth = 0;
982 }
983
984 static arc_buf_hdr_t *
985 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
986 {
987 const dva_t *dva = BP_IDENTITY(bp);
988 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
989 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
990 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
991 arc_buf_hdr_t *hdr;
992
993 mutex_enter(hash_lock);
994 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
995 hdr = hdr->b_hash_next) {
996 if (BUF_EQUAL(spa, dva, birth, hdr)) {
997 *lockp = hash_lock;
998 return (hdr);
999 }
1000 }
1001 mutex_exit(hash_lock);
1002 *lockp = NULL;
1003 return (NULL);
1004 }
1005
1006 /*
1007 * Insert an entry into the hash table. If there is already an element
1008 * equal to elem in the hash table, then the already existing element
1009 * will be returned and the new element will not be inserted.
1010 * Otherwise returns NULL.
1011 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1012 */
1013 static arc_buf_hdr_t *
1014 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1015 {
1016 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1017 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1018 arc_buf_hdr_t *fhdr;
1019 uint32_t i;
1020
1021 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1022 ASSERT(hdr->b_birth != 0);
1023 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1024
1025 if (lockp != NULL) {
1026 *lockp = hash_lock;
1027 mutex_enter(hash_lock);
1028 } else {
1029 ASSERT(MUTEX_HELD(hash_lock));
1030 }
1031
1032 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1033 fhdr = fhdr->b_hash_next, i++) {
1034 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1035 return (fhdr);
1036 }
1037
1038 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1039 buf_hash_table.ht_table[idx] = hdr;
1040 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
1041
1042 /* collect some hash table performance data */
1043 if (i > 0) {
1044 ARCSTAT_BUMP(arcstat_hash_collisions);
1045 if (i == 1)
1046 ARCSTAT_BUMP(arcstat_hash_chains);
1047
1048 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1049 }
1050
1051 ARCSTAT_BUMP(arcstat_hash_elements);
1052 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1053
1054 return (NULL);
1055 }
1056
1057 static void
1058 buf_hash_remove(arc_buf_hdr_t *hdr)
1059 {
1060 arc_buf_hdr_t *fhdr, **hdrp;
1061 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1062
1063 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1064 ASSERT(HDR_IN_HASH_TABLE(hdr));
1065
1066 hdrp = &buf_hash_table.ht_table[idx];
1067 while ((fhdr = *hdrp) != hdr) {
1068 ASSERT(fhdr != NULL);
1069 hdrp = &fhdr->b_hash_next;
1070 }
1071 *hdrp = hdr->b_hash_next;
1072 hdr->b_hash_next = NULL;
1073 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE;
1074
1075 /* collect some hash table performance data */
1076 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1077
1078 if (buf_hash_table.ht_table[idx] &&
1079 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1080 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1081 }
1082
1083 /*
1084 * Global data structures and functions for the buf kmem cache.
1085 */
1086 static kmem_cache_t *hdr_full_cache;
1087 static kmem_cache_t *hdr_l2only_cache;
1088 static kmem_cache_t *buf_cache;
1089
1090 static void
1091 buf_fini(void)
1092 {
1093 int i;
1094
1095 kmem_free(buf_hash_table.ht_table,
1096 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1097 for (i = 0; i < BUF_LOCKS; i++)
1098 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1099 kmem_cache_destroy(hdr_full_cache);
1100 kmem_cache_destroy(hdr_l2only_cache);
1101 kmem_cache_destroy(buf_cache);
1102 }
1103
1104 /*
1105 * Constructor callback - called when the cache is empty
1106 * and a new buf is requested.
1107 */
1108 /* ARGSUSED */
1109 static int
1110 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1111 {
1112 arc_buf_hdr_t *hdr = vbuf;
1113
1114 bzero(hdr, HDR_FULL_SIZE);
1115 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1116 refcount_create(&hdr->b_l1hdr.b_refcnt);
1117 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1118 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1119 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1120
1121 return (0);
1122 }
1123
1124 /* ARGSUSED */
1125 static int
1126 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1127 {
1128 arc_buf_hdr_t *hdr = vbuf;
1129
1130 bzero(hdr, HDR_L2ONLY_SIZE);
1131 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1132
1133 return (0);
1134 }
1135
1136 /* ARGSUSED */
1137 static int
1138 buf_cons(void *vbuf, void *unused, int kmflag)
1139 {
1140 arc_buf_t *buf = vbuf;
1141
1142 bzero(buf, sizeof (arc_buf_t));
1143 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1144 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1145
1146 return (0);
1147 }
1148
1149 /*
1150 * Destructor callback - called when a cached buf is
1151 * no longer required.
1152 */
1153 /* ARGSUSED */
1154 static void
1155 hdr_full_dest(void *vbuf, void *unused)
1156 {
1157 arc_buf_hdr_t *hdr = vbuf;
1158
1159 ASSERT(BUF_EMPTY(hdr));
1160 cv_destroy(&hdr->b_l1hdr.b_cv);
1161 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1162 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1163 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1164 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1165 }
1166
1167 /* ARGSUSED */
1168 static void
1169 hdr_l2only_dest(void *vbuf, void *unused)
1170 {
1171 arc_buf_hdr_t *hdr = vbuf;
1172
1173 ASSERT(BUF_EMPTY(hdr));
1174 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1175 }
1176
1177 /* ARGSUSED */
1178 static void
1179 buf_dest(void *vbuf, void *unused)
1180 {
1181 arc_buf_t *buf = vbuf;
1182
1183 mutex_destroy(&buf->b_evict_lock);
1184 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1185 }
1186
1187 /*
1188 * Reclaim callback -- invoked when memory is low.
1189 */
1190 /* ARGSUSED */
1191 static void
1192 hdr_recl(void *unused)
1193 {
1194 dprintf("hdr_recl called\n");
1195 /*
1196 * umem calls the reclaim func when we destroy the buf cache,
1197 * which is after we do arc_fini().
1198 */
1199 if (!arc_dead)
1200 cv_signal(&arc_reclaim_thread_cv);
1201 }
1202
1203 static void
1204 buf_init(void)
1205 {
1206 uint64_t *ct;
1207 uint64_t hsize = 1ULL << 12;
1208 int i, j;
1209
1210 /*
1211 * The hash table is big enough to fill all of physical memory
1212 * with an average block size of zfs_arc_average_blocksize (default 8K).
1213 * By default, the table will take up
1214 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1215 */
1216 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1217 hsize <<= 1;
1218 retry:
1219 buf_hash_table.ht_mask = hsize - 1;
1220 buf_hash_table.ht_table =
1221 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1222 if (buf_hash_table.ht_table == NULL) {
1223 ASSERT(hsize > (1ULL << 8));
1224 hsize >>= 1;
1225 goto retry;
1226 }
1227
1228 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1229 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1230 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1231 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1232 NULL, NULL, 0);
1233 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1234 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1235
1236 for (i = 0; i < 256; i++)
1237 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1238 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1239
1240 for (i = 0; i < BUF_LOCKS; i++) {
1241 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1242 NULL, MUTEX_DEFAULT, NULL);
1243 }
1244 }
1245
1246 /*
1247 * Transition between the two allocation states for the arc_buf_hdr struct.
1248 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1249 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1250 * version is used when a cache buffer is only in the L2ARC in order to reduce
1251 * memory usage.
1252 */
1253 static arc_buf_hdr_t *
1254 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
1255 {
1256 ASSERT(HDR_HAS_L2HDR(hdr));
1257
1258 arc_buf_hdr_t *nhdr;
1259 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1260
1261 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
1262 (old == hdr_l2only_cache && new == hdr_full_cache));
1263
1264 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
1265
1266 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
1267 buf_hash_remove(hdr);
1268
1269 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
1270
1271 if (new == hdr_full_cache) {
1272 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1273 /*
1274 * arc_access and arc_change_state need to be aware that a
1275 * header has just come out of L2ARC, so we set its state to
1276 * l2c_only even though it's about to change.
1277 */
1278 nhdr->b_l1hdr.b_state = arc_l2c_only;
1279
1280 /* Verify previous threads set to NULL before freeing */
1281 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1282 } else {
1283 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1284 ASSERT0(hdr->b_l1hdr.b_datacnt);
1285
1286 /*
1287 * If we've reached here, We must have been called from
1288 * arc_evict_hdr(), as such we should have already been
1289 * removed from any ghost list we were previously on
1290 * (which protects us from racing with arc_evict_state),
1291 * thus no locking is needed during this check.
1292 */
1293 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1294
1295 /*
1296 * A buffer must not be moved into the arc_l2c_only
1297 * state if it's not finished being written out to the
1298 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1299 * might try to be accessed, even though it was removed.
1300 */
1301 VERIFY(!HDR_L2_WRITING(hdr));
1302 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1303
1304 #ifdef ZFS_DEBUG
1305 if (hdr->b_l1hdr.b_thawed != NULL) {
1306 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1307 hdr->b_l1hdr.b_thawed = NULL;
1308 }
1309 #endif
1310
1311 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR;
1312 }
1313 /*
1314 * The header has been reallocated so we need to re-insert it into any
1315 * lists it was on.
1316 */
1317 (void) buf_hash_insert(nhdr, NULL);
1318
1319 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
1320
1321 mutex_enter(&dev->l2ad_mtx);
1322
1323 /*
1324 * We must place the realloc'ed header back into the list at
1325 * the same spot. Otherwise, if it's placed earlier in the list,
1326 * l2arc_write_buffers() could find it during the function's
1327 * write phase, and try to write it out to the l2arc.
1328 */
1329 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
1330 list_remove(&dev->l2ad_buflist, hdr);
1331
1332 mutex_exit(&dev->l2ad_mtx);
1333
1334 /*
1335 * Since we're using the pointer address as the tag when
1336 * incrementing and decrementing the l2ad_alloc refcount, we
1337 * must remove the old pointer (that we're about to destroy) and
1338 * add the new pointer to the refcount. Otherwise we'd remove
1339 * the wrong pointer address when calling arc_hdr_destroy() later.
1340 */
1341
1342 (void) refcount_remove_many(&dev->l2ad_alloc,
1343 hdr->b_l2hdr.b_asize, hdr);
1344
1345 (void) refcount_add_many(&dev->l2ad_alloc,
1346 nhdr->b_l2hdr.b_asize, nhdr);
1347
1348 buf_discard_identity(hdr);
1349 hdr->b_freeze_cksum = NULL;
1350 kmem_cache_free(old, hdr);
1351
1352 return (nhdr);
1353 }
1354
1355
1356 #define ARC_MINTIME (hz>>4) /* 62 ms */
1357
1358 static void
1359 arc_cksum_verify(arc_buf_t *buf)
1360 {
1361 zio_cksum_t zc;
1362
1363 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1364 return;
1365
1366 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1367 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) {
1368 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1369 return;
1370 }
1371 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1372 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1373 panic("buffer modified while frozen!");
1374 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1375 }
1376
1377 static int
1378 arc_cksum_equal(arc_buf_t *buf)
1379 {
1380 zio_cksum_t zc;
1381 int equal;
1382
1383 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1384 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1385 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1386 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1387
1388 return (equal);
1389 }
1390
1391 static void
1392 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1393 {
1394 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1395 return;
1396
1397 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1398 if (buf->b_hdr->b_freeze_cksum != NULL) {
1399 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1400 return;
1401 }
1402 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1403 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1404 buf->b_hdr->b_freeze_cksum);
1405 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1406 arc_buf_watch(buf);
1407 }
1408
1409 #ifndef _KERNEL
1410 typedef struct procctl {
1411 long cmd;
1412 prwatch_t prwatch;
1413 } procctl_t;
1414 #endif
1415
1416 /* ARGSUSED */
1417 static void
1418 arc_buf_unwatch(arc_buf_t *buf)
1419 {
1420 #ifndef _KERNEL
1421 if (arc_watch) {
1422 int result;
1423 procctl_t ctl;
1424 ctl.cmd = PCWATCH;
1425 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1426 ctl.prwatch.pr_size = 0;
1427 ctl.prwatch.pr_wflags = 0;
1428 result = write(arc_procfd, &ctl, sizeof (ctl));
1429 ASSERT3U(result, ==, sizeof (ctl));
1430 }
1431 #endif
1432 }
1433
1434 /* ARGSUSED */
1435 static void
1436 arc_buf_watch(arc_buf_t *buf)
1437 {
1438 #ifndef _KERNEL
1439 if (arc_watch) {
1440 int result;
1441 procctl_t ctl;
1442 ctl.cmd = PCWATCH;
1443 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1444 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1445 ctl.prwatch.pr_wflags = WA_WRITE;
1446 result = write(arc_procfd, &ctl, sizeof (ctl));
1447 ASSERT3U(result, ==, sizeof (ctl));
1448 }
1449 #endif
1450 }
1451
1452 static arc_buf_contents_t
1453 arc_buf_type(arc_buf_hdr_t *hdr)
1454 {
1455 if (HDR_ISTYPE_METADATA(hdr)) {
1456 return (ARC_BUFC_METADATA);
1457 } else {
1458 return (ARC_BUFC_DATA);
1459 }
1460 }
1461
1462 static uint32_t
1463 arc_bufc_to_flags(arc_buf_contents_t type)
1464 {
1465 switch (type) {
1466 case ARC_BUFC_DATA:
1467 /* metadata field is 0 if buffer contains normal data */
1468 return (0);
1469 case ARC_BUFC_METADATA:
1470 return (ARC_FLAG_BUFC_METADATA);
1471 default:
1472 break;
1473 }
1474 panic("undefined ARC buffer type!");
1475 return ((uint32_t)-1);
1476 }
1477
1478 void
1479 arc_buf_thaw(arc_buf_t *buf)
1480 {
1481 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1482 if (buf->b_hdr->b_l1hdr.b_state != arc_anon)
1483 panic("modifying non-anon buffer!");
1484 if (HDR_IO_IN_PROGRESS(buf->b_hdr))
1485 panic("modifying buffer while i/o in progress!");
1486 arc_cksum_verify(buf);
1487 }
1488
1489 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1490 if (buf->b_hdr->b_freeze_cksum != NULL) {
1491 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1492 buf->b_hdr->b_freeze_cksum = NULL;
1493 }
1494
1495 #ifdef ZFS_DEBUG
1496 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1497 if (buf->b_hdr->b_l1hdr.b_thawed != NULL)
1498 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1);
1499 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1500 }
1501 #endif
1502
1503 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1504
1505 arc_buf_unwatch(buf);
1506 }
1507
1508 void
1509 arc_buf_freeze(arc_buf_t *buf)
1510 {
1511 kmutex_t *hash_lock;
1512
1513 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1514 return;
1515
1516 hash_lock = HDR_LOCK(buf->b_hdr);
1517 mutex_enter(hash_lock);
1518
1519 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1520 buf->b_hdr->b_l1hdr.b_state == arc_anon);
1521 arc_cksum_compute(buf, B_FALSE);
1522 mutex_exit(hash_lock);
1523
1524 }
1525
1526 static void
1527 add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1528 {
1529 ASSERT(HDR_HAS_L1HDR(hdr));
1530 ASSERT(MUTEX_HELD(hash_lock));
1531 arc_state_t *state = hdr->b_l1hdr.b_state;
1532
1533 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
1534 (state != arc_anon)) {
1535 /* We don't use the L2-only state list. */
1536 if (state != arc_l2c_only) {
1537 arc_buf_contents_t type = arc_buf_type(hdr);
1538 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt;
1539 multilist_t *list = &state->arcs_list[type];
1540 uint64_t *size = &state->arcs_lsize[type];
1541
1542 multilist_remove(list, hdr);
1543
1544 if (GHOST_STATE(state)) {
1545 ASSERT0(hdr->b_l1hdr.b_datacnt);
1546 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
1547 delta = hdr->b_size;
1548 }
1549 ASSERT(delta > 0);
1550 ASSERT3U(*size, >=, delta);
1551 atomic_add_64(size, -delta);
1552 }
1553 /* remove the prefetch flag if we get a reference */
1554 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
1555 }
1556 }
1557
1558 static int
1559 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1560 {
1561 int cnt;
1562 arc_state_t *state = hdr->b_l1hdr.b_state;
1563
1564 ASSERT(HDR_HAS_L1HDR(hdr));
1565 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1566 ASSERT(!GHOST_STATE(state));
1567
1568 /*
1569 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1570 * check to prevent usage of the arc_l2c_only list.
1571 */
1572 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
1573 (state != arc_anon)) {
1574 arc_buf_contents_t type = arc_buf_type(hdr);
1575 multilist_t *list = &state->arcs_list[type];
1576 uint64_t *size = &state->arcs_lsize[type];
1577
1578 multilist_insert(list, hdr);
1579
1580 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
1581 atomic_add_64(size, hdr->b_size *
1582 hdr->b_l1hdr.b_datacnt);
1583 }
1584 return (cnt);
1585 }
1586
1587 /*
1588 * Move the supplied buffer to the indicated state. The hash lock
1589 * for the buffer must be held by the caller.
1590 */
1591 static void
1592 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
1593 kmutex_t *hash_lock)
1594 {
1595 arc_state_t *old_state;
1596 int64_t refcnt;
1597 uint32_t datacnt;
1598 uint64_t from_delta, to_delta;
1599 arc_buf_contents_t buftype = arc_buf_type(hdr);
1600
1601 /*
1602 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1603 * in arc_read() when bringing a buffer out of the L2ARC. However, the
1604 * L1 hdr doesn't always exist when we change state to arc_anon before
1605 * destroying a header, in which case reallocating to add the L1 hdr is
1606 * pointless.
1607 */
1608 if (HDR_HAS_L1HDR(hdr)) {
1609 old_state = hdr->b_l1hdr.b_state;
1610 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
1611 datacnt = hdr->b_l1hdr.b_datacnt;
1612 } else {
1613 old_state = arc_l2c_only;
1614 refcnt = 0;
1615 datacnt = 0;
1616 }
1617
1618 ASSERT(MUTEX_HELD(hash_lock));
1619 ASSERT3P(new_state, !=, old_state);
1620 ASSERT(refcnt == 0 || datacnt > 0);
1621 ASSERT(!GHOST_STATE(new_state) || datacnt == 0);
1622 ASSERT(old_state != arc_anon || datacnt <= 1);
1623
1624 from_delta = to_delta = datacnt * hdr->b_size;
1625
1626 /*
1627 * If this buffer is evictable, transfer it from the
1628 * old state list to the new state list.
1629 */
1630 if (refcnt == 0) {
1631 if (old_state != arc_anon && old_state != arc_l2c_only) {
1632 uint64_t *size = &old_state->arcs_lsize[buftype];
1633
1634 ASSERT(HDR_HAS_L1HDR(hdr));
1635 multilist_remove(&old_state->arcs_list[buftype], hdr);
1636
1637 /*
1638 * If prefetching out of the ghost cache,
1639 * we will have a non-zero datacnt.
1640 */
1641 if (GHOST_STATE(old_state) && datacnt == 0) {
1642 /* ghost elements have a ghost size */
1643 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1644 from_delta = hdr->b_size;
1645 }
1646 ASSERT3U(*size, >=, from_delta);
1647 atomic_add_64(size, -from_delta);
1648 }
1649 if (new_state != arc_anon && new_state != arc_l2c_only) {
1650 uint64_t *size = &new_state->arcs_lsize[buftype];
1651
1652 /*
1653 * An L1 header always exists here, since if we're
1654 * moving to some L1-cached state (i.e. not l2c_only or
1655 * anonymous), we realloc the header to add an L1hdr
1656 * beforehand.
1657 */
1658 ASSERT(HDR_HAS_L1HDR(hdr));
1659 multilist_insert(&new_state->arcs_list[buftype], hdr);
1660
1661 /* ghost elements have a ghost size */
1662 if (GHOST_STATE(new_state)) {
1663 ASSERT0(datacnt);
1664 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1665 to_delta = hdr->b_size;
1666 }
1667 atomic_add_64(size, to_delta);
1668 }
1669 }
1670
1671 ASSERT(!BUF_EMPTY(hdr));
1672 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
1673 buf_hash_remove(hdr);
1674
1675 /* adjust state sizes (ignore arc_l2c_only) */
1676
1677 if (to_delta && new_state != arc_l2c_only) {
1678 ASSERT(HDR_HAS_L1HDR(hdr));
1679 if (GHOST_STATE(new_state)) {
1680 ASSERT0(datacnt);
1681
1682 /*
1683 * We moving a header to a ghost state, we first
1684 * remove all arc buffers. Thus, we'll have a
1685 * datacnt of zero, and no arc buffer to use for
1686 * the reference. As a result, we use the arc
1687 * header pointer for the reference.
1688 */
1689 (void) refcount_add_many(&new_state->arcs_size,
1690 hdr->b_size, hdr);
1691 } else {
1692 ASSERT3U(datacnt, !=, 0);
1693
1694 /*
1695 * Each individual buffer holds a unique reference,
1696 * thus we must remove each of these references one
1697 * at a time.
1698 */
1699 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1700 buf = buf->b_next) {
1701 (void) refcount_add_many(&new_state->arcs_size,
1702 hdr->b_size, buf);
1703 }
1704 }
1705 }
1706
1707 if (from_delta && old_state != arc_l2c_only) {
1708 ASSERT(HDR_HAS_L1HDR(hdr));
1709 if (GHOST_STATE(old_state)) {
1710 /*
1711 * When moving a header off of a ghost state,
1712 * there's the possibility for datacnt to be
1713 * non-zero. This is because we first add the
1714 * arc buffer to the header prior to changing
1715 * the header's state. Since we used the header
1716 * for the reference when putting the header on
1717 * the ghost state, we must balance that and use
1718 * the header when removing off the ghost state
1719 * (even though datacnt is non zero).
1720 */
1721
1722 IMPLY(datacnt == 0, new_state == arc_anon ||
1723 new_state == arc_l2c_only);
1724
1725 (void) refcount_remove_many(&old_state->arcs_size,
1726 hdr->b_size, hdr);
1727 } else {
1728 ASSERT3P(datacnt, !=, 0);
1729
1730 /*
1731 * Each individual buffer holds a unique reference,
1732 * thus we must remove each of these references one
1733 * at a time.
1734 */
1735 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1736 buf = buf->b_next) {
1737 (void) refcount_remove_many(
1738 &old_state->arcs_size, hdr->b_size, buf);
1739 }
1740 }
1741 }
1742
1743 if (HDR_HAS_L1HDR(hdr))
1744 hdr->b_l1hdr.b_state = new_state;
1745
1746 /*
1747 * L2 headers should never be on the L2 state list since they don't
1748 * have L1 headers allocated.
1749 */
1750 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
1751 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
1752 }
1753
1754 void
1755 arc_space_consume(uint64_t space, arc_space_type_t type)
1756 {
1757 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1758
1759 switch (type) {
1760 case ARC_SPACE_DATA:
1761 ARCSTAT_INCR(arcstat_data_size, space);
1762 break;
1763 case ARC_SPACE_META:
1764 ARCSTAT_INCR(arcstat_metadata_size, space);
1765 break;
1766 case ARC_SPACE_OTHER:
1767 ARCSTAT_INCR(arcstat_other_size, space);
1768 break;
1769 case ARC_SPACE_HDRS:
1770 ARCSTAT_INCR(arcstat_hdr_size, space);
1771 break;
1772 case ARC_SPACE_L2HDRS:
1773 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1774 break;
1775 }
1776
1777 if (type != ARC_SPACE_DATA)
1778 ARCSTAT_INCR(arcstat_meta_used, space);
1779
1780 atomic_add_64(&arc_size, space);
1781 }
1782
1783 void
1784 arc_space_return(uint64_t space, arc_space_type_t type)
1785 {
1786 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1787
1788 switch (type) {
1789 case ARC_SPACE_DATA:
1790 ARCSTAT_INCR(arcstat_data_size, -space);
1791 break;
1792 case ARC_SPACE_META:
1793 ARCSTAT_INCR(arcstat_metadata_size, -space);
1794 break;
1795 case ARC_SPACE_OTHER:
1796 ARCSTAT_INCR(arcstat_other_size, -space);
1797 break;
1798 case ARC_SPACE_HDRS:
1799 ARCSTAT_INCR(arcstat_hdr_size, -space);
1800 break;
1801 case ARC_SPACE_L2HDRS:
1802 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1803 break;
1804 }
1805
1806 if (type != ARC_SPACE_DATA) {
1807 ASSERT(arc_meta_used >= space);
1808 if (arc_meta_max < arc_meta_used)
1809 arc_meta_max = arc_meta_used;
1810 ARCSTAT_INCR(arcstat_meta_used, -space);
1811 }
1812
1813 ASSERT(arc_size >= space);
1814 atomic_add_64(&arc_size, -space);
1815 }
1816
1817 arc_buf_t *
1818 arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
1819 {
1820 arc_buf_hdr_t *hdr;
1821 arc_buf_t *buf;
1822
1823 ASSERT3U(size, >, 0);
1824 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
1825 ASSERT(BUF_EMPTY(hdr));
1826 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
1827 hdr->b_size = size;
1828 hdr->b_spa = spa_load_guid(spa);
1829
1830 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1831 buf->b_hdr = hdr;
1832 buf->b_data = NULL;
1833 buf->b_efunc = NULL;
1834 buf->b_private = NULL;
1835 buf->b_next = NULL;
1836
1837 hdr->b_flags = arc_bufc_to_flags(type);
1838 hdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1839
1840 hdr->b_l1hdr.b_buf = buf;
1841 hdr->b_l1hdr.b_state = arc_anon;
1842 hdr->b_l1hdr.b_arc_access = 0;
1843 hdr->b_l1hdr.b_datacnt = 1;
1844 hdr->b_l1hdr.b_tmp_cdata = NULL;
1845
1846 arc_get_data_buf(buf);
1847 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
1848 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
1849
1850 return (buf);
1851 }
1852
1853 static char *arc_onloan_tag = "onloan";
1854
1855 /*
1856 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1857 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1858 * buffers must be returned to the arc before they can be used by the DMU or
1859 * freed.
1860 */
1861 arc_buf_t *
1862 arc_loan_buf(spa_t *spa, int size)
1863 {
1864 arc_buf_t *buf;
1865
1866 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1867
1868 atomic_add_64(&arc_loaned_bytes, size);
1869 return (buf);
1870 }
1871
1872 /*
1873 * Return a loaned arc buffer to the arc.
1874 */
1875 void
1876 arc_return_buf(arc_buf_t *buf, void *tag)
1877 {
1878 arc_buf_hdr_t *hdr = buf->b_hdr;
1879
1880 ASSERT(buf->b_data != NULL);
1881 ASSERT(HDR_HAS_L1HDR(hdr));
1882 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
1883 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
1884
1885 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1886 }
1887
1888 /* Detach an arc_buf from a dbuf (tag) */
1889 void
1890 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1891 {
1892 arc_buf_hdr_t *hdr = buf->b_hdr;
1893
1894 ASSERT(buf->b_data != NULL);
1895 ASSERT(HDR_HAS_L1HDR(hdr));
1896 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
1897 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
1898 buf->b_efunc = NULL;
1899 buf->b_private = NULL;
1900
1901 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1902 }
1903
1904 static arc_buf_t *
1905 arc_buf_clone(arc_buf_t *from)
1906 {
1907 arc_buf_t *buf;
1908 arc_buf_hdr_t *hdr = from->b_hdr;
1909 uint64_t size = hdr->b_size;
1910
1911 ASSERT(HDR_HAS_L1HDR(hdr));
1912 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
1913
1914 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1915 buf->b_hdr = hdr;
1916 buf->b_data = NULL;
1917 buf->b_efunc = NULL;
1918 buf->b_private = NULL;
1919 buf->b_next = hdr->b_l1hdr.b_buf;
1920 hdr->b_l1hdr.b_buf = buf;
1921 arc_get_data_buf(buf);
1922 bcopy(from->b_data, buf->b_data, size);
1923
1924 /*
1925 * This buffer already exists in the arc so create a duplicate
1926 * copy for the caller. If the buffer is associated with user data
1927 * then track the size and number of duplicates. These stats will be
1928 * updated as duplicate buffers are created and destroyed.
1929 */
1930 if (HDR_ISTYPE_DATA(hdr)) {
1931 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1932 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1933 }
1934 hdr->b_l1hdr.b_datacnt += 1;
1935 return (buf);
1936 }
1937
1938 void
1939 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1940 {
1941 arc_buf_hdr_t *hdr;
1942 kmutex_t *hash_lock;
1943
1944 /*
1945 * Check to see if this buffer is evicted. Callers
1946 * must verify b_data != NULL to know if the add_ref
1947 * was successful.
1948 */
1949 mutex_enter(&buf->b_evict_lock);
1950 if (buf->b_data == NULL) {
1951 mutex_exit(&buf->b_evict_lock);
1952 return;
1953 }
1954 hash_lock = HDR_LOCK(buf->b_hdr);
1955 mutex_enter(hash_lock);
1956 hdr = buf->b_hdr;
1957 ASSERT(HDR_HAS_L1HDR(hdr));
1958 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1959 mutex_exit(&buf->b_evict_lock);
1960
1961 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
1962 hdr->b_l1hdr.b_state == arc_mfu);
1963
1964 add_reference(hdr, hash_lock, tag);
1965 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1966 arc_access(hdr, hash_lock);
1967 mutex_exit(hash_lock);
1968 ARCSTAT_BUMP(arcstat_hits);
1969 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
1970 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
1971 data, metadata, hits);
1972 }
1973
1974 static void
1975 arc_buf_free_on_write(void *data, size_t size,
1976 void (*free_func)(void *, size_t))
1977 {
1978 l2arc_data_free_t *df;
1979
1980 df = kmem_alloc(sizeof (*df), KM_SLEEP);
1981 df->l2df_data = data;
1982 df->l2df_size = size;
1983 df->l2df_func = free_func;
1984 mutex_enter(&l2arc_free_on_write_mtx);
1985 list_insert_head(l2arc_free_on_write, df);
1986 mutex_exit(&l2arc_free_on_write_mtx);
1987 }
1988
1989 /*
1990 * Free the arc data buffer. If it is an l2arc write in progress,
1991 * the buffer is placed on l2arc_free_on_write to be freed later.
1992 */
1993 static void
1994 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1995 {
1996 arc_buf_hdr_t *hdr = buf->b_hdr;
1997
1998 if (HDR_L2_WRITING(hdr)) {
1999 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func);
2000 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2001 } else {
2002 free_func(buf->b_data, hdr->b_size);
2003 }
2004 }
2005
2006 static void
2007 arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr)
2008 {
2009 ASSERT(HDR_HAS_L2HDR(hdr));
2010 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx));
2011
2012 /*
2013 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
2014 * that doesn't exist, the header is in the arc_l2c_only state,
2015 * and there isn't anything to free (it's already been freed).
2016 */
2017 if (!HDR_HAS_L1HDR(hdr))
2018 return;
2019
2020 /*
2021 * The header isn't being written to the l2arc device, thus it
2022 * shouldn't have a b_tmp_cdata to free.
2023 */
2024 if (!HDR_L2_WRITING(hdr)) {
2025 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2026 return;
2027 }
2028
2029 /*
2030 * The header does not have compression enabled. This can be due
2031 * to the buffer not being compressible, or because we're
2032 * freeing the buffer before the second phase of
2033 * l2arc_write_buffer() has started (which does the compression
2034 * step). In either case, b_tmp_cdata does not point to a
2035 * separately compressed buffer, so there's nothing to free (it
2036 * points to the same buffer as the arc_buf_t's b_data field).
2037 */
2038 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_OFF) {
2039 hdr->b_l1hdr.b_tmp_cdata = NULL;
2040 return;
2041 }
2042
2043 /*
2044 * There's nothing to free since the buffer was all zero's and
2045 * compressed to a zero length buffer.
2046 */
2047 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_EMPTY) {
2048 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2049 return;
2050 }
2051
2052 ASSERT(L2ARC_IS_VALID_COMPRESS(hdr->b_l2hdr.b_compress));
2053
2054 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata,
2055 hdr->b_size, zio_data_buf_free);
2056
2057 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write);
2058 hdr->b_l1hdr.b_tmp_cdata = NULL;
2059 }
2060
2061 /*
2062 * Free up buf->b_data and if 'remove' is set, then pull the
2063 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2064 */
2065 static void
2066 arc_buf_destroy(arc_buf_t *buf, boolean_t remove)
2067 {
2068 arc_buf_t **bufp;
2069
2070 /* free up data associated with the buf */
2071 if (buf->b_data != NULL) {
2072 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
2073 uint64_t size = buf->b_hdr->b_size;
2074 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
2075
2076 arc_cksum_verify(buf);
2077 arc_buf_unwatch(buf);
2078
2079 if (type == ARC_BUFC_METADATA) {
2080 arc_buf_data_free(buf, zio_buf_free);
2081 arc_space_return(size, ARC_SPACE_META);
2082 } else {
2083 ASSERT(type == ARC_BUFC_DATA);
2084 arc_buf_data_free(buf, zio_data_buf_free);
2085 arc_space_return(size, ARC_SPACE_DATA);
2086 }
2087
2088 /* protected by hash lock, if in the hash table */
2089 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) {
2090 uint64_t *cnt = &state->arcs_lsize[type];
2091
2092 ASSERT(refcount_is_zero(
2093 &buf->b_hdr->b_l1hdr.b_refcnt));
2094 ASSERT(state != arc_anon && state != arc_l2c_only);
2095
2096 ASSERT3U(*cnt, >=, size);
2097 atomic_add_64(cnt, -size);
2098 }
2099
2100 (void) refcount_remove_many(&state->arcs_size, size, buf);
2101 buf->b_data = NULL;
2102
2103 /*
2104 * If we're destroying a duplicate buffer make sure
2105 * that the appropriate statistics are updated.
2106 */
2107 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 &&
2108 HDR_ISTYPE_DATA(buf->b_hdr)) {
2109 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
2110 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
2111 }
2112 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0);
2113 buf->b_hdr->b_l1hdr.b_datacnt -= 1;
2114 }
2115
2116 /* only remove the buf if requested */
2117 if (!remove)
2118 return;
2119
2120 /* remove the buf from the hdr list */
2121 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf;
2122 bufp = &(*bufp)->b_next)
2123 continue;
2124 *bufp = buf->b_next;
2125 buf->b_next = NULL;
2126
2127 ASSERT(buf->b_efunc == NULL);
2128
2129 /* clean up the buf */
2130 buf->b_hdr = NULL;
2131 kmem_cache_free(buf_cache, buf);
2132 }
2133
2134 static void
2135 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
2136 {
2137 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
2138 l2arc_dev_t *dev = l2hdr->b_dev;
2139
2140 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
2141 ASSERT(HDR_HAS_L2HDR(hdr));
2142
2143 list_remove(&dev->l2ad_buflist, hdr);
2144
2145 /*
2146 * We don't want to leak the b_tmp_cdata buffer that was
2147 * allocated in l2arc_write_buffers()
2148 */
2149 arc_buf_l2_cdata_free(hdr);
2150
2151 /*
2152 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2153 * this header is being processed by l2arc_write_buffers() (i.e.
2154 * it's in the first stage of l2arc_write_buffers()).
2155 * Re-affirming that truth here, just to serve as a reminder. If
2156 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2157 * may not have its HDR_L2_WRITING flag set. (the write may have
2158 * completed, in which case HDR_L2_WRITING will be false and the
2159 * b_daddr field will point to the address of the buffer on disk).
2160 */
2161 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr));
2162
2163 /*
2164 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2165 * l2arc_write_buffers(). Since we've just removed this header
2166 * from the l2arc buffer list, this header will never reach the
2167 * second stage of l2arc_write_buffers(), which increments the
2168 * accounting stats for this header. Thus, we must be careful
2169 * not to decrement them for this header either.
2170 */
2171 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) {
2172 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
2173 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
2174
2175 vdev_space_update(dev->l2ad_vdev,
2176 -l2hdr->b_asize, 0, 0);
2177
2178 (void) refcount_remove_many(&dev->l2ad_alloc,
2179 l2hdr->b_asize, hdr);
2180 }
2181
2182 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
2183 }
2184
2185 static void
2186 arc_hdr_destroy(arc_buf_hdr_t *hdr)
2187 {
2188 if (HDR_HAS_L1HDR(hdr)) {
2189 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
2190 hdr->b_l1hdr.b_datacnt > 0);
2191 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2192 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2193 }
2194 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2195 ASSERT(!HDR_IN_HASH_TABLE(hdr));
2196
2197 if (HDR_HAS_L2HDR(hdr)) {
2198 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2199 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
2200
2201 if (!buflist_held)
2202 mutex_enter(&dev->l2ad_mtx);
2203
2204 /*
2205 * Even though we checked this conditional above, we
2206 * need to check this again now that we have the
2207 * l2ad_mtx. This is because we could be racing with
2208 * another thread calling l2arc_evict() which might have
2209 * destroyed this header's L2 portion as we were waiting
2210 * to acquire the l2ad_mtx. If that happens, we don't
2211 * want to re-destroy the header's L2 portion.
2212 */
2213 if (HDR_HAS_L2HDR(hdr))
2214 arc_hdr_l2hdr_destroy(hdr);
2215
2216 if (!buflist_held)
2217 mutex_exit(&dev->l2ad_mtx);
2218 }
2219
2220 if (!BUF_EMPTY(hdr))
2221 buf_discard_identity(hdr);
2222
2223 if (hdr->b_freeze_cksum != NULL) {
2224 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
2225 hdr->b_freeze_cksum = NULL;
2226 }
2227
2228 if (HDR_HAS_L1HDR(hdr)) {
2229 while (hdr->b_l1hdr.b_buf) {
2230 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2231
2232 if (buf->b_efunc != NULL) {
2233 mutex_enter(&arc_user_evicts_lock);
2234 mutex_enter(&buf->b_evict_lock);
2235 ASSERT(buf->b_hdr != NULL);
2236 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE);
2237 hdr->b_l1hdr.b_buf = buf->b_next;
2238 buf->b_hdr = &arc_eviction_hdr;
2239 buf->b_next = arc_eviction_list;
2240 arc_eviction_list = buf;
2241 mutex_exit(&buf->b_evict_lock);
2242 cv_signal(&arc_user_evicts_cv);
2243 mutex_exit(&arc_user_evicts_lock);
2244 } else {
2245 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE);
2246 }
2247 }
2248 #ifdef ZFS_DEBUG
2249 if (hdr->b_l1hdr.b_thawed != NULL) {
2250 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2251 hdr->b_l1hdr.b_thawed = NULL;
2252 }
2253 #endif
2254 }
2255
2256 ASSERT3P(hdr->b_hash_next, ==, NULL);
2257 if (HDR_HAS_L1HDR(hdr)) {
2258 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2259 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
2260 kmem_cache_free(hdr_full_cache, hdr);
2261 } else {
2262 kmem_cache_free(hdr_l2only_cache, hdr);
2263 }
2264 }
2265
2266 void
2267 arc_buf_free(arc_buf_t *buf, void *tag)
2268 {
2269 arc_buf_hdr_t *hdr = buf->b_hdr;
2270 int hashed = hdr->b_l1hdr.b_state != arc_anon;
2271
2272 ASSERT(buf->b_efunc == NULL);
2273 ASSERT(buf->b_data != NULL);
2274
2275 if (hashed) {
2276 kmutex_t *hash_lock = HDR_LOCK(hdr);
2277
2278 mutex_enter(hash_lock);
2279 hdr = buf->b_hdr;
2280 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2281
2282 (void) remove_reference(hdr, hash_lock, tag);
2283 if (hdr->b_l1hdr.b_datacnt > 1) {
2284 arc_buf_destroy(buf, TRUE);
2285 } else {
2286 ASSERT(buf == hdr->b_l1hdr.b_buf);
2287 ASSERT(buf->b_efunc == NULL);
2288 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2289 }
2290 mutex_exit(hash_lock);
2291 } else if (HDR_IO_IN_PROGRESS(hdr)) {
2292 int destroy_hdr;
2293 /*
2294 * We are in the middle of an async write. Don't destroy
2295 * this buffer unless the write completes before we finish
2296 * decrementing the reference count.
2297 */
2298 mutex_enter(&arc_user_evicts_lock);
2299 (void) remove_reference(hdr, NULL, tag);
2300 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2301 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
2302 mutex_exit(&arc_user_evicts_lock);
2303 if (destroy_hdr)
2304 arc_hdr_destroy(hdr);
2305 } else {
2306 if (remove_reference(hdr, NULL, tag) > 0)
2307 arc_buf_destroy(buf, TRUE);
2308 else
2309 arc_hdr_destroy(hdr);
2310 }
2311 }
2312
2313 boolean_t
2314 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
2315 {
2316 arc_buf_hdr_t *hdr = buf->b_hdr;
2317 kmutex_t *hash_lock = HDR_LOCK(hdr);
2318 boolean_t no_callback = (buf->b_efunc == NULL);
2319
2320 if (hdr->b_l1hdr.b_state == arc_anon) {
2321 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
2322 arc_buf_free(buf, tag);
2323 return (no_callback);
2324 }
2325
2326 mutex_enter(hash_lock);
2327 hdr = buf->b_hdr;
2328 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
2329 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2330 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2331 ASSERT(buf->b_data != NULL);
2332
2333 (void) remove_reference(hdr, hash_lock, tag);
2334 if (hdr->b_l1hdr.b_datacnt > 1) {
2335 if (no_callback)
2336 arc_buf_destroy(buf, TRUE);
2337 } else if (no_callback) {
2338 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL);
2339 ASSERT(buf->b_efunc == NULL);
2340 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2341 }
2342 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 ||
2343 refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2344 mutex_exit(hash_lock);
2345 return (no_callback);
2346 }
2347
2348 int32_t
2349 arc_buf_size(arc_buf_t *buf)
2350 {
2351 return (buf->b_hdr->b_size);
2352 }
2353
2354 /*
2355 * Called from the DMU to determine if the current buffer should be
2356 * evicted. In order to ensure proper locking, the eviction must be initiated
2357 * from the DMU. Return true if the buffer is associated with user data and
2358 * duplicate buffers still exist.
2359 */
2360 boolean_t
2361 arc_buf_eviction_needed(arc_buf_t *buf)
2362 {
2363 arc_buf_hdr_t *hdr;
2364 boolean_t evict_needed = B_FALSE;
2365
2366 if (zfs_disable_dup_eviction)
2367 return (B_FALSE);
2368
2369 mutex_enter(&buf->b_evict_lock);
2370 hdr = buf->b_hdr;
2371 if (hdr == NULL) {
2372 /*
2373 * We are in arc_do_user_evicts(); let that function
2374 * perform the eviction.
2375 */
2376 ASSERT(buf->b_data == NULL);
2377 mutex_exit(&buf->b_evict_lock);
2378 return (B_FALSE);
2379 } else if (buf->b_data == NULL) {
2380 /*
2381 * We have already been added to the arc eviction list;
2382 * recommend eviction.
2383 */
2384 ASSERT3P(hdr, ==, &arc_eviction_hdr);
2385 mutex_exit(&buf->b_evict_lock);
2386 return (B_TRUE);
2387 }
2388
2389 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr))
2390 evict_needed = B_TRUE;
2391
2392 mutex_exit(&buf->b_evict_lock);
2393 return (evict_needed);
2394 }
2395
2396 /*
2397 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2398 * state of the header is dependent on it's state prior to entering this
2399 * function. The following transitions are possible:
2400 *
2401 * - arc_mru -> arc_mru_ghost
2402 * - arc_mfu -> arc_mfu_ghost
2403 * - arc_mru_ghost -> arc_l2c_only
2404 * - arc_mru_ghost -> deleted
2405 * - arc_mfu_ghost -> arc_l2c_only
2406 * - arc_mfu_ghost -> deleted
2407 */
2408 static int64_t
2409 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
2410 {
2411 arc_state_t *evicted_state, *state;
2412 int64_t bytes_evicted = 0;
2413
2414 ASSERT(MUTEX_HELD(hash_lock));
2415 ASSERT(HDR_HAS_L1HDR(hdr));
2416
2417 state = hdr->b_l1hdr.b_state;
2418 if (GHOST_STATE(state)) {
2419 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2420 ASSERT(hdr->b_l1hdr.b_buf == NULL);
2421
2422 /*
2423 * l2arc_write_buffers() relies on a header's L1 portion
2424 * (i.e. it's b_tmp_cdata field) during it's write phase.
2425 * Thus, we cannot push a header onto the arc_l2c_only
2426 * state (removing it's L1 piece) until the header is
2427 * done being written to the l2arc.
2428 */
2429 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
2430 ARCSTAT_BUMP(arcstat_evict_l2_skip);
2431 return (bytes_evicted);
2432 }
2433
2434 ARCSTAT_BUMP(arcstat_deleted);
2435 bytes_evicted += hdr->b_size;
2436
2437 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
2438
2439 if (HDR_HAS_L2HDR(hdr)) {
2440 /*
2441 * This buffer is cached on the 2nd Level ARC;
2442 * don't destroy the header.
2443 */
2444 arc_change_state(arc_l2c_only, hdr, hash_lock);
2445 /*
2446 * dropping from L1+L2 cached to L2-only,
2447 * realloc to remove the L1 header.
2448 */
2449 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
2450 hdr_l2only_cache);
2451 } else {
2452 arc_change_state(arc_anon, hdr, hash_lock);
2453 arc_hdr_destroy(hdr);
2454 }
2455 return (bytes_evicted);
2456 }
2457
2458 ASSERT(state == arc_mru || state == arc_mfu);
2459 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2460
2461 /* prefetch buffers have a minimum lifespan */
2462 if (HDR_IO_IN_PROGRESS(hdr) ||
2463 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
2464 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
2465 arc_min_prefetch_lifespan)) {
2466 ARCSTAT_BUMP(arcstat_evict_skip);
2467 return (bytes_evicted);
2468 }
2469
2470 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
2471 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0);
2472 while (hdr->b_l1hdr.b_buf) {
2473 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2474 if (!mutex_tryenter(&buf->b_evict_lock)) {
2475 ARCSTAT_BUMP(arcstat_mutex_miss);
2476 break;
2477 }
2478 if (buf->b_data != NULL)
2479 bytes_evicted += hdr->b_size;
2480 if (buf->b_efunc != NULL) {
2481 mutex_enter(&arc_user_evicts_lock);
2482 arc_buf_destroy(buf, FALSE);
2483 hdr->b_l1hdr.b_buf = buf->b_next;
2484 buf->b_hdr = &arc_eviction_hdr;
2485 buf->b_next = arc_eviction_list;
2486 arc_eviction_list = buf;
2487 cv_signal(&arc_user_evicts_cv);
2488 mutex_exit(&arc_user_evicts_lock);
2489 mutex_exit(&buf->b_evict_lock);
2490 } else {
2491 mutex_exit(&buf->b_evict_lock);
2492 arc_buf_destroy(buf, TRUE);
2493 }
2494 }
2495
2496 if (HDR_HAS_L2HDR(hdr)) {
2497 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size);
2498 } else {
2499 if (l2arc_write_eligible(hdr->b_spa, hdr))
2500 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size);
2501 else
2502 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size);
2503 }
2504
2505 if (hdr->b_l1hdr.b_datacnt == 0) {
2506 arc_change_state(evicted_state, hdr, hash_lock);
2507 ASSERT(HDR_IN_HASH_TABLE(hdr));
2508 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
2509 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
2510 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
2511 }
2512
2513 return (bytes_evicted);
2514 }
2515
2516 static uint64_t
2517 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
2518 uint64_t spa, int64_t bytes)
2519 {
2520 multilist_sublist_t *mls;
2521 uint64_t bytes_evicted = 0;
2522 arc_buf_hdr_t *hdr;
2523 kmutex_t *hash_lock;
2524 int evict_count = 0;
2525
2526 ASSERT3P(marker, !=, NULL);
2527 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2528
2529 mls = multilist_sublist_lock(ml, idx);
2530
2531 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
2532 hdr = multilist_sublist_prev(mls, marker)) {
2533 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
2534 (evict_count >= zfs_arc_evict_batch_limit))
2535 break;
2536
2537 /*
2538 * To keep our iteration location, move the marker
2539 * forward. Since we're not holding hdr's hash lock, we
2540 * must be very careful and not remove 'hdr' from the
2541 * sublist. Otherwise, other consumers might mistake the
2542 * 'hdr' as not being on a sublist when they call the
2543 * multilist_link_active() function (they all rely on
2544 * the hash lock protecting concurrent insertions and
2545 * removals). multilist_sublist_move_forward() was
2546 * specifically implemented to ensure this is the case
2547 * (only 'marker' will be removed and re-inserted).
2548 */
2549 multilist_sublist_move_forward(mls, marker);
2550
2551 /*
2552 * The only case where the b_spa field should ever be
2553 * zero, is the marker headers inserted by
2554 * arc_evict_state(). It's possible for multiple threads
2555 * to be calling arc_evict_state() concurrently (e.g.
2556 * dsl_pool_close() and zio_inject_fault()), so we must
2557 * skip any markers we see from these other threads.
2558 */
2559 if (hdr->b_spa == 0)
2560 continue;
2561
2562 /* we're only interested in evicting buffers of a certain spa */
2563 if (spa != 0 && hdr->b_spa != spa) {
2564 ARCSTAT_BUMP(arcstat_evict_skip);
2565 continue;
2566 }
2567
2568 hash_lock = HDR_LOCK(hdr);
2569
2570 /*
2571 * We aren't calling this function from any code path
2572 * that would already be holding a hash lock, so we're
2573 * asserting on this assumption to be defensive in case
2574 * this ever changes. Without this check, it would be
2575 * possible to incorrectly increment arcstat_mutex_miss
2576 * below (e.g. if the code changed such that we called
2577 * this function with a hash lock held).
2578 */
2579 ASSERT(!MUTEX_HELD(hash_lock));
2580
2581 if (mutex_tryenter(hash_lock)) {
2582 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
2583 mutex_exit(hash_lock);
2584
2585 bytes_evicted += evicted;
2586
2587 /*
2588 * If evicted is zero, arc_evict_hdr() must have
2589 * decided to skip this header, don't increment
2590 * evict_count in this case.
2591 */
2592 if (evicted != 0)
2593 evict_count++;
2594
2595 /*
2596 * If arc_size isn't overflowing, signal any
2597 * threads that might happen to be waiting.
2598 *
2599 * For each header evicted, we wake up a single
2600 * thread. If we used cv_broadcast, we could
2601 * wake up "too many" threads causing arc_size
2602 * to significantly overflow arc_c; since
2603 * arc_get_data_buf() doesn't check for overflow
2604 * when it's woken up (it doesn't because it's
2605 * possible for the ARC to be overflowing while
2606 * full of un-evictable buffers, and the
2607 * function should proceed in this case).
2608 *
2609 * If threads are left sleeping, due to not
2610 * using cv_broadcast, they will be woken up
2611 * just before arc_reclaim_thread() sleeps.
2612 */
2613 mutex_enter(&arc_reclaim_lock);
2614 if (!arc_is_overflowing())
2615 cv_signal(&arc_reclaim_waiters_cv);
2616 mutex_exit(&arc_reclaim_lock);
2617 } else {
2618 ARCSTAT_BUMP(arcstat_mutex_miss);
2619 }
2620 }
2621
2622 multilist_sublist_unlock(mls);
2623
2624 return (bytes_evicted);
2625 }
2626
2627 /*
2628 * Evict buffers from the given arc state, until we've removed the
2629 * specified number of bytes. Move the removed buffers to the
2630 * appropriate evict state.
2631 *
2632 * This function makes a "best effort". It skips over any buffers
2633 * it can't get a hash_lock on, and so, may not catch all candidates.
2634 * It may also return without evicting as much space as requested.
2635 *
2636 * If bytes is specified using the special value ARC_EVICT_ALL, this
2637 * will evict all available (i.e. unlocked and evictable) buffers from
2638 * the given arc state; which is used by arc_flush().
2639 */
2640 static uint64_t
2641 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
2642 arc_buf_contents_t type)
2643 {
2644 uint64_t total_evicted = 0;
2645 multilist_t *ml = &state->arcs_list[type];
2646 int num_sublists;
2647 arc_buf_hdr_t **markers;
2648
2649 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2650
2651 num_sublists = multilist_get_num_sublists(ml);
2652
2653 /*
2654 * If we've tried to evict from each sublist, made some
2655 * progress, but still have not hit the target number of bytes
2656 * to evict, we want to keep trying. The markers allow us to
2657 * pick up where we left off for each individual sublist, rather
2658 * than starting from the tail each time.
2659 */
2660 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
2661 for (int i = 0; i < num_sublists; i++) {
2662 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
2663
2664 /*
2665 * A b_spa of 0 is used to indicate that this header is
2666 * a marker. This fact is used in arc_adjust_type() and
2667 * arc_evict_state_impl().
2668 */
2669 markers[i]->b_spa = 0;
2670
2671 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2672 multilist_sublist_insert_tail(mls, markers[i]);
2673 multilist_sublist_unlock(mls);
2674 }
2675
2676 /*
2677 * While we haven't hit our target number of bytes to evict, or
2678 * we're evicting all available buffers.
2679 */
2680 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
2681 /*
2682 * Start eviction using a randomly selected sublist,
2683 * this is to try and evenly balance eviction across all
2684 * sublists. Always starting at the same sublist
2685 * (e.g. index 0) would cause evictions to favor certain
2686 * sublists over others.
2687 */
2688 int sublist_idx = multilist_get_random_index(ml);
2689 uint64_t scan_evicted = 0;
2690
2691 for (int i = 0; i < num_sublists; i++) {
2692 uint64_t bytes_remaining;
2693 uint64_t bytes_evicted;
2694
2695 if (bytes == ARC_EVICT_ALL)
2696 bytes_remaining = ARC_EVICT_ALL;
2697 else if (total_evicted < bytes)
2698 bytes_remaining = bytes - total_evicted;
2699 else
2700 break;
2701
2702 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
2703 markers[sublist_idx], spa, bytes_remaining);
2704
2705 scan_evicted += bytes_evicted;
2706 total_evicted += bytes_evicted;
2707
2708 /* we've reached the end, wrap to the beginning */
2709 if (++sublist_idx >= num_sublists)
2710 sublist_idx = 0;
2711 }
2712
2713 /*
2714 * If we didn't evict anything during this scan, we have
2715 * no reason to believe we'll evict more during another
2716 * scan, so break the loop.
2717 */
2718 if (scan_evicted == 0) {
2719 /* This isn't possible, let's make that obvious */
2720 ASSERT3S(bytes, !=, 0);
2721
2722 /*
2723 * When bytes is ARC_EVICT_ALL, the only way to
2724 * break the loop is when scan_evicted is zero.
2725 * In that case, we actually have evicted enough,
2726 * so we don't want to increment the kstat.
2727 */
2728 if (bytes != ARC_EVICT_ALL) {
2729 ASSERT3S(total_evicted, <, bytes);
2730 ARCSTAT_BUMP(arcstat_evict_not_enough);
2731 }
2732
2733 break;
2734 }
2735 }
2736
2737 for (int i = 0; i < num_sublists; i++) {
2738 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2739 multilist_sublist_remove(mls, markers[i]);
2740 multilist_sublist_unlock(mls);
2741
2742 kmem_cache_free(hdr_full_cache, markers[i]);
2743 }
2744 kmem_free(markers, sizeof (*markers) * num_sublists);
2745
2746 return (total_evicted);
2747 }
2748
2749 /*
2750 * Flush all "evictable" data of the given type from the arc state
2751 * specified. This will not evict any "active" buffers (i.e. referenced).
2752 *
2753 * When 'retry' is set to FALSE, the function will make a single pass
2754 * over the state and evict any buffers that it can. Since it doesn't
2755 * continually retry the eviction, it might end up leaving some buffers
2756 * in the ARC due to lock misses.
2757 *
2758 * When 'retry' is set to TRUE, the function will continually retry the
2759 * eviction until *all* evictable buffers have been removed from the
2760 * state. As a result, if concurrent insertions into the state are
2761 * allowed (e.g. if the ARC isn't shutting down), this function might
2762 * wind up in an infinite loop, continually trying to evict buffers.
2763 */
2764 static uint64_t
2765 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
2766 boolean_t retry)
2767 {
2768 uint64_t evicted = 0;
2769
2770 while (state->arcs_lsize[type] != 0) {
2771 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
2772
2773 if (!retry)
2774 break;
2775 }
2776
2777 return (evicted);
2778 }
2779
2780 /*
2781 * Evict the specified number of bytes from the state specified,
2782 * restricting eviction to the spa and type given. This function
2783 * prevents us from trying to evict more from a state's list than
2784 * is "evictable", and to skip evicting altogether when passed a
2785 * negative value for "bytes". In contrast, arc_evict_state() will
2786 * evict everything it can, when passed a negative value for "bytes".
2787 */
2788 static uint64_t
2789 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
2790 arc_buf_contents_t type)
2791 {
2792 int64_t delta;
2793
2794 if (bytes > 0 && state->arcs_lsize[type] > 0) {
2795 delta = MIN(state->arcs_lsize[type], bytes);
2796 return (arc_evict_state(state, spa, delta, type));
2797 }
2798
2799 return (0);
2800 }
2801
2802 /*
2803 * Evict metadata buffers from the cache, such that arc_meta_used is
2804 * capped by the arc_meta_limit tunable.
2805 */
2806 static uint64_t
2807 arc_adjust_meta(void)
2808 {
2809 uint64_t total_evicted = 0;
2810 int64_t target;
2811
2812 /*
2813 * If we're over the meta limit, we want to evict enough
2814 * metadata to get back under the meta limit. We don't want to
2815 * evict so much that we drop the MRU below arc_p, though. If
2816 * we're over the meta limit more than we're over arc_p, we
2817 * evict some from the MRU here, and some from the MFU below.
2818 */
2819 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
2820 (int64_t)(refcount_count(&arc_anon->arcs_size) +
2821 refcount_count(&arc_mru->arcs_size) - arc_p));
2822
2823 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2824
2825 /*
2826 * Similar to the above, we want to evict enough bytes to get us
2827 * below the meta limit, but not so much as to drop us below the
2828 * space alloted to the MFU (which is defined as arc_c - arc_p).
2829 */
2830 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
2831 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
2832
2833 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
2834
2835 return (total_evicted);
2836 }
2837
2838 /*
2839 * Return the type of the oldest buffer in the given arc state
2840 *
2841 * This function will select a random sublist of type ARC_BUFC_DATA and
2842 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
2843 * is compared, and the type which contains the "older" buffer will be
2844 * returned.
2845 */
2846 static arc_buf_contents_t
2847 arc_adjust_type(arc_state_t *state)
2848 {
2849 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
2850 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
2851 int data_idx = multilist_get_random_index(data_ml);
2852 int meta_idx = multilist_get_random_index(meta_ml);
2853 multilist_sublist_t *data_mls;
2854 multilist_sublist_t *meta_mls;
2855 arc_buf_contents_t type;
2856 arc_buf_hdr_t *data_hdr;
2857 arc_buf_hdr_t *meta_hdr;
2858
2859 /*
2860 * We keep the sublist lock until we're finished, to prevent
2861 * the headers from being destroyed via arc_evict_state().
2862 */
2863 data_mls = multilist_sublist_lock(data_ml, data_idx);
2864 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
2865
2866 /*
2867 * These two loops are to ensure we skip any markers that
2868 * might be at the tail of the lists due to arc_evict_state().
2869 */
2870
2871 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
2872 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
2873 if (data_hdr->b_spa != 0)
2874 break;
2875 }
2876
2877 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
2878 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
2879 if (meta_hdr->b_spa != 0)
2880 break;
2881 }
2882
2883 if (data_hdr == NULL && meta_hdr == NULL) {
2884 type = ARC_BUFC_DATA;
2885 } else if (data_hdr == NULL) {
2886 ASSERT3P(meta_hdr, !=, NULL);
2887 type = ARC_BUFC_METADATA;
2888 } else if (meta_hdr == NULL) {
2889 ASSERT3P(data_hdr, !=, NULL);
2890 type = ARC_BUFC_DATA;
2891 } else {
2892 ASSERT3P(data_hdr, !=, NULL);
2893 ASSERT3P(meta_hdr, !=, NULL);
2894
2895 /* The headers can't be on the sublist without an L1 header */
2896 ASSERT(HDR_HAS_L1HDR(data_hdr));
2897 ASSERT(HDR_HAS_L1HDR(meta_hdr));
2898
2899 if (data_hdr->b_l1hdr.b_arc_access <
2900 meta_hdr->b_l1hdr.b_arc_access) {
2901 type = ARC_BUFC_DATA;
2902 } else {
2903 type = ARC_BUFC_METADATA;
2904 }
2905 }
2906
2907 multilist_sublist_unlock(meta_mls);
2908 multilist_sublist_unlock(data_mls);
2909
2910 return (type);
2911 }
2912
2913 /*
2914 * Evict buffers from the cache, such that arc_size is capped by arc_c.
2915 */
2916 static uint64_t
2917 arc_adjust(void)
2918 {
2919 uint64_t total_evicted = 0;
2920 uint64_t bytes;
2921 int64_t target;
2922
2923 /*
2924 * If we're over arc_meta_limit, we want to correct that before
2925 * potentially evicting data buffers below.
2926 */
2927 total_evicted += arc_adjust_meta();
2928
2929 /*
2930 * Adjust MRU size
2931 *
2932 * If we're over the target cache size, we want to evict enough
2933 * from the list to get back to our target size. We don't want
2934 * to evict too much from the MRU, such that it drops below
2935 * arc_p. So, if we're over our target cache size more than
2936 * the MRU is over arc_p, we'll evict enough to get back to
2937 * arc_p here, and then evict more from the MFU below.
2938 */
2939 target = MIN((int64_t)(arc_size - arc_c),
2940 (int64_t)(refcount_count(&arc_anon->arcs_size) +
2941 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
2942
2943 /*
2944 * If we're below arc_meta_min, always prefer to evict data.
2945 * Otherwise, try to satisfy the requested number of bytes to
2946 * evict from the type which contains older buffers; in an
2947 * effort to keep newer buffers in the cache regardless of their
2948 * type. If we cannot satisfy the number of bytes from this
2949 * type, spill over into the next type.
2950 */
2951 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
2952 arc_meta_used > arc_meta_min) {
2953 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2954 total_evicted += bytes;
2955
2956 /*
2957 * If we couldn't evict our target number of bytes from
2958 * metadata, we try to get the rest from data.
2959 */
2960 target -= bytes;
2961
2962 total_evicted +=
2963 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
2964 } else {
2965 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
2966 total_evicted += bytes;
2967
2968 /*
2969 * If we couldn't evict our target number of bytes from
2970 * data, we try to get the rest from metadata.
2971 */
2972 target -= bytes;
2973
2974 total_evicted +=
2975 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2976 }
2977
2978 /*
2979 * Adjust MFU size
2980 *
2981 * Now that we've tried to evict enough from the MRU to get its
2982 * size back to arc_p, if we're still above the target cache
2983 * size, we evict the rest from the MFU.
2984 */
2985 target = arc_size - arc_c;
2986
2987 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
2988 arc_meta_used > arc_meta_min) {
2989 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
2990 total_evicted += bytes;
2991
2992 /*
2993 * If we couldn't evict our target number of bytes from
2994 * metadata, we try to get the rest from data.
2995 */
2996 target -= bytes;
2997
2998 total_evicted +=
2999 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3000 } else {
3001 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3002 total_evicted += bytes;
3003
3004 /*
3005 * If we couldn't evict our target number of bytes from
3006 * data, we try to get the rest from data.
3007 */
3008 target -= bytes;
3009
3010 total_evicted +=
3011 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3012 }
3013
3014 /*
3015 * Adjust ghost lists
3016 *
3017 * In addition to the above, the ARC also defines target values
3018 * for the ghost lists. The sum of the mru list and mru ghost
3019 * list should never exceed the target size of the cache, and
3020 * the sum of the mru list, mfu list, mru ghost list, and mfu
3021 * ghost list should never exceed twice the target size of the
3022 * cache. The following logic enforces these limits on the ghost
3023 * caches, and evicts from them as needed.
3024 */
3025 target = refcount_count(&arc_mru->arcs_size) +
3026 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3027
3028 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3029 total_evicted += bytes;
3030
3031 target -= bytes;
3032
3033 total_evicted +=
3034 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3035
3036 /*
3037 * We assume the sum of the mru list and mfu list is less than
3038 * or equal to arc_c (we enforced this above), which means we
3039 * can use the simpler of the two equations below:
3040 *
3041 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3042 * mru ghost + mfu ghost <= arc_c
3043 */
3044 target = refcount_count(&arc_mru_ghost->arcs_size) +
3045 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3046
3047 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3048 total_evicted += bytes;
3049
3050 target -= bytes;
3051
3052 total_evicted +=
3053 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3054
3055 return (total_evicted);
3056 }
3057
3058 static void
3059 arc_do_user_evicts(void)
3060 {
3061 mutex_enter(&arc_user_evicts_lock);
3062 while (arc_eviction_list != NULL) {
3063 arc_buf_t *buf = arc_eviction_list;
3064 arc_eviction_list = buf->b_next;
3065 mutex_enter(&buf->b_evict_lock);
3066 buf->b_hdr = NULL;
3067 mutex_exit(&buf->b_evict_lock);
3068 mutex_exit(&arc_user_evicts_lock);
3069
3070 if (buf->b_efunc != NULL)
3071 VERIFY0(buf->b_efunc(buf->b_private));
3072
3073 buf->b_efunc = NULL;
3074 buf->b_private = NULL;
3075 kmem_cache_free(buf_cache, buf);
3076 mutex_enter(&arc_user_evicts_lock);
3077 }
3078 mutex_exit(&arc_user_evicts_lock);
3079 }
3080
3081 void
3082 arc_flush(spa_t *spa, boolean_t retry)
3083 {
3084 uint64_t guid = 0;
3085
3086 /*
3087 * If retry is TRUE, a spa must not be specified since we have
3088 * no good way to determine if all of a spa's buffers have been
3089 * evicted from an arc state.
3090 */
3091 ASSERT(!retry || spa == 0);
3092
3093 if (spa != NULL)
3094 guid = spa_load_guid(spa);
3095
3096 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3097 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3098
3099 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3100 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3101
3102 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3103 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3104
3105 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3106 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3107
3108 arc_do_user_evicts();
3109 ASSERT(spa || arc_eviction_list == NULL);
3110 }
3111
3112 void
3113 arc_shrink(int64_t to_free)
3114 {
3115 if (arc_c > arc_c_min) {
3116
3117 if (arc_c > arc_c_min + to_free)
3118 atomic_add_64(&arc_c, -to_free);
3119 else
3120 arc_c = arc_c_min;
3121
3122 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3123 if (arc_c > arc_size)
3124 arc_c = MAX(arc_size, arc_c_min);
3125 if (arc_p > arc_c)
3126 arc_p = (arc_c >> 1);
3127 ASSERT(arc_c >= arc_c_min);
3128 ASSERT((int64_t)arc_p >= 0);
3129 }
3130
3131 if (arc_size > arc_c)
3132 (void) arc_adjust();
3133 }
3134
3135 typedef enum free_memory_reason_t {
3136 FMR_UNKNOWN,
3137 FMR_NEEDFREE,
3138 FMR_LOTSFREE,
3139 FMR_SWAPFS_MINFREE,
3140 FMR_PAGES_PP_MAXIMUM,
3141 FMR_HEAP_ARENA,
3142 FMR_ZIO_ARENA,
3143 } free_memory_reason_t;
3144
3145 int64_t last_free_memory;
3146 free_memory_reason_t last_free_reason;
3147
3148 /*
3149 * Additional reserve of pages for pp_reserve.
3150 */
3151 int64_t arc_pages_pp_reserve = 64;
3152
3153 /*
3154 * Additional reserve of pages for swapfs.
3155 */
3156 int64_t arc_swapfs_reserve = 64;
3157
3158 /*
3159 * Return the amount of memory that can be consumed before reclaim will be
3160 * needed. Positive if there is sufficient free memory, negative indicates
3161 * the amount of memory that needs to be freed up.
3162 */
3163 static int64_t
3164 arc_available_memory(void)
3165 {
3166 int64_t lowest = INT64_MAX;
3167 int64_t n;
3168 free_memory_reason_t r = FMR_UNKNOWN;
3169
3170 #ifdef _KERNEL
3171 if (needfree > 0) {
3172 n = PAGESIZE * (-needfree);
3173 if (n < lowest) {
3174 lowest = n;
3175 r = FMR_NEEDFREE;
3176 }
3177 }
3178
3179 /*
3180 * check that we're out of range of the pageout scanner. It starts to
3181 * schedule paging if freemem is less than lotsfree and needfree.
3182 * lotsfree is the high-water mark for pageout, and needfree is the
3183 * number of needed free pages. We add extra pages here to make sure
3184 * the scanner doesn't start up while we're freeing memory.
3185 */
3186 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3187 if (n < lowest) {
3188 lowest = n;
3189 r = FMR_LOTSFREE;
3190 }
3191
3192 /*
3193 * check to make sure that swapfs has enough space so that anon
3194 * reservations can still succeed. anon_resvmem() checks that the
3195 * availrmem is greater than swapfs_minfree, and the number of reserved
3196 * swap pages. We also add a bit of extra here just to prevent
3197 * circumstances from getting really dire.
3198 */
3199 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3200 desfree - arc_swapfs_reserve);
3201 if (n < lowest) {
3202 lowest = n;
3203 r = FMR_SWAPFS_MINFREE;
3204 }
3205
3206
3207 /*
3208 * Check that we have enough availrmem that memory locking (e.g., via
3209 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3210 * stores the number of pages that cannot be locked; when availrmem
3211 * drops below pages_pp_maximum, page locking mechanisms such as
3212 * page_pp_lock() will fail.)
3213 */
3214 n = PAGESIZE * (availrmem - pages_pp_maximum -
3215 arc_pages_pp_reserve);
3216 if (n < lowest) {
3217 lowest = n;
3218 r = FMR_PAGES_PP_MAXIMUM;
3219 }
3220
3221 #if defined(__i386)
3222 /*
3223 * If we're on an i386 platform, it's possible that we'll exhaust the
3224 * kernel heap space before we ever run out of available physical
3225 * memory. Most checks of the size of the heap_area compare against
3226 * tune.t_minarmem, which is the minimum available real memory that we
3227 * can have in the system. However, this is generally fixed at 25 pages
3228 * which is so low that it's useless. In this comparison, we seek to
3229 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3230 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3231 * free)
3232 */
3233 n = vmem_size(heap_arena, VMEM_FREE) -
3234 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3235 if (n < lowest) {
3236 lowest = n;
3237 r = FMR_HEAP_ARENA;
3238 }
3239 #endif
3240
3241 /*
3242 * If zio data pages are being allocated out of a separate heap segment,
3243 * then enforce that the size of available vmem for this arena remains
3244 * above about 1/16th free.
3245 *
3246 * Note: The 1/16th arena free requirement was put in place
3247 * to aggressively evict memory from the arc in order to avoid
3248 * memory fragmentation issues.
3249 */
3250 if (zio_arena != NULL) {
3251 n = vmem_size(zio_arena, VMEM_FREE) -
3252 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
3253 if (n < lowest) {
3254 lowest = n;
3255 r = FMR_ZIO_ARENA;
3256 }
3257 }
3258 #else
3259 /* Every 100 calls, free a small amount */
3260 if (spa_get_random(100) == 0)
3261 lowest = -1024;
3262 #endif
3263
3264 last_free_memory = lowest;
3265 last_free_reason = r;
3266
3267 return (lowest);
3268 }
3269
3270
3271 /*
3272 * Determine if the system is under memory pressure and is asking
3273 * to reclaim memory. A return value of TRUE indicates that the system
3274 * is under memory pressure and that the arc should adjust accordingly.
3275 */
3276 static boolean_t
3277 arc_reclaim_needed(void)
3278 {
3279 return (arc_available_memory() < 0);
3280 }
3281
3282 static void
3283 arc_kmem_reap_now(void)
3284 {
3285 size_t i;
3286 kmem_cache_t *prev_cache = NULL;
3287 kmem_cache_t *prev_data_cache = NULL;
3288 extern kmem_cache_t *zio_buf_cache[];
3289 extern kmem_cache_t *zio_data_buf_cache[];
3290 extern kmem_cache_t *range_seg_cache;
3291
3292 #ifdef _KERNEL
3293 if (arc_meta_used >= arc_meta_limit) {
3294 /*
3295 * We are exceeding our meta-data cache limit.
3296 * Purge some DNLC entries to release holds on meta-data.
3297 */
3298 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
3299 }
3300 #if defined(__i386)
3301 /*
3302 * Reclaim unused memory from all kmem caches.
3303 */
3304 kmem_reap();
3305 #endif
3306 #endif
3307
3308 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
3309 if (zio_buf_cache[i] != prev_cache) {
3310 prev_cache = zio_buf_cache[i];
3311 kmem_cache_reap_now(zio_buf_cache[i]);
3312 }
3313 if (zio_data_buf_cache[i] != prev_data_cache) {
3314 prev_data_cache = zio_data_buf_cache[i];
3315 kmem_cache_reap_now(zio_data_buf_cache[i]);
3316 }
3317 }
3318 kmem_cache_reap_now(buf_cache);
3319 kmem_cache_reap_now(hdr_full_cache);
3320 kmem_cache_reap_now(hdr_l2only_cache);
3321 kmem_cache_reap_now(range_seg_cache);
3322
3323 if (zio_arena != NULL) {
3324 /*
3325 * Ask the vmem arena to reclaim unused memory from its
3326 * quantum caches.
3327 */
3328 vmem_qcache_reap(zio_arena);
3329 }
3330 }
3331
3332 /*
3333 * Threads can block in arc_get_data_buf() waiting for this thread to evict
3334 * enough data and signal them to proceed. When this happens, the threads in
3335 * arc_get_data_buf() are sleeping while holding the hash lock for their
3336 * particular arc header. Thus, we must be careful to never sleep on a
3337 * hash lock in this thread. This is to prevent the following deadlock:
3338 *
3339 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3340 * waiting for the reclaim thread to signal it.
3341 *
3342 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3343 * fails, and goes to sleep forever.
3344 *
3345 * This possible deadlock is avoided by always acquiring a hash lock
3346 * using mutex_tryenter() from arc_reclaim_thread().
3347 */
3348 static void
3349 arc_reclaim_thread(void)
3350 {
3351 clock_t growtime = 0;
3352 callb_cpr_t cpr;
3353
3354 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
3355
3356 mutex_enter(&arc_reclaim_lock);
3357 while (!arc_reclaim_thread_exit) {
3358 int64_t free_memory = arc_available_memory();
3359 uint64_t evicted = 0;
3360
3361 mutex_exit(&arc_reclaim_lock);
3362
3363 if (free_memory < 0) {
3364
3365 arc_no_grow = B_TRUE;
3366 arc_warm = B_TRUE;
3367
3368 /*
3369 * Wait at least zfs_grow_retry (default 60) seconds
3370 * before considering growing.
3371 */
3372 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
3373
3374 arc_kmem_reap_now();
3375
3376 /*
3377 * If we are still low on memory, shrink the ARC
3378 * so that we have arc_shrink_min free space.
3379 */
3380 free_memory = arc_available_memory();
3381
3382 int64_t to_free =
3383 (arc_c >> arc_shrink_shift) - free_memory;
3384 if (to_free > 0) {
3385 #ifdef _KERNEL
3386 to_free = MAX(to_free, ptob(needfree));
3387 #endif
3388 arc_shrink(to_free);
3389 }
3390 } else if (free_memory < arc_c >> arc_no_grow_shift) {
3391 arc_no_grow = B_TRUE;
3392 } else if (ddi_get_lbolt() >= growtime) {
3393 arc_no_grow = B_FALSE;
3394 }
3395
3396 evicted = arc_adjust();
3397
3398 mutex_enter(&arc_reclaim_lock);
3399
3400 /*
3401 * If evicted is zero, we couldn't evict anything via
3402 * arc_adjust(). This could be due to hash lock
3403 * collisions, but more likely due to the majority of
3404 * arc buffers being unevictable. Therefore, even if
3405 * arc_size is above arc_c, another pass is unlikely to
3406 * be helpful and could potentially cause us to enter an
3407 * infinite loop.
3408 */
3409 if (arc_size <= arc_c || evicted == 0) {
3410 /*
3411 * We're either no longer overflowing, or we
3412 * can't evict anything more, so we should wake
3413 * up any threads before we go to sleep.
3414 */
3415 cv_broadcast(&arc_reclaim_waiters_cv);
3416
3417 /*
3418 * Block until signaled, or after one second (we
3419 * might need to perform arc_kmem_reap_now()
3420 * even if we aren't being signalled)
3421 */
3422 CALLB_CPR_SAFE_BEGIN(&cpr);
3423 (void) cv_timedwait(&arc_reclaim_thread_cv,
3424 &arc_reclaim_lock, ddi_get_lbolt() + hz);
3425 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
3426 }
3427 }
3428
3429 arc_reclaim_thread_exit = FALSE;
3430 cv_broadcast(&arc_reclaim_thread_cv);
3431 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
3432 thread_exit();
3433 }
3434
3435 static void
3436 arc_user_evicts_thread(void)
3437 {
3438 callb_cpr_t cpr;
3439
3440 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG);
3441
3442 mutex_enter(&arc_user_evicts_lock);
3443 while (!arc_user_evicts_thread_exit) {
3444 mutex_exit(&arc_user_evicts_lock);
3445
3446 arc_do_user_evicts();
3447
3448 /*
3449 * This is necessary in order for the mdb ::arc dcmd to
3450 * show up to date information. Since the ::arc command
3451 * does not call the kstat's update function, without
3452 * this call, the command may show stale stats for the
3453 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3454 * with this change, the data might be up to 1 second
3455 * out of date; but that should suffice. The arc_state_t
3456 * structures can be queried directly if more accurate
3457 * information is needed.
3458 */
3459 if (arc_ksp != NULL)
3460 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
3461
3462 mutex_enter(&arc_user_evicts_lock);
3463
3464 /*
3465 * Block until signaled, or after one second (we need to
3466 * call the arc's kstat update function regularly).
3467 */
3468 CALLB_CPR_SAFE_BEGIN(&cpr);
3469 (void) cv_timedwait(&arc_user_evicts_cv,
3470 &arc_user_evicts_lock, ddi_get_lbolt() + hz);
3471 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock);
3472 }
3473
3474 arc_user_evicts_thread_exit = FALSE;
3475 cv_broadcast(&arc_user_evicts_cv);
3476 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */
3477 thread_exit();
3478 }
3479
3480 /*
3481 * Adapt arc info given the number of bytes we are trying to add and
3482 * the state that we are comming from. This function is only called
3483 * when we are adding new content to the cache.
3484 */
3485 static void
3486 arc_adapt(int bytes, arc_state_t *state)
3487 {
3488 int mult;
3489 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
3490 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
3491 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
3492
3493 if (state == arc_l2c_only)
3494 return;
3495
3496 ASSERT(bytes > 0);
3497 /*
3498 * Adapt the target size of the MRU list:
3499 * - if we just hit in the MRU ghost list, then increase
3500 * the target size of the MRU list.
3501 * - if we just hit in the MFU ghost list, then increase
3502 * the target size of the MFU list by decreasing the
3503 * target size of the MRU list.
3504 */
3505 if (state == arc_mru_ghost) {
3506 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
3507 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
3508
3509 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
3510 } else if (state == arc_mfu_ghost) {
3511 uint64_t delta;
3512
3513 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
3514 mult = MIN(mult, 10);
3515
3516 delta = MIN(bytes * mult, arc_p);
3517 arc_p = MAX(arc_p_min, arc_p - delta);
3518 }
3519 ASSERT((int64_t)arc_p >= 0);
3520
3521 if (arc_reclaim_needed()) {
3522 cv_signal(&arc_reclaim_thread_cv);
3523 return;
3524 }
3525
3526 if (arc_no_grow)
3527 return;
3528
3529 if (arc_c >= arc_c_max)
3530 return;
3531
3532 /*
3533 * If we're within (2 * maxblocksize) bytes of the target
3534 * cache size, increment the target cache size
3535 */
3536 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
3537 atomic_add_64(&arc_c, (int64_t)bytes);
3538 if (arc_c > arc_c_max)
3539 arc_c = arc_c_max;
3540 else if (state == arc_anon)
3541 atomic_add_64(&arc_p, (int64_t)bytes);
3542 if (arc_p > arc_c)
3543 arc_p = arc_c;
3544 }
3545 ASSERT((int64_t)arc_p >= 0);
3546 }
3547
3548 /*
3549 * Check if arc_size has grown past our upper threshold, determined by
3550 * zfs_arc_overflow_shift.
3551 */
3552 static boolean_t
3553 arc_is_overflowing(void)
3554 {
3555 /* Always allow at least one block of overflow */
3556 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
3557 arc_c >> zfs_arc_overflow_shift);
3558
3559 return (arc_size >= arc_c + overflow);
3560 }
3561
3562 /*
3563 * The buffer, supplied as the first argument, needs a data block. If we
3564 * are hitting the hard limit for the cache size, we must sleep, waiting
3565 * for the eviction thread to catch up. If we're past the target size
3566 * but below the hard limit, we'll only signal the reclaim thread and
3567 * continue on.
3568 */
3569 static void
3570 arc_get_data_buf(arc_buf_t *buf)
3571 {
3572 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
3573 uint64_t size = buf->b_hdr->b_size;
3574 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
3575
3576 arc_adapt(size, state);
3577
3578 /*
3579 * If arc_size is currently overflowing, and has grown past our
3580 * upper limit, we must be adding data faster than the evict
3581 * thread can evict. Thus, to ensure we don't compound the
3582 * problem by adding more data and forcing arc_size to grow even
3583 * further past it's target size, we halt and wait for the
3584 * eviction thread to catch up.
3585 *
3586 * It's also possible that the reclaim thread is unable to evict
3587 * enough buffers to get arc_size below the overflow limit (e.g.
3588 * due to buffers being un-evictable, or hash lock collisions).
3589 * In this case, we want to proceed regardless if we're
3590 * overflowing; thus we don't use a while loop here.
3591 */
3592 if (arc_is_overflowing()) {
3593 mutex_enter(&arc_reclaim_lock);
3594
3595 /*
3596 * Now that we've acquired the lock, we may no longer be
3597 * over the overflow limit, lets check.
3598 *
3599 * We're ignoring the case of spurious wake ups. If that
3600 * were to happen, it'd let this thread consume an ARC
3601 * buffer before it should have (i.e. before we're under
3602 * the overflow limit and were signalled by the reclaim
3603 * thread). As long as that is a rare occurrence, it
3604 * shouldn't cause any harm.
3605 */
3606 if (arc_is_overflowing()) {
3607 cv_signal(&arc_reclaim_thread_cv);
3608 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
3609 }
3610
3611 mutex_exit(&arc_reclaim_lock);
3612 }
3613
3614 if (type == ARC_BUFC_METADATA) {
3615 buf->b_data = zio_buf_alloc(size);
3616 arc_space_consume(size, ARC_SPACE_META);
3617 } else {
3618 ASSERT(type == ARC_BUFC_DATA);
3619 buf->b_data = zio_data_buf_alloc(size);
3620 arc_space_consume(size, ARC_SPACE_DATA);
3621 }
3622
3623 /*
3624 * Update the state size. Note that ghost states have a
3625 * "ghost size" and so don't need to be updated.
3626 */
3627 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) {
3628 arc_buf_hdr_t *hdr = buf->b_hdr;
3629 arc_state_t *state = hdr->b_l1hdr.b_state;
3630
3631 (void) refcount_add_many(&state->arcs_size, size, buf);
3632
3633 /*
3634 * If this is reached via arc_read, the link is
3635 * protected by the hash lock. If reached via
3636 * arc_buf_alloc, the header should not be accessed by
3637 * any other thread. And, if reached via arc_read_done,
3638 * the hash lock will protect it if it's found in the
3639 * hash table; otherwise no other thread should be
3640 * trying to [add|remove]_reference it.
3641 */
3642 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3643 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3644 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type],
3645 size);
3646 }
3647 /*
3648 * If we are growing the cache, and we are adding anonymous
3649 * data, and we have outgrown arc_p, update arc_p
3650 */
3651 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
3652 (refcount_count(&arc_anon->arcs_size) +
3653 refcount_count(&arc_mru->arcs_size) > arc_p))
3654 arc_p = MIN(arc_c, arc_p + size);
3655 }
3656 }
3657
3658 /*
3659 * This routine is called whenever a buffer is accessed.
3660 * NOTE: the hash lock is dropped in this function.
3661 */
3662 static void
3663 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3664 {
3665 clock_t now;
3666
3667 ASSERT(MUTEX_HELD(hash_lock));
3668 ASSERT(HDR_HAS_L1HDR(hdr));
3669
3670 if (hdr->b_l1hdr.b_state == arc_anon) {
3671 /*
3672 * This buffer is not in the cache, and does not
3673 * appear in our "ghost" list. Add the new buffer
3674 * to the MRU state.
3675 */
3676
3677 ASSERT0(hdr->b_l1hdr.b_arc_access);
3678 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3679 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3680 arc_change_state(arc_mru, hdr, hash_lock);
3681
3682 } else if (hdr->b_l1hdr.b_state == arc_mru) {
3683 now = ddi_get_lbolt();
3684
3685 /*
3686 * If this buffer is here because of a prefetch, then either:
3687 * - clear the flag if this is a "referencing" read
3688 * (any subsequent access will bump this into the MFU state).
3689 * or
3690 * - move the buffer to the head of the list if this is
3691 * another prefetch (to make it less likely to be evicted).
3692 */
3693 if (HDR_PREFETCH(hdr)) {
3694 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
3695 /* link protected by hash lock */
3696 ASSERT(multilist_link_active(
3697 &hdr->b_l1hdr.b_arc_node));
3698 } else {
3699 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3700 ARCSTAT_BUMP(arcstat_mru_hits);
3701 }
3702 hdr->b_l1hdr.b_arc_access = now;
3703 return;
3704 }
3705
3706 /*
3707 * This buffer has been "accessed" only once so far,
3708 * but it is still in the cache. Move it to the MFU
3709 * state.
3710 */
3711 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
3712 /*
3713 * More than 125ms have passed since we
3714 * instantiated this buffer. Move it to the
3715 * most frequently used state.
3716 */
3717 hdr->b_l1hdr.b_arc_access = now;
3718 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3719 arc_change_state(arc_mfu, hdr, hash_lock);
3720 }
3721 ARCSTAT_BUMP(arcstat_mru_hits);
3722 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
3723 arc_state_t *new_state;
3724 /*
3725 * This buffer has been "accessed" recently, but
3726 * was evicted from the cache. Move it to the
3727 * MFU state.
3728 */
3729
3730 if (HDR_PREFETCH(hdr)) {
3731 new_state = arc_mru;
3732 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
3733 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3734 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3735 } else {
3736 new_state = arc_mfu;
3737 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3738 }
3739
3740 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3741 arc_change_state(new_state, hdr, hash_lock);
3742
3743 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
3744 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
3745 /*
3746 * This buffer has been accessed more than once and is
3747 * still in the cache. Keep it in the MFU state.
3748 *
3749 * NOTE: an add_reference() that occurred when we did
3750 * the arc_read() will have kicked this off the list.
3751 * If it was a prefetch, we will explicitly move it to
3752 * the head of the list now.
3753 */
3754 if ((HDR_PREFETCH(hdr)) != 0) {
3755 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3756 /* link protected by hash_lock */
3757 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3758 }
3759 ARCSTAT_BUMP(arcstat_mfu_hits);
3760 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3761 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
3762 arc_state_t *new_state = arc_mfu;
3763 /*
3764 * This buffer has been accessed more than once but has
3765 * been evicted from the cache. Move it back to the
3766 * MFU state.
3767 */
3768
3769 if (HDR_PREFETCH(hdr)) {
3770 /*
3771 * This is a prefetch access...
3772 * move this block back to the MRU state.
3773 */
3774 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3775 new_state = arc_mru;
3776 }
3777
3778 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3779 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3780 arc_change_state(new_state, hdr, hash_lock);
3781
3782 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
3783 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
3784 /*
3785 * This buffer is on the 2nd Level ARC.
3786 */
3787
3788 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3789 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3790 arc_change_state(arc_mfu, hdr, hash_lock);
3791 } else {
3792 ASSERT(!"invalid arc state");
3793 }
3794 }
3795
3796 /* a generic arc_done_func_t which you can use */
3797 /* ARGSUSED */
3798 void
3799 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
3800 {
3801 if (zio == NULL || zio->io_error == 0)
3802 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
3803 VERIFY(arc_buf_remove_ref(buf, arg));
3804 }
3805
3806 /* a generic arc_done_func_t */
3807 void
3808 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
3809 {
3810 arc_buf_t **bufp = arg;
3811 if (zio && zio->io_error) {
3812 VERIFY(arc_buf_remove_ref(buf, arg));
3813 *bufp = NULL;
3814 } else {
3815 *bufp = buf;
3816 ASSERT(buf->b_data);
3817 }
3818 }
3819
3820 static void
3821 arc_read_done(zio_t *zio)
3822 {
3823 arc_buf_hdr_t *hdr;
3824 arc_buf_t *buf;
3825 arc_buf_t *abuf; /* buffer we're assigning to callback */
3826 kmutex_t *hash_lock = NULL;
3827 arc_callback_t *callback_list, *acb;
3828 int freeable = FALSE;
3829
3830 buf = zio->io_private;
3831 hdr = buf->b_hdr;
3832
3833 /*
3834 * The hdr was inserted into hash-table and removed from lists
3835 * prior to starting I/O. We should find this header, since
3836 * it's in the hash table, and it should be legit since it's
3837 * not possible to evict it during the I/O. The only possible
3838 * reason for it not to be found is if we were freed during the
3839 * read.
3840 */
3841 if (HDR_IN_HASH_TABLE(hdr)) {
3842 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
3843 ASSERT3U(hdr->b_dva.dva_word[0], ==,
3844 BP_IDENTITY(zio->io_bp)->dva_word[0]);
3845 ASSERT3U(hdr->b_dva.dva_word[1], ==,
3846 BP_IDENTITY(zio->io_bp)->dva_word[1]);
3847
3848 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
3849 &hash_lock);
3850
3851 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
3852 hash_lock == NULL) ||
3853 (found == hdr &&
3854 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
3855 (found == hdr && HDR_L2_READING(hdr)));
3856 }
3857
3858 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED;
3859 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
3860 hdr->b_flags &= ~ARC_FLAG_L2CACHE;
3861
3862 /* byteswap if necessary */
3863 callback_list = hdr->b_l1hdr.b_acb;
3864 ASSERT(callback_list != NULL);
3865 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
3866 dmu_object_byteswap_t bswap =
3867 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
3868 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
3869 byteswap_uint64_array :
3870 dmu_ot_byteswap[bswap].ob_func;
3871 func(buf->b_data, hdr->b_size);
3872 }
3873
3874 arc_cksum_compute(buf, B_FALSE);
3875 arc_buf_watch(buf);
3876
3877 if (hash_lock && zio->io_error == 0 &&
3878 hdr->b_l1hdr.b_state == arc_anon) {
3879 /*
3880 * Only call arc_access on anonymous buffers. This is because
3881 * if we've issued an I/O for an evicted buffer, we've already
3882 * called arc_access (to prevent any simultaneous readers from
3883 * getting confused).
3884 */
3885 arc_access(hdr, hash_lock);
3886 }
3887
3888 /* create copies of the data buffer for the callers */
3889 abuf = buf;
3890 for (acb = callback_list; acb; acb = acb->acb_next) {
3891 if (acb->acb_done) {
3892 if (abuf == NULL) {
3893 ARCSTAT_BUMP(arcstat_duplicate_reads);
3894 abuf = arc_buf_clone(buf);
3895 }
3896 acb->acb_buf = abuf;
3897 abuf = NULL;
3898 }
3899 }
3900 hdr->b_l1hdr.b_acb = NULL;
3901 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
3902 ASSERT(!HDR_BUF_AVAILABLE(hdr));
3903 if (abuf == buf) {
3904 ASSERT(buf->b_efunc == NULL);
3905 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
3906 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
3907 }
3908
3909 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
3910 callback_list != NULL);
3911
3912 if (zio->io_error != 0) {
3913 hdr->b_flags |= ARC_FLAG_IO_ERROR;
3914 if (hdr->b_l1hdr.b_state != arc_anon)
3915 arc_change_state(arc_anon, hdr, hash_lock);
3916 if (HDR_IN_HASH_TABLE(hdr))
3917 buf_hash_remove(hdr);
3918 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
3919 }
3920
3921 /*
3922 * Broadcast before we drop the hash_lock to avoid the possibility
3923 * that the hdr (and hence the cv) might be freed before we get to
3924 * the cv_broadcast().
3925 */
3926 cv_broadcast(&hdr->b_l1hdr.b_cv);
3927
3928 if (hash_lock != NULL) {
3929 mutex_exit(hash_lock);
3930 } else {
3931 /*
3932 * This block was freed while we waited for the read to
3933 * complete. It has been removed from the hash table and
3934 * moved to the anonymous state (so that it won't show up
3935 * in the cache).
3936 */
3937 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3938 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
3939 }
3940
3941 /* execute each callback and free its structure */
3942 while ((acb = callback_list) != NULL) {
3943 if (acb->acb_done)
3944 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3945
3946 if (acb->acb_zio_dummy != NULL) {
3947 acb->acb_zio_dummy->io_error = zio->io_error;
3948 zio_nowait(acb->acb_zio_dummy);
3949 }
3950
3951 callback_list = acb->acb_next;
3952 kmem_free(acb, sizeof (arc_callback_t));
3953 }
3954
3955 if (freeable)
3956 arc_hdr_destroy(hdr);
3957 }
3958
3959 /*
3960 * "Read" the block at the specified DVA (in bp) via the
3961 * cache. If the block is found in the cache, invoke the provided
3962 * callback immediately and return. Note that the `zio' parameter
3963 * in the callback will be NULL in this case, since no IO was
3964 * required. If the block is not in the cache pass the read request
3965 * on to the spa with a substitute callback function, so that the
3966 * requested block will be added to the cache.
3967 *
3968 * If a read request arrives for a block that has a read in-progress,
3969 * either wait for the in-progress read to complete (and return the
3970 * results); or, if this is a read with a "done" func, add a record
3971 * to the read to invoke the "done" func when the read completes,
3972 * and return; or just return.
3973 *
3974 * arc_read_done() will invoke all the requested "done" functions
3975 * for readers of this block.
3976 */
3977 int
3978 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
3979 void *private, zio_priority_t priority, int zio_flags,
3980 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
3981 {
3982 arc_buf_hdr_t *hdr = NULL;
3983 arc_buf_t *buf = NULL;
3984 kmutex_t *hash_lock = NULL;
3985 zio_t *rzio;
3986 uint64_t guid = spa_load_guid(spa);
3987
3988 ASSERT(!BP_IS_EMBEDDED(bp) ||
3989 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
3990
3991 top:
3992 if (!BP_IS_EMBEDDED(bp)) {
3993 /*
3994 * Embedded BP's have no DVA and require no I/O to "read".
3995 * Create an anonymous arc buf to back it.
3996 */
3997 hdr = buf_hash_find(guid, bp, &hash_lock);
3998 }
3999
4000 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) {
4001
4002 *arc_flags |= ARC_FLAG_CACHED;
4003
4004 if (HDR_IO_IN_PROGRESS(hdr)) {
4005
4006 if (*arc_flags & ARC_FLAG_WAIT) {
4007 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4008 mutex_exit(hash_lock);
4009 goto top;
4010 }
4011 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4012
4013 if (done) {
4014 arc_callback_t *acb = NULL;
4015
4016 acb = kmem_zalloc(sizeof (arc_callback_t),
4017 KM_SLEEP);
4018 acb->acb_done = done;
4019 acb->acb_private = private;
4020 if (pio != NULL)
4021 acb->acb_zio_dummy = zio_null(pio,
4022 spa, NULL, NULL, NULL, zio_flags);
4023
4024 ASSERT(acb->acb_done != NULL);
4025 acb->acb_next = hdr->b_l1hdr.b_acb;
4026 hdr->b_l1hdr.b_acb = acb;
4027 add_reference(hdr, hash_lock, private);
4028 mutex_exit(hash_lock);
4029 return (0);
4030 }
4031 mutex_exit(hash_lock);
4032 return (0);
4033 }
4034
4035 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4036 hdr->b_l1hdr.b_state == arc_mfu);
4037
4038 if (done) {
4039 add_reference(hdr, hash_lock, private);
4040 /*
4041 * If this block is already in use, create a new
4042 * copy of the data so that we will be guaranteed
4043 * that arc_release() will always succeed.
4044 */
4045 buf = hdr->b_l1hdr.b_buf;
4046 ASSERT(buf);
4047 ASSERT(buf->b_data);
4048 if (HDR_BUF_AVAILABLE(hdr)) {
4049 ASSERT(buf->b_efunc == NULL);
4050 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4051 } else {
4052 buf = arc_buf_clone(buf);
4053 }
4054
4055 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4056 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4057 hdr->b_flags |= ARC_FLAG_PREFETCH;
4058 }
4059 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4060 arc_access(hdr, hash_lock);
4061 if (*arc_flags & ARC_FLAG_L2CACHE)
4062 hdr->b_flags |= ARC_FLAG_L2CACHE;
4063 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4064 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4065 mutex_exit(hash_lock);
4066 ARCSTAT_BUMP(arcstat_hits);
4067 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4068 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4069 data, metadata, hits);
4070
4071 if (done)
4072 done(NULL, buf, private);
4073 } else {
4074 uint64_t size = BP_GET_LSIZE(bp);
4075 arc_callback_t *acb;
4076 vdev_t *vd = NULL;
4077 uint64_t addr = 0;
4078 boolean_t devw = B_FALSE;
4079 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
4080 int32_t b_asize = 0;
4081
4082 if (hdr == NULL) {
4083 /* this block is not in the cache */
4084 arc_buf_hdr_t *exists = NULL;
4085 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4086 buf = arc_buf_alloc(spa, size, private, type);
4087 hdr = buf->b_hdr;
4088 if (!BP_IS_EMBEDDED(bp)) {
4089 hdr->b_dva = *BP_IDENTITY(bp);
4090 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4091 exists = buf_hash_insert(hdr, &hash_lock);
4092 }
4093 if (exists != NULL) {
4094 /* somebody beat us to the hash insert */
4095 mutex_exit(hash_lock);
4096 buf_discard_identity(hdr);
4097 (void) arc_buf_remove_ref(buf, private);
4098 goto top; /* restart the IO request */
4099 }
4100
4101 /* if this is a prefetch, we don't have a reference */
4102 if (*arc_flags & ARC_FLAG_PREFETCH) {
4103 (void) remove_reference(hdr, hash_lock,
4104 private);
4105 hdr->b_flags |= ARC_FLAG_PREFETCH;
4106 }
4107 if (*arc_flags & ARC_FLAG_L2CACHE)
4108 hdr->b_flags |= ARC_FLAG_L2CACHE;
4109 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4110 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4111 if (BP_GET_LEVEL(bp) > 0)
4112 hdr->b_flags |= ARC_FLAG_INDIRECT;
4113 } else {
4114 /*
4115 * This block is in the ghost cache. If it was L2-only
4116 * (and thus didn't have an L1 hdr), we realloc the
4117 * header to add an L1 hdr.
4118 */
4119 if (!HDR_HAS_L1HDR(hdr)) {
4120 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4121 hdr_full_cache);
4122 }
4123
4124 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4125 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4126 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4127 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
4128
4129 /* if this is a prefetch, we don't have a reference */
4130 if (*arc_flags & ARC_FLAG_PREFETCH)
4131 hdr->b_flags |= ARC_FLAG_PREFETCH;
4132 else
4133 add_reference(hdr, hash_lock, private);
4134 if (*arc_flags & ARC_FLAG_L2CACHE)
4135 hdr->b_flags |= ARC_FLAG_L2CACHE;
4136 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4137 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4138 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
4139 buf->b_hdr = hdr;
4140 buf->b_data = NULL;
4141 buf->b_efunc = NULL;
4142 buf->b_private = NULL;
4143 buf->b_next = NULL;
4144 hdr->b_l1hdr.b_buf = buf;
4145 ASSERT0(hdr->b_l1hdr.b_datacnt);
4146 hdr->b_l1hdr.b_datacnt = 1;
4147 arc_get_data_buf(buf);
4148 arc_access(hdr, hash_lock);
4149 }
4150
4151 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
4152
4153 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
4154 acb->acb_done = done;
4155 acb->acb_private = private;
4156
4157 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4158 hdr->b_l1hdr.b_acb = acb;
4159 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4160
4161 if (HDR_HAS_L2HDR(hdr) &&
4162 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
4163 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
4164 addr = hdr->b_l2hdr.b_daddr;
4165 b_compress = hdr->b_l2hdr.b_compress;
4166 b_asize = hdr->b_l2hdr.b_asize;
4167 /*
4168 * Lock out device removal.
4169 */
4170 if (vdev_is_dead(vd) ||
4171 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
4172 vd = NULL;
4173 }
4174
4175 if (hash_lock != NULL)
4176 mutex_exit(hash_lock);
4177
4178 /*
4179 * At this point, we have a level 1 cache miss. Try again in
4180 * L2ARC if possible.
4181 */
4182 ASSERT3U(hdr->b_size, ==, size);
4183 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
4184 uint64_t, size, zbookmark_phys_t *, zb);
4185 ARCSTAT_BUMP(arcstat_misses);
4186 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4187 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4188 data, metadata, misses);
4189
4190 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
4191 /*
4192 * Read from the L2ARC if the following are true:
4193 * 1. The L2ARC vdev was previously cached.
4194 * 2. This buffer still has L2ARC metadata.
4195 * 3. This buffer isn't currently writing to the L2ARC.
4196 * 4. The L2ARC entry wasn't evicted, which may
4197 * also have invalidated the vdev.
4198 * 5. This isn't prefetch and l2arc_noprefetch is set.
4199 */
4200 if (HDR_HAS_L2HDR(hdr) &&
4201 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
4202 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
4203 l2arc_read_callback_t *cb;
4204
4205 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
4206 ARCSTAT_BUMP(arcstat_l2_hits);
4207
4208 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
4209 KM_SLEEP);
4210 cb->l2rcb_buf = buf;
4211 cb->l2rcb_spa = spa;
4212 cb->l2rcb_bp = *bp;
4213 cb->l2rcb_zb = *zb;
4214 cb->l2rcb_flags = zio_flags;
4215 cb->l2rcb_compress = b_compress;
4216
4217 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
4218 addr + size < vd->vdev_psize -
4219 VDEV_LABEL_END_SIZE);
4220
4221 /*
4222 * l2arc read. The SCL_L2ARC lock will be
4223 * released by l2arc_read_done().
4224 * Issue a null zio if the underlying buffer
4225 * was squashed to zero size by compression.
4226 */
4227 if (b_compress == ZIO_COMPRESS_EMPTY) {
4228 rzio = zio_null(pio, spa, vd,
4229 l2arc_read_done, cb,
4230 zio_flags | ZIO_FLAG_DONT_CACHE |
4231 ZIO_FLAG_CANFAIL |
4232 ZIO_FLAG_DONT_PROPAGATE |
4233 ZIO_FLAG_DONT_RETRY);
4234 } else {
4235 rzio = zio_read_phys(pio, vd, addr,
4236 b_asize, buf->b_data,
4237 ZIO_CHECKSUM_OFF,
4238 l2arc_read_done, cb, priority,
4239 zio_flags | ZIO_FLAG_DONT_CACHE |
4240 ZIO_FLAG_CANFAIL |
4241 ZIO_FLAG_DONT_PROPAGATE |
4242 ZIO_FLAG_DONT_RETRY, B_FALSE);
4243 }
4244 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
4245 zio_t *, rzio);
4246 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
4247
4248 if (*arc_flags & ARC_FLAG_NOWAIT) {
4249 zio_nowait(rzio);
4250 return (0);
4251 }
4252
4253 ASSERT(*arc_flags & ARC_FLAG_WAIT);
4254 if (zio_wait(rzio) == 0)
4255 return (0);
4256
4257 /* l2arc read error; goto zio_read() */
4258 } else {
4259 DTRACE_PROBE1(l2arc__miss,
4260 arc_buf_hdr_t *, hdr);
4261 ARCSTAT_BUMP(arcstat_l2_misses);
4262 if (HDR_L2_WRITING(hdr))
4263 ARCSTAT_BUMP(arcstat_l2_rw_clash);
4264 spa_config_exit(spa, SCL_L2ARC, vd);
4265 }
4266 } else {
4267 if (vd != NULL)
4268 spa_config_exit(spa, SCL_L2ARC, vd);
4269 if (l2arc_ndev != 0) {
4270 DTRACE_PROBE1(l2arc__miss,
4271 arc_buf_hdr_t *, hdr);
4272 ARCSTAT_BUMP(arcstat_l2_misses);
4273 }
4274 }
4275
4276 rzio = zio_read(pio, spa, bp, buf->b_data, size,
4277 arc_read_done, buf, priority, zio_flags, zb);
4278
4279 if (*arc_flags & ARC_FLAG_WAIT)
4280 return (zio_wait(rzio));
4281
4282 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4283 zio_nowait(rzio);
4284 }
4285 return (0);
4286 }
4287
4288 void
4289 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
4290 {
4291 ASSERT(buf->b_hdr != NULL);
4292 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon);
4293 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) ||
4294 func == NULL);
4295 ASSERT(buf->b_efunc == NULL);
4296 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
4297
4298 buf->b_efunc = func;
4299 buf->b_private = private;
4300 }
4301
4302 /*
4303 * Notify the arc that a block was freed, and thus will never be used again.
4304 */
4305 void
4306 arc_freed(spa_t *spa, const blkptr_t *bp)
4307 {
4308 arc_buf_hdr_t *hdr;
4309 kmutex_t *hash_lock;
4310 uint64_t guid = spa_load_guid(spa);
4311
4312 ASSERT(!BP_IS_EMBEDDED(bp));
4313
4314 hdr = buf_hash_find(guid, bp, &hash_lock);
4315 if (hdr == NULL)
4316 return;
4317 if (HDR_BUF_AVAILABLE(hdr)) {
4318 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
4319 add_reference(hdr, hash_lock, FTAG);
4320 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4321 mutex_exit(hash_lock);
4322
4323 arc_release(buf, FTAG);
4324 (void) arc_buf_remove_ref(buf, FTAG);
4325 } else {
4326 mutex_exit(hash_lock);
4327 }
4328
4329 }
4330
4331 /*
4332 * Clear the user eviction callback set by arc_set_callback(), first calling
4333 * it if it exists. Because the presence of a callback keeps an arc_buf cached
4334 * clearing the callback may result in the arc_buf being destroyed. However,
4335 * it will not result in the *last* arc_buf being destroyed, hence the data
4336 * will remain cached in the ARC. We make a copy of the arc buffer here so
4337 * that we can process the callback without holding any locks.
4338 *
4339 * It's possible that the callback is already in the process of being cleared
4340 * by another thread. In this case we can not clear the callback.
4341 *
4342 * Returns B_TRUE if the callback was successfully called and cleared.
4343 */
4344 boolean_t
4345 arc_clear_callback(arc_buf_t *buf)
4346 {
4347 arc_buf_hdr_t *hdr;
4348 kmutex_t *hash_lock;
4349 arc_evict_func_t *efunc = buf->b_efunc;
4350 void *private = buf->b_private;
4351
4352 mutex_enter(&buf->b_evict_lock);
4353 hdr = buf->b_hdr;
4354 if (hdr == NULL) {
4355 /*
4356 * We are in arc_do_user_evicts().
4357 */
4358 ASSERT(buf->b_data == NULL);
4359 mutex_exit(&buf->b_evict_lock);
4360 return (B_FALSE);
4361 } else if (buf->b_data == NULL) {
4362 /*
4363 * We are on the eviction list; process this buffer now
4364 * but let arc_do_user_evicts() do the reaping.
4365 */
4366 buf->b_efunc = NULL;
4367 mutex_exit(&buf->b_evict_lock);
4368 VERIFY0(efunc(private));
4369 return (B_TRUE);
4370 }
4371 hash_lock = HDR_LOCK(hdr);
4372 mutex_enter(hash_lock);
4373 hdr = buf->b_hdr;
4374 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4375
4376 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <,
4377 hdr->b_l1hdr.b_datacnt);
4378 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4379 hdr->b_l1hdr.b_state == arc_mfu);
4380
4381 buf->b_efunc = NULL;
4382 buf->b_private = NULL;
4383
4384 if (hdr->b_l1hdr.b_datacnt > 1) {
4385 mutex_exit(&buf->b_evict_lock);
4386 arc_buf_destroy(buf, TRUE);
4387 } else {
4388 ASSERT(buf == hdr->b_l1hdr.b_buf);
4389 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4390 mutex_exit(&buf->b_evict_lock);
4391 }
4392
4393 mutex_exit(hash_lock);
4394 VERIFY0(efunc(private));
4395 return (B_TRUE);
4396 }
4397
4398 /*
4399 * Release this buffer from the cache, making it an anonymous buffer. This
4400 * must be done after a read and prior to modifying the buffer contents.
4401 * If the buffer has more than one reference, we must make
4402 * a new hdr for the buffer.
4403 */
4404 void
4405 arc_release(arc_buf_t *buf, void *tag)
4406 {
4407 arc_buf_hdr_t *hdr = buf->b_hdr;
4408
4409 /*
4410 * It would be nice to assert that if it's DMU metadata (level >
4411 * 0 || it's the dnode file), then it must be syncing context.
4412 * But we don't know that information at this level.
4413 */
4414
4415 mutex_enter(&buf->b_evict_lock);
4416
4417 ASSERT(HDR_HAS_L1HDR(hdr));
4418
4419 /*
4420 * We don't grab the hash lock prior to this check, because if
4421 * the buffer's header is in the arc_anon state, it won't be
4422 * linked into the hash table.
4423 */
4424 if (hdr->b_l1hdr.b_state == arc_anon) {
4425 mutex_exit(&buf->b_evict_lock);
4426 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4427 ASSERT(!HDR_IN_HASH_TABLE(hdr));
4428 ASSERT(!HDR_HAS_L2HDR(hdr));
4429 ASSERT(BUF_EMPTY(hdr));
4430
4431 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1);
4432 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
4433 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
4434
4435 ASSERT3P(buf->b_efunc, ==, NULL);
4436 ASSERT3P(buf->b_private, ==, NULL);
4437
4438 hdr->b_l1hdr.b_arc_access = 0;
4439 arc_buf_thaw(buf);
4440
4441 return;
4442 }
4443
4444 kmutex_t *hash_lock = HDR_LOCK(hdr);
4445 mutex_enter(hash_lock);
4446
4447 /*
4448 * This assignment is only valid as long as the hash_lock is
4449 * held, we must be careful not to reference state or the
4450 * b_state field after dropping the lock.
4451 */
4452 arc_state_t *state = hdr->b_l1hdr.b_state;
4453 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4454 ASSERT3P(state, !=, arc_anon);
4455
4456 /* this buffer is not on any list */
4457 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
4458
4459 if (HDR_HAS_L2HDR(hdr)) {
4460 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4461
4462 /*
4463 * We have to recheck this conditional again now that
4464 * we're holding the l2ad_mtx to prevent a race with
4465 * another thread which might be concurrently calling
4466 * l2arc_evict(). In that case, l2arc_evict() might have
4467 * destroyed the header's L2 portion as we were waiting
4468 * to acquire the l2ad_mtx.
4469 */
4470 if (HDR_HAS_L2HDR(hdr))
4471 arc_hdr_l2hdr_destroy(hdr);
4472
4473 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4474 }
4475
4476 /*
4477 * Do we have more than one buf?
4478 */
4479 if (hdr->b_l1hdr.b_datacnt > 1) {
4480 arc_buf_hdr_t *nhdr;
4481 arc_buf_t **bufp;
4482 uint64_t blksz = hdr->b_size;
4483 uint64_t spa = hdr->b_spa;
4484 arc_buf_contents_t type = arc_buf_type(hdr);
4485 uint32_t flags = hdr->b_flags;
4486
4487 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
4488 /*
4489 * Pull the data off of this hdr and attach it to
4490 * a new anonymous hdr.
4491 */
4492 (void) remove_reference(hdr, hash_lock, tag);
4493 bufp = &hdr->b_l1hdr.b_buf;
4494 while (*bufp != buf)
4495 bufp = &(*bufp)->b_next;
4496 *bufp = buf->b_next;
4497 buf->b_next = NULL;
4498
4499 ASSERT3P(state, !=, arc_l2c_only);
4500
4501 (void) refcount_remove_many(
4502 &state->arcs_size, hdr->b_size, buf);
4503
4504 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
4505 ASSERT3P(state, !=, arc_l2c_only);
4506 uint64_t *size = &state->arcs_lsize[type];
4507 ASSERT3U(*size, >=, hdr->b_size);
4508 atomic_add_64(size, -hdr->b_size);
4509 }
4510
4511 /*
4512 * We're releasing a duplicate user data buffer, update
4513 * our statistics accordingly.
4514 */
4515 if (HDR_ISTYPE_DATA(hdr)) {
4516 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
4517 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
4518 -hdr->b_size);
4519 }
4520 hdr->b_l1hdr.b_datacnt -= 1;
4521 arc_cksum_verify(buf);
4522 arc_buf_unwatch(buf);
4523
4524 mutex_exit(hash_lock);
4525
4526 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
4527 nhdr->b_size = blksz;
4528 nhdr->b_spa = spa;
4529
4530 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING;
4531 nhdr->b_flags |= arc_bufc_to_flags(type);
4532 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
4533
4534 nhdr->b_l1hdr.b_buf = buf;
4535 nhdr->b_l1hdr.b_datacnt = 1;
4536 nhdr->b_l1hdr.b_state = arc_anon;
4537 nhdr->b_l1hdr.b_arc_access = 0;
4538 nhdr->b_l1hdr.b_tmp_cdata = NULL;
4539 nhdr->b_freeze_cksum = NULL;
4540
4541 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
4542 buf->b_hdr = nhdr;
4543 mutex_exit(&buf->b_evict_lock);
4544 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf);
4545 } else {
4546 mutex_exit(&buf->b_evict_lock);
4547 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
4548 /* protected by hash lock, or hdr is on arc_anon */
4549 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4550 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4551 arc_change_state(arc_anon, hdr, hash_lock);
4552 hdr->b_l1hdr.b_arc_access = 0;
4553 mutex_exit(hash_lock);
4554
4555 buf_discard_identity(hdr);
4556 arc_buf_thaw(buf);
4557 }
4558 buf->b_efunc = NULL;
4559 buf->b_private = NULL;
4560 }
4561
4562 int
4563 arc_released(arc_buf_t *buf)
4564 {
4565 int released;
4566
4567 mutex_enter(&buf->b_evict_lock);
4568 released = (buf->b_data != NULL &&
4569 buf->b_hdr->b_l1hdr.b_state == arc_anon);
4570 mutex_exit(&buf->b_evict_lock);
4571 return (released);
4572 }
4573
4574 #ifdef ZFS_DEBUG
4575 int
4576 arc_referenced(arc_buf_t *buf)
4577 {
4578 int referenced;
4579
4580 mutex_enter(&buf->b_evict_lock);
4581 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
4582 mutex_exit(&buf->b_evict_lock);
4583 return (referenced);
4584 }
4585 #endif
4586
4587 static void
4588 arc_write_ready(zio_t *zio)
4589 {
4590 arc_write_callback_t *callback = zio->io_private;
4591 arc_buf_t *buf = callback->awcb_buf;
4592 arc_buf_hdr_t *hdr = buf->b_hdr;
4593
4594 ASSERT(HDR_HAS_L1HDR(hdr));
4595 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
4596 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4597 callback->awcb_ready(zio, buf, callback->awcb_private);
4598
4599 /*
4600 * If the IO is already in progress, then this is a re-write
4601 * attempt, so we need to thaw and re-compute the cksum.
4602 * It is the responsibility of the callback to handle the
4603 * accounting for any re-write attempt.
4604 */
4605 if (HDR_IO_IN_PROGRESS(hdr)) {
4606 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
4607 if (hdr->b_freeze_cksum != NULL) {
4608 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
4609 hdr->b_freeze_cksum = NULL;
4610 }
4611 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
4612 }
4613 arc_cksum_compute(buf, B_FALSE);
4614 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4615 }
4616
4617 /*
4618 * The SPA calls this callback for each physical write that happens on behalf
4619 * of a logical write. See the comment in dbuf_write_physdone() for details.
4620 */
4621 static void
4622 arc_write_physdone(zio_t *zio)
4623 {
4624 arc_write_callback_t *cb = zio->io_private;
4625 if (cb->awcb_physdone != NULL)
4626 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
4627 }
4628
4629 static void
4630 arc_write_done(zio_t *zio)
4631 {
4632 arc_write_callback_t *callback = zio->io_private;
4633 arc_buf_t *buf = callback->awcb_buf;
4634 arc_buf_hdr_t *hdr = buf->b_hdr;
4635
4636 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4637
4638 if (zio->io_error == 0) {
4639 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
4640 buf_discard_identity(hdr);
4641 } else {
4642 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
4643 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
4644 }
4645 } else {
4646 ASSERT(BUF_EMPTY(hdr));
4647 }
4648
4649 /*
4650 * If the block to be written was all-zero or compressed enough to be
4651 * embedded in the BP, no write was performed so there will be no
4652 * dva/birth/checksum. The buffer must therefore remain anonymous
4653 * (and uncached).
4654 */
4655 if (!BUF_EMPTY(hdr)) {
4656 arc_buf_hdr_t *exists;
4657 kmutex_t *hash_lock;
4658
4659 ASSERT(zio->io_error == 0);
4660
4661 arc_cksum_verify(buf);
4662
4663 exists = buf_hash_insert(hdr, &hash_lock);
4664 if (exists != NULL) {
4665 /*
4666 * This can only happen if we overwrite for
4667 * sync-to-convergence, because we remove
4668 * buffers from the hash table when we arc_free().
4669 */
4670 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
4671 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4672 panic("bad overwrite, hdr=%p exists=%p",
4673 (void *)hdr, (void *)exists);
4674 ASSERT(refcount_is_zero(
4675 &exists->b_l1hdr.b_refcnt));
4676 arc_change_state(arc_anon, exists, hash_lock);
4677 mutex_exit(hash_lock);
4678 arc_hdr_destroy(exists);
4679 exists = buf_hash_insert(hdr, &hash_lock);
4680 ASSERT3P(exists, ==, NULL);
4681 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
4682 /* nopwrite */
4683 ASSERT(zio->io_prop.zp_nopwrite);
4684 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4685 panic("bad nopwrite, hdr=%p exists=%p",
4686 (void *)hdr, (void *)exists);
4687 } else {
4688 /* Dedup */
4689 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
4690 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
4691 ASSERT(BP_GET_DEDUP(zio->io_bp));
4692 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
4693 }
4694 }
4695 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4696 /* if it's not anon, we are doing a scrub */
4697 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
4698 arc_access(hdr, hash_lock);
4699 mutex_exit(hash_lock);
4700 } else {
4701 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4702 }
4703
4704 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4705 callback->awcb_done(zio, buf, callback->awcb_private);
4706
4707 kmem_free(callback, sizeof (arc_write_callback_t));
4708 }
4709
4710 zio_t *
4711 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
4712 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
4713 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
4714 arc_done_func_t *done, void *private, zio_priority_t priority,
4715 int zio_flags, const zbookmark_phys_t *zb)
4716 {
4717 arc_buf_hdr_t *hdr = buf->b_hdr;
4718 arc_write_callback_t *callback;
4719 zio_t *zio;
4720
4721 ASSERT(ready != NULL);
4722 ASSERT(done != NULL);
4723 ASSERT(!HDR_IO_ERROR(hdr));
4724 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4725 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4726 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4727 if (l2arc)
4728 hdr->b_flags |= ARC_FLAG_L2CACHE;
4729 if (l2arc_compress)
4730 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4731 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
4732 callback->awcb_ready = ready;
4733 callback->awcb_physdone = physdone;
4734 callback->awcb_done = done;
4735 callback->awcb_private = private;
4736 callback->awcb_buf = buf;
4737
4738 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
4739 arc_write_ready, arc_write_physdone, arc_write_done, callback,
4740 priority, zio_flags, zb);
4741
4742 return (zio);
4743 }
4744
4745 static int
4746 arc_memory_throttle(uint64_t reserve, uint64_t txg)
4747 {
4748 #ifdef _KERNEL
4749 uint64_t available_memory = ptob(freemem);
4750 static uint64_t page_load = 0;
4751 static uint64_t last_txg = 0;
4752
4753 #if defined(__i386)
4754 available_memory =
4755 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
4756 #endif
4757
4758 if (freemem > physmem * arc_lotsfree_percent / 100)
4759 return (0);
4760
4761 if (txg > last_txg) {
4762 last_txg = txg;
4763 page_load = 0;
4764 }
4765 /*
4766 * If we are in pageout, we know that memory is already tight,
4767 * the arc is already going to be evicting, so we just want to
4768 * continue to let page writes occur as quickly as possible.
4769 */
4770 if (curproc == proc_pageout) {
4771 if (page_load > MAX(ptob(minfree), available_memory) / 4)
4772 return (SET_ERROR(ERESTART));
4773 /* Note: reserve is inflated, so we deflate */
4774 page_load += reserve / 8;
4775 return (0);
4776 } else if (page_load > 0 && arc_reclaim_needed()) {
4777 /* memory is low, delay before restarting */
4778 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
4779 return (SET_ERROR(EAGAIN));
4780 }
4781 page_load = 0;
4782 #endif
4783 return (0);
4784 }
4785
4786 void
4787 arc_tempreserve_clear(uint64_t reserve)
4788 {
4789 atomic_add_64(&arc_tempreserve, -reserve);
4790 ASSERT((int64_t)arc_tempreserve >= 0);
4791 }
4792
4793 int
4794 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
4795 {
4796 int error;
4797 uint64_t anon_size;
4798
4799 if (reserve > arc_c/4 && !arc_no_grow)
4800 arc_c = MIN(arc_c_max, reserve * 4);
4801 if (reserve > arc_c)
4802 return (SET_ERROR(ENOMEM));
4803
4804 /*
4805 * Don't count loaned bufs as in flight dirty data to prevent long
4806 * network delays from blocking transactions that are ready to be
4807 * assigned to a txg.
4808 */
4809 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
4810 arc_loaned_bytes), 0);
4811
4812 /*
4813 * Writes will, almost always, require additional memory allocations
4814 * in order to compress/encrypt/etc the data. We therefore need to
4815 * make sure that there is sufficient available memory for this.
4816 */
4817 error = arc_memory_throttle(reserve, txg);
4818 if (error != 0)
4819 return (error);
4820
4821 /*
4822 * Throttle writes when the amount of dirty data in the cache
4823 * gets too large. We try to keep the cache less than half full
4824 * of dirty blocks so that our sync times don't grow too large.
4825 * Note: if two requests come in concurrently, we might let them
4826 * both succeed, when one of them should fail. Not a huge deal.
4827 */
4828
4829 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
4830 anon_size > arc_c / 4) {
4831 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
4832 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
4833 arc_tempreserve>>10,
4834 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
4835 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
4836 reserve>>10, arc_c>>10);
4837 return (SET_ERROR(ERESTART));
4838 }
4839 atomic_add_64(&arc_tempreserve, reserve);
4840 return (0);
4841 }
4842
4843 static void
4844 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
4845 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
4846 {
4847 size->value.ui64 = refcount_count(&state->arcs_size);
4848 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
4849 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
4850 }
4851
4852 static int
4853 arc_kstat_update(kstat_t *ksp, int rw)
4854 {
4855 arc_stats_t *as = ksp->ks_data;
4856
4857 if (rw == KSTAT_WRITE) {
4858 return (EACCES);
4859 } else {
4860 arc_kstat_update_state(arc_anon,
4861 &as->arcstat_anon_size,
4862 &as->arcstat_anon_evictable_data,
4863 &as->arcstat_anon_evictable_metadata);
4864 arc_kstat_update_state(arc_mru,
4865 &as->arcstat_mru_size,
4866 &as->arcstat_mru_evictable_data,
4867 &as->arcstat_mru_evictable_metadata);
4868 arc_kstat_update_state(arc_mru_ghost,
4869 &as->arcstat_mru_ghost_size,
4870 &as->arcstat_mru_ghost_evictable_data,
4871 &as->arcstat_mru_ghost_evictable_metadata);
4872 arc_kstat_update_state(arc_mfu,
4873 &as->arcstat_mfu_size,
4874 &as->arcstat_mfu_evictable_data,
4875 &as->arcstat_mfu_evictable_metadata);
4876 arc_kstat_update_state(arc_mfu_ghost,
4877 &as->arcstat_mfu_ghost_size,
4878 &as->arcstat_mfu_ghost_evictable_data,
4879 &as->arcstat_mfu_ghost_evictable_metadata);
4880 }
4881
4882 return (0);
4883 }
4884
4885 /*
4886 * This function *must* return indices evenly distributed between all
4887 * sublists of the multilist. This is needed due to how the ARC eviction
4888 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
4889 * distributed between all sublists and uses this assumption when
4890 * deciding which sublist to evict from and how much to evict from it.
4891 */
4892 unsigned int
4893 arc_state_multilist_index_func(multilist_t *ml, void *obj)
4894 {
4895 arc_buf_hdr_t *hdr = obj;
4896
4897 /*
4898 * We rely on b_dva to generate evenly distributed index
4899 * numbers using buf_hash below. So, as an added precaution,
4900 * let's make sure we never add empty buffers to the arc lists.
4901 */
4902 ASSERT(!BUF_EMPTY(hdr));
4903
4904 /*
4905 * The assumption here, is the hash value for a given
4906 * arc_buf_hdr_t will remain constant throughout it's lifetime
4907 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
4908 * Thus, we don't need to store the header's sublist index
4909 * on insertion, as this index can be recalculated on removal.
4910 *
4911 * Also, the low order bits of the hash value are thought to be
4912 * distributed evenly. Otherwise, in the case that the multilist
4913 * has a power of two number of sublists, each sublists' usage
4914 * would not be evenly distributed.
4915 */
4916 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
4917 multilist_get_num_sublists(ml));
4918 }
4919
4920 void
4921 arc_init(void)
4922 {
4923 /*
4924 * allmem is "all memory that we could possibly use".
4925 */
4926 #ifdef _KERNEL
4927 uint64_t allmem = ptob(physmem - swapfs_minfree);
4928 #else
4929 uint64_t allmem = (physmem * PAGESIZE) / 2;
4930 #endif
4931
4932 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
4933 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
4934 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
4935
4936 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
4937 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL);
4938
4939 /* Convert seconds to clock ticks */
4940 arc_min_prefetch_lifespan = 1 * hz;
4941
4942 /* Start out with 1/8 of all memory */
4943 arc_c = allmem / 8;
4944
4945 #ifdef _KERNEL
4946 /*
4947 * On architectures where the physical memory can be larger
4948 * than the addressable space (intel in 32-bit mode), we may
4949 * need to limit the cache to 1/8 of VM size.
4950 */
4951 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
4952 #endif
4953
4954 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
4955 arc_c_min = MAX(allmem / 32, 64 << 20);
4956 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
4957 if (allmem >= 1 << 30)
4958 arc_c_max = allmem - (1 << 30);
4959 else
4960 arc_c_max = arc_c_min;
4961 arc_c_max = MAX(allmem * 3 / 4, arc_c_max);
4962
4963 /*
4964 * Allow the tunables to override our calculations if they are
4965 * reasonable (ie. over 64MB)
4966 */
4967 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem)
4968 arc_c_max = zfs_arc_max;
4969 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
4970 arc_c_min = zfs_arc_min;
4971
4972 arc_c = arc_c_max;
4973 arc_p = (arc_c >> 1);
4974
4975 /* limit meta-data to 1/4 of the arc capacity */
4976 arc_meta_limit = arc_c_max / 4;
4977
4978 /* Allow the tunable to override if it is reasonable */
4979 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
4980 arc_meta_limit = zfs_arc_meta_limit;
4981
4982 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
4983 arc_c_min = arc_meta_limit / 2;
4984
4985 if (zfs_arc_meta_min > 0) {
4986 arc_meta_min = zfs_arc_meta_min;
4987 } else {
4988 arc_meta_min = arc_c_min / 2;
4989 }
4990
4991 if (zfs_arc_grow_retry > 0)
4992 arc_grow_retry = zfs_arc_grow_retry;
4993
4994 if (zfs_arc_shrink_shift > 0)
4995 arc_shrink_shift = zfs_arc_shrink_shift;
4996
4997 /*
4998 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
4999 */
5000 if (arc_no_grow_shift >= arc_shrink_shift)
5001 arc_no_grow_shift = arc_shrink_shift - 1;
5002
5003 if (zfs_arc_p_min_shift > 0)
5004 arc_p_min_shift = zfs_arc_p_min_shift;
5005
5006 if (zfs_arc_num_sublists_per_state < 1)
5007 zfs_arc_num_sublists_per_state = MAX(boot_ncpus, 1);
5008
5009 /* if kmem_flags are set, lets try to use less memory */
5010 if (kmem_debugging())
5011 arc_c = arc_c / 2;
5012 if (arc_c < arc_c_min)
5013 arc_c = arc_c_min;
5014
5015 arc_anon = &ARC_anon;
5016 arc_mru = &ARC_mru;
5017 arc_mru_ghost = &ARC_mru_ghost;
5018 arc_mfu = &ARC_mfu;
5019 arc_mfu_ghost = &ARC_mfu_ghost;
5020 arc_l2c_only = &ARC_l2c_only;
5021 arc_size = 0;
5022
5023 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
5024 sizeof (arc_buf_hdr_t),
5025 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5026 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5027 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
5028 sizeof (arc_buf_hdr_t),
5029 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5030 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5031 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
5032 sizeof (arc_buf_hdr_t),
5033 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5034 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5035 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
5036 sizeof (arc_buf_hdr_t),
5037 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5038 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5039 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
5040 sizeof (arc_buf_hdr_t),
5041 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5042 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5043 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
5044 sizeof (arc_buf_hdr_t),
5045 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5046 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5047 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
5048 sizeof (arc_buf_hdr_t),
5049 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5050 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5051 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
5052 sizeof (arc_buf_hdr_t),
5053 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5054 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5055 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
5056 sizeof (arc_buf_hdr_t),
5057 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5058 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5059 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
5060 sizeof (arc_buf_hdr_t),
5061 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5062 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5063
5064 refcount_create(&arc_anon->arcs_size);
5065 refcount_create(&arc_mru->arcs_size);
5066 refcount_create(&arc_mru_ghost->arcs_size);
5067 refcount_create(&arc_mfu->arcs_size);
5068 refcount_create(&arc_mfu_ghost->arcs_size);
5069 refcount_create(&arc_l2c_only->arcs_size);
5070
5071 buf_init();
5072
5073 arc_reclaim_thread_exit = FALSE;
5074 arc_user_evicts_thread_exit = FALSE;
5075 arc_eviction_list = NULL;
5076 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
5077
5078 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
5079 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
5080
5081 if (arc_ksp != NULL) {
5082 arc_ksp->ks_data = &arc_stats;
5083 arc_ksp->ks_update = arc_kstat_update;
5084 kstat_install(arc_ksp);
5085 }
5086
5087 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
5088 TS_RUN, minclsyspri);
5089
5090 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0,
5091 TS_RUN, minclsyspri);
5092
5093 arc_dead = FALSE;
5094 arc_warm = B_FALSE;
5095
5096 /*
5097 * Calculate maximum amount of dirty data per pool.
5098 *
5099 * If it has been set by /etc/system, take that.
5100 * Otherwise, use a percentage of physical memory defined by
5101 * zfs_dirty_data_max_percent (default 10%) with a cap at
5102 * zfs_dirty_data_max_max (default 4GB).
5103 */
5104 if (zfs_dirty_data_max == 0) {
5105 zfs_dirty_data_max = physmem * PAGESIZE *
5106 zfs_dirty_data_max_percent / 100;
5107 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
5108 zfs_dirty_data_max_max);
5109 }
5110 }
5111
5112 void
5113 arc_fini(void)
5114 {
5115 mutex_enter(&arc_reclaim_lock);
5116 arc_reclaim_thread_exit = TRUE;
5117 /*
5118 * The reclaim thread will set arc_reclaim_thread_exit back to
5119 * FALSE when it is finished exiting; we're waiting for that.
5120 */
5121 while (arc_reclaim_thread_exit) {
5122 cv_signal(&arc_reclaim_thread_cv);
5123 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
5124 }
5125 mutex_exit(&arc_reclaim_lock);
5126
5127 mutex_enter(&arc_user_evicts_lock);
5128 arc_user_evicts_thread_exit = TRUE;
5129 /*
5130 * The user evicts thread will set arc_user_evicts_thread_exit
5131 * to FALSE when it is finished exiting; we're waiting for that.
5132 */
5133 while (arc_user_evicts_thread_exit) {
5134 cv_signal(&arc_user_evicts_cv);
5135 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock);
5136 }
5137 mutex_exit(&arc_user_evicts_lock);
5138
5139 /* Use TRUE to ensure *all* buffers are evicted */
5140 arc_flush(NULL, TRUE);
5141
5142 arc_dead = TRUE;
5143
5144 if (arc_ksp != NULL) {
5145 kstat_delete(arc_ksp);
5146 arc_ksp = NULL;
5147 }
5148
5149 mutex_destroy(&arc_reclaim_lock);
5150 cv_destroy(&arc_reclaim_thread_cv);
5151 cv_destroy(&arc_reclaim_waiters_cv);
5152
5153 mutex_destroy(&arc_user_evicts_lock);
5154 cv_destroy(&arc_user_evicts_cv);
5155
5156 refcount_destroy(&arc_anon->arcs_size);
5157 refcount_destroy(&arc_mru->arcs_size);
5158 refcount_destroy(&arc_mru_ghost->arcs_size);
5159 refcount_destroy(&arc_mfu->arcs_size);
5160 refcount_destroy(&arc_mfu_ghost->arcs_size);
5161 refcount_destroy(&arc_l2c_only->arcs_size);
5162
5163 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
5164 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
5165 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
5166 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
5167 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
5168 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
5169 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
5170 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
5171
5172 buf_fini();
5173
5174 ASSERT0(arc_loaned_bytes);
5175 }
5176
5177 /*
5178 * Level 2 ARC
5179 *
5180 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5181 * It uses dedicated storage devices to hold cached data, which are populated
5182 * using large infrequent writes. The main role of this cache is to boost
5183 * the performance of random read workloads. The intended L2ARC devices
5184 * include short-stroked disks, solid state disks, and other media with
5185 * substantially faster read latency than disk.
5186 *
5187 * +-----------------------+
5188 * | ARC |
5189 * +-----------------------+
5190 * | ^ ^
5191 * | | |
5192 * l2arc_feed_thread() arc_read()
5193 * | | |
5194 * | l2arc read |
5195 * V | |
5196 * +---------------+ |
5197 * | L2ARC | |
5198 * +---------------+ |
5199 * | ^ |
5200 * l2arc_write() | |
5201 * | | |
5202 * V | |
5203 * +-------+ +-------+
5204 * | vdev | | vdev |
5205 * | cache | | cache |
5206 * +-------+ +-------+
5207 * +=========+ .-----.
5208 * : L2ARC : |-_____-|
5209 * : devices : | Disks |
5210 * +=========+ `-_____-'
5211 *
5212 * Read requests are satisfied from the following sources, in order:
5213 *
5214 * 1) ARC
5215 * 2) vdev cache of L2ARC devices
5216 * 3) L2ARC devices
5217 * 4) vdev cache of disks
5218 * 5) disks
5219 *
5220 * Some L2ARC device types exhibit extremely slow write performance.
5221 * To accommodate for this there are some significant differences between
5222 * the L2ARC and traditional cache design:
5223 *
5224 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
5225 * the ARC behave as usual, freeing buffers and placing headers on ghost
5226 * lists. The ARC does not send buffers to the L2ARC during eviction as
5227 * this would add inflated write latencies for all ARC memory pressure.
5228 *
5229 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5230 * It does this by periodically scanning buffers from the eviction-end of
5231 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5232 * not already there. It scans until a headroom of buffers is satisfied,
5233 * which itself is a buffer for ARC eviction. If a compressible buffer is
5234 * found during scanning and selected for writing to an L2ARC device, we
5235 * temporarily boost scanning headroom during the next scan cycle to make
5236 * sure we adapt to compression effects (which might significantly reduce
5237 * the data volume we write to L2ARC). The thread that does this is
5238 * l2arc_feed_thread(), illustrated below; example sizes are included to
5239 * provide a better sense of ratio than this diagram:
5240 *
5241 * head --> tail
5242 * +---------------------+----------+
5243 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
5244 * +---------------------+----------+ | o L2ARC eligible
5245 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
5246 * +---------------------+----------+ |
5247 * 15.9 Gbytes ^ 32 Mbytes |
5248 * headroom |
5249 * l2arc_feed_thread()
5250 * |
5251 * l2arc write hand <--[oooo]--'
5252 * | 8 Mbyte
5253 * | write max
5254 * V
5255 * +==============================+
5256 * L2ARC dev |####|#|###|###| |####| ... |
5257 * +==============================+
5258 * 32 Gbytes
5259 *
5260 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5261 * evicted, then the L2ARC has cached a buffer much sooner than it probably
5262 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
5263 * safe to say that this is an uncommon case, since buffers at the end of
5264 * the ARC lists have moved there due to inactivity.
5265 *
5266 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5267 * then the L2ARC simply misses copying some buffers. This serves as a
5268 * pressure valve to prevent heavy read workloads from both stalling the ARC
5269 * with waits and clogging the L2ARC with writes. This also helps prevent
5270 * the potential for the L2ARC to churn if it attempts to cache content too
5271 * quickly, such as during backups of the entire pool.
5272 *
5273 * 5. After system boot and before the ARC has filled main memory, there are
5274 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5275 * lists can remain mostly static. Instead of searching from tail of these
5276 * lists as pictured, the l2arc_feed_thread() will search from the list heads
5277 * for eligible buffers, greatly increasing its chance of finding them.
5278 *
5279 * The L2ARC device write speed is also boosted during this time so that
5280 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
5281 * there are no L2ARC reads, and no fear of degrading read performance
5282 * through increased writes.
5283 *
5284 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5285 * the vdev queue can aggregate them into larger and fewer writes. Each
5286 * device is written to in a rotor fashion, sweeping writes through
5287 * available space then repeating.
5288 *
5289 * 7. The L2ARC does not store dirty content. It never needs to flush
5290 * write buffers back to disk based storage.
5291 *
5292 * 8. If an ARC buffer is written (and dirtied) which also exists in the
5293 * L2ARC, the now stale L2ARC buffer is immediately dropped.
5294 *
5295 * The performance of the L2ARC can be tweaked by a number of tunables, which
5296 * may be necessary for different workloads:
5297 *
5298 * l2arc_write_max max write bytes per interval
5299 * l2arc_write_boost extra write bytes during device warmup
5300 * l2arc_noprefetch skip caching prefetched buffers
5301 * l2arc_headroom number of max device writes to precache
5302 * l2arc_headroom_boost when we find compressed buffers during ARC
5303 * scanning, we multiply headroom by this
5304 * percentage factor for the next scan cycle,
5305 * since more compressed buffers are likely to
5306 * be present
5307 * l2arc_feed_secs seconds between L2ARC writing
5308 *
5309 * Tunables may be removed or added as future performance improvements are
5310 * integrated, and also may become zpool properties.
5311 *
5312 * There are three key functions that control how the L2ARC warms up:
5313 *
5314 * l2arc_write_eligible() check if a buffer is eligible to cache
5315 * l2arc_write_size() calculate how much to write
5316 * l2arc_write_interval() calculate sleep delay between writes
5317 *
5318 * These three functions determine what to write, how much, and how quickly
5319 * to send writes.
5320 */
5321
5322 static boolean_t
5323 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
5324 {
5325 /*
5326 * A buffer is *not* eligible for the L2ARC if it:
5327 * 1. belongs to a different spa.
5328 * 2. is already cached on the L2ARC.
5329 * 3. has an I/O in progress (it may be an incomplete read).
5330 * 4. is flagged not eligible (zfs property).
5331 */
5332 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
5333 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
5334 return (B_FALSE);
5335
5336 return (B_TRUE);
5337 }
5338
5339 static uint64_t
5340 l2arc_write_size(void)
5341 {
5342 uint64_t size;
5343
5344 /*
5345 * Make sure our globals have meaningful values in case the user
5346 * altered them.
5347 */
5348 size = l2arc_write_max;
5349 if (size == 0) {
5350 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
5351 "be greater than zero, resetting it to the default (%d)",
5352 L2ARC_WRITE_SIZE);
5353 size = l2arc_write_max = L2ARC_WRITE_SIZE;
5354 }
5355
5356 if (arc_warm == B_FALSE)
5357 size += l2arc_write_boost;
5358
5359 return (size);
5360
5361 }
5362
5363 static clock_t
5364 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
5365 {
5366 clock_t interval, next, now;
5367
5368 /*
5369 * If the ARC lists are busy, increase our write rate; if the
5370 * lists are stale, idle back. This is achieved by checking
5371 * how much we previously wrote - if it was more than half of
5372 * what we wanted, schedule the next write much sooner.
5373 */
5374 if (l2arc_feed_again && wrote > (wanted / 2))
5375 interval = (hz * l2arc_feed_min_ms) / 1000;
5376 else
5377 interval = hz * l2arc_feed_secs;
5378
5379 now = ddi_get_lbolt();
5380 next = MAX(now, MIN(now + interval, began + interval));
5381
5382 return (next);
5383 }
5384
5385 /*
5386 * Cycle through L2ARC devices. This is how L2ARC load balances.
5387 * If a device is returned, this also returns holding the spa config lock.
5388 */
5389 static l2arc_dev_t *
5390 l2arc_dev_get_next(void)
5391 {
5392 l2arc_dev_t *first, *next = NULL;
5393
5394 /*
5395 * Lock out the removal of spas (spa_namespace_lock), then removal
5396 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
5397 * both locks will be dropped and a spa config lock held instead.
5398 */
5399 mutex_enter(&spa_namespace_lock);
5400 mutex_enter(&l2arc_dev_mtx);
5401
5402 /* if there are no vdevs, there is nothing to do */
5403 if (l2arc_ndev == 0)
5404 goto out;
5405
5406 first = NULL;
5407 next = l2arc_dev_last;
5408 do {
5409 /* loop around the list looking for a non-faulted vdev */
5410 if (next == NULL) {
5411 next = list_head(l2arc_dev_list);
5412 } else {
5413 next = list_next(l2arc_dev_list, next);
5414 if (next == NULL)
5415 next = list_head(l2arc_dev_list);
5416 }
5417
5418 /* if we have come back to the start, bail out */
5419 if (first == NULL)
5420 first = next;
5421 else if (next == first)
5422 break;
5423
5424 } while (vdev_is_dead(next->l2ad_vdev));
5425
5426 /* if we were unable to find any usable vdevs, return NULL */
5427 if (vdev_is_dead(next->l2ad_vdev))
5428 next = NULL;
5429
5430 l2arc_dev_last = next;
5431
5432 out:
5433 mutex_exit(&l2arc_dev_mtx);
5434
5435 /*
5436 * Grab the config lock to prevent the 'next' device from being
5437 * removed while we are writing to it.
5438 */
5439 if (next != NULL)
5440 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
5441 mutex_exit(&spa_namespace_lock);
5442
5443 return (next);
5444 }
5445
5446 /*
5447 * Free buffers that were tagged for destruction.
5448 */
5449 static void
5450 l2arc_do_free_on_write()
5451 {
5452 list_t *buflist;
5453 l2arc_data_free_t *df, *df_prev;
5454
5455 mutex_enter(&l2arc_free_on_write_mtx);
5456 buflist = l2arc_free_on_write;
5457
5458 for (df = list_tail(buflist); df; df = df_prev) {
5459 df_prev = list_prev(buflist, df);
5460 ASSERT(df->l2df_data != NULL);
5461 ASSERT(df->l2df_func != NULL);
5462 df->l2df_func(df->l2df_data, df->l2df_size);
5463 list_remove(buflist, df);
5464 kmem_free(df, sizeof (l2arc_data_free_t));
5465 }
5466
5467 mutex_exit(&l2arc_free_on_write_mtx);
5468 }
5469
5470 /*
5471 * A write to a cache device has completed. Update all headers to allow
5472 * reads from these buffers to begin.
5473 */
5474 static void
5475 l2arc_write_done(zio_t *zio)
5476 {
5477 l2arc_write_callback_t *cb;
5478 l2arc_dev_t *dev;
5479 list_t *buflist;
5480 arc_buf_hdr_t *head, *hdr, *hdr_prev;
5481 kmutex_t *hash_lock;
5482 int64_t bytes_dropped = 0;
5483
5484 cb = zio->io_private;
5485 ASSERT(cb != NULL);
5486 dev = cb->l2wcb_dev;
5487 ASSERT(dev != NULL);
5488 head = cb->l2wcb_head;
5489 ASSERT(head != NULL);
5490 buflist = &dev->l2ad_buflist;
5491 ASSERT(buflist != NULL);
5492 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
5493 l2arc_write_callback_t *, cb);
5494
5495 if (zio->io_error != 0)
5496 ARCSTAT_BUMP(arcstat_l2_writes_error);
5497
5498 /*
5499 * All writes completed, or an error was hit.
5500 */
5501 top:
5502 mutex_enter(&dev->l2ad_mtx);
5503 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
5504 hdr_prev = list_prev(buflist, hdr);
5505
5506 hash_lock = HDR_LOCK(hdr);
5507
5508 /*
5509 * We cannot use mutex_enter or else we can deadlock
5510 * with l2arc_write_buffers (due to swapping the order
5511 * the hash lock and l2ad_mtx are taken).
5512 */
5513 if (!mutex_tryenter(hash_lock)) {
5514 /*
5515 * Missed the hash lock. We must retry so we
5516 * don't leave the ARC_FLAG_L2_WRITING bit set.
5517 */
5518 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
5519
5520 /*
5521 * We don't want to rescan the headers we've
5522 * already marked as having been written out, so
5523 * we reinsert the head node so we can pick up
5524 * where we left off.
5525 */
5526 list_remove(buflist, head);
5527 list_insert_after(buflist, hdr, head);
5528
5529 mutex_exit(&dev->l2ad_mtx);
5530
5531 /*
5532 * We wait for the hash lock to become available
5533 * to try and prevent busy waiting, and increase
5534 * the chance we'll be able to acquire the lock
5535 * the next time around.
5536 */
5537 mutex_enter(hash_lock);
5538 mutex_exit(hash_lock);
5539 goto top;
5540 }
5541
5542 /*
5543 * We could not have been moved into the arc_l2c_only
5544 * state while in-flight due to our ARC_FLAG_L2_WRITING
5545 * bit being set. Let's just ensure that's being enforced.
5546 */
5547 ASSERT(HDR_HAS_L1HDR(hdr));
5548
5549 /*
5550 * We may have allocated a buffer for L2ARC compression,
5551 * we must release it to avoid leaking this data.
5552 */
5553 l2arc_release_cdata_buf(hdr);
5554
5555 if (zio->io_error != 0) {
5556 /*
5557 * Error - drop L2ARC entry.
5558 */
5559 list_remove(buflist, hdr);
5560 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
5561
5562 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize);
5563 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
5564
5565 bytes_dropped += hdr->b_l2hdr.b_asize;
5566 (void) refcount_remove_many(&dev->l2ad_alloc,
5567 hdr->b_l2hdr.b_asize, hdr);
5568 }
5569
5570 /*
5571 * Allow ARC to begin reads and ghost list evictions to
5572 * this L2ARC entry.
5573 */
5574 hdr->b_flags &= ~ARC_FLAG_L2_WRITING;
5575
5576 mutex_exit(hash_lock);
5577 }
5578
5579 atomic_inc_64(&l2arc_writes_done);
5580 list_remove(buflist, head);
5581 ASSERT(!HDR_HAS_L1HDR(head));
5582 kmem_cache_free(hdr_l2only_cache, head);
5583 mutex_exit(&dev->l2ad_mtx);
5584
5585 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
5586
5587 l2arc_do_free_on_write();
5588
5589 kmem_free(cb, sizeof (l2arc_write_callback_t));
5590 }
5591
5592 /*
5593 * A read to a cache device completed. Validate buffer contents before
5594 * handing over to the regular ARC routines.
5595 */
5596 static void
5597 l2arc_read_done(zio_t *zio)
5598 {
5599 l2arc_read_callback_t *cb;
5600 arc_buf_hdr_t *hdr;
5601 arc_buf_t *buf;
5602 kmutex_t *hash_lock;
5603 int equal;
5604
5605 ASSERT(zio->io_vd != NULL);
5606 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
5607
5608 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
5609
5610 cb = zio->io_private;
5611 ASSERT(cb != NULL);
5612 buf = cb->l2rcb_buf;
5613 ASSERT(buf != NULL);
5614
5615 hash_lock = HDR_LOCK(buf->b_hdr);
5616 mutex_enter(hash_lock);
5617 hdr = buf->b_hdr;
5618 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5619
5620 /*
5621 * If the buffer was compressed, decompress it first.
5622 */
5623 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
5624 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
5625 ASSERT(zio->io_data != NULL);
5626 ASSERT3U(zio->io_size, ==, hdr->b_size);
5627 ASSERT3U(BP_GET_LSIZE(&cb->l2rcb_bp), ==, hdr->b_size);
5628
5629 /*
5630 * Check this survived the L2ARC journey.
5631 */
5632 equal = arc_cksum_equal(buf);
5633 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
5634 mutex_exit(hash_lock);
5635 zio->io_private = buf;
5636 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
5637 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
5638 arc_read_done(zio);
5639 } else {
5640 mutex_exit(hash_lock);
5641 /*
5642 * Buffer didn't survive caching. Increment stats and
5643 * reissue to the original storage device.
5644 */
5645 if (zio->io_error != 0) {
5646 ARCSTAT_BUMP(arcstat_l2_io_error);
5647 } else {
5648 zio->io_error = SET_ERROR(EIO);
5649 }
5650 if (!equal)
5651 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
5652
5653 /*
5654 * If there's no waiter, issue an async i/o to the primary
5655 * storage now. If there *is* a waiter, the caller must
5656 * issue the i/o in a context where it's OK to block.
5657 */
5658 if (zio->io_waiter == NULL) {
5659 zio_t *pio = zio_unique_parent(zio);
5660
5661 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
5662
5663 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
5664 buf->b_data, hdr->b_size, arc_read_done, buf,
5665 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
5666 }
5667 }
5668
5669 kmem_free(cb, sizeof (l2arc_read_callback_t));
5670 }
5671
5672 /*
5673 * This is the list priority from which the L2ARC will search for pages to
5674 * cache. This is used within loops (0..3) to cycle through lists in the
5675 * desired order. This order can have a significant effect on cache
5676 * performance.
5677 *
5678 * Currently the metadata lists are hit first, MFU then MRU, followed by
5679 * the data lists. This function returns a locked list, and also returns
5680 * the lock pointer.
5681 */
5682 static multilist_sublist_t *
5683 l2arc_sublist_lock(int list_num)
5684 {
5685 multilist_t *ml = NULL;
5686 unsigned int idx;
5687
5688 ASSERT(list_num >= 0 && list_num <= 3);
5689
5690 switch (list_num) {
5691 case 0:
5692 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
5693 break;
5694 case 1:
5695 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
5696 break;
5697 case 2:
5698 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
5699 break;
5700 case 3:
5701 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
5702 break;
5703 }
5704
5705 /*
5706 * Return a randomly-selected sublist. This is acceptable
5707 * because the caller feeds only a little bit of data for each
5708 * call (8MB). Subsequent calls will result in different
5709 * sublists being selected.
5710 */
5711 idx = multilist_get_random_index(ml);
5712 return (multilist_sublist_lock(ml, idx));
5713 }
5714
5715 /*
5716 * Evict buffers from the device write hand to the distance specified in
5717 * bytes. This distance may span populated buffers, it may span nothing.
5718 * This is clearing a region on the L2ARC device ready for writing.
5719 * If the 'all' boolean is set, every buffer is evicted.
5720 */
5721 static void
5722 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
5723 {
5724 list_t *buflist;
5725 arc_buf_hdr_t *hdr, *hdr_prev;
5726 kmutex_t *hash_lock;
5727 uint64_t taddr;
5728
5729 buflist = &dev->l2ad_buflist;
5730
5731 if (!all && dev->l2ad_first) {
5732 /*
5733 * This is the first sweep through the device. There is
5734 * nothing to evict.
5735 */
5736 return;
5737 }
5738
5739 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
5740 /*
5741 * When nearing the end of the device, evict to the end
5742 * before the device write hand jumps to the start.
5743 */
5744 taddr = dev->l2ad_end;
5745 } else {
5746 taddr = dev->l2ad_hand + distance;
5747 }
5748 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
5749 uint64_t, taddr, boolean_t, all);
5750
5751 top:
5752 mutex_enter(&dev->l2ad_mtx);
5753 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
5754 hdr_prev = list_prev(buflist, hdr);
5755
5756 hash_lock = HDR_LOCK(hdr);
5757
5758 /*
5759 * We cannot use mutex_enter or else we can deadlock
5760 * with l2arc_write_buffers (due to swapping the order
5761 * the hash lock and l2ad_mtx are taken).
5762 */
5763 if (!mutex_tryenter(hash_lock)) {
5764 /*
5765 * Missed the hash lock. Retry.
5766 */
5767 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
5768 mutex_exit(&dev->l2ad_mtx);
5769 mutex_enter(hash_lock);
5770 mutex_exit(hash_lock);
5771 goto top;
5772 }
5773
5774 if (HDR_L2_WRITE_HEAD(hdr)) {
5775 /*
5776 * We hit a write head node. Leave it for
5777 * l2arc_write_done().
5778 */
5779 list_remove(buflist, hdr);
5780 mutex_exit(hash_lock);
5781 continue;
5782 }
5783
5784 if (!all && HDR_HAS_L2HDR(hdr) &&
5785 (hdr->b_l2hdr.b_daddr > taddr ||
5786 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
5787 /*
5788 * We've evicted to the target address,
5789 * or the end of the device.
5790 */
5791 mutex_exit(hash_lock);
5792 break;
5793 }
5794
5795 ASSERT(HDR_HAS_L2HDR(hdr));
5796 if (!HDR_HAS_L1HDR(hdr)) {
5797 ASSERT(!HDR_L2_READING(hdr));
5798 /*
5799 * This doesn't exist in the ARC. Destroy.
5800 * arc_hdr_destroy() will call list_remove()
5801 * and decrement arcstat_l2_size.
5802 */
5803 arc_change_state(arc_anon, hdr, hash_lock);
5804 arc_hdr_destroy(hdr);
5805 } else {
5806 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
5807 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
5808 /*
5809 * Invalidate issued or about to be issued
5810 * reads, since we may be about to write
5811 * over this location.
5812 */
5813 if (HDR_L2_READING(hdr)) {
5814 ARCSTAT_BUMP(arcstat_l2_evict_reading);
5815 hdr->b_flags |= ARC_FLAG_L2_EVICTED;
5816 }
5817
5818 /* Ensure this header has finished being written */
5819 ASSERT(!HDR_L2_WRITING(hdr));
5820 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
5821
5822 arc_hdr_l2hdr_destroy(hdr);
5823 }
5824 mutex_exit(hash_lock);
5825 }
5826 mutex_exit(&dev->l2ad_mtx);
5827 }
5828
5829 /*
5830 * Find and write ARC buffers to the L2ARC device.
5831 *
5832 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
5833 * for reading until they have completed writing.
5834 * The headroom_boost is an in-out parameter used to maintain headroom boost
5835 * state between calls to this function.
5836 *
5837 * Returns the number of bytes actually written (which may be smaller than
5838 * the delta by which the device hand has changed due to alignment).
5839 */
5840 static uint64_t
5841 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
5842 boolean_t *headroom_boost)
5843 {
5844 arc_buf_hdr_t *hdr, *hdr_prev, *head;
5845 uint64_t write_asize, write_psize, write_sz, headroom,
5846 buf_compress_minsz;
5847 void *buf_data;
5848 boolean_t full;
5849 l2arc_write_callback_t *cb;
5850 zio_t *pio, *wzio;
5851 uint64_t guid = spa_load_guid(spa);
5852 const boolean_t do_headroom_boost = *headroom_boost;
5853
5854 ASSERT(dev->l2ad_vdev != NULL);
5855
5856 /* Lower the flag now, we might want to raise it again later. */
5857 *headroom_boost = B_FALSE;
5858
5859 pio = NULL;
5860 write_sz = write_asize = write_psize = 0;
5861 full = B_FALSE;
5862 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
5863 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD;
5864 head->b_flags |= ARC_FLAG_HAS_L2HDR;
5865
5866 /*
5867 * We will want to try to compress buffers that are at least 2x the
5868 * device sector size.
5869 */
5870 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
5871
5872 /*
5873 * Copy buffers for L2ARC writing.
5874 */
5875 for (int try = 0; try <= 3; try++) {
5876 multilist_sublist_t *mls = l2arc_sublist_lock(try);
5877 uint64_t passed_sz = 0;
5878
5879 /*
5880 * L2ARC fast warmup.
5881 *
5882 * Until the ARC is warm and starts to evict, read from the
5883 * head of the ARC lists rather than the tail.
5884 */
5885 if (arc_warm == B_FALSE)
5886 hdr = multilist_sublist_head(mls);
5887 else
5888 hdr = multilist_sublist_tail(mls);
5889
5890 headroom = target_sz * l2arc_headroom;
5891 if (do_headroom_boost)
5892 headroom = (headroom * l2arc_headroom_boost) / 100;
5893
5894 for (; hdr; hdr = hdr_prev) {
5895 kmutex_t *hash_lock;
5896 uint64_t buf_sz;
5897
5898 if (arc_warm == B_FALSE)
5899 hdr_prev = multilist_sublist_next(mls, hdr);
5900 else
5901 hdr_prev = multilist_sublist_prev(mls, hdr);
5902
5903 hash_lock = HDR_LOCK(hdr);
5904 if (!mutex_tryenter(hash_lock)) {
5905 /*
5906 * Skip this buffer rather than waiting.
5907 */
5908 continue;
5909 }
5910
5911 passed_sz += hdr->b_size;
5912 if (passed_sz > headroom) {
5913 /*
5914 * Searched too far.
5915 */
5916 mutex_exit(hash_lock);
5917 break;
5918 }
5919
5920 if (!l2arc_write_eligible(guid, hdr)) {
5921 mutex_exit(hash_lock);
5922 continue;
5923 }
5924
5925 if ((write_sz + hdr->b_size) > target_sz) {
5926 full = B_TRUE;
5927 mutex_exit(hash_lock);
5928 break;
5929 }
5930
5931 if (pio == NULL) {
5932 /*
5933 * Insert a dummy header on the buflist so
5934 * l2arc_write_done() can find where the
5935 * write buffers begin without searching.
5936 */
5937 mutex_enter(&dev->l2ad_mtx);
5938 list_insert_head(&dev->l2ad_buflist, head);
5939 mutex_exit(&dev->l2ad_mtx);
5940
5941 cb = kmem_alloc(
5942 sizeof (l2arc_write_callback_t), KM_SLEEP);
5943 cb->l2wcb_dev = dev;
5944 cb->l2wcb_head = head;
5945 pio = zio_root(spa, l2arc_write_done, cb,
5946 ZIO_FLAG_CANFAIL);
5947 }
5948
5949 /*
5950 * Create and add a new L2ARC header.
5951 */
5952 hdr->b_l2hdr.b_dev = dev;
5953 hdr->b_flags |= ARC_FLAG_L2_WRITING;
5954 /*
5955 * Temporarily stash the data buffer in b_tmp_cdata.
5956 * The subsequent write step will pick it up from
5957 * there. This is because can't access b_l1hdr.b_buf
5958 * without holding the hash_lock, which we in turn
5959 * can't access without holding the ARC list locks
5960 * (which we want to avoid during compression/writing).
5961 */
5962 hdr->b_l2hdr.b_compress = ZIO_COMPRESS_OFF;
5963 hdr->b_l2hdr.b_asize = hdr->b_size;
5964 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data;
5965
5966 /*
5967 * Explicitly set the b_daddr field to a known
5968 * value which means "invalid address". This
5969 * enables us to differentiate which stage of
5970 * l2arc_write_buffers() the particular header
5971 * is in (e.g. this loop, or the one below).
5972 * ARC_FLAG_L2_WRITING is not enough to make
5973 * this distinction, and we need to know in
5974 * order to do proper l2arc vdev accounting in
5975 * arc_release() and arc_hdr_destroy().
5976 *
5977 * Note, we can't use a new flag to distinguish
5978 * the two stages because we don't hold the
5979 * header's hash_lock below, in the second stage
5980 * of this function. Thus, we can't simply
5981 * change the b_flags field to denote that the
5982 * IO has been sent. We can change the b_daddr
5983 * field of the L2 portion, though, since we'll
5984 * be holding the l2ad_mtx; which is why we're
5985 * using it to denote the header's state change.
5986 */
5987 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET;
5988
5989 buf_sz = hdr->b_size;
5990 hdr->b_flags |= ARC_FLAG_HAS_L2HDR;
5991
5992 mutex_enter(&dev->l2ad_mtx);
5993 list_insert_head(&dev->l2ad_buflist, hdr);
5994 mutex_exit(&dev->l2ad_mtx);
5995
5996 /*
5997 * Compute and store the buffer cksum before
5998 * writing. On debug the cksum is verified first.
5999 */
6000 arc_cksum_verify(hdr->b_l1hdr.b_buf);
6001 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE);
6002
6003 mutex_exit(hash_lock);
6004
6005 write_sz += buf_sz;
6006 }
6007
6008 multilist_sublist_unlock(mls);
6009
6010 if (full == B_TRUE)
6011 break;
6012 }
6013
6014 /* No buffers selected for writing? */
6015 if (pio == NULL) {
6016 ASSERT0(write_sz);
6017 ASSERT(!HDR_HAS_L1HDR(head));
6018 kmem_cache_free(hdr_l2only_cache, head);
6019 return (0);
6020 }
6021
6022 mutex_enter(&dev->l2ad_mtx);
6023
6024 /*
6025 * Now start writing the buffers. We're starting at the write head
6026 * and work backwards, retracing the course of the buffer selector
6027 * loop above.
6028 */
6029 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr;
6030 hdr = list_prev(&dev->l2ad_buflist, hdr)) {
6031 uint64_t buf_sz;
6032
6033 /*
6034 * We rely on the L1 portion of the header below, so
6035 * it's invalid for this header to have been evicted out
6036 * of the ghost cache, prior to being written out. The
6037 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6038 */
6039 ASSERT(HDR_HAS_L1HDR(hdr));
6040
6041 /*
6042 * We shouldn't need to lock the buffer here, since we flagged
6043 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
6044 * take care to only access its L2 cache parameters. In
6045 * particular, hdr->l1hdr.b_buf may be invalid by now due to
6046 * ARC eviction.
6047 */
6048 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
6049
6050 if ((HDR_L2COMPRESS(hdr)) &&
6051 hdr->b_l2hdr.b_asize >= buf_compress_minsz) {
6052 if (l2arc_compress_buf(hdr)) {
6053 /*
6054 * If compression succeeded, enable headroom
6055 * boost on the next scan cycle.
6056 */
6057 *headroom_boost = B_TRUE;
6058 }
6059 }
6060
6061 /*
6062 * Pick up the buffer data we had previously stashed away
6063 * (and now potentially also compressed).
6064 */
6065 buf_data = hdr->b_l1hdr.b_tmp_cdata;
6066 buf_sz = hdr->b_l2hdr.b_asize;
6067
6068 /*
6069 * We need to do this regardless if buf_sz is zero or
6070 * not, otherwise, when this l2hdr is evicted we'll
6071 * remove a reference that was never added.
6072 */
6073 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr);
6074
6075 /* Compression may have squashed the buffer to zero length. */
6076 if (buf_sz != 0) {
6077 uint64_t buf_p_sz;
6078
6079 wzio = zio_write_phys(pio, dev->l2ad_vdev,
6080 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
6081 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
6082 ZIO_FLAG_CANFAIL, B_FALSE);
6083
6084 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
6085 zio_t *, wzio);
6086 (void) zio_nowait(wzio);
6087
6088 write_asize += buf_sz;
6089
6090 /*
6091 * Keep the clock hand suitably device-aligned.
6092 */
6093 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
6094 write_psize += buf_p_sz;
6095 dev->l2ad_hand += buf_p_sz;
6096 }
6097 }
6098
6099 mutex_exit(&dev->l2ad_mtx);
6100
6101 ASSERT3U(write_asize, <=, target_sz);
6102 ARCSTAT_BUMP(arcstat_l2_writes_sent);
6103 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
6104 ARCSTAT_INCR(arcstat_l2_size, write_sz);
6105 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
6106 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
6107
6108 /*
6109 * Bump device hand to the device start if it is approaching the end.
6110 * l2arc_evict() will already have evicted ahead for this case.
6111 */
6112 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
6113 dev->l2ad_hand = dev->l2ad_start;
6114 dev->l2ad_first = B_FALSE;
6115 }
6116
6117 dev->l2ad_writing = B_TRUE;
6118 (void) zio_wait(pio);
6119 dev->l2ad_writing = B_FALSE;
6120
6121 return (write_asize);
6122 }
6123
6124 /*
6125 * Compresses an L2ARC buffer.
6126 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6127 * size in l2hdr->b_asize. This routine tries to compress the data and
6128 * depending on the compression result there are three possible outcomes:
6129 * *) The buffer was incompressible. The original l2hdr contents were left
6130 * untouched and are ready for writing to an L2 device.
6131 * *) The buffer was all-zeros, so there is no need to write it to an L2
6132 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6133 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6134 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6135 * data buffer which holds the compressed data to be written, and b_asize
6136 * tells us how much data there is. b_compress is set to the appropriate
6137 * compression algorithm. Once writing is done, invoke
6138 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6139 *
6140 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6141 * buffer was incompressible).
6142 */
6143 static boolean_t
6144 l2arc_compress_buf(arc_buf_hdr_t *hdr)
6145 {
6146 void *cdata;
6147 size_t csize, len, rounded;
6148 ASSERT(HDR_HAS_L2HDR(hdr));
6149 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
6150
6151 ASSERT(HDR_HAS_L1HDR(hdr));
6152 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF);
6153 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6154
6155 len = l2hdr->b_asize;
6156 cdata = zio_data_buf_alloc(len);
6157 ASSERT3P(cdata, !=, NULL);
6158 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata,
6159 cdata, l2hdr->b_asize);
6160
6161 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE);
6162 if (rounded > csize) {
6163 bzero((char *)cdata + csize, rounded - csize);
6164 csize = rounded;
6165 }
6166
6167 if (csize == 0) {
6168 /* zero block, indicate that there's nothing to write */
6169 zio_data_buf_free(cdata, len);
6170 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
6171 l2hdr->b_asize = 0;
6172 hdr->b_l1hdr.b_tmp_cdata = NULL;
6173 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
6174 return (B_TRUE);
6175 } else if (csize > 0 && csize < len) {
6176 /*
6177 * Compression succeeded, we'll keep the cdata around for
6178 * writing and release it afterwards.
6179 */
6180 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
6181 l2hdr->b_asize = csize;
6182 hdr->b_l1hdr.b_tmp_cdata = cdata;
6183 ARCSTAT_BUMP(arcstat_l2_compress_successes);
6184 return (B_TRUE);
6185 } else {
6186 /*
6187 * Compression failed, release the compressed buffer.
6188 * l2hdr will be left unmodified.
6189 */
6190 zio_data_buf_free(cdata, len);
6191 ARCSTAT_BUMP(arcstat_l2_compress_failures);
6192 return (B_FALSE);
6193 }
6194 }
6195
6196 /*
6197 * Decompresses a zio read back from an l2arc device. On success, the
6198 * underlying zio's io_data buffer is overwritten by the uncompressed
6199 * version. On decompression error (corrupt compressed stream), the
6200 * zio->io_error value is set to signal an I/O error.
6201 *
6202 * Please note that the compressed data stream is not checksummed, so
6203 * if the underlying device is experiencing data corruption, we may feed
6204 * corrupt data to the decompressor, so the decompressor needs to be
6205 * able to handle this situation (LZ4 does).
6206 */
6207 static void
6208 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
6209 {
6210 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
6211
6212 if (zio->io_error != 0) {
6213 /*
6214 * An io error has occured, just restore the original io
6215 * size in preparation for a main pool read.
6216 */
6217 zio->io_orig_size = zio->io_size = hdr->b_size;
6218 return;
6219 }
6220
6221 if (c == ZIO_COMPRESS_EMPTY) {
6222 /*
6223 * An empty buffer results in a null zio, which means we
6224 * need to fill its io_data after we're done restoring the
6225 * buffer's contents.
6226 */
6227 ASSERT(hdr->b_l1hdr.b_buf != NULL);
6228 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size);
6229 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data;
6230 } else {
6231 ASSERT(zio->io_data != NULL);
6232 /*
6233 * We copy the compressed data from the start of the arc buffer
6234 * (the zio_read will have pulled in only what we need, the
6235 * rest is garbage which we will overwrite at decompression)
6236 * and then decompress back to the ARC data buffer. This way we
6237 * can minimize copying by simply decompressing back over the
6238 * original compressed data (rather than decompressing to an
6239 * aux buffer and then copying back the uncompressed buffer,
6240 * which is likely to be much larger).
6241 */
6242 uint64_t csize;
6243 void *cdata;
6244
6245 csize = zio->io_size;
6246 cdata = zio_data_buf_alloc(csize);
6247 bcopy(zio->io_data, cdata, csize);
6248 if (zio_decompress_data(c, cdata, zio->io_data, csize,
6249 hdr->b_size) != 0)
6250 zio->io_error = EIO;
6251 zio_data_buf_free(cdata, csize);
6252 }
6253
6254 /* Restore the expected uncompressed IO size. */
6255 zio->io_orig_size = zio->io_size = hdr->b_size;
6256 }
6257
6258 /*
6259 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6260 * This buffer serves as a temporary holder of compressed data while
6261 * the buffer entry is being written to an l2arc device. Once that is
6262 * done, we can dispose of it.
6263 */
6264 static void
6265 l2arc_release_cdata_buf(arc_buf_hdr_t *hdr)
6266 {
6267 ASSERT(HDR_HAS_L2HDR(hdr));
6268 enum zio_compress comp = hdr->b_l2hdr.b_compress;
6269
6270 ASSERT(HDR_HAS_L1HDR(hdr));
6271 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp));
6272
6273 if (comp == ZIO_COMPRESS_OFF) {
6274 /*
6275 * In this case, b_tmp_cdata points to the same buffer
6276 * as the arc_buf_t's b_data field. We don't want to
6277 * free it, since the arc_buf_t will handle that.
6278 */
6279 hdr->b_l1hdr.b_tmp_cdata = NULL;
6280 } else if (comp == ZIO_COMPRESS_EMPTY) {
6281 /*
6282 * In this case, b_tmp_cdata was compressed to an empty
6283 * buffer, thus there's nothing to free and b_tmp_cdata
6284 * should have been set to NULL in l2arc_write_buffers().
6285 */
6286 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
6287 } else {
6288 /*
6289 * If the data was compressed, then we've allocated a
6290 * temporary buffer for it, so now we need to release it.
6291 */
6292 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6293 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata,
6294 hdr->b_size);
6295 hdr->b_l1hdr.b_tmp_cdata = NULL;
6296 }
6297
6298 }
6299
6300 /*
6301 * This thread feeds the L2ARC at regular intervals. This is the beating
6302 * heart of the L2ARC.
6303 */
6304 static void
6305 l2arc_feed_thread(void)
6306 {
6307 callb_cpr_t cpr;
6308 l2arc_dev_t *dev;
6309 spa_t *spa;
6310 uint64_t size, wrote;
6311 clock_t begin, next = ddi_get_lbolt();
6312 boolean_t headroom_boost = B_FALSE;
6313
6314 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
6315
6316 mutex_enter(&l2arc_feed_thr_lock);
6317
6318 while (l2arc_thread_exit == 0) {
6319 CALLB_CPR_SAFE_BEGIN(&cpr);
6320 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
6321 next);
6322 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
6323 next = ddi_get_lbolt() + hz;
6324
6325 /*
6326 * Quick check for L2ARC devices.
6327 */
6328 mutex_enter(&l2arc_dev_mtx);
6329 if (l2arc_ndev == 0) {
6330 mutex_exit(&l2arc_dev_mtx);
6331 continue;
6332 }
6333 mutex_exit(&l2arc_dev_mtx);
6334 begin = ddi_get_lbolt();
6335
6336 /*
6337 * This selects the next l2arc device to write to, and in
6338 * doing so the next spa to feed from: dev->l2ad_spa. This
6339 * will return NULL if there are now no l2arc devices or if
6340 * they are all faulted.
6341 *
6342 * If a device is returned, its spa's config lock is also
6343 * held to prevent device removal. l2arc_dev_get_next()
6344 * will grab and release l2arc_dev_mtx.
6345 */
6346 if ((dev = l2arc_dev_get_next()) == NULL)
6347 continue;
6348
6349 spa = dev->l2ad_spa;
6350 ASSERT(spa != NULL);
6351
6352 /*
6353 * If the pool is read-only then force the feed thread to
6354 * sleep a little longer.
6355 */
6356 if (!spa_writeable(spa)) {
6357 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
6358 spa_config_exit(spa, SCL_L2ARC, dev);
6359 continue;
6360 }
6361
6362 /*
6363 * Avoid contributing to memory pressure.
6364 */
6365 if (arc_reclaim_needed()) {
6366 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
6367 spa_config_exit(spa, SCL_L2ARC, dev);
6368 continue;
6369 }
6370
6371 ARCSTAT_BUMP(arcstat_l2_feeds);
6372
6373 size = l2arc_write_size();
6374
6375 /*
6376 * Evict L2ARC buffers that will be overwritten.
6377 */
6378 l2arc_evict(dev, size, B_FALSE);
6379
6380 /*
6381 * Write ARC buffers.
6382 */
6383 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
6384
6385 /*
6386 * Calculate interval between writes.
6387 */
6388 next = l2arc_write_interval(begin, size, wrote);
6389 spa_config_exit(spa, SCL_L2ARC, dev);
6390 }
6391
6392 l2arc_thread_exit = 0;
6393 cv_broadcast(&l2arc_feed_thr_cv);
6394 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
6395 thread_exit();
6396 }
6397
6398 boolean_t
6399 l2arc_vdev_present(vdev_t *vd)
6400 {
6401 l2arc_dev_t *dev;
6402
6403 mutex_enter(&l2arc_dev_mtx);
6404 for (dev = list_head(l2arc_dev_list); dev != NULL;
6405 dev = list_next(l2arc_dev_list, dev)) {
6406 if (dev->l2ad_vdev == vd)
6407 break;
6408 }
6409 mutex_exit(&l2arc_dev_mtx);
6410
6411 return (dev != NULL);
6412 }
6413
6414 /*
6415 * Add a vdev for use by the L2ARC. By this point the spa has already
6416 * validated the vdev and opened it.
6417 */
6418 void
6419 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
6420 {
6421 l2arc_dev_t *adddev;
6422
6423 ASSERT(!l2arc_vdev_present(vd));
6424
6425 /*
6426 * Create a new l2arc device entry.
6427 */
6428 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
6429 adddev->l2ad_spa = spa;
6430 adddev->l2ad_vdev = vd;
6431 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
6432 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
6433 adddev->l2ad_hand = adddev->l2ad_start;
6434 adddev->l2ad_first = B_TRUE;
6435 adddev->l2ad_writing = B_FALSE;
6436
6437 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
6438 /*
6439 * This is a list of all ARC buffers that are still valid on the
6440 * device.
6441 */
6442 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
6443 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
6444
6445 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
6446 refcount_create(&adddev->l2ad_alloc);
6447
6448 /*
6449 * Add device to global list
6450 */
6451 mutex_enter(&l2arc_dev_mtx);
6452 list_insert_head(l2arc_dev_list, adddev);
6453 atomic_inc_64(&l2arc_ndev);
6454 mutex_exit(&l2arc_dev_mtx);
6455 }
6456
6457 /*
6458 * Remove a vdev from the L2ARC.
6459 */
6460 void
6461 l2arc_remove_vdev(vdev_t *vd)
6462 {
6463 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
6464
6465 /*
6466 * Find the device by vdev
6467 */
6468 mutex_enter(&l2arc_dev_mtx);
6469 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
6470 nextdev = list_next(l2arc_dev_list, dev);
6471 if (vd == dev->l2ad_vdev) {
6472 remdev = dev;
6473 break;
6474 }
6475 }
6476 ASSERT(remdev != NULL);
6477
6478 /*
6479 * Remove device from global list
6480 */
6481 list_remove(l2arc_dev_list, remdev);
6482 l2arc_dev_last = NULL; /* may have been invalidated */
6483 atomic_dec_64(&l2arc_ndev);
6484 mutex_exit(&l2arc_dev_mtx);
6485
6486 /*
6487 * Clear all buflists and ARC references. L2ARC device flush.
6488 */
6489 l2arc_evict(remdev, 0, B_TRUE);
6490 list_destroy(&remdev->l2ad_buflist);
6491 mutex_destroy(&remdev->l2ad_mtx);
6492 refcount_destroy(&remdev->l2ad_alloc);
6493 kmem_free(remdev, sizeof (l2arc_dev_t));
6494 }
6495
6496 void
6497 l2arc_init(void)
6498 {
6499 l2arc_thread_exit = 0;
6500 l2arc_ndev = 0;
6501 l2arc_writes_sent = 0;
6502 l2arc_writes_done = 0;
6503
6504 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
6505 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
6506 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
6507 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
6508
6509 l2arc_dev_list = &L2ARC_dev_list;
6510 l2arc_free_on_write = &L2ARC_free_on_write;
6511 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
6512 offsetof(l2arc_dev_t, l2ad_node));
6513 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
6514 offsetof(l2arc_data_free_t, l2df_list_node));
6515 }
6516
6517 void
6518 l2arc_fini(void)
6519 {
6520 /*
6521 * This is called from dmu_fini(), which is called from spa_fini();
6522 * Because of this, we can assume that all l2arc devices have
6523 * already been removed when the pools themselves were removed.
6524 */
6525
6526 l2arc_do_free_on_write();
6527
6528 mutex_destroy(&l2arc_feed_thr_lock);
6529 cv_destroy(&l2arc_feed_thr_cv);
6530 mutex_destroy(&l2arc_dev_mtx);
6531 mutex_destroy(&l2arc_free_on_write_mtx);
6532
6533 list_destroy(l2arc_dev_list);
6534 list_destroy(l2arc_free_on_write);
6535 }
6536
6537 void
6538 l2arc_start(void)
6539 {
6540 if (!(spa_mode_global & FWRITE))
6541 return;
6542
6543 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
6544 TS_RUN, minclsyspri);
6545 }
6546
6547 void
6548 l2arc_stop(void)
6549 {
6550 if (!(spa_mode_global & FWRITE))
6551 return;
6552
6553 mutex_enter(&l2arc_feed_thr_lock);
6554 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
6555 l2arc_thread_exit = 1;
6556 while (l2arc_thread_exit != 0)
6557 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
6558 mutex_exit(&l2arc_feed_thr_lock);
6559 }