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) 2018, Joyent, Inc.
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2017 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 * Note that the majority of the performance stats are manipulated
103 * with atomic operations.
104 *
105 * The L2ARC uses the l2ad_mtx on each vdev for the following:
106 *
107 * - L2ARC buflist creation
108 * - L2ARC buflist eviction
109 * - L2ARC write completion, which walks L2ARC buflists
110 * - ARC header destruction, as it removes from L2ARC buflists
111 * - ARC header release, as it removes from L2ARC buflists
112 */
113
114 /*
115 * ARC operation:
116 *
117 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
118 * This structure can point either to a block that is still in the cache or to
119 * one that is only accessible in an L2 ARC device, or it can provide
120 * information about a block that was recently evicted. If a block is
121 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
122 * information to retrieve it from the L2ARC device. This information is
123 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
124 * that is in this state cannot access the data directly.
125 *
126 * Blocks that are actively being referenced or have not been evicted
127 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
128 * the arc_buf_hdr_t that will point to the data block in memory. A block can
129 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
130 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
131 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
132 *
133 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
134 * ability to store the physical data (b_pabd) associated with the DVA of the
135 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
136 * it will match its on-disk compression characteristics. This behavior can be
137 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
138 * compressed ARC functionality is disabled, the b_pabd will point to an
139 * uncompressed version of the on-disk data.
140 *
141 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
142 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
143 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
144 * consumer. The ARC will provide references to this data and will keep it
145 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
146 * data block and will evict any arc_buf_t that is no longer referenced. The
147 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
148 * "overhead_size" kstat.
149 *
150 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
151 * compressed form. The typical case is that consumers will want uncompressed
152 * data, and when that happens a new data buffer is allocated where the data is
153 * decompressed for them to use. Currently the only consumer who wants
154 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
155 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
156 * with the arc_buf_hdr_t.
157 *
158 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
159 * first one is owned by a compressed send consumer (and therefore references
160 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
161 * used by any other consumer (and has its own uncompressed copy of the data
162 * buffer).
163 *
164 * arc_buf_hdr_t
165 * +-----------+
166 * | fields |
167 * | common to |
168 * | L1- and |
169 * | L2ARC |
170 * +-----------+
171 * | l2arc_buf_hdr_t
172 * | |
173 * +-----------+
174 * | l1arc_buf_hdr_t
175 * | | arc_buf_t
176 * | b_buf +------------>+-----------+ arc_buf_t
177 * | b_pabd +-+ |b_next +---->+-----------+
178 * +-----------+ | |-----------| |b_next +-->NULL
179 * | |b_comp = T | +-----------+
180 * | |b_data +-+ |b_comp = F |
181 * | +-----------+ | |b_data +-+
182 * +->+------+ | +-----------+ |
183 * compressed | | | |
184 * data | |<--------------+ | uncompressed
185 * +------+ compressed, | data
186 * shared +-->+------+
187 * data | |
188 * | |
189 * +------+
190 *
191 * When a consumer reads a block, the ARC must first look to see if the
192 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
193 * arc_buf_t and either copies uncompressed data into a new data buffer from an
194 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
195 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
196 * hdr is compressed and the desired compression characteristics of the
197 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
198 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
199 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
200 * be anywhere in the hdr's list.
201 *
202 * The diagram below shows an example of an uncompressed ARC hdr that is
203 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
204 * the last element in the buf list):
205 *
206 * arc_buf_hdr_t
207 * +-----------+
208 * | |
209 * | |
210 * | |
211 * +-----------+
212 * l2arc_buf_hdr_t| |
213 * | |
214 * +-----------+
215 * l1arc_buf_hdr_t| |
216 * | | arc_buf_t (shared)
217 * | b_buf +------------>+---------+ arc_buf_t
218 * | | |b_next +---->+---------+
219 * | b_pabd +-+ |---------| |b_next +-->NULL
220 * +-----------+ | | | +---------+
221 * | |b_data +-+ | |
222 * | +---------+ | |b_data +-+
223 * +->+------+ | +---------+ |
224 * | | | |
225 * uncompressed | | | |
226 * data +------+ | |
227 * ^ +->+------+ |
228 * | uncompressed | | |
229 * | data | | |
230 * | +------+ |
231 * +---------------------------------+
232 *
233 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
234 * since the physical block is about to be rewritten. The new data contents
235 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
236 * it may compress the data before writing it to disk. The ARC will be called
237 * with the transformed data and will bcopy the transformed on-disk block into
238 * a newly allocated b_pabd. Writes are always done into buffers which have
239 * either been loaned (and hence are new and don't have other readers) or
240 * buffers which have been released (and hence have their own hdr, if there
241 * were originally other readers of the buf's original hdr). This ensures that
242 * the ARC only needs to update a single buf and its hdr after a write occurs.
243 *
244 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
245 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
246 * that when compressed ARC is enabled that the L2ARC blocks are identical
247 * to the on-disk block in the main data pool. This provides a significant
248 * advantage since the ARC can leverage the bp's checksum when reading from the
249 * L2ARC to determine if the contents are valid. However, if the compressed
250 * ARC is disabled, then the L2ARC's block must be transformed to look
251 * like the physical block in the main data pool before comparing the
252 * checksum and determining its validity.
253 */
254
255 #include <sys/spa.h>
256 #include <sys/zio.h>
257 #include <sys/spa_impl.h>
258 #include <sys/zio_compress.h>
259 #include <sys/zio_checksum.h>
260 #include <sys/zfs_context.h>
261 #include <sys/arc.h>
262 #include <sys/refcount.h>
263 #include <sys/vdev.h>
264 #include <sys/vdev_impl.h>
265 #include <sys/dsl_pool.h>
266 #include <sys/zio_checksum.h>
267 #include <sys/multilist.h>
268 #include <sys/abd.h>
269 #ifdef _KERNEL
270 #include <sys/vmsystm.h>
271 #include <vm/anon.h>
272 #include <sys/fs/swapnode.h>
273 #include <sys/dnlc.h>
274 #endif
275 #include <sys/callb.h>
276 #include <sys/kstat.h>
277 #include <zfs_fletcher.h>
278 #include <sys/aggsum.h>
279 #include <sys/cityhash.h>
280
281 #ifndef _KERNEL
282 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
283 boolean_t arc_watch = B_FALSE;
284 int arc_procfd;
285 #endif
286
287 static kmutex_t arc_reclaim_lock;
288 static kcondvar_t arc_reclaim_thread_cv;
289 static boolean_t arc_reclaim_thread_exit;
290 static kcondvar_t arc_reclaim_waiters_cv;
291
292 uint_t arc_reduce_dnlc_percent = 3;
293
294 /*
295 * The number of headers to evict in arc_evict_state_impl() before
296 * dropping the sublist lock and evicting from another sublist. A lower
297 * value means we're more likely to evict the "correct" header (i.e. the
298 * oldest header in the arc state), but comes with higher overhead
299 * (i.e. more invocations of arc_evict_state_impl()).
300 */
301 int zfs_arc_evict_batch_limit = 10;
302
303 /* number of seconds before growing cache again */
304 static int arc_grow_retry = 60;
305
306 /* number of milliseconds before attempting a kmem-cache-reap */
307 static int arc_kmem_cache_reap_retry_ms = 1000;
308
309 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
310 int zfs_arc_overflow_shift = 8;
311
312 /* shift of arc_c for calculating both min and max arc_p */
313 static int arc_p_min_shift = 4;
314
315 /* log2(fraction of arc to reclaim) */
316 static int arc_shrink_shift = 7;
317
318 /*
319 * log2(fraction of ARC which must be free to allow growing).
320 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
321 * when reading a new block into the ARC, we will evict an equal-sized block
322 * from the ARC.
323 *
324 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
325 * we will still not allow it to grow.
326 */
327 int arc_no_grow_shift = 5;
328
329
330 /*
331 * minimum lifespan of a prefetch block in clock ticks
332 * (initialized in arc_init())
333 */
334 static int arc_min_prefetch_lifespan;
335
336 /*
337 * If this percent of memory is free, don't throttle.
338 */
339 int arc_lotsfree_percent = 10;
340
341 static int arc_dead;
342
343 /*
344 * The arc has filled available memory and has now warmed up.
345 */
346 static boolean_t arc_warm;
347
348 /*
349 * log2 fraction of the zio arena to keep free.
350 */
351 int arc_zio_arena_free_shift = 2;
352
353 /*
354 * These tunables are for performance analysis.
355 */
356 uint64_t zfs_arc_max;
357 uint64_t zfs_arc_min;
358 uint64_t zfs_arc_meta_limit = 0;
359 uint64_t zfs_arc_meta_min = 0;
360 int zfs_arc_grow_retry = 0;
361 int zfs_arc_shrink_shift = 0;
362 int zfs_arc_p_min_shift = 0;
363 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
364
365 boolean_t zfs_compressed_arc_enabled = B_TRUE;
366
367 /*
368 * Note that buffers can be in one of 6 states:
369 * ARC_anon - anonymous (discussed below)
370 * ARC_mru - recently used, currently cached
371 * ARC_mru_ghost - recentely used, no longer in cache
372 * ARC_mfu - frequently used, currently cached
373 * ARC_mfu_ghost - frequently used, no longer in cache
374 * ARC_l2c_only - exists in L2ARC but not other states
375 * When there are no active references to the buffer, they are
376 * are linked onto a list in one of these arc states. These are
377 * the only buffers that can be evicted or deleted. Within each
378 * state there are multiple lists, one for meta-data and one for
379 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
380 * etc.) is tracked separately so that it can be managed more
381 * explicitly: favored over data, limited explicitly.
382 *
383 * Anonymous buffers are buffers that are not associated with
384 * a DVA. These are buffers that hold dirty block copies
385 * before they are written to stable storage. By definition,
386 * they are "ref'd" and are considered part of arc_mru
387 * that cannot be freed. Generally, they will aquire a DVA
388 * as they are written and migrate onto the arc_mru list.
389 *
390 * The ARC_l2c_only state is for buffers that are in the second
391 * level ARC but no longer in any of the ARC_m* lists. The second
392 * level ARC itself may also contain buffers that are in any of
393 * the ARC_m* states - meaning that a buffer can exist in two
394 * places. The reason for the ARC_l2c_only state is to keep the
395 * buffer header in the hash table, so that reads that hit the
396 * second level ARC benefit from these fast lookups.
397 */
398
399 typedef struct arc_state {
400 /*
401 * list of evictable buffers
402 */
403 multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
404 /*
405 * total amount of evictable data in this state
406 */
407 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
408 /*
409 * total amount of data in this state; this includes: evictable,
410 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
411 */
412 refcount_t arcs_size;
413 } arc_state_t;
414
415 /* The 6 states: */
416 static arc_state_t ARC_anon;
417 static arc_state_t ARC_mru;
418 static arc_state_t ARC_mru_ghost;
419 static arc_state_t ARC_mfu;
420 static arc_state_t ARC_mfu_ghost;
421 static arc_state_t ARC_l2c_only;
422
423 typedef struct arc_stats {
424 kstat_named_t arcstat_hits;
425 kstat_named_t arcstat_misses;
426 kstat_named_t arcstat_demand_data_hits;
427 kstat_named_t arcstat_demand_data_misses;
428 kstat_named_t arcstat_demand_metadata_hits;
429 kstat_named_t arcstat_demand_metadata_misses;
430 kstat_named_t arcstat_prefetch_data_hits;
431 kstat_named_t arcstat_prefetch_data_misses;
432 kstat_named_t arcstat_prefetch_metadata_hits;
433 kstat_named_t arcstat_prefetch_metadata_misses;
434 kstat_named_t arcstat_mru_hits;
435 kstat_named_t arcstat_mru_ghost_hits;
436 kstat_named_t arcstat_mfu_hits;
437 kstat_named_t arcstat_mfu_ghost_hits;
438 kstat_named_t arcstat_deleted;
439 /*
440 * Number of buffers that could not be evicted because the hash lock
441 * was held by another thread. The lock may not necessarily be held
442 * by something using the same buffer, since hash locks are shared
443 * by multiple buffers.
444 */
445 kstat_named_t arcstat_mutex_miss;
446 /*
447 * Number of buffers skipped because they have I/O in progress, are
448 * indrect prefetch buffers that have not lived long enough, or are
449 * not from the spa we're trying to evict from.
450 */
451 kstat_named_t arcstat_evict_skip;
452 /*
453 * Number of times arc_evict_state() was unable to evict enough
454 * buffers to reach it's target amount.
455 */
456 kstat_named_t arcstat_evict_not_enough;
457 kstat_named_t arcstat_evict_l2_cached;
458 kstat_named_t arcstat_evict_l2_eligible;
459 kstat_named_t arcstat_evict_l2_ineligible;
460 kstat_named_t arcstat_evict_l2_skip;
461 kstat_named_t arcstat_hash_elements;
462 kstat_named_t arcstat_hash_elements_max;
463 kstat_named_t arcstat_hash_collisions;
464 kstat_named_t arcstat_hash_chains;
465 kstat_named_t arcstat_hash_chain_max;
466 kstat_named_t arcstat_p;
467 kstat_named_t arcstat_c;
468 kstat_named_t arcstat_c_min;
469 kstat_named_t arcstat_c_max;
470 /* Not updated directly; only synced in arc_kstat_update. */
471 kstat_named_t arcstat_size;
472 /*
473 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
474 * Note that the compressed bytes may match the uncompressed bytes
475 * if the block is either not compressed or compressed arc is disabled.
476 */
477 kstat_named_t arcstat_compressed_size;
478 /*
479 * Uncompressed size of the data stored in b_pabd. If compressed
480 * arc is disabled then this value will be identical to the stat
481 * above.
482 */
483 kstat_named_t arcstat_uncompressed_size;
484 /*
485 * Number of bytes stored in all the arc_buf_t's. This is classified
486 * as "overhead" since this data is typically short-lived and will
487 * be evicted from the arc when it becomes unreferenced unless the
488 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
489 * values have been set (see comment in dbuf.c for more information).
490 */
491 kstat_named_t arcstat_overhead_size;
492 /*
493 * Number of bytes consumed by internal ARC structures necessary
494 * for tracking purposes; these structures are not actually
495 * backed by ARC buffers. This includes arc_buf_hdr_t structures
496 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
497 * caches), and arc_buf_t structures (allocated via arc_buf_t
498 * cache).
499 * Not updated directly; only synced in arc_kstat_update.
500 */
501 kstat_named_t arcstat_hdr_size;
502 /*
503 * Number of bytes consumed by ARC buffers of type equal to
504 * ARC_BUFC_DATA. This is generally consumed by buffers backing
505 * on disk user data (e.g. plain file contents).
506 * Not updated directly; only synced in arc_kstat_update.
507 */
508 kstat_named_t arcstat_data_size;
509 /*
510 * Number of bytes consumed by ARC buffers of type equal to
511 * ARC_BUFC_METADATA. This is generally consumed by buffers
512 * backing on disk data that is used for internal ZFS
513 * structures (e.g. ZAP, dnode, indirect blocks, etc).
514 * Not updated directly; only synced in arc_kstat_update.
515 */
516 kstat_named_t arcstat_metadata_size;
517 /*
518 * Number of bytes consumed by various buffers and structures
519 * not actually backed with ARC buffers. This includes bonus
520 * buffers (allocated directly via zio_buf_* functions),
521 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
522 * cache), and dnode_t structures (allocated via dnode_t cache).
523 * Not updated directly; only synced in arc_kstat_update.
524 */
525 kstat_named_t arcstat_other_size;
526 /*
527 * Total number of bytes consumed by ARC buffers residing in the
528 * arc_anon state. This includes *all* buffers in the arc_anon
529 * state; e.g. data, metadata, evictable, and unevictable buffers
530 * are all included in this value.
531 * Not updated directly; only synced in arc_kstat_update.
532 */
533 kstat_named_t arcstat_anon_size;
534 /*
535 * Number of bytes consumed by ARC buffers that meet the
536 * following criteria: backing buffers of type ARC_BUFC_DATA,
537 * residing in the arc_anon state, and are eligible for eviction
538 * (e.g. have no outstanding holds on the buffer).
539 * Not updated directly; only synced in arc_kstat_update.
540 */
541 kstat_named_t arcstat_anon_evictable_data;
542 /*
543 * Number of bytes consumed by ARC buffers that meet the
544 * following criteria: backing buffers of type ARC_BUFC_METADATA,
545 * residing in the arc_anon state, and are eligible for eviction
546 * (e.g. have no outstanding holds on the buffer).
547 * Not updated directly; only synced in arc_kstat_update.
548 */
549 kstat_named_t arcstat_anon_evictable_metadata;
550 /*
551 * Total number of bytes consumed by ARC buffers residing in the
552 * arc_mru state. This includes *all* buffers in the arc_mru
553 * state; e.g. data, metadata, evictable, and unevictable buffers
554 * are all included in this value.
555 * Not updated directly; only synced in arc_kstat_update.
556 */
557 kstat_named_t arcstat_mru_size;
558 /*
559 * Number of bytes consumed by ARC buffers that meet the
560 * following criteria: backing buffers of type ARC_BUFC_DATA,
561 * residing in the arc_mru state, and are eligible for eviction
562 * (e.g. have no outstanding holds on the buffer).
563 * Not updated directly; only synced in arc_kstat_update.
564 */
565 kstat_named_t arcstat_mru_evictable_data;
566 /*
567 * Number of bytes consumed by ARC buffers that meet the
568 * following criteria: backing buffers of type ARC_BUFC_METADATA,
569 * residing in the arc_mru state, and are eligible for eviction
570 * (e.g. have no outstanding holds on the buffer).
571 * Not updated directly; only synced in arc_kstat_update.
572 */
573 kstat_named_t arcstat_mru_evictable_metadata;
574 /*
575 * Total number of bytes that *would have been* consumed by ARC
576 * buffers in the arc_mru_ghost state. The key thing to note
577 * here, is the fact that this size doesn't actually indicate
578 * RAM consumption. The ghost lists only consist of headers and
579 * don't actually have ARC buffers linked off of these headers.
580 * Thus, *if* the headers had associated ARC buffers, these
581 * buffers *would have* consumed this number of bytes.
582 * Not updated directly; only synced in arc_kstat_update.
583 */
584 kstat_named_t arcstat_mru_ghost_size;
585 /*
586 * Number of bytes that *would have been* consumed by ARC
587 * buffers that are eligible for eviction, of type
588 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
589 * Not updated directly; only synced in arc_kstat_update.
590 */
591 kstat_named_t arcstat_mru_ghost_evictable_data;
592 /*
593 * Number of bytes that *would have been* consumed by ARC
594 * buffers that are eligible for eviction, of type
595 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
596 * Not updated directly; only synced in arc_kstat_update.
597 */
598 kstat_named_t arcstat_mru_ghost_evictable_metadata;
599 /*
600 * Total number of bytes consumed by ARC buffers residing in the
601 * arc_mfu state. This includes *all* buffers in the arc_mfu
602 * state; e.g. data, metadata, evictable, and unevictable buffers
603 * are all included in this value.
604 * Not updated directly; only synced in arc_kstat_update.
605 */
606 kstat_named_t arcstat_mfu_size;
607 /*
608 * Number of bytes consumed by ARC buffers that are eligible for
609 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
610 * state.
611 * Not updated directly; only synced in arc_kstat_update.
612 */
613 kstat_named_t arcstat_mfu_evictable_data;
614 /*
615 * Number of bytes consumed by ARC buffers that are eligible for
616 * eviction, of type ARC_BUFC_METADATA, and reside in the
617 * arc_mfu state.
618 * Not updated directly; only synced in arc_kstat_update.
619 */
620 kstat_named_t arcstat_mfu_evictable_metadata;
621 /*
622 * Total number of bytes that *would have been* consumed by ARC
623 * buffers in the arc_mfu_ghost state. See the comment above
624 * arcstat_mru_ghost_size for more details.
625 * Not updated directly; only synced in arc_kstat_update.
626 */
627 kstat_named_t arcstat_mfu_ghost_size;
628 /*
629 * Number of bytes that *would have been* consumed by ARC
630 * buffers that are eligible for eviction, of type
631 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
632 * Not updated directly; only synced in arc_kstat_update.
633 */
634 kstat_named_t arcstat_mfu_ghost_evictable_data;
635 /*
636 * Number of bytes that *would have been* consumed by ARC
637 * buffers that are eligible for eviction, of type
638 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
639 * Not updated directly; only synced in arc_kstat_update.
640 */
641 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
642 kstat_named_t arcstat_l2_hits;
643 kstat_named_t arcstat_l2_misses;
644 kstat_named_t arcstat_l2_feeds;
645 kstat_named_t arcstat_l2_rw_clash;
646 kstat_named_t arcstat_l2_read_bytes;
647 kstat_named_t arcstat_l2_write_bytes;
648 kstat_named_t arcstat_l2_writes_sent;
649 kstat_named_t arcstat_l2_writes_done;
650 kstat_named_t arcstat_l2_writes_error;
651 kstat_named_t arcstat_l2_writes_lock_retry;
652 kstat_named_t arcstat_l2_evict_lock_retry;
653 kstat_named_t arcstat_l2_evict_reading;
654 kstat_named_t arcstat_l2_evict_l1cached;
655 kstat_named_t arcstat_l2_free_on_write;
656 kstat_named_t arcstat_l2_abort_lowmem;
657 kstat_named_t arcstat_l2_cksum_bad;
658 kstat_named_t arcstat_l2_io_error;
659 kstat_named_t arcstat_l2_lsize;
660 kstat_named_t arcstat_l2_psize;
661 /* Not updated directly; only synced in arc_kstat_update. */
662 kstat_named_t arcstat_l2_hdr_size;
663 kstat_named_t arcstat_memory_throttle_count;
664 /* Not updated directly; only synced in arc_kstat_update. */
665 kstat_named_t arcstat_meta_used;
666 kstat_named_t arcstat_meta_limit;
667 kstat_named_t arcstat_meta_max;
668 kstat_named_t arcstat_meta_min;
669 kstat_named_t arcstat_sync_wait_for_async;
670 kstat_named_t arcstat_demand_hit_predictive_prefetch;
671 } arc_stats_t;
672
673 static arc_stats_t arc_stats = {
674 { "hits", KSTAT_DATA_UINT64 },
675 { "misses", KSTAT_DATA_UINT64 },
676 { "demand_data_hits", KSTAT_DATA_UINT64 },
677 { "demand_data_misses", KSTAT_DATA_UINT64 },
678 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
679 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
680 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
681 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
682 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
683 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
684 { "mru_hits", KSTAT_DATA_UINT64 },
685 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
686 { "mfu_hits", KSTAT_DATA_UINT64 },
687 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
688 { "deleted", KSTAT_DATA_UINT64 },
689 { "mutex_miss", KSTAT_DATA_UINT64 },
690 { "evict_skip", KSTAT_DATA_UINT64 },
691 { "evict_not_enough", KSTAT_DATA_UINT64 },
692 { "evict_l2_cached", KSTAT_DATA_UINT64 },
693 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
694 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
695 { "evict_l2_skip", KSTAT_DATA_UINT64 },
696 { "hash_elements", KSTAT_DATA_UINT64 },
697 { "hash_elements_max", KSTAT_DATA_UINT64 },
698 { "hash_collisions", KSTAT_DATA_UINT64 },
699 { "hash_chains", KSTAT_DATA_UINT64 },
700 { "hash_chain_max", KSTAT_DATA_UINT64 },
701 { "p", KSTAT_DATA_UINT64 },
702 { "c", KSTAT_DATA_UINT64 },
703 { "c_min", KSTAT_DATA_UINT64 },
704 { "c_max", KSTAT_DATA_UINT64 },
705 { "size", KSTAT_DATA_UINT64 },
706 { "compressed_size", KSTAT_DATA_UINT64 },
707 { "uncompressed_size", KSTAT_DATA_UINT64 },
708 { "overhead_size", KSTAT_DATA_UINT64 },
709 { "hdr_size", KSTAT_DATA_UINT64 },
710 { "data_size", KSTAT_DATA_UINT64 },
711 { "metadata_size", KSTAT_DATA_UINT64 },
712 { "other_size", KSTAT_DATA_UINT64 },
713 { "anon_size", KSTAT_DATA_UINT64 },
714 { "anon_evictable_data", KSTAT_DATA_UINT64 },
715 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
716 { "mru_size", KSTAT_DATA_UINT64 },
717 { "mru_evictable_data", KSTAT_DATA_UINT64 },
718 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
719 { "mru_ghost_size", KSTAT_DATA_UINT64 },
720 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
721 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
722 { "mfu_size", KSTAT_DATA_UINT64 },
723 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
724 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
725 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
726 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
727 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
728 { "l2_hits", KSTAT_DATA_UINT64 },
729 { "l2_misses", KSTAT_DATA_UINT64 },
730 { "l2_feeds", KSTAT_DATA_UINT64 },
731 { "l2_rw_clash", KSTAT_DATA_UINT64 },
732 { "l2_read_bytes", KSTAT_DATA_UINT64 },
733 { "l2_write_bytes", KSTAT_DATA_UINT64 },
734 { "l2_writes_sent", KSTAT_DATA_UINT64 },
735 { "l2_writes_done", KSTAT_DATA_UINT64 },
736 { "l2_writes_error", KSTAT_DATA_UINT64 },
737 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
738 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
739 { "l2_evict_reading", KSTAT_DATA_UINT64 },
740 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
741 { "l2_free_on_write", KSTAT_DATA_UINT64 },
742 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
743 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
744 { "l2_io_error", KSTAT_DATA_UINT64 },
745 { "l2_size", KSTAT_DATA_UINT64 },
746 { "l2_asize", KSTAT_DATA_UINT64 },
747 { "l2_hdr_size", KSTAT_DATA_UINT64 },
748 { "memory_throttle_count", KSTAT_DATA_UINT64 },
749 { "arc_meta_used", KSTAT_DATA_UINT64 },
750 { "arc_meta_limit", KSTAT_DATA_UINT64 },
751 { "arc_meta_max", KSTAT_DATA_UINT64 },
752 { "arc_meta_min", KSTAT_DATA_UINT64 },
753 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
754 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
755 };
756
757 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
758
759 #define ARCSTAT_INCR(stat, val) \
760 atomic_add_64(&arc_stats.stat.value.ui64, (val))
761
762 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
763 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
764
765 #define ARCSTAT_MAX(stat, val) { \
766 uint64_t m; \
767 while ((val) > (m = arc_stats.stat.value.ui64) && \
768 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
769 continue; \
770 }
771
772 #define ARCSTAT_MAXSTAT(stat) \
773 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
774
775 /*
776 * We define a macro to allow ARC hits/misses to be easily broken down by
777 * two separate conditions, giving a total of four different subtypes for
778 * each of hits and misses (so eight statistics total).
779 */
780 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
781 if (cond1) { \
782 if (cond2) { \
783 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
784 } else { \
785 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
786 } \
787 } else { \
788 if (cond2) { \
789 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
790 } else { \
791 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
792 } \
793 }
794
795 kstat_t *arc_ksp;
796 static arc_state_t *arc_anon;
797 static arc_state_t *arc_mru;
798 static arc_state_t *arc_mru_ghost;
799 static arc_state_t *arc_mfu;
800 static arc_state_t *arc_mfu_ghost;
801 static arc_state_t *arc_l2c_only;
802
803 /*
804 * There are several ARC variables that are critical to export as kstats --
805 * but we don't want to have to grovel around in the kstat whenever we wish to
806 * manipulate them. For these variables, we therefore define them to be in
807 * terms of the statistic variable. This assures that we are not introducing
808 * the possibility of inconsistency by having shadow copies of the variables,
809 * while still allowing the code to be readable.
810 */
811 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
812 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
813 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
814 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
815 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
816 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
817 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
818
819 /* compressed size of entire arc */
820 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
821 /* uncompressed size of entire arc */
822 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
823 /* number of bytes in the arc from arc_buf_t's */
824 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
825
826 /*
827 * There are also some ARC variables that we want to export, but that are
828 * updated so often that having the canonical representation be the statistic
829 * variable causes a performance bottleneck. We want to use aggsum_t's for these
830 * instead, but still be able to export the kstat in the same way as before.
831 * The solution is to always use the aggsum version, except in the kstat update
832 * callback.
833 */
834 aggsum_t arc_size;
835 aggsum_t arc_meta_used;
836 aggsum_t astat_data_size;
837 aggsum_t astat_metadata_size;
838 aggsum_t astat_hdr_size;
839 aggsum_t astat_other_size;
840 aggsum_t astat_l2_hdr_size;
841
842 static int arc_no_grow; /* Don't try to grow cache size */
843 static uint64_t arc_tempreserve;
844 static uint64_t arc_loaned_bytes;
845
846 typedef struct arc_callback arc_callback_t;
847
848 struct arc_callback {
849 void *acb_private;
850 arc_done_func_t *acb_done;
851 arc_buf_t *acb_buf;
852 boolean_t acb_compressed;
853 zio_t *acb_zio_dummy;
854 arc_callback_t *acb_next;
855 };
856
857 typedef struct arc_write_callback arc_write_callback_t;
858
859 struct arc_write_callback {
860 void *awcb_private;
861 arc_done_func_t *awcb_ready;
862 arc_done_func_t *awcb_children_ready;
863 arc_done_func_t *awcb_physdone;
864 arc_done_func_t *awcb_done;
865 arc_buf_t *awcb_buf;
866 };
867
868 /*
869 * ARC buffers are separated into multiple structs as a memory saving measure:
870 * - Common fields struct, always defined, and embedded within it:
871 * - L2-only fields, always allocated but undefined when not in L2ARC
872 * - L1-only fields, only allocated when in L1ARC
873 *
874 * Buffer in L1 Buffer only in L2
875 * +------------------------+ +------------------------+
876 * | arc_buf_hdr_t | | arc_buf_hdr_t |
877 * | | | |
878 * | | | |
879 * | | | |
880 * +------------------------+ +------------------------+
881 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
882 * | (undefined if L1-only) | | |
883 * +------------------------+ +------------------------+
884 * | l1arc_buf_hdr_t |
885 * | |
886 * | |
887 * | |
888 * | |
889 * +------------------------+
890 *
891 * Because it's possible for the L2ARC to become extremely large, we can wind
892 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
893 * is minimized by only allocating the fields necessary for an L1-cached buffer
894 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
895 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
896 * words in pointers. arc_hdr_realloc() is used to switch a header between
897 * these two allocation states.
898 */
899 typedef struct l1arc_buf_hdr {
900 kmutex_t b_freeze_lock;
901 zio_cksum_t *b_freeze_cksum;
902 #ifdef ZFS_DEBUG
903 /*
904 * Used for debugging with kmem_flags - by allocating and freeing
905 * b_thawed when the buffer is thawed, we get a record of the stack
906 * trace that thawed it.
907 */
908 void *b_thawed;
909 #endif
910
911 arc_buf_t *b_buf;
912 uint32_t b_bufcnt;
913 /* for waiting on writes to complete */
914 kcondvar_t b_cv;
915 uint8_t b_byteswap;
916
917 /* protected by arc state mutex */
918 arc_state_t *b_state;
919 multilist_node_t b_arc_node;
920
921 /* updated atomically */
922 clock_t b_arc_access;
923
924 /* self protecting */
925 refcount_t b_refcnt;
926
927 arc_callback_t *b_acb;
928 abd_t *b_pabd;
929 } l1arc_buf_hdr_t;
930
931 typedef struct l2arc_dev l2arc_dev_t;
932
933 typedef struct l2arc_buf_hdr {
934 /* protected by arc_buf_hdr mutex */
935 l2arc_dev_t *b_dev; /* L2ARC device */
936 uint64_t b_daddr; /* disk address, offset byte */
937
938 list_node_t b_l2node;
939 } l2arc_buf_hdr_t;
940
941 struct arc_buf_hdr {
942 /* protected by hash lock */
943 dva_t b_dva;
944 uint64_t b_birth;
945
946 arc_buf_contents_t b_type;
947 arc_buf_hdr_t *b_hash_next;
948 arc_flags_t b_flags;
949
950 /*
951 * This field stores the size of the data buffer after
952 * compression, and is set in the arc's zio completion handlers.
953 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
954 *
955 * While the block pointers can store up to 32MB in their psize
956 * field, we can only store up to 32MB minus 512B. This is due
957 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
958 * a field of zeros represents 512B in the bp). We can't use a
959 * bias of 1 since we need to reserve a psize of zero, here, to
960 * represent holes and embedded blocks.
961 *
962 * This isn't a problem in practice, since the maximum size of a
963 * buffer is limited to 16MB, so we never need to store 32MB in
964 * this field. Even in the upstream illumos code base, the
965 * maximum size of a buffer is limited to 16MB.
966 */
967 uint16_t b_psize;
968
969 /*
970 * This field stores the size of the data buffer before
971 * compression, and cannot change once set. It is in units
972 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
973 */
974 uint16_t b_lsize; /* immutable */
975 uint64_t b_spa; /* immutable */
976
977 /* L2ARC fields. Undefined when not in L2ARC. */
978 l2arc_buf_hdr_t b_l2hdr;
979 /* L1ARC fields. Undefined when in l2arc_only state */
980 l1arc_buf_hdr_t b_l1hdr;
981 };
982
983 #define GHOST_STATE(state) \
984 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
985 (state) == arc_l2c_only)
986
987 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
988 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
989 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
990 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
991 #define HDR_COMPRESSION_ENABLED(hdr) \
992 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
993
994 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
995 #define HDR_L2_READING(hdr) \
996 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
997 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
998 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
999 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1000 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1001 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1002
1003 #define HDR_ISTYPE_METADATA(hdr) \
1004 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1005 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1006
1007 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1008 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1009
1010 /* For storing compression mode in b_flags */
1011 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1012
1013 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1014 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1015 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1016 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1017
1018 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1019 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1020 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1021
1022 /*
1023 * Other sizes
1024 */
1025
1026 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1027 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1028
1029 /*
1030 * Hash table routines
1031 */
1032
1033 #define HT_LOCK_PAD 64
1034
1035 struct ht_lock {
1036 kmutex_t ht_lock;
1037 #ifdef _KERNEL
1038 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1039 #endif
1040 };
1041
1042 #define BUF_LOCKS 256
1043 typedef struct buf_hash_table {
1044 uint64_t ht_mask;
1045 arc_buf_hdr_t **ht_table;
1046 struct ht_lock ht_locks[BUF_LOCKS];
1047 } buf_hash_table_t;
1048
1049 static buf_hash_table_t buf_hash_table;
1050
1051 #define BUF_HASH_INDEX(spa, dva, birth) \
1052 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1053 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1054 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1055 #define HDR_LOCK(hdr) \
1056 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1057
1058 uint64_t zfs_crc64_table[256];
1059
1060 /*
1061 * Level 2 ARC
1062 */
1063
1064 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1065 #define L2ARC_HEADROOM 2 /* num of writes */
1066 /*
1067 * If we discover during ARC scan any buffers to be compressed, we boost
1068 * our headroom for the next scanning cycle by this percentage multiple.
1069 */
1070 #define L2ARC_HEADROOM_BOOST 200
1071 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1072 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1073
1074 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1075 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1076
1077 /* L2ARC Performance Tunables */
1078 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1079 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1080 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1081 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1082 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1083 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1084 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1085 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1086 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1087
1088 /*
1089 * L2ARC Internals
1090 */
1091 struct l2arc_dev {
1092 vdev_t *l2ad_vdev; /* vdev */
1093 spa_t *l2ad_spa; /* spa */
1094 uint64_t l2ad_hand; /* next write location */
1095 uint64_t l2ad_start; /* first addr on device */
1096 uint64_t l2ad_end; /* last addr on device */
1097 boolean_t l2ad_first; /* first sweep through */
1098 boolean_t l2ad_writing; /* currently writing */
1099 kmutex_t l2ad_mtx; /* lock for buffer list */
1100 list_t l2ad_buflist; /* buffer list */
1101 list_node_t l2ad_node; /* device list node */
1102 refcount_t l2ad_alloc; /* allocated bytes */
1103 };
1104
1105 static list_t L2ARC_dev_list; /* device list */
1106 static list_t *l2arc_dev_list; /* device list pointer */
1107 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1108 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1109 static list_t L2ARC_free_on_write; /* free after write buf list */
1110 static list_t *l2arc_free_on_write; /* free after write list ptr */
1111 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1112 static uint64_t l2arc_ndev; /* number of devices */
1113
1114 typedef struct l2arc_read_callback {
1115 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1116 blkptr_t l2rcb_bp; /* original blkptr */
1117 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1118 int l2rcb_flags; /* original flags */
1119 abd_t *l2rcb_abd; /* temporary buffer */
1120 } l2arc_read_callback_t;
1121
1122 typedef struct l2arc_write_callback {
1123 l2arc_dev_t *l2wcb_dev; /* device info */
1124 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1125 } l2arc_write_callback_t;
1126
1127 typedef struct l2arc_data_free {
1128 /* protected by l2arc_free_on_write_mtx */
1129 abd_t *l2df_abd;
1130 size_t l2df_size;
1131 arc_buf_contents_t l2df_type;
1132 list_node_t l2df_list_node;
1133 } l2arc_data_free_t;
1134
1135 static kmutex_t l2arc_feed_thr_lock;
1136 static kcondvar_t l2arc_feed_thr_cv;
1137 static uint8_t l2arc_thread_exit;
1138
1139 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1140 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1141 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1142 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1143 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1144 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1145 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1146 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1147 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1148 static boolean_t arc_is_overflowing();
1149 static void arc_buf_watch(arc_buf_t *);
1150
1151 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1152 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1153 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1154 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1155
1156 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1157 static void l2arc_read_done(zio_t *);
1158
1159
1160 /*
1161 * We use Cityhash for this. It's fast, and has good hash properties without
1162 * requiring any large static buffers.
1163 */
1164 static uint64_t
1165 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1166 {
1167 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1168 }
1169
1170 #define HDR_EMPTY(hdr) \
1171 ((hdr)->b_dva.dva_word[0] == 0 && \
1172 (hdr)->b_dva.dva_word[1] == 0)
1173
1174 #define HDR_EQUAL(spa, dva, birth, hdr) \
1175 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1176 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1177 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1178
1179 static void
1180 buf_discard_identity(arc_buf_hdr_t *hdr)
1181 {
1182 hdr->b_dva.dva_word[0] = 0;
1183 hdr->b_dva.dva_word[1] = 0;
1184 hdr->b_birth = 0;
1185 }
1186
1187 static arc_buf_hdr_t *
1188 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1189 {
1190 const dva_t *dva = BP_IDENTITY(bp);
1191 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1192 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1193 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1194 arc_buf_hdr_t *hdr;
1195
1196 mutex_enter(hash_lock);
1197 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1198 hdr = hdr->b_hash_next) {
1199 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1200 *lockp = hash_lock;
1201 return (hdr);
1202 }
1203 }
1204 mutex_exit(hash_lock);
1205 *lockp = NULL;
1206 return (NULL);
1207 }
1208
1209 /*
1210 * Insert an entry into the hash table. If there is already an element
1211 * equal to elem in the hash table, then the already existing element
1212 * will be returned and the new element will not be inserted.
1213 * Otherwise returns NULL.
1214 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1215 */
1216 static arc_buf_hdr_t *
1217 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1218 {
1219 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1220 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1221 arc_buf_hdr_t *fhdr;
1222 uint32_t i;
1223
1224 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1225 ASSERT(hdr->b_birth != 0);
1226 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1227
1228 if (lockp != NULL) {
1229 *lockp = hash_lock;
1230 mutex_enter(hash_lock);
1231 } else {
1232 ASSERT(MUTEX_HELD(hash_lock));
1233 }
1234
1235 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1236 fhdr = fhdr->b_hash_next, i++) {
1237 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1238 return (fhdr);
1239 }
1240
1241 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1242 buf_hash_table.ht_table[idx] = hdr;
1243 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1244
1245 /* collect some hash table performance data */
1246 if (i > 0) {
1247 ARCSTAT_BUMP(arcstat_hash_collisions);
1248 if (i == 1)
1249 ARCSTAT_BUMP(arcstat_hash_chains);
1250
1251 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1252 }
1253
1254 ARCSTAT_BUMP(arcstat_hash_elements);
1255 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1256
1257 return (NULL);
1258 }
1259
1260 static void
1261 buf_hash_remove(arc_buf_hdr_t *hdr)
1262 {
1263 arc_buf_hdr_t *fhdr, **hdrp;
1264 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1265
1266 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1267 ASSERT(HDR_IN_HASH_TABLE(hdr));
1268
1269 hdrp = &buf_hash_table.ht_table[idx];
1270 while ((fhdr = *hdrp) != hdr) {
1271 ASSERT3P(fhdr, !=, NULL);
1272 hdrp = &fhdr->b_hash_next;
1273 }
1274 *hdrp = hdr->b_hash_next;
1275 hdr->b_hash_next = NULL;
1276 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1277
1278 /* collect some hash table performance data */
1279 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1280
1281 if (buf_hash_table.ht_table[idx] &&
1282 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1283 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1284 }
1285
1286 /*
1287 * Global data structures and functions for the buf kmem cache.
1288 */
1289 static kmem_cache_t *hdr_full_cache;
1290 static kmem_cache_t *hdr_l2only_cache;
1291 static kmem_cache_t *buf_cache;
1292
1293 static void
1294 buf_fini(void)
1295 {
1296 int i;
1297
1298 kmem_free(buf_hash_table.ht_table,
1299 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1300 for (i = 0; i < BUF_LOCKS; i++)
1301 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1302 kmem_cache_destroy(hdr_full_cache);
1303 kmem_cache_destroy(hdr_l2only_cache);
1304 kmem_cache_destroy(buf_cache);
1305 }
1306
1307 /*
1308 * Constructor callback - called when the cache is empty
1309 * and a new buf is requested.
1310 */
1311 /* ARGSUSED */
1312 static int
1313 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1314 {
1315 arc_buf_hdr_t *hdr = vbuf;
1316
1317 bzero(hdr, HDR_FULL_SIZE);
1318 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1319 refcount_create(&hdr->b_l1hdr.b_refcnt);
1320 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1321 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1322 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1323
1324 return (0);
1325 }
1326
1327 /* ARGSUSED */
1328 static int
1329 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1330 {
1331 arc_buf_hdr_t *hdr = vbuf;
1332
1333 bzero(hdr, HDR_L2ONLY_SIZE);
1334 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1335
1336 return (0);
1337 }
1338
1339 /* ARGSUSED */
1340 static int
1341 buf_cons(void *vbuf, void *unused, int kmflag)
1342 {
1343 arc_buf_t *buf = vbuf;
1344
1345 bzero(buf, sizeof (arc_buf_t));
1346 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1347 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1348
1349 return (0);
1350 }
1351
1352 /*
1353 * Destructor callback - called when a cached buf is
1354 * no longer required.
1355 */
1356 /* ARGSUSED */
1357 static void
1358 hdr_full_dest(void *vbuf, void *unused)
1359 {
1360 arc_buf_hdr_t *hdr = vbuf;
1361
1362 ASSERT(HDR_EMPTY(hdr));
1363 cv_destroy(&hdr->b_l1hdr.b_cv);
1364 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1365 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1366 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1367 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1368 }
1369
1370 /* ARGSUSED */
1371 static void
1372 hdr_l2only_dest(void *vbuf, void *unused)
1373 {
1374 arc_buf_hdr_t *hdr = vbuf;
1375
1376 ASSERT(HDR_EMPTY(hdr));
1377 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1378 }
1379
1380 /* ARGSUSED */
1381 static void
1382 buf_dest(void *vbuf, void *unused)
1383 {
1384 arc_buf_t *buf = vbuf;
1385
1386 mutex_destroy(&buf->b_evict_lock);
1387 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1388 }
1389
1390 /*
1391 * Reclaim callback -- invoked when memory is low.
1392 */
1393 /* ARGSUSED */
1394 static void
1395 hdr_recl(void *unused)
1396 {
1397 dprintf("hdr_recl called\n");
1398 /*
1399 * umem calls the reclaim func when we destroy the buf cache,
1400 * which is after we do arc_fini().
1401 */
1402 if (!arc_dead)
1403 cv_signal(&arc_reclaim_thread_cv);
1404 }
1405
1406 static void
1407 buf_init(void)
1408 {
1409 uint64_t *ct;
1410 uint64_t hsize = 1ULL << 12;
1411 int i, j;
1412
1413 /*
1414 * The hash table is big enough to fill all of physical memory
1415 * with an average block size of zfs_arc_average_blocksize (default 8K).
1416 * By default, the table will take up
1417 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1418 */
1419 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1420 hsize <<= 1;
1421 retry:
1422 buf_hash_table.ht_mask = hsize - 1;
1423 buf_hash_table.ht_table =
1424 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1425 if (buf_hash_table.ht_table == NULL) {
1426 ASSERT(hsize > (1ULL << 8));
1427 hsize >>= 1;
1428 goto retry;
1429 }
1430
1431 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1432 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1433 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1434 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1435 NULL, NULL, 0);
1436 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1437 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1438
1439 for (i = 0; i < 256; i++)
1440 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1441 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1442
1443 for (i = 0; i < BUF_LOCKS; i++) {
1444 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1445 NULL, MUTEX_DEFAULT, NULL);
1446 }
1447 }
1448
1449 /*
1450 * This is the size that the buf occupies in memory. If the buf is compressed,
1451 * it will correspond to the compressed size. You should use this method of
1452 * getting the buf size unless you explicitly need the logical size.
1453 */
1454 int32_t
1455 arc_buf_size(arc_buf_t *buf)
1456 {
1457 return (ARC_BUF_COMPRESSED(buf) ?
1458 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1459 }
1460
1461 int32_t
1462 arc_buf_lsize(arc_buf_t *buf)
1463 {
1464 return (HDR_GET_LSIZE(buf->b_hdr));
1465 }
1466
1467 enum zio_compress
1468 arc_get_compression(arc_buf_t *buf)
1469 {
1470 return (ARC_BUF_COMPRESSED(buf) ?
1471 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1472 }
1473
1474 #define ARC_MINTIME (hz>>4) /* 62 ms */
1475
1476 static inline boolean_t
1477 arc_buf_is_shared(arc_buf_t *buf)
1478 {
1479 boolean_t shared = (buf->b_data != NULL &&
1480 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1481 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1482 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1483 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1484 IMPLY(shared, ARC_BUF_SHARED(buf));
1485 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1486
1487 /*
1488 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1489 * already being shared" requirement prevents us from doing that.
1490 */
1491
1492 return (shared);
1493 }
1494
1495 /*
1496 * Free the checksum associated with this header. If there is no checksum, this
1497 * is a no-op.
1498 */
1499 static inline void
1500 arc_cksum_free(arc_buf_hdr_t *hdr)
1501 {
1502 ASSERT(HDR_HAS_L1HDR(hdr));
1503 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1504 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1505 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1506 hdr->b_l1hdr.b_freeze_cksum = NULL;
1507 }
1508 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1509 }
1510
1511 /*
1512 * Return true iff at least one of the bufs on hdr is not compressed.
1513 */
1514 static boolean_t
1515 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1516 {
1517 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1518 if (!ARC_BUF_COMPRESSED(b)) {
1519 return (B_TRUE);
1520 }
1521 }
1522 return (B_FALSE);
1523 }
1524
1525 /*
1526 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1527 * matches the checksum that is stored in the hdr. If there is no checksum,
1528 * or if the buf is compressed, this is a no-op.
1529 */
1530 static void
1531 arc_cksum_verify(arc_buf_t *buf)
1532 {
1533 arc_buf_hdr_t *hdr = buf->b_hdr;
1534 zio_cksum_t zc;
1535
1536 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1537 return;
1538
1539 if (ARC_BUF_COMPRESSED(buf)) {
1540 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1541 arc_hdr_has_uncompressed_buf(hdr));
1542 return;
1543 }
1544
1545 ASSERT(HDR_HAS_L1HDR(hdr));
1546
1547 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1548 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1549 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1550 return;
1551 }
1552
1553 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1554 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1555 panic("buffer modified while frozen!");
1556 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1557 }
1558
1559 static boolean_t
1560 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1561 {
1562 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1563 boolean_t valid_cksum;
1564
1565 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1566 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1567
1568 /*
1569 * We rely on the blkptr's checksum to determine if the block
1570 * is valid or not. When compressed arc is enabled, the l2arc
1571 * writes the block to the l2arc just as it appears in the pool.
1572 * This allows us to use the blkptr's checksum to validate the
1573 * data that we just read off of the l2arc without having to store
1574 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1575 * arc is disabled, then the data written to the l2arc is always
1576 * uncompressed and won't match the block as it exists in the main
1577 * pool. When this is the case, we must first compress it if it is
1578 * compressed on the main pool before we can validate the checksum.
1579 */
1580 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1581 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1582 uint64_t lsize = HDR_GET_LSIZE(hdr);
1583 uint64_t csize;
1584
1585 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1586 csize = zio_compress_data(compress, zio->io_abd,
1587 abd_to_buf(cdata), lsize);
1588
1589 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1590 if (csize < HDR_GET_PSIZE(hdr)) {
1591 /*
1592 * Compressed blocks are always a multiple of the
1593 * smallest ashift in the pool. Ideally, we would
1594 * like to round up the csize to the next
1595 * spa_min_ashift but that value may have changed
1596 * since the block was last written. Instead,
1597 * we rely on the fact that the hdr's psize
1598 * was set to the psize of the block when it was
1599 * last written. We set the csize to that value
1600 * and zero out any part that should not contain
1601 * data.
1602 */
1603 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
1604 csize = HDR_GET_PSIZE(hdr);
1605 }
1606 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
1607 }
1608
1609 /*
1610 * Block pointers always store the checksum for the logical data.
1611 * If the block pointer has the gang bit set, then the checksum
1612 * it represents is for the reconstituted data and not for an
1613 * individual gang member. The zio pipeline, however, must be able to
1614 * determine the checksum of each of the gang constituents so it
1615 * treats the checksum comparison differently than what we need
1616 * for l2arc blocks. This prevents us from using the
1617 * zio_checksum_error() interface directly. Instead we must call the
1618 * zio_checksum_error_impl() so that we can ensure the checksum is
1619 * generated using the correct checksum algorithm and accounts for the
1620 * logical I/O size and not just a gang fragment.
1621 */
1622 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1623 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1624 zio->io_offset, NULL) == 0);
1625 zio_pop_transforms(zio);
1626 return (valid_cksum);
1627 }
1628
1629 /*
1630 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1631 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1632 * isn't modified later on. If buf is compressed or there is already a checksum
1633 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1634 */
1635 static void
1636 arc_cksum_compute(arc_buf_t *buf)
1637 {
1638 arc_buf_hdr_t *hdr = buf->b_hdr;
1639
1640 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1641 return;
1642
1643 ASSERT(HDR_HAS_L1HDR(hdr));
1644
1645 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1646 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1647 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
1648 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1649 return;
1650 } else if (ARC_BUF_COMPRESSED(buf)) {
1651 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1652 return;
1653 }
1654
1655 ASSERT(!ARC_BUF_COMPRESSED(buf));
1656 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1657 KM_SLEEP);
1658 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1659 hdr->b_l1hdr.b_freeze_cksum);
1660 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1661 arc_buf_watch(buf);
1662 }
1663
1664 #ifndef _KERNEL
1665 typedef struct procctl {
1666 long cmd;
1667 prwatch_t prwatch;
1668 } procctl_t;
1669 #endif
1670
1671 /* ARGSUSED */
1672 static void
1673 arc_buf_unwatch(arc_buf_t *buf)
1674 {
1675 #ifndef _KERNEL
1676 if (arc_watch) {
1677 int result;
1678 procctl_t ctl;
1679 ctl.cmd = PCWATCH;
1680 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1681 ctl.prwatch.pr_size = 0;
1682 ctl.prwatch.pr_wflags = 0;
1683 result = write(arc_procfd, &ctl, sizeof (ctl));
1684 ASSERT3U(result, ==, sizeof (ctl));
1685 }
1686 #endif
1687 }
1688
1689 /* ARGSUSED */
1690 static void
1691 arc_buf_watch(arc_buf_t *buf)
1692 {
1693 #ifndef _KERNEL
1694 if (arc_watch) {
1695 int result;
1696 procctl_t ctl;
1697 ctl.cmd = PCWATCH;
1698 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1699 ctl.prwatch.pr_size = arc_buf_size(buf);
1700 ctl.prwatch.pr_wflags = WA_WRITE;
1701 result = write(arc_procfd, &ctl, sizeof (ctl));
1702 ASSERT3U(result, ==, sizeof (ctl));
1703 }
1704 #endif
1705 }
1706
1707 static arc_buf_contents_t
1708 arc_buf_type(arc_buf_hdr_t *hdr)
1709 {
1710 arc_buf_contents_t type;
1711 if (HDR_ISTYPE_METADATA(hdr)) {
1712 type = ARC_BUFC_METADATA;
1713 } else {
1714 type = ARC_BUFC_DATA;
1715 }
1716 VERIFY3U(hdr->b_type, ==, type);
1717 return (type);
1718 }
1719
1720 boolean_t
1721 arc_is_metadata(arc_buf_t *buf)
1722 {
1723 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1724 }
1725
1726 static uint32_t
1727 arc_bufc_to_flags(arc_buf_contents_t type)
1728 {
1729 switch (type) {
1730 case ARC_BUFC_DATA:
1731 /* metadata field is 0 if buffer contains normal data */
1732 return (0);
1733 case ARC_BUFC_METADATA:
1734 return (ARC_FLAG_BUFC_METADATA);
1735 default:
1736 break;
1737 }
1738 panic("undefined ARC buffer type!");
1739 return ((uint32_t)-1);
1740 }
1741
1742 void
1743 arc_buf_thaw(arc_buf_t *buf)
1744 {
1745 arc_buf_hdr_t *hdr = buf->b_hdr;
1746
1747 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1748 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1749
1750 arc_cksum_verify(buf);
1751
1752 /*
1753 * Compressed buffers do not manipulate the b_freeze_cksum or
1754 * allocate b_thawed.
1755 */
1756 if (ARC_BUF_COMPRESSED(buf)) {
1757 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1758 arc_hdr_has_uncompressed_buf(hdr));
1759 return;
1760 }
1761
1762 ASSERT(HDR_HAS_L1HDR(hdr));
1763 arc_cksum_free(hdr);
1764
1765 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1766 #ifdef ZFS_DEBUG
1767 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1768 if (hdr->b_l1hdr.b_thawed != NULL)
1769 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1770 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1771 }
1772 #endif
1773
1774 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1775
1776 arc_buf_unwatch(buf);
1777 }
1778
1779 void
1780 arc_buf_freeze(arc_buf_t *buf)
1781 {
1782 arc_buf_hdr_t *hdr = buf->b_hdr;
1783 kmutex_t *hash_lock;
1784
1785 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1786 return;
1787
1788 if (ARC_BUF_COMPRESSED(buf)) {
1789 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1790 arc_hdr_has_uncompressed_buf(hdr));
1791 return;
1792 }
1793
1794 hash_lock = HDR_LOCK(hdr);
1795 mutex_enter(hash_lock);
1796
1797 ASSERT(HDR_HAS_L1HDR(hdr));
1798 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
1799 hdr->b_l1hdr.b_state == arc_anon);
1800 arc_cksum_compute(buf);
1801 mutex_exit(hash_lock);
1802 }
1803
1804 /*
1805 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1806 * the following functions should be used to ensure that the flags are
1807 * updated in a thread-safe way. When manipulating the flags either
1808 * the hash_lock must be held or the hdr must be undiscoverable. This
1809 * ensures that we're not racing with any other threads when updating
1810 * the flags.
1811 */
1812 static inline void
1813 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1814 {
1815 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1816 hdr->b_flags |= flags;
1817 }
1818
1819 static inline void
1820 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1821 {
1822 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1823 hdr->b_flags &= ~flags;
1824 }
1825
1826 /*
1827 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1828 * done in a special way since we have to clear and set bits
1829 * at the same time. Consumers that wish to set the compression bits
1830 * must use this function to ensure that the flags are updated in
1831 * thread-safe manner.
1832 */
1833 static void
1834 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1835 {
1836 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1837
1838 /*
1839 * Holes and embedded blocks will always have a psize = 0 so
1840 * we ignore the compression of the blkptr and set the
1841 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
1842 * Holes and embedded blocks remain anonymous so we don't
1843 * want to uncompress them. Mark them as uncompressed.
1844 */
1845 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1846 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1847 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
1848 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1849 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1850 } else {
1851 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1852 HDR_SET_COMPRESS(hdr, cmp);
1853 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1854 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1855 }
1856 }
1857
1858 /*
1859 * Looks for another buf on the same hdr which has the data decompressed, copies
1860 * from it, and returns true. If no such buf exists, returns false.
1861 */
1862 static boolean_t
1863 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1864 {
1865 arc_buf_hdr_t *hdr = buf->b_hdr;
1866 boolean_t copied = B_FALSE;
1867
1868 ASSERT(HDR_HAS_L1HDR(hdr));
1869 ASSERT3P(buf->b_data, !=, NULL);
1870 ASSERT(!ARC_BUF_COMPRESSED(buf));
1871
1872 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1873 from = from->b_next) {
1874 /* can't use our own data buffer */
1875 if (from == buf) {
1876 continue;
1877 }
1878
1879 if (!ARC_BUF_COMPRESSED(from)) {
1880 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1881 copied = B_TRUE;
1882 break;
1883 }
1884 }
1885
1886 /*
1887 * There were no decompressed bufs, so there should not be a
1888 * checksum on the hdr either.
1889 */
1890 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1891
1892 return (copied);
1893 }
1894
1895 /*
1896 * Given a buf that has a data buffer attached to it, this function will
1897 * efficiently fill the buf with data of the specified compression setting from
1898 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1899 * are already sharing a data buf, no copy is performed.
1900 *
1901 * If the buf is marked as compressed but uncompressed data was requested, this
1902 * will allocate a new data buffer for the buf, remove that flag, and fill the
1903 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1904 * uncompressed data, and (since we haven't added support for it yet) if you
1905 * want compressed data your buf must already be marked as compressed and have
1906 * the correct-sized data buffer.
1907 */
1908 static int
1909 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
1910 {
1911 arc_buf_hdr_t *hdr = buf->b_hdr;
1912 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
1913 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
1914
1915 ASSERT3P(buf->b_data, !=, NULL);
1916 IMPLY(compressed, hdr_compressed);
1917 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
1918
1919 if (hdr_compressed == compressed) {
1920 if (!arc_buf_is_shared(buf)) {
1921 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
1922 arc_buf_size(buf));
1923 }
1924 } else {
1925 ASSERT(hdr_compressed);
1926 ASSERT(!compressed);
1927 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
1928
1929 /*
1930 * If the buf is sharing its data with the hdr, unlink it and
1931 * allocate a new data buffer for the buf.
1932 */
1933 if (arc_buf_is_shared(buf)) {
1934 ASSERT(ARC_BUF_COMPRESSED(buf));
1935
1936 /* We need to give the buf it's own b_data */
1937 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
1938 buf->b_data =
1939 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
1940 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
1941
1942 /* Previously overhead was 0; just add new overhead */
1943 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
1944 } else if (ARC_BUF_COMPRESSED(buf)) {
1945 /* We need to reallocate the buf's b_data */
1946 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
1947 buf);
1948 buf->b_data =
1949 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
1950
1951 /* We increased the size of b_data; update overhead */
1952 ARCSTAT_INCR(arcstat_overhead_size,
1953 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
1954 }
1955
1956 /*
1957 * Regardless of the buf's previous compression settings, it
1958 * should not be compressed at the end of this function.
1959 */
1960 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1961
1962 /*
1963 * Try copying the data from another buf which already has a
1964 * decompressed version. If that's not possible, it's time to
1965 * bite the bullet and decompress the data from the hdr.
1966 */
1967 if (arc_buf_try_copy_decompressed_data(buf)) {
1968 /* Skip byteswapping and checksumming (already done) */
1969 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
1970 return (0);
1971 } else {
1972 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1973 hdr->b_l1hdr.b_pabd, buf->b_data,
1974 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
1975
1976 /*
1977 * Absent hardware errors or software bugs, this should
1978 * be impossible, but log it anyway so we can debug it.
1979 */
1980 if (error != 0) {
1981 zfs_dbgmsg(
1982 "hdr %p, compress %d, psize %d, lsize %d",
1983 hdr, HDR_GET_COMPRESS(hdr),
1984 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
1985 return (SET_ERROR(EIO));
1986 }
1987 }
1988 }
1989
1990 /* Byteswap the buf's data if necessary */
1991 if (bswap != DMU_BSWAP_NUMFUNCS) {
1992 ASSERT(!HDR_SHARED_DATA(hdr));
1993 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
1994 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
1995 }
1996
1997 /* Compute the hdr's checksum if necessary */
1998 arc_cksum_compute(buf);
1999
2000 return (0);
2001 }
2002
2003 int
2004 arc_decompress(arc_buf_t *buf)
2005 {
2006 return (arc_buf_fill(buf, B_FALSE));
2007 }
2008
2009 /*
2010 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
2011 */
2012 static uint64_t
2013 arc_hdr_size(arc_buf_hdr_t *hdr)
2014 {
2015 uint64_t size;
2016
2017 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
2018 HDR_GET_PSIZE(hdr) > 0) {
2019 size = HDR_GET_PSIZE(hdr);
2020 } else {
2021 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
2022 size = HDR_GET_LSIZE(hdr);
2023 }
2024 return (size);
2025 }
2026
2027 /*
2028 * Increment the amount of evictable space in the arc_state_t's refcount.
2029 * We account for the space used by the hdr and the arc buf individually
2030 * so that we can add and remove them from the refcount individually.
2031 */
2032 static void
2033 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2034 {
2035 arc_buf_contents_t type = arc_buf_type(hdr);
2036
2037 ASSERT(HDR_HAS_L1HDR(hdr));
2038
2039 if (GHOST_STATE(state)) {
2040 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2041 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2042 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2043 (void) refcount_add_many(&state->arcs_esize[type],
2044 HDR_GET_LSIZE(hdr), hdr);
2045 return;
2046 }
2047
2048 ASSERT(!GHOST_STATE(state));
2049 if (hdr->b_l1hdr.b_pabd != NULL) {
2050 (void) refcount_add_many(&state->arcs_esize[type],
2051 arc_hdr_size(hdr), hdr);
2052 }
2053 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2054 buf = buf->b_next) {
2055 if (arc_buf_is_shared(buf))
2056 continue;
2057 (void) refcount_add_many(&state->arcs_esize[type],
2058 arc_buf_size(buf), buf);
2059 }
2060 }
2061
2062 /*
2063 * Decrement the amount of evictable space in the arc_state_t's refcount.
2064 * We account for the space used by the hdr and the arc buf individually
2065 * so that we can add and remove them from the refcount individually.
2066 */
2067 static void
2068 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2069 {
2070 arc_buf_contents_t type = arc_buf_type(hdr);
2071
2072 ASSERT(HDR_HAS_L1HDR(hdr));
2073
2074 if (GHOST_STATE(state)) {
2075 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2076 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2077 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2078 (void) refcount_remove_many(&state->arcs_esize[type],
2079 HDR_GET_LSIZE(hdr), hdr);
2080 return;
2081 }
2082
2083 ASSERT(!GHOST_STATE(state));
2084 if (hdr->b_l1hdr.b_pabd != NULL) {
2085 (void) refcount_remove_many(&state->arcs_esize[type],
2086 arc_hdr_size(hdr), hdr);
2087 }
2088 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2089 buf = buf->b_next) {
2090 if (arc_buf_is_shared(buf))
2091 continue;
2092 (void) refcount_remove_many(&state->arcs_esize[type],
2093 arc_buf_size(buf), buf);
2094 }
2095 }
2096
2097 /*
2098 * Add a reference to this hdr indicating that someone is actively
2099 * referencing that memory. When the refcount transitions from 0 to 1,
2100 * we remove it from the respective arc_state_t list to indicate that
2101 * it is not evictable.
2102 */
2103 static void
2104 add_reference(arc_buf_hdr_t *hdr, void *tag)
2105 {
2106 ASSERT(HDR_HAS_L1HDR(hdr));
2107 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2108 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2109 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2110 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2111 }
2112
2113 arc_state_t *state = hdr->b_l1hdr.b_state;
2114
2115 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2116 (state != arc_anon)) {
2117 /* We don't use the L2-only state list. */
2118 if (state != arc_l2c_only) {
2119 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2120 hdr);
2121 arc_evictable_space_decrement(hdr, state);
2122 }
2123 /* remove the prefetch flag if we get a reference */
2124 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2125 }
2126 }
2127
2128 /*
2129 * Remove a reference from this hdr. When the reference transitions from
2130 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2131 * list making it eligible for eviction.
2132 */
2133 static int
2134 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2135 {
2136 int cnt;
2137 arc_state_t *state = hdr->b_l1hdr.b_state;
2138
2139 ASSERT(HDR_HAS_L1HDR(hdr));
2140 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2141 ASSERT(!GHOST_STATE(state));
2142
2143 /*
2144 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2145 * check to prevent usage of the arc_l2c_only list.
2146 */
2147 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2148 (state != arc_anon)) {
2149 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2150 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2151 arc_evictable_space_increment(hdr, state);
2152 }
2153 return (cnt);
2154 }
2155
2156 /*
2157 * Move the supplied buffer to the indicated state. The hash lock
2158 * for the buffer must be held by the caller.
2159 */
2160 static void
2161 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2162 kmutex_t *hash_lock)
2163 {
2164 arc_state_t *old_state;
2165 int64_t refcnt;
2166 uint32_t bufcnt;
2167 boolean_t update_old, update_new;
2168 arc_buf_contents_t buftype = arc_buf_type(hdr);
2169
2170 /*
2171 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2172 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2173 * L1 hdr doesn't always exist when we change state to arc_anon before
2174 * destroying a header, in which case reallocating to add the L1 hdr is
2175 * pointless.
2176 */
2177 if (HDR_HAS_L1HDR(hdr)) {
2178 old_state = hdr->b_l1hdr.b_state;
2179 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2180 bufcnt = hdr->b_l1hdr.b_bufcnt;
2181 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2182 } else {
2183 old_state = arc_l2c_only;
2184 refcnt = 0;
2185 bufcnt = 0;
2186 update_old = B_FALSE;
2187 }
2188 update_new = update_old;
2189
2190 ASSERT(MUTEX_HELD(hash_lock));
2191 ASSERT3P(new_state, !=, old_state);
2192 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2193 ASSERT(old_state != arc_anon || bufcnt <= 1);
2194
2195 /*
2196 * If this buffer is evictable, transfer it from the
2197 * old state list to the new state list.
2198 */
2199 if (refcnt == 0) {
2200 if (old_state != arc_anon && old_state != arc_l2c_only) {
2201 ASSERT(HDR_HAS_L1HDR(hdr));
2202 multilist_remove(old_state->arcs_list[buftype], hdr);
2203
2204 if (GHOST_STATE(old_state)) {
2205 ASSERT0(bufcnt);
2206 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2207 update_old = B_TRUE;
2208 }
2209 arc_evictable_space_decrement(hdr, old_state);
2210 }
2211 if (new_state != arc_anon && new_state != arc_l2c_only) {
2212
2213 /*
2214 * An L1 header always exists here, since if we're
2215 * moving to some L1-cached state (i.e. not l2c_only or
2216 * anonymous), we realloc the header to add an L1hdr
2217 * beforehand.
2218 */
2219 ASSERT(HDR_HAS_L1HDR(hdr));
2220 multilist_insert(new_state->arcs_list[buftype], hdr);
2221
2222 if (GHOST_STATE(new_state)) {
2223 ASSERT0(bufcnt);
2224 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2225 update_new = B_TRUE;
2226 }
2227 arc_evictable_space_increment(hdr, new_state);
2228 }
2229 }
2230
2231 ASSERT(!HDR_EMPTY(hdr));
2232 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2233 buf_hash_remove(hdr);
2234
2235 /* adjust state sizes (ignore arc_l2c_only) */
2236
2237 if (update_new && new_state != arc_l2c_only) {
2238 ASSERT(HDR_HAS_L1HDR(hdr));
2239 if (GHOST_STATE(new_state)) {
2240 ASSERT0(bufcnt);
2241
2242 /*
2243 * When moving a header to a ghost state, we first
2244 * remove all arc buffers. Thus, we'll have a
2245 * bufcnt of zero, and no arc buffer to use for
2246 * the reference. As a result, we use the arc
2247 * header pointer for the reference.
2248 */
2249 (void) refcount_add_many(&new_state->arcs_size,
2250 HDR_GET_LSIZE(hdr), hdr);
2251 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2252 } else {
2253 uint32_t buffers = 0;
2254
2255 /*
2256 * Each individual buffer holds a unique reference,
2257 * thus we must remove each of these references one
2258 * at a time.
2259 */
2260 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2261 buf = buf->b_next) {
2262 ASSERT3U(bufcnt, !=, 0);
2263 buffers++;
2264
2265 /*
2266 * When the arc_buf_t is sharing the data
2267 * block with the hdr, the owner of the
2268 * reference belongs to the hdr. Only
2269 * add to the refcount if the arc_buf_t is
2270 * not shared.
2271 */
2272 if (arc_buf_is_shared(buf))
2273 continue;
2274
2275 (void) refcount_add_many(&new_state->arcs_size,
2276 arc_buf_size(buf), buf);
2277 }
2278 ASSERT3U(bufcnt, ==, buffers);
2279
2280 if (hdr->b_l1hdr.b_pabd != NULL) {
2281 (void) refcount_add_many(&new_state->arcs_size,
2282 arc_hdr_size(hdr), hdr);
2283 } else {
2284 ASSERT(GHOST_STATE(old_state));
2285 }
2286 }
2287 }
2288
2289 if (update_old && old_state != arc_l2c_only) {
2290 ASSERT(HDR_HAS_L1HDR(hdr));
2291 if (GHOST_STATE(old_state)) {
2292 ASSERT0(bufcnt);
2293 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2294
2295 /*
2296 * When moving a header off of a ghost state,
2297 * the header will not contain any arc buffers.
2298 * We use the arc header pointer for the reference
2299 * which is exactly what we did when we put the
2300 * header on the ghost state.
2301 */
2302
2303 (void) refcount_remove_many(&old_state->arcs_size,
2304 HDR_GET_LSIZE(hdr), hdr);
2305 } else {
2306 uint32_t buffers = 0;
2307
2308 /*
2309 * Each individual buffer holds a unique reference,
2310 * thus we must remove each of these references one
2311 * at a time.
2312 */
2313 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2314 buf = buf->b_next) {
2315 ASSERT3U(bufcnt, !=, 0);
2316 buffers++;
2317
2318 /*
2319 * When the arc_buf_t is sharing the data
2320 * block with the hdr, the owner of the
2321 * reference belongs to the hdr. Only
2322 * add to the refcount if the arc_buf_t is
2323 * not shared.
2324 */
2325 if (arc_buf_is_shared(buf))
2326 continue;
2327
2328 (void) refcount_remove_many(
2329 &old_state->arcs_size, arc_buf_size(buf),
2330 buf);
2331 }
2332 ASSERT3U(bufcnt, ==, buffers);
2333 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2334 (void) refcount_remove_many(
2335 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2336 }
2337 }
2338
2339 if (HDR_HAS_L1HDR(hdr))
2340 hdr->b_l1hdr.b_state = new_state;
2341
2342 /*
2343 * L2 headers should never be on the L2 state list since they don't
2344 * have L1 headers allocated.
2345 */
2346 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2347 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2348 }
2349
2350 void
2351 arc_space_consume(uint64_t space, arc_space_type_t type)
2352 {
2353 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2354
2355 switch (type) {
2356 case ARC_SPACE_DATA:
2357 aggsum_add(&astat_data_size, space);
2358 break;
2359 case ARC_SPACE_META:
2360 aggsum_add(&astat_metadata_size, space);
2361 break;
2362 case ARC_SPACE_OTHER:
2363 aggsum_add(&astat_other_size, space);
2364 break;
2365 case ARC_SPACE_HDRS:
2366 aggsum_add(&astat_hdr_size, space);
2367 break;
2368 case ARC_SPACE_L2HDRS:
2369 aggsum_add(&astat_l2_hdr_size, space);
2370 break;
2371 }
2372
2373 if (type != ARC_SPACE_DATA)
2374 aggsum_add(&arc_meta_used, space);
2375
2376 aggsum_add(&arc_size, space);
2377 }
2378
2379 void
2380 arc_space_return(uint64_t space, arc_space_type_t type)
2381 {
2382 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2383
2384 switch (type) {
2385 case ARC_SPACE_DATA:
2386 aggsum_add(&astat_data_size, -space);
2387 break;
2388 case ARC_SPACE_META:
2389 aggsum_add(&astat_metadata_size, -space);
2390 break;
2391 case ARC_SPACE_OTHER:
2392 aggsum_add(&astat_other_size, -space);
2393 break;
2394 case ARC_SPACE_HDRS:
2395 aggsum_add(&astat_hdr_size, -space);
2396 break;
2397 case ARC_SPACE_L2HDRS:
2398 aggsum_add(&astat_l2_hdr_size, -space);
2399 break;
2400 }
2401
2402 if (type != ARC_SPACE_DATA) {
2403 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2404 /*
2405 * We use the upper bound here rather than the precise value
2406 * because the arc_meta_max value doesn't need to be
2407 * precise. It's only consumed by humans via arcstats.
2408 */
2409 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2410 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2411 aggsum_add(&arc_meta_used, -space);
2412 }
2413
2414 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2415 aggsum_add(&arc_size, -space);
2416 }
2417
2418 /*
2419 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2420 * with the hdr's b_pabd.
2421 */
2422 static boolean_t
2423 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2424 {
2425 /*
2426 * The criteria for sharing a hdr's data are:
2427 * 1. the hdr's compression matches the buf's compression
2428 * 2. the hdr doesn't need to be byteswapped
2429 * 3. the hdr isn't already being shared
2430 * 4. the buf is either compressed or it is the last buf in the hdr list
2431 *
2432 * Criterion #4 maintains the invariant that shared uncompressed
2433 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2434 * might ask, "if a compressed buf is allocated first, won't that be the
2435 * last thing in the list?", but in that case it's impossible to create
2436 * a shared uncompressed buf anyway (because the hdr must be compressed
2437 * to have the compressed buf). You might also think that #3 is
2438 * sufficient to make this guarantee, however it's possible
2439 * (specifically in the rare L2ARC write race mentioned in
2440 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2441 * is sharable, but wasn't at the time of its allocation. Rather than
2442 * allow a new shared uncompressed buf to be created and then shuffle
2443 * the list around to make it the last element, this simply disallows
2444 * sharing if the new buf isn't the first to be added.
2445 */
2446 ASSERT3P(buf->b_hdr, ==, hdr);
2447 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2448 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2449 return (buf_compressed == hdr_compressed &&
2450 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2451 !HDR_SHARED_DATA(hdr) &&
2452 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2453 }
2454
2455 /*
2456 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2457 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2458 * copy was made successfully, or an error code otherwise.
2459 */
2460 static int
2461 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2462 boolean_t fill, arc_buf_t **ret)
2463 {
2464 arc_buf_t *buf;
2465
2466 ASSERT(HDR_HAS_L1HDR(hdr));
2467 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2468 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2469 hdr->b_type == ARC_BUFC_METADATA);
2470 ASSERT3P(ret, !=, NULL);
2471 ASSERT3P(*ret, ==, NULL);
2472
2473 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2474 buf->b_hdr = hdr;
2475 buf->b_data = NULL;
2476 buf->b_next = hdr->b_l1hdr.b_buf;
2477 buf->b_flags = 0;
2478
2479 add_reference(hdr, tag);
2480
2481 /*
2482 * We're about to change the hdr's b_flags. We must either
2483 * hold the hash_lock or be undiscoverable.
2484 */
2485 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2486
2487 /*
2488 * Only honor requests for compressed bufs if the hdr is actually
2489 * compressed.
2490 */
2491 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2492 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2493
2494 /*
2495 * If the hdr's data can be shared then we share the data buffer and
2496 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2497 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2498 * buffer to store the buf's data.
2499 *
2500 * There are two additional restrictions here because we're sharing
2501 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2502 * actively involved in an L2ARC write, because if this buf is used by
2503 * an arc_write() then the hdr's data buffer will be released when the
2504 * write completes, even though the L2ARC write might still be using it.
2505 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2506 * need to be ABD-aware.
2507 */
2508 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2509 abd_is_linear(hdr->b_l1hdr.b_pabd);
2510
2511 /* Set up b_data and sharing */
2512 if (can_share) {
2513 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2514 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2515 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2516 } else {
2517 buf->b_data =
2518 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2519 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2520 }
2521 VERIFY3P(buf->b_data, !=, NULL);
2522
2523 hdr->b_l1hdr.b_buf = buf;
2524 hdr->b_l1hdr.b_bufcnt += 1;
2525
2526 /*
2527 * If the user wants the data from the hdr, we need to either copy or
2528 * decompress the data.
2529 */
2530 if (fill) {
2531 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
2532 }
2533
2534 return (0);
2535 }
2536
2537 static char *arc_onloan_tag = "onloan";
2538
2539 static inline void
2540 arc_loaned_bytes_update(int64_t delta)
2541 {
2542 atomic_add_64(&arc_loaned_bytes, delta);
2543
2544 /* assert that it did not wrap around */
2545 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2546 }
2547
2548 /*
2549 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2550 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2551 * buffers must be returned to the arc before they can be used by the DMU or
2552 * freed.
2553 */
2554 arc_buf_t *
2555 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2556 {
2557 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2558 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2559
2560 arc_loaned_bytes_update(size);
2561
2562 return (buf);
2563 }
2564
2565 arc_buf_t *
2566 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2567 enum zio_compress compression_type)
2568 {
2569 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2570 psize, lsize, compression_type);
2571
2572 arc_loaned_bytes_update(psize);
2573
2574 return (buf);
2575 }
2576
2577
2578 /*
2579 * Return a loaned arc buffer to the arc.
2580 */
2581 void
2582 arc_return_buf(arc_buf_t *buf, void *tag)
2583 {
2584 arc_buf_hdr_t *hdr = buf->b_hdr;
2585
2586 ASSERT3P(buf->b_data, !=, NULL);
2587 ASSERT(HDR_HAS_L1HDR(hdr));
2588 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2589 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2590
2591 arc_loaned_bytes_update(-arc_buf_size(buf));
2592 }
2593
2594 /* Detach an arc_buf from a dbuf (tag) */
2595 void
2596 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2597 {
2598 arc_buf_hdr_t *hdr = buf->b_hdr;
2599
2600 ASSERT3P(buf->b_data, !=, NULL);
2601 ASSERT(HDR_HAS_L1HDR(hdr));
2602 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2603 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2604
2605 arc_loaned_bytes_update(arc_buf_size(buf));
2606 }
2607
2608 static void
2609 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2610 {
2611 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2612
2613 df->l2df_abd = abd;
2614 df->l2df_size = size;
2615 df->l2df_type = type;
2616 mutex_enter(&l2arc_free_on_write_mtx);
2617 list_insert_head(l2arc_free_on_write, df);
2618 mutex_exit(&l2arc_free_on_write_mtx);
2619 }
2620
2621 static void
2622 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2623 {
2624 arc_state_t *state = hdr->b_l1hdr.b_state;
2625 arc_buf_contents_t type = arc_buf_type(hdr);
2626 uint64_t size = arc_hdr_size(hdr);
2627
2628 /* protected by hash lock, if in the hash table */
2629 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2630 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2631 ASSERT(state != arc_anon && state != arc_l2c_only);
2632
2633 (void) refcount_remove_many(&state->arcs_esize[type],
2634 size, hdr);
2635 }
2636 (void) refcount_remove_many(&state->arcs_size, size, hdr);
2637 if (type == ARC_BUFC_METADATA) {
2638 arc_space_return(size, ARC_SPACE_META);
2639 } else {
2640 ASSERT(type == ARC_BUFC_DATA);
2641 arc_space_return(size, ARC_SPACE_DATA);
2642 }
2643
2644 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2645 }
2646
2647 /*
2648 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2649 * data buffer, we transfer the refcount ownership to the hdr and update
2650 * the appropriate kstats.
2651 */
2652 static void
2653 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2654 {
2655 arc_state_t *state = hdr->b_l1hdr.b_state;
2656
2657 ASSERT(arc_can_share(hdr, buf));
2658 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2659 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2660
2661 /*
2662 * Start sharing the data buffer. We transfer the
2663 * refcount ownership to the hdr since it always owns
2664 * the refcount whenever an arc_buf_t is shared.
2665 */
2666 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2667 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2668 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2669 HDR_ISTYPE_METADATA(hdr));
2670 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2671 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2672
2673 /*
2674 * Since we've transferred ownership to the hdr we need
2675 * to increment its compressed and uncompressed kstats and
2676 * decrement the overhead size.
2677 */
2678 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2679 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2680 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2681 }
2682
2683 static void
2684 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2685 {
2686 arc_state_t *state = hdr->b_l1hdr.b_state;
2687
2688 ASSERT(arc_buf_is_shared(buf));
2689 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2690 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2691
2692 /*
2693 * We are no longer sharing this buffer so we need
2694 * to transfer its ownership to the rightful owner.
2695 */
2696 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
2697 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2698 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
2699 abd_put(hdr->b_l1hdr.b_pabd);
2700 hdr->b_l1hdr.b_pabd = NULL;
2701 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2702
2703 /*
2704 * Since the buffer is no longer shared between
2705 * the arc buf and the hdr, count it as overhead.
2706 */
2707 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2708 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2709 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2710 }
2711
2712 /*
2713 * Remove an arc_buf_t from the hdr's buf list and return the last
2714 * arc_buf_t on the list. If no buffers remain on the list then return
2715 * NULL.
2716 */
2717 static arc_buf_t *
2718 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2719 {
2720 ASSERT(HDR_HAS_L1HDR(hdr));
2721 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2722
2723 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
2724 arc_buf_t *lastbuf = NULL;
2725
2726 /*
2727 * Remove the buf from the hdr list and locate the last
2728 * remaining buffer on the list.
2729 */
2730 while (*bufp != NULL) {
2731 if (*bufp == buf)
2732 *bufp = buf->b_next;
2733
2734 /*
2735 * If we've removed a buffer in the middle of
2736 * the list then update the lastbuf and update
2737 * bufp.
2738 */
2739 if (*bufp != NULL) {
2740 lastbuf = *bufp;
2741 bufp = &(*bufp)->b_next;
2742 }
2743 }
2744 buf->b_next = NULL;
2745 ASSERT3P(lastbuf, !=, buf);
2746 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
2747 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
2748 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
2749
2750 return (lastbuf);
2751 }
2752
2753 /*
2754 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
2755 * list and free it.
2756 */
2757 static void
2758 arc_buf_destroy_impl(arc_buf_t *buf)
2759 {
2760 arc_buf_hdr_t *hdr = buf->b_hdr;
2761
2762 /*
2763 * Free up the data associated with the buf but only if we're not
2764 * sharing this with the hdr. If we are sharing it with the hdr, the
2765 * hdr is responsible for doing the free.
2766 */
2767 if (buf->b_data != NULL) {
2768 /*
2769 * We're about to change the hdr's b_flags. We must either
2770 * hold the hash_lock or be undiscoverable.
2771 */
2772 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2773
2774 arc_cksum_verify(buf);
2775 arc_buf_unwatch(buf);
2776
2777 if (arc_buf_is_shared(buf)) {
2778 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2779 } else {
2780 uint64_t size = arc_buf_size(buf);
2781 arc_free_data_buf(hdr, buf->b_data, size, buf);
2782 ARCSTAT_INCR(arcstat_overhead_size, -size);
2783 }
2784 buf->b_data = NULL;
2785
2786 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
2787 hdr->b_l1hdr.b_bufcnt -= 1;
2788 }
2789
2790 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
2791
2792 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
2793 /*
2794 * If the current arc_buf_t is sharing its data buffer with the
2795 * hdr, then reassign the hdr's b_pabd to share it with the new
2796 * buffer at the end of the list. The shared buffer is always
2797 * the last one on the hdr's buffer list.
2798 *
2799 * There is an equivalent case for compressed bufs, but since
2800 * they aren't guaranteed to be the last buf in the list and
2801 * that is an exceedingly rare case, we just allow that space be
2802 * wasted temporarily.
2803 */
2804 if (lastbuf != NULL) {
2805 /* Only one buf can be shared at once */
2806 VERIFY(!arc_buf_is_shared(lastbuf));
2807 /* hdr is uncompressed so can't have compressed buf */
2808 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
2809
2810 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2811 arc_hdr_free_pabd(hdr);
2812
2813 /*
2814 * We must setup a new shared block between the
2815 * last buffer and the hdr. The data would have
2816 * been allocated by the arc buf so we need to transfer
2817 * ownership to the hdr since it's now being shared.
2818 */
2819 arc_share_buf(hdr, lastbuf);
2820 }
2821 } else if (HDR_SHARED_DATA(hdr)) {
2822 /*
2823 * Uncompressed shared buffers are always at the end
2824 * of the list. Compressed buffers don't have the
2825 * same requirements. This makes it hard to
2826 * simply assert that the lastbuf is shared so
2827 * we rely on the hdr's compression flags to determine
2828 * if we have a compressed, shared buffer.
2829 */
2830 ASSERT3P(lastbuf, !=, NULL);
2831 ASSERT(arc_buf_is_shared(lastbuf) ||
2832 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2833 }
2834
2835 /*
2836 * Free the checksum if we're removing the last uncompressed buf from
2837 * this hdr.
2838 */
2839 if (!arc_hdr_has_uncompressed_buf(hdr)) {
2840 arc_cksum_free(hdr);
2841 }
2842
2843 /* clean up the buf */
2844 buf->b_hdr = NULL;
2845 kmem_cache_free(buf_cache, buf);
2846 }
2847
2848 static void
2849 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
2850 {
2851 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2852 ASSERT(HDR_HAS_L1HDR(hdr));
2853 ASSERT(!HDR_SHARED_DATA(hdr));
2854
2855 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2856 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
2857 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2858 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2859
2860 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2861 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2862 }
2863
2864 static void
2865 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
2866 {
2867 ASSERT(HDR_HAS_L1HDR(hdr));
2868 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2869
2870 /*
2871 * If the hdr is currently being written to the l2arc then
2872 * we defer freeing the data by adding it to the l2arc_free_on_write
2873 * list. The l2arc will free the data once it's finished
2874 * writing it to the l2arc device.
2875 */
2876 if (HDR_L2_WRITING(hdr)) {
2877 arc_hdr_free_on_write(hdr);
2878 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2879 } else {
2880 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
2881 arc_hdr_size(hdr), hdr);
2882 }
2883 hdr->b_l1hdr.b_pabd = NULL;
2884 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2885
2886 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2887 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2888 }
2889
2890 static arc_buf_hdr_t *
2891 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
2892 enum zio_compress compression_type, arc_buf_contents_t type)
2893 {
2894 arc_buf_hdr_t *hdr;
2895
2896 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
2897
2898 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
2899 ASSERT(HDR_EMPTY(hdr));
2900 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2901 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
2902 HDR_SET_PSIZE(hdr, psize);
2903 HDR_SET_LSIZE(hdr, lsize);
2904 hdr->b_spa = spa;
2905 hdr->b_type = type;
2906 hdr->b_flags = 0;
2907 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
2908 arc_hdr_set_compress(hdr, compression_type);
2909
2910 hdr->b_l1hdr.b_state = arc_anon;
2911 hdr->b_l1hdr.b_arc_access = 0;
2912 hdr->b_l1hdr.b_bufcnt = 0;
2913 hdr->b_l1hdr.b_buf = NULL;
2914
2915 /*
2916 * Allocate the hdr's buffer. This will contain either
2917 * the compressed or uncompressed data depending on the block
2918 * it references and compressed arc enablement.
2919 */
2920 arc_hdr_alloc_pabd(hdr);
2921 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2922
2923 return (hdr);
2924 }
2925
2926 /*
2927 * Transition between the two allocation states for the arc_buf_hdr struct.
2928 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
2929 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
2930 * version is used when a cache buffer is only in the L2ARC in order to reduce
2931 * memory usage.
2932 */
2933 static arc_buf_hdr_t *
2934 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
2935 {
2936 ASSERT(HDR_HAS_L2HDR(hdr));
2937
2938 arc_buf_hdr_t *nhdr;
2939 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2940
2941 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
2942 (old == hdr_l2only_cache && new == hdr_full_cache));
2943
2944 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
2945
2946 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
2947 buf_hash_remove(hdr);
2948
2949 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
2950
2951 if (new == hdr_full_cache) {
2952 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
2953 /*
2954 * arc_access and arc_change_state need to be aware that a
2955 * header has just come out of L2ARC, so we set its state to
2956 * l2c_only even though it's about to change.
2957 */
2958 nhdr->b_l1hdr.b_state = arc_l2c_only;
2959
2960 /* Verify previous threads set to NULL before freeing */
2961 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
2962 } else {
2963 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2964 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2965 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2966
2967 /*
2968 * If we've reached here, We must have been called from
2969 * arc_evict_hdr(), as such we should have already been
2970 * removed from any ghost list we were previously on
2971 * (which protects us from racing with arc_evict_state),
2972 * thus no locking is needed during this check.
2973 */
2974 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2975
2976 /*
2977 * A buffer must not be moved into the arc_l2c_only
2978 * state if it's not finished being written out to the
2979 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
2980 * might try to be accessed, even though it was removed.
2981 */
2982 VERIFY(!HDR_L2_WRITING(hdr));
2983 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2984
2985 #ifdef ZFS_DEBUG
2986 if (hdr->b_l1hdr.b_thawed != NULL) {
2987 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2988 hdr->b_l1hdr.b_thawed = NULL;
2989 }
2990 #endif
2991
2992 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
2993 }
2994 /*
2995 * The header has been reallocated so we need to re-insert it into any
2996 * lists it was on.
2997 */
2998 (void) buf_hash_insert(nhdr, NULL);
2999
3000 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3001
3002 mutex_enter(&dev->l2ad_mtx);
3003
3004 /*
3005 * We must place the realloc'ed header back into the list at
3006 * the same spot. Otherwise, if it's placed earlier in the list,
3007 * l2arc_write_buffers() could find it during the function's
3008 * write phase, and try to write it out to the l2arc.
3009 */
3010 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3011 list_remove(&dev->l2ad_buflist, hdr);
3012
3013 mutex_exit(&dev->l2ad_mtx);
3014
3015 /*
3016 * Since we're using the pointer address as the tag when
3017 * incrementing and decrementing the l2ad_alloc refcount, we
3018 * must remove the old pointer (that we're about to destroy) and
3019 * add the new pointer to the refcount. Otherwise we'd remove
3020 * the wrong pointer address when calling arc_hdr_destroy() later.
3021 */
3022
3023 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
3024 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
3025
3026 buf_discard_identity(hdr);
3027 kmem_cache_free(old, hdr);
3028
3029 return (nhdr);
3030 }
3031
3032 /*
3033 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3034 * The buf is returned thawed since we expect the consumer to modify it.
3035 */
3036 arc_buf_t *
3037 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3038 {
3039 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3040 ZIO_COMPRESS_OFF, type);
3041 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3042
3043 arc_buf_t *buf = NULL;
3044 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3045 arc_buf_thaw(buf);
3046
3047 return (buf);
3048 }
3049
3050 /*
3051 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3052 * for bufs containing metadata.
3053 */
3054 arc_buf_t *
3055 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3056 enum zio_compress compression_type)
3057 {
3058 ASSERT3U(lsize, >, 0);
3059 ASSERT3U(lsize, >=, psize);
3060 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3061 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3062
3063 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3064 compression_type, ARC_BUFC_DATA);
3065 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3066
3067 arc_buf_t *buf = NULL;
3068 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3069 arc_buf_thaw(buf);
3070 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3071
3072 if (!arc_buf_is_shared(buf)) {
3073 /*
3074 * To ensure that the hdr has the correct data in it if we call
3075 * arc_decompress() on this buf before it's been written to
3076 * disk, it's easiest if we just set up sharing between the
3077 * buf and the hdr.
3078 */
3079 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3080 arc_hdr_free_pabd(hdr);
3081 arc_share_buf(hdr, buf);
3082 }
3083
3084 return (buf);
3085 }
3086
3087 static void
3088 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3089 {
3090 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3091 l2arc_dev_t *dev = l2hdr->b_dev;
3092 uint64_t psize = arc_hdr_size(hdr);
3093
3094 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3095 ASSERT(HDR_HAS_L2HDR(hdr));
3096
3097 list_remove(&dev->l2ad_buflist, hdr);
3098
3099 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3100 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3101
3102 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3103
3104 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3105 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3106 }
3107
3108 static void
3109 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3110 {
3111 if (HDR_HAS_L1HDR(hdr)) {
3112 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3113 hdr->b_l1hdr.b_bufcnt > 0);
3114 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3115 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3116 }
3117 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3118 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3119
3120 if (!HDR_EMPTY(hdr))
3121 buf_discard_identity(hdr);
3122
3123 if (HDR_HAS_L2HDR(hdr)) {
3124 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3125 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3126
3127 if (!buflist_held)
3128 mutex_enter(&dev->l2ad_mtx);
3129
3130 /*
3131 * Even though we checked this conditional above, we
3132 * need to check this again now that we have the
3133 * l2ad_mtx. This is because we could be racing with
3134 * another thread calling l2arc_evict() which might have
3135 * destroyed this header's L2 portion as we were waiting
3136 * to acquire the l2ad_mtx. If that happens, we don't
3137 * want to re-destroy the header's L2 portion.
3138 */
3139 if (HDR_HAS_L2HDR(hdr))
3140 arc_hdr_l2hdr_destroy(hdr);
3141
3142 if (!buflist_held)
3143 mutex_exit(&dev->l2ad_mtx);
3144 }
3145
3146 if (HDR_HAS_L1HDR(hdr)) {
3147 arc_cksum_free(hdr);
3148
3149 while (hdr->b_l1hdr.b_buf != NULL)
3150 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3151
3152 #ifdef ZFS_DEBUG
3153 if (hdr->b_l1hdr.b_thawed != NULL) {
3154 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3155 hdr->b_l1hdr.b_thawed = NULL;
3156 }
3157 #endif
3158
3159 if (hdr->b_l1hdr.b_pabd != NULL) {
3160 arc_hdr_free_pabd(hdr);
3161 }
3162 }
3163
3164 ASSERT3P(hdr->b_hash_next, ==, NULL);
3165 if (HDR_HAS_L1HDR(hdr)) {
3166 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3167 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3168 kmem_cache_free(hdr_full_cache, hdr);
3169 } else {
3170 kmem_cache_free(hdr_l2only_cache, hdr);
3171 }
3172 }
3173
3174 void
3175 arc_buf_destroy(arc_buf_t *buf, void* tag)
3176 {
3177 arc_buf_hdr_t *hdr = buf->b_hdr;
3178 kmutex_t *hash_lock = HDR_LOCK(hdr);
3179
3180 if (hdr->b_l1hdr.b_state == arc_anon) {
3181 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3182 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3183 VERIFY0(remove_reference(hdr, NULL, tag));
3184 arc_hdr_destroy(hdr);
3185 return;
3186 }
3187
3188 mutex_enter(hash_lock);
3189 ASSERT3P(hdr, ==, buf->b_hdr);
3190 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3191 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3192 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3193 ASSERT3P(buf->b_data, !=, NULL);
3194
3195 (void) remove_reference(hdr, hash_lock, tag);
3196 arc_buf_destroy_impl(buf);
3197 mutex_exit(hash_lock);
3198 }
3199
3200 /*
3201 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3202 * state of the header is dependent on it's state prior to entering this
3203 * function. The following transitions are possible:
3204 *
3205 * - arc_mru -> arc_mru_ghost
3206 * - arc_mfu -> arc_mfu_ghost
3207 * - arc_mru_ghost -> arc_l2c_only
3208 * - arc_mru_ghost -> deleted
3209 * - arc_mfu_ghost -> arc_l2c_only
3210 * - arc_mfu_ghost -> deleted
3211 */
3212 static int64_t
3213 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3214 {
3215 arc_state_t *evicted_state, *state;
3216 int64_t bytes_evicted = 0;
3217
3218 ASSERT(MUTEX_HELD(hash_lock));
3219 ASSERT(HDR_HAS_L1HDR(hdr));
3220
3221 state = hdr->b_l1hdr.b_state;
3222 if (GHOST_STATE(state)) {
3223 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3224 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3225
3226 /*
3227 * l2arc_write_buffers() relies on a header's L1 portion
3228 * (i.e. its b_pabd field) during it's write phase.
3229 * Thus, we cannot push a header onto the arc_l2c_only
3230 * state (removing it's L1 piece) until the header is
3231 * done being written to the l2arc.
3232 */
3233 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3234 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3235 return (bytes_evicted);
3236 }
3237
3238 ARCSTAT_BUMP(arcstat_deleted);
3239 bytes_evicted += HDR_GET_LSIZE(hdr);
3240
3241 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3242
3243 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3244 if (HDR_HAS_L2HDR(hdr)) {
3245 /*
3246 * This buffer is cached on the 2nd Level ARC;
3247 * don't destroy the header.
3248 */
3249 arc_change_state(arc_l2c_only, hdr, hash_lock);
3250 /*
3251 * dropping from L1+L2 cached to L2-only,
3252 * realloc to remove the L1 header.
3253 */
3254 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3255 hdr_l2only_cache);
3256 } else {
3257 arc_change_state(arc_anon, hdr, hash_lock);
3258 arc_hdr_destroy(hdr);
3259 }
3260 return (bytes_evicted);
3261 }
3262
3263 ASSERT(state == arc_mru || state == arc_mfu);
3264 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3265
3266 /* prefetch buffers have a minimum lifespan */
3267 if (HDR_IO_IN_PROGRESS(hdr) ||
3268 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3269 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3270 arc_min_prefetch_lifespan)) {
3271 ARCSTAT_BUMP(arcstat_evict_skip);
3272 return (bytes_evicted);
3273 }
3274
3275 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3276 while (hdr->b_l1hdr.b_buf) {
3277 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3278 if (!mutex_tryenter(&buf->b_evict_lock)) {
3279 ARCSTAT_BUMP(arcstat_mutex_miss);
3280 break;
3281 }
3282 if (buf->b_data != NULL)
3283 bytes_evicted += HDR_GET_LSIZE(hdr);
3284 mutex_exit(&buf->b_evict_lock);
3285 arc_buf_destroy_impl(buf);
3286 }
3287
3288 if (HDR_HAS_L2HDR(hdr)) {
3289 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3290 } else {
3291 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3292 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3293 HDR_GET_LSIZE(hdr));
3294 } else {
3295 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3296 HDR_GET_LSIZE(hdr));
3297 }
3298 }
3299
3300 if (hdr->b_l1hdr.b_bufcnt == 0) {
3301 arc_cksum_free(hdr);
3302
3303 bytes_evicted += arc_hdr_size(hdr);
3304
3305 /*
3306 * If this hdr is being evicted and has a compressed
3307 * buffer then we discard it here before we change states.
3308 * This ensures that the accounting is updated correctly
3309 * in arc_free_data_impl().
3310 */
3311 arc_hdr_free_pabd(hdr);
3312
3313 arc_change_state(evicted_state, hdr, hash_lock);
3314 ASSERT(HDR_IN_HASH_TABLE(hdr));
3315 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3316 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3317 }
3318
3319 return (bytes_evicted);
3320 }
3321
3322 static uint64_t
3323 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3324 uint64_t spa, int64_t bytes)
3325 {
3326 multilist_sublist_t *mls;
3327 uint64_t bytes_evicted = 0;
3328 arc_buf_hdr_t *hdr;
3329 kmutex_t *hash_lock;
3330 int evict_count = 0;
3331
3332 ASSERT3P(marker, !=, NULL);
3333 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3334
3335 mls = multilist_sublist_lock(ml, idx);
3336
3337 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3338 hdr = multilist_sublist_prev(mls, marker)) {
3339 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3340 (evict_count >= zfs_arc_evict_batch_limit))
3341 break;
3342
3343 /*
3344 * To keep our iteration location, move the marker
3345 * forward. Since we're not holding hdr's hash lock, we
3346 * must be very careful and not remove 'hdr' from the
3347 * sublist. Otherwise, other consumers might mistake the
3348 * 'hdr' as not being on a sublist when they call the
3349 * multilist_link_active() function (they all rely on
3350 * the hash lock protecting concurrent insertions and
3351 * removals). multilist_sublist_move_forward() was
3352 * specifically implemented to ensure this is the case
3353 * (only 'marker' will be removed and re-inserted).
3354 */
3355 multilist_sublist_move_forward(mls, marker);
3356
3357 /*
3358 * The only case where the b_spa field should ever be
3359 * zero, is the marker headers inserted by
3360 * arc_evict_state(). It's possible for multiple threads
3361 * to be calling arc_evict_state() concurrently (e.g.
3362 * dsl_pool_close() and zio_inject_fault()), so we must
3363 * skip any markers we see from these other threads.
3364 */
3365 if (hdr->b_spa == 0)
3366 continue;
3367
3368 /* we're only interested in evicting buffers of a certain spa */
3369 if (spa != 0 && hdr->b_spa != spa) {
3370 ARCSTAT_BUMP(arcstat_evict_skip);
3371 continue;
3372 }
3373
3374 hash_lock = HDR_LOCK(hdr);
3375
3376 /*
3377 * We aren't calling this function from any code path
3378 * that would already be holding a hash lock, so we're
3379 * asserting on this assumption to be defensive in case
3380 * this ever changes. Without this check, it would be
3381 * possible to incorrectly increment arcstat_mutex_miss
3382 * below (e.g. if the code changed such that we called
3383 * this function with a hash lock held).
3384 */
3385 ASSERT(!MUTEX_HELD(hash_lock));
3386
3387 if (mutex_tryenter(hash_lock)) {
3388 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3389 mutex_exit(hash_lock);
3390
3391 bytes_evicted += evicted;
3392
3393 /*
3394 * If evicted is zero, arc_evict_hdr() must have
3395 * decided to skip this header, don't increment
3396 * evict_count in this case.
3397 */
3398 if (evicted != 0)
3399 evict_count++;
3400
3401 /*
3402 * If arc_size isn't overflowing, signal any
3403 * threads that might happen to be waiting.
3404 *
3405 * For each header evicted, we wake up a single
3406 * thread. If we used cv_broadcast, we could
3407 * wake up "too many" threads causing arc_size
3408 * to significantly overflow arc_c; since
3409 * arc_get_data_impl() doesn't check for overflow
3410 * when it's woken up (it doesn't because it's
3411 * possible for the ARC to be overflowing while
3412 * full of un-evictable buffers, and the
3413 * function should proceed in this case).
3414 *
3415 * If threads are left sleeping, due to not
3416 * using cv_broadcast, they will be woken up
3417 * just before arc_reclaim_thread() sleeps.
3418 */
3419 mutex_enter(&arc_reclaim_lock);
3420 if (!arc_is_overflowing())
3421 cv_signal(&arc_reclaim_waiters_cv);
3422 mutex_exit(&arc_reclaim_lock);
3423 } else {
3424 ARCSTAT_BUMP(arcstat_mutex_miss);
3425 }
3426 }
3427
3428 multilist_sublist_unlock(mls);
3429
3430 return (bytes_evicted);
3431 }
3432
3433 /*
3434 * Evict buffers from the given arc state, until we've removed the
3435 * specified number of bytes. Move the removed buffers to the
3436 * appropriate evict state.
3437 *
3438 * This function makes a "best effort". It skips over any buffers
3439 * it can't get a hash_lock on, and so, may not catch all candidates.
3440 * It may also return without evicting as much space as requested.
3441 *
3442 * If bytes is specified using the special value ARC_EVICT_ALL, this
3443 * will evict all available (i.e. unlocked and evictable) buffers from
3444 * the given arc state; which is used by arc_flush().
3445 */
3446 static uint64_t
3447 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3448 arc_buf_contents_t type)
3449 {
3450 uint64_t total_evicted = 0;
3451 multilist_t *ml = state->arcs_list[type];
3452 int num_sublists;
3453 arc_buf_hdr_t **markers;
3454
3455 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3456
3457 num_sublists = multilist_get_num_sublists(ml);
3458
3459 /*
3460 * If we've tried to evict from each sublist, made some
3461 * progress, but still have not hit the target number of bytes
3462 * to evict, we want to keep trying. The markers allow us to
3463 * pick up where we left off for each individual sublist, rather
3464 * than starting from the tail each time.
3465 */
3466 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3467 for (int i = 0; i < num_sublists; i++) {
3468 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3469
3470 /*
3471 * A b_spa of 0 is used to indicate that this header is
3472 * a marker. This fact is used in arc_adjust_type() and
3473 * arc_evict_state_impl().
3474 */
3475 markers[i]->b_spa = 0;
3476
3477 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3478 multilist_sublist_insert_tail(mls, markers[i]);
3479 multilist_sublist_unlock(mls);
3480 }
3481
3482 /*
3483 * While we haven't hit our target number of bytes to evict, or
3484 * we're evicting all available buffers.
3485 */
3486 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3487 /*
3488 * Start eviction using a randomly selected sublist,
3489 * this is to try and evenly balance eviction across all
3490 * sublists. Always starting at the same sublist
3491 * (e.g. index 0) would cause evictions to favor certain
3492 * sublists over others.
3493 */
3494 int sublist_idx = multilist_get_random_index(ml);
3495 uint64_t scan_evicted = 0;
3496
3497 for (int i = 0; i < num_sublists; i++) {
3498 uint64_t bytes_remaining;
3499 uint64_t bytes_evicted;
3500
3501 if (bytes == ARC_EVICT_ALL)
3502 bytes_remaining = ARC_EVICT_ALL;
3503 else if (total_evicted < bytes)
3504 bytes_remaining = bytes - total_evicted;
3505 else
3506 break;
3507
3508 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3509 markers[sublist_idx], spa, bytes_remaining);
3510
3511 scan_evicted += bytes_evicted;
3512 total_evicted += bytes_evicted;
3513
3514 /* we've reached the end, wrap to the beginning */
3515 if (++sublist_idx >= num_sublists)
3516 sublist_idx = 0;
3517 }
3518
3519 /*
3520 * If we didn't evict anything during this scan, we have
3521 * no reason to believe we'll evict more during another
3522 * scan, so break the loop.
3523 */
3524 if (scan_evicted == 0) {
3525 /* This isn't possible, let's make that obvious */
3526 ASSERT3S(bytes, !=, 0);
3527
3528 /*
3529 * When bytes is ARC_EVICT_ALL, the only way to
3530 * break the loop is when scan_evicted is zero.
3531 * In that case, we actually have evicted enough,
3532 * so we don't want to increment the kstat.
3533 */
3534 if (bytes != ARC_EVICT_ALL) {
3535 ASSERT3S(total_evicted, <, bytes);
3536 ARCSTAT_BUMP(arcstat_evict_not_enough);
3537 }
3538
3539 break;
3540 }
3541 }
3542
3543 for (int i = 0; i < num_sublists; i++) {
3544 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3545 multilist_sublist_remove(mls, markers[i]);
3546 multilist_sublist_unlock(mls);
3547
3548 kmem_cache_free(hdr_full_cache, markers[i]);
3549 }
3550 kmem_free(markers, sizeof (*markers) * num_sublists);
3551
3552 return (total_evicted);
3553 }
3554
3555 /*
3556 * Flush all "evictable" data of the given type from the arc state
3557 * specified. This will not evict any "active" buffers (i.e. referenced).
3558 *
3559 * When 'retry' is set to B_FALSE, the function will make a single pass
3560 * over the state and evict any buffers that it can. Since it doesn't
3561 * continually retry the eviction, it might end up leaving some buffers
3562 * in the ARC due to lock misses.
3563 *
3564 * When 'retry' is set to B_TRUE, the function will continually retry the
3565 * eviction until *all* evictable buffers have been removed from the
3566 * state. As a result, if concurrent insertions into the state are
3567 * allowed (e.g. if the ARC isn't shutting down), this function might
3568 * wind up in an infinite loop, continually trying to evict buffers.
3569 */
3570 static uint64_t
3571 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3572 boolean_t retry)
3573 {
3574 uint64_t evicted = 0;
3575
3576 while (refcount_count(&state->arcs_esize[type]) != 0) {
3577 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3578
3579 if (!retry)
3580 break;
3581 }
3582
3583 return (evicted);
3584 }
3585
3586 /*
3587 * Evict the specified number of bytes from the state specified,
3588 * restricting eviction to the spa and type given. This function
3589 * prevents us from trying to evict more from a state's list than
3590 * is "evictable", and to skip evicting altogether when passed a
3591 * negative value for "bytes". In contrast, arc_evict_state() will
3592 * evict everything it can, when passed a negative value for "bytes".
3593 */
3594 static uint64_t
3595 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3596 arc_buf_contents_t type)
3597 {
3598 int64_t delta;
3599
3600 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3601 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3602 return (arc_evict_state(state, spa, delta, type));
3603 }
3604
3605 return (0);
3606 }
3607
3608 /*
3609 * Evict metadata buffers from the cache, such that arc_meta_used is
3610 * capped by the arc_meta_limit tunable.
3611 */
3612 static uint64_t
3613 arc_adjust_meta(uint64_t meta_used)
3614 {
3615 uint64_t total_evicted = 0;
3616 int64_t target;
3617
3618 /*
3619 * If we're over the meta limit, we want to evict enough
3620 * metadata to get back under the meta limit. We don't want to
3621 * evict so much that we drop the MRU below arc_p, though. If
3622 * we're over the meta limit more than we're over arc_p, we
3623 * evict some from the MRU here, and some from the MFU below.
3624 */
3625 target = MIN((int64_t)(meta_used - arc_meta_limit),
3626 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3627 refcount_count(&arc_mru->arcs_size) - arc_p));
3628
3629 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3630
3631 /*
3632 * Similar to the above, we want to evict enough bytes to get us
3633 * below the meta limit, but not so much as to drop us below the
3634 * space allotted to the MFU (which is defined as arc_c - arc_p).
3635 */
3636 target = MIN((int64_t)(meta_used - arc_meta_limit),
3637 (int64_t)(refcount_count(&arc_mfu->arcs_size) -
3638 (arc_c - arc_p)));
3639
3640 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3641
3642 return (total_evicted);
3643 }
3644
3645 /*
3646 * Return the type of the oldest buffer in the given arc state
3647 *
3648 * This function will select a random sublist of type ARC_BUFC_DATA and
3649 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3650 * is compared, and the type which contains the "older" buffer will be
3651 * returned.
3652 */
3653 static arc_buf_contents_t
3654 arc_adjust_type(arc_state_t *state)
3655 {
3656 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
3657 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
3658 int data_idx = multilist_get_random_index(data_ml);
3659 int meta_idx = multilist_get_random_index(meta_ml);
3660 multilist_sublist_t *data_mls;
3661 multilist_sublist_t *meta_mls;
3662 arc_buf_contents_t type;
3663 arc_buf_hdr_t *data_hdr;
3664 arc_buf_hdr_t *meta_hdr;
3665
3666 /*
3667 * We keep the sublist lock until we're finished, to prevent
3668 * the headers from being destroyed via arc_evict_state().
3669 */
3670 data_mls = multilist_sublist_lock(data_ml, data_idx);
3671 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3672
3673 /*
3674 * These two loops are to ensure we skip any markers that
3675 * might be at the tail of the lists due to arc_evict_state().
3676 */
3677
3678 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3679 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3680 if (data_hdr->b_spa != 0)
3681 break;
3682 }
3683
3684 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3685 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3686 if (meta_hdr->b_spa != 0)
3687 break;
3688 }
3689
3690 if (data_hdr == NULL && meta_hdr == NULL) {
3691 type = ARC_BUFC_DATA;
3692 } else if (data_hdr == NULL) {
3693 ASSERT3P(meta_hdr, !=, NULL);
3694 type = ARC_BUFC_METADATA;
3695 } else if (meta_hdr == NULL) {
3696 ASSERT3P(data_hdr, !=, NULL);
3697 type = ARC_BUFC_DATA;
3698 } else {
3699 ASSERT3P(data_hdr, !=, NULL);
3700 ASSERT3P(meta_hdr, !=, NULL);
3701
3702 /* The headers can't be on the sublist without an L1 header */
3703 ASSERT(HDR_HAS_L1HDR(data_hdr));
3704 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3705
3706 if (data_hdr->b_l1hdr.b_arc_access <
3707 meta_hdr->b_l1hdr.b_arc_access) {
3708 type = ARC_BUFC_DATA;
3709 } else {
3710 type = ARC_BUFC_METADATA;
3711 }
3712 }
3713
3714 multilist_sublist_unlock(meta_mls);
3715 multilist_sublist_unlock(data_mls);
3716
3717 return (type);
3718 }
3719
3720 /*
3721 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3722 */
3723 static uint64_t
3724 arc_adjust(void)
3725 {
3726 uint64_t total_evicted = 0;
3727 uint64_t bytes;
3728 int64_t target;
3729 uint64_t asize = aggsum_value(&arc_size);
3730 uint64_t ameta = aggsum_value(&arc_meta_used);
3731
3732 /*
3733 * If we're over arc_meta_limit, we want to correct that before
3734 * potentially evicting data buffers below.
3735 */
3736 total_evicted += arc_adjust_meta(ameta);
3737
3738 /*
3739 * Adjust MRU size
3740 *
3741 * If we're over the target cache size, we want to evict enough
3742 * from the list to get back to our target size. We don't want
3743 * to evict too much from the MRU, such that it drops below
3744 * arc_p. So, if we're over our target cache size more than
3745 * the MRU is over arc_p, we'll evict enough to get back to
3746 * arc_p here, and then evict more from the MFU below.
3747 */
3748 target = MIN((int64_t)(asize - arc_c),
3749 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3750 refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
3751
3752 /*
3753 * If we're below arc_meta_min, always prefer to evict data.
3754 * Otherwise, try to satisfy the requested number of bytes to
3755 * evict from the type which contains older buffers; in an
3756 * effort to keep newer buffers in the cache regardless of their
3757 * type. If we cannot satisfy the number of bytes from this
3758 * type, spill over into the next type.
3759 */
3760 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
3761 ameta > arc_meta_min) {
3762 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3763 total_evicted += bytes;
3764
3765 /*
3766 * If we couldn't evict our target number of bytes from
3767 * metadata, we try to get the rest from data.
3768 */
3769 target -= bytes;
3770
3771 total_evicted +=
3772 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3773 } else {
3774 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3775 total_evicted += bytes;
3776
3777 /*
3778 * If we couldn't evict our target number of bytes from
3779 * data, we try to get the rest from metadata.
3780 */
3781 target -= bytes;
3782
3783 total_evicted +=
3784 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3785 }
3786
3787 /*
3788 * Adjust MFU size
3789 *
3790 * Now that we've tried to evict enough from the MRU to get its
3791 * size back to arc_p, if we're still above the target cache
3792 * size, we evict the rest from the MFU.
3793 */
3794 target = asize - arc_c;
3795
3796 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
3797 ameta > arc_meta_min) {
3798 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3799 total_evicted += bytes;
3800
3801 /*
3802 * If we couldn't evict our target number of bytes from
3803 * metadata, we try to get the rest from data.
3804 */
3805 target -= bytes;
3806
3807 total_evicted +=
3808 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3809 } else {
3810 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3811 total_evicted += bytes;
3812
3813 /*
3814 * If we couldn't evict our target number of bytes from
3815 * data, we try to get the rest from data.
3816 */
3817 target -= bytes;
3818
3819 total_evicted +=
3820 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3821 }
3822
3823 /*
3824 * Adjust ghost lists
3825 *
3826 * In addition to the above, the ARC also defines target values
3827 * for the ghost lists. The sum of the mru list and mru ghost
3828 * list should never exceed the target size of the cache, and
3829 * the sum of the mru list, mfu list, mru ghost list, and mfu
3830 * ghost list should never exceed twice the target size of the
3831 * cache. The following logic enforces these limits on the ghost
3832 * caches, and evicts from them as needed.
3833 */
3834 target = refcount_count(&arc_mru->arcs_size) +
3835 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3836
3837 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3838 total_evicted += bytes;
3839
3840 target -= bytes;
3841
3842 total_evicted +=
3843 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3844
3845 /*
3846 * We assume the sum of the mru list and mfu list is less than
3847 * or equal to arc_c (we enforced this above), which means we
3848 * can use the simpler of the two equations below:
3849 *
3850 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3851 * mru ghost + mfu ghost <= arc_c
3852 */
3853 target = refcount_count(&arc_mru_ghost->arcs_size) +
3854 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3855
3856 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3857 total_evicted += bytes;
3858
3859 target -= bytes;
3860
3861 total_evicted +=
3862 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3863
3864 return (total_evicted);
3865 }
3866
3867 void
3868 arc_flush(spa_t *spa, boolean_t retry)
3869 {
3870 uint64_t guid = 0;
3871
3872 /*
3873 * If retry is B_TRUE, a spa must not be specified since we have
3874 * no good way to determine if all of a spa's buffers have been
3875 * evicted from an arc state.
3876 */
3877 ASSERT(!retry || spa == 0);
3878
3879 if (spa != NULL)
3880 guid = spa_load_guid(spa);
3881
3882 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3883 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3884
3885 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3886 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3887
3888 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3889 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3890
3891 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3892 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3893 }
3894
3895 void
3896 arc_shrink(int64_t to_free)
3897 {
3898 uint64_t asize = aggsum_value(&arc_size);
3899 if (arc_c > arc_c_min) {
3900
3901 if (arc_c > arc_c_min + to_free)
3902 atomic_add_64(&arc_c, -to_free);
3903 else
3904 arc_c = arc_c_min;
3905
3906 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3907 if (asize < arc_c)
3908 arc_c = MAX(asize, arc_c_min);
3909 if (arc_p > arc_c)
3910 arc_p = (arc_c >> 1);
3911 ASSERT(arc_c >= arc_c_min);
3912 ASSERT((int64_t)arc_p >= 0);
3913 }
3914
3915 if (asize > arc_c)
3916 (void) arc_adjust();
3917 }
3918
3919 typedef enum free_memory_reason_t {
3920 FMR_UNKNOWN,
3921 FMR_NEEDFREE,
3922 FMR_LOTSFREE,
3923 FMR_SWAPFS_MINFREE,
3924 FMR_PAGES_PP_MAXIMUM,
3925 FMR_HEAP_ARENA,
3926 FMR_ZIO_ARENA,
3927 } free_memory_reason_t;
3928
3929 int64_t last_free_memory;
3930 free_memory_reason_t last_free_reason;
3931
3932 /*
3933 * Additional reserve of pages for pp_reserve.
3934 */
3935 int64_t arc_pages_pp_reserve = 64;
3936
3937 /*
3938 * Additional reserve of pages for swapfs.
3939 */
3940 int64_t arc_swapfs_reserve = 64;
3941
3942 /*
3943 * Return the amount of memory that can be consumed before reclaim will be
3944 * needed. Positive if there is sufficient free memory, negative indicates
3945 * the amount of memory that needs to be freed up.
3946 */
3947 static int64_t
3948 arc_available_memory(void)
3949 {
3950 int64_t lowest = INT64_MAX;
3951 int64_t n;
3952 free_memory_reason_t r = FMR_UNKNOWN;
3953
3954 #ifdef _KERNEL
3955 if (needfree > 0) {
3956 n = PAGESIZE * (-needfree);
3957 if (n < lowest) {
3958 lowest = n;
3959 r = FMR_NEEDFREE;
3960 }
3961 }
3962
3963 /*
3964 * check that we're out of range of the pageout scanner. It starts to
3965 * schedule paging if freemem is less than lotsfree and needfree.
3966 * lotsfree is the high-water mark for pageout, and needfree is the
3967 * number of needed free pages. We add extra pages here to make sure
3968 * the scanner doesn't start up while we're freeing memory.
3969 */
3970 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3971 if (n < lowest) {
3972 lowest = n;
3973 r = FMR_LOTSFREE;
3974 }
3975
3976 /*
3977 * check to make sure that swapfs has enough space so that anon
3978 * reservations can still succeed. anon_resvmem() checks that the
3979 * availrmem is greater than swapfs_minfree, and the number of reserved
3980 * swap pages. We also add a bit of extra here just to prevent
3981 * circumstances from getting really dire.
3982 */
3983 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3984 desfree - arc_swapfs_reserve);
3985 if (n < lowest) {
3986 lowest = n;
3987 r = FMR_SWAPFS_MINFREE;
3988 }
3989
3990
3991 /*
3992 * Check that we have enough availrmem that memory locking (e.g., via
3993 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3994 * stores the number of pages that cannot be locked; when availrmem
3995 * drops below pages_pp_maximum, page locking mechanisms such as
3996 * page_pp_lock() will fail.)
3997 */
3998 n = PAGESIZE * (availrmem - pages_pp_maximum -
3999 arc_pages_pp_reserve);
4000 if (n < lowest) {
4001 lowest = n;
4002 r = FMR_PAGES_PP_MAXIMUM;
4003 }
4004
4005 #if defined(__i386)
4006 /*
4007 * If we're on an i386 platform, it's possible that we'll exhaust the
4008 * kernel heap space before we ever run out of available physical
4009 * memory. Most checks of the size of the heap_area compare against
4010 * tune.t_minarmem, which is the minimum available real memory that we
4011 * can have in the system. However, this is generally fixed at 25 pages
4012 * which is so low that it's useless. In this comparison, we seek to
4013 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4014 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4015 * free)
4016 */
4017 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
4018 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
4019 if (n < lowest) {
4020 lowest = n;
4021 r = FMR_HEAP_ARENA;
4022 }
4023 #endif
4024
4025 /*
4026 * If zio data pages are being allocated out of a separate heap segment,
4027 * then enforce that the size of available vmem for this arena remains
4028 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4029 *
4030 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4031 * memory (in the zio_arena) free, which can avoid memory
4032 * fragmentation issues.
4033 */
4034 if (zio_arena != NULL) {
4035 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4036 (vmem_size(zio_arena, VMEM_ALLOC) >>
4037 arc_zio_arena_free_shift);
4038 if (n < lowest) {
4039 lowest = n;
4040 r = FMR_ZIO_ARENA;
4041 }
4042 }
4043 #else
4044 /* Every 100 calls, free a small amount */
4045 if (spa_get_random(100) == 0)
4046 lowest = -1024;
4047 #endif
4048
4049 last_free_memory = lowest;
4050 last_free_reason = r;
4051
4052 return (lowest);
4053 }
4054
4055
4056 /*
4057 * Determine if the system is under memory pressure and is asking
4058 * to reclaim memory. A return value of B_TRUE indicates that the system
4059 * is under memory pressure and that the arc should adjust accordingly.
4060 */
4061 static boolean_t
4062 arc_reclaim_needed(void)
4063 {
4064 return (arc_available_memory() < 0);
4065 }
4066
4067 static void
4068 arc_kmem_reap_now(void)
4069 {
4070 size_t i;
4071 kmem_cache_t *prev_cache = NULL;
4072 kmem_cache_t *prev_data_cache = NULL;
4073 extern kmem_cache_t *zio_buf_cache[];
4074 extern kmem_cache_t *zio_data_buf_cache[];
4075 extern kmem_cache_t *range_seg_cache;
4076 extern kmem_cache_t *abd_chunk_cache;
4077
4078 #ifdef _KERNEL
4079 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4080 /*
4081 * We are exceeding our meta-data cache limit.
4082 * Purge some DNLC entries to release holds on meta-data.
4083 */
4084 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4085 }
4086 #if defined(__i386)
4087 /*
4088 * Reclaim unused memory from all kmem caches.
4089 */
4090 kmem_reap();
4091 #endif
4092 #endif
4093
4094 /*
4095 * If a kmem reap is already active, don't schedule more. We must
4096 * check for this because kmem_cache_reap_soon() won't actually
4097 * block on the cache being reaped (this is to prevent callers from
4098 * becoming implicitly blocked by a system-wide kmem reap -- which,
4099 * on a system with many, many full magazines, can take minutes).
4100 */
4101 if (kmem_cache_reap_active())
4102 return;
4103
4104 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4105 if (zio_buf_cache[i] != prev_cache) {
4106 prev_cache = zio_buf_cache[i];
4107 kmem_cache_reap_soon(zio_buf_cache[i]);
4108 }
4109 if (zio_data_buf_cache[i] != prev_data_cache) {
4110 prev_data_cache = zio_data_buf_cache[i];
4111 kmem_cache_reap_soon(zio_data_buf_cache[i]);
4112 }
4113 }
4114 kmem_cache_reap_soon(abd_chunk_cache);
4115 kmem_cache_reap_soon(buf_cache);
4116 kmem_cache_reap_soon(hdr_full_cache);
4117 kmem_cache_reap_soon(hdr_l2only_cache);
4118 kmem_cache_reap_soon(range_seg_cache);
4119
4120 if (zio_arena != NULL) {
4121 /*
4122 * Ask the vmem arena to reclaim unused memory from its
4123 * quantum caches.
4124 */
4125 vmem_qcache_reap(zio_arena);
4126 }
4127 }
4128
4129 /*
4130 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4131 * enough data and signal them to proceed. When this happens, the threads in
4132 * arc_get_data_impl() are sleeping while holding the hash lock for their
4133 * particular arc header. Thus, we must be careful to never sleep on a
4134 * hash lock in this thread. This is to prevent the following deadlock:
4135 *
4136 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4137 * waiting for the reclaim thread to signal it.
4138 *
4139 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4140 * fails, and goes to sleep forever.
4141 *
4142 * This possible deadlock is avoided by always acquiring a hash lock
4143 * using mutex_tryenter() from arc_reclaim_thread().
4144 */
4145 /* ARGSUSED */
4146 static void
4147 arc_reclaim_thread(void *unused)
4148 {
4149 hrtime_t growtime = 0;
4150 hrtime_t kmem_reap_time = 0;
4151 callb_cpr_t cpr;
4152
4153 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4154
4155 mutex_enter(&arc_reclaim_lock);
4156 while (!arc_reclaim_thread_exit) {
4157 uint64_t evicted = 0;
4158
4159 /*
4160 * This is necessary in order for the mdb ::arc dcmd to
4161 * show up to date information. Since the ::arc command
4162 * does not call the kstat's update function, without
4163 * this call, the command may show stale stats for the
4164 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4165 * with this change, the data might be up to 1 second
4166 * out of date; but that should suffice. The arc_state_t
4167 * structures can be queried directly if more accurate
4168 * information is needed.
4169 */
4170 if (arc_ksp != NULL)
4171 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4172
4173 mutex_exit(&arc_reclaim_lock);
4174
4175 /*
4176 * We call arc_adjust() before (possibly) calling
4177 * arc_kmem_reap_now(), so that we can wake up
4178 * arc_get_data_impl() sooner.
4179 */
4180 evicted = arc_adjust();
4181
4182 int64_t free_memory = arc_available_memory();
4183 if (free_memory < 0) {
4184 hrtime_t curtime = gethrtime();
4185 arc_no_grow = B_TRUE;
4186 arc_warm = B_TRUE;
4187
4188 /*
4189 * Wait at least zfs_grow_retry (default 60) seconds
4190 * before considering growing.
4191 */
4192 growtime = curtime + SEC2NSEC(arc_grow_retry);
4193
4194 /*
4195 * Wait at least arc_kmem_cache_reap_retry_ms
4196 * between arc_kmem_reap_now() calls. Without
4197 * this check it is possible to end up in a
4198 * situation where we spend lots of time
4199 * reaping caches, while we're near arc_c_min.
4200 */
4201 if (curtime >= kmem_reap_time) {
4202 arc_kmem_reap_now();
4203 kmem_reap_time = gethrtime() +
4204 MSEC2NSEC(arc_kmem_cache_reap_retry_ms);
4205 }
4206
4207 /*
4208 * If we are still low on memory, shrink the ARC
4209 * so that we have arc_shrink_min free space.
4210 */
4211 free_memory = arc_available_memory();
4212
4213 int64_t to_free =
4214 (arc_c >> arc_shrink_shift) - free_memory;
4215 if (to_free > 0) {
4216 #ifdef _KERNEL
4217 to_free = MAX(to_free, ptob(needfree));
4218 #endif
4219 arc_shrink(to_free);
4220 }
4221 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4222 arc_no_grow = B_TRUE;
4223 } else if (gethrtime() >= growtime) {
4224 arc_no_grow = B_FALSE;
4225 }
4226
4227 mutex_enter(&arc_reclaim_lock);
4228
4229 /*
4230 * If evicted is zero, we couldn't evict anything via
4231 * arc_adjust(). This could be due to hash lock
4232 * collisions, but more likely due to the majority of
4233 * arc buffers being unevictable. Therefore, even if
4234 * arc_size is above arc_c, another pass is unlikely to
4235 * be helpful and could potentially cause us to enter an
4236 * infinite loop.
4237 */
4238 if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) {
4239 /*
4240 * We're either no longer overflowing, or we
4241 * can't evict anything more, so we should wake
4242 * up any threads before we go to sleep.
4243 */
4244 cv_broadcast(&arc_reclaim_waiters_cv);
4245
4246 /*
4247 * Block until signaled, or after one second (we
4248 * might need to perform arc_kmem_reap_now()
4249 * even if we aren't being signalled)
4250 */
4251 CALLB_CPR_SAFE_BEGIN(&cpr);
4252 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4253 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4254 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4255 }
4256 }
4257
4258 arc_reclaim_thread_exit = B_FALSE;
4259 cv_broadcast(&arc_reclaim_thread_cv);
4260 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4261 thread_exit();
4262 }
4263
4264 /*
4265 * Adapt arc info given the number of bytes we are trying to add and
4266 * the state that we are comming from. This function is only called
4267 * when we are adding new content to the cache.
4268 */
4269 static void
4270 arc_adapt(int bytes, arc_state_t *state)
4271 {
4272 int mult;
4273 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4274 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
4275 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
4276
4277 if (state == arc_l2c_only)
4278 return;
4279
4280 ASSERT(bytes > 0);
4281 /*
4282 * Adapt the target size of the MRU list:
4283 * - if we just hit in the MRU ghost list, then increase
4284 * the target size of the MRU list.
4285 * - if we just hit in the MFU ghost list, then increase
4286 * the target size of the MFU list by decreasing the
4287 * target size of the MRU list.
4288 */
4289 if (state == arc_mru_ghost) {
4290 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4291 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4292
4293 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4294 } else if (state == arc_mfu_ghost) {
4295 uint64_t delta;
4296
4297 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4298 mult = MIN(mult, 10);
4299
4300 delta = MIN(bytes * mult, arc_p);
4301 arc_p = MAX(arc_p_min, arc_p - delta);
4302 }
4303 ASSERT((int64_t)arc_p >= 0);
4304
4305 if (arc_reclaim_needed()) {
4306 cv_signal(&arc_reclaim_thread_cv);
4307 return;
4308 }
4309
4310 if (arc_no_grow)
4311 return;
4312
4313 if (arc_c >= arc_c_max)
4314 return;
4315
4316 /*
4317 * If we're within (2 * maxblocksize) bytes of the target
4318 * cache size, increment the target cache size
4319 */
4320 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
4321 0) {
4322 atomic_add_64(&arc_c, (int64_t)bytes);
4323 if (arc_c > arc_c_max)
4324 arc_c = arc_c_max;
4325 else if (state == arc_anon)
4326 atomic_add_64(&arc_p, (int64_t)bytes);
4327 if (arc_p > arc_c)
4328 arc_p = arc_c;
4329 }
4330 ASSERT((int64_t)arc_p >= 0);
4331 }
4332
4333 /*
4334 * Check if arc_size has grown past our upper threshold, determined by
4335 * zfs_arc_overflow_shift.
4336 */
4337 static boolean_t
4338 arc_is_overflowing(void)
4339 {
4340 /* Always allow at least one block of overflow */
4341 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4342 arc_c >> zfs_arc_overflow_shift);
4343
4344 /*
4345 * We just compare the lower bound here for performance reasons. Our
4346 * primary goals are to make sure that the arc never grows without
4347 * bound, and that it can reach its maximum size. This check
4348 * accomplishes both goals. The maximum amount we could run over by is
4349 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4350 * in the ARC. In practice, that's in the tens of MB, which is low
4351 * enough to be safe.
4352 */
4353 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
4354 }
4355
4356 static abd_t *
4357 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4358 {
4359 arc_buf_contents_t type = arc_buf_type(hdr);
4360
4361 arc_get_data_impl(hdr, size, tag);
4362 if (type == ARC_BUFC_METADATA) {
4363 return (abd_alloc(size, B_TRUE));
4364 } else {
4365 ASSERT(type == ARC_BUFC_DATA);
4366 return (abd_alloc(size, B_FALSE));
4367 }
4368 }
4369
4370 static void *
4371 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4372 {
4373 arc_buf_contents_t type = arc_buf_type(hdr);
4374
4375 arc_get_data_impl(hdr, size, tag);
4376 if (type == ARC_BUFC_METADATA) {
4377 return (zio_buf_alloc(size));
4378 } else {
4379 ASSERT(type == ARC_BUFC_DATA);
4380 return (zio_data_buf_alloc(size));
4381 }
4382 }
4383
4384 /*
4385 * Allocate a block and return it to the caller. If we are hitting the
4386 * hard limit for the cache size, we must sleep, waiting for the eviction
4387 * thread to catch up. If we're past the target size but below the hard
4388 * limit, we'll only signal the reclaim thread and continue on.
4389 */
4390 static void
4391 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4392 {
4393 arc_state_t *state = hdr->b_l1hdr.b_state;
4394 arc_buf_contents_t type = arc_buf_type(hdr);
4395
4396 arc_adapt(size, state);
4397
4398 /*
4399 * If arc_size is currently overflowing, and has grown past our
4400 * upper limit, we must be adding data faster than the evict
4401 * thread can evict. Thus, to ensure we don't compound the
4402 * problem by adding more data and forcing arc_size to grow even
4403 * further past it's target size, we halt and wait for the
4404 * eviction thread to catch up.
4405 *
4406 * It's also possible that the reclaim thread is unable to evict
4407 * enough buffers to get arc_size below the overflow limit (e.g.
4408 * due to buffers being un-evictable, or hash lock collisions).
4409 * In this case, we want to proceed regardless if we're
4410 * overflowing; thus we don't use a while loop here.
4411 */
4412 if (arc_is_overflowing()) {
4413 mutex_enter(&arc_reclaim_lock);
4414
4415 /*
4416 * Now that we've acquired the lock, we may no longer be
4417 * over the overflow limit, lets check.
4418 *
4419 * We're ignoring the case of spurious wake ups. If that
4420 * were to happen, it'd let this thread consume an ARC
4421 * buffer before it should have (i.e. before we're under
4422 * the overflow limit and were signalled by the reclaim
4423 * thread). As long as that is a rare occurrence, it
4424 * shouldn't cause any harm.
4425 */
4426 if (arc_is_overflowing()) {
4427 cv_signal(&arc_reclaim_thread_cv);
4428 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4429 }
4430
4431 mutex_exit(&arc_reclaim_lock);
4432 }
4433
4434 VERIFY3U(hdr->b_type, ==, type);
4435 if (type == ARC_BUFC_METADATA) {
4436 arc_space_consume(size, ARC_SPACE_META);
4437 } else {
4438 arc_space_consume(size, ARC_SPACE_DATA);
4439 }
4440
4441 /*
4442 * Update the state size. Note that ghost states have a
4443 * "ghost size" and so don't need to be updated.
4444 */
4445 if (!GHOST_STATE(state)) {
4446
4447 (void) refcount_add_many(&state->arcs_size, size, tag);
4448
4449 /*
4450 * If this is reached via arc_read, the link is
4451 * protected by the hash lock. If reached via
4452 * arc_buf_alloc, the header should not be accessed by
4453 * any other thread. And, if reached via arc_read_done,
4454 * the hash lock will protect it if it's found in the
4455 * hash table; otherwise no other thread should be
4456 * trying to [add|remove]_reference it.
4457 */
4458 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4459 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4460 (void) refcount_add_many(&state->arcs_esize[type],
4461 size, tag);
4462 }
4463
4464 /*
4465 * If we are growing the cache, and we are adding anonymous
4466 * data, and we have outgrown arc_p, update arc_p
4467 */
4468 if (aggsum_compare(&arc_size, arc_c) < 0 &&
4469 hdr->b_l1hdr.b_state == arc_anon &&
4470 (refcount_count(&arc_anon->arcs_size) +
4471 refcount_count(&arc_mru->arcs_size) > arc_p))
4472 arc_p = MIN(arc_c, arc_p + size);
4473 }
4474 }
4475
4476 static void
4477 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
4478 {
4479 arc_free_data_impl(hdr, size, tag);
4480 abd_free(abd);
4481 }
4482
4483 static void
4484 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
4485 {
4486 arc_buf_contents_t type = arc_buf_type(hdr);
4487
4488 arc_free_data_impl(hdr, size, tag);
4489 if (type == ARC_BUFC_METADATA) {
4490 zio_buf_free(buf, size);
4491 } else {
4492 ASSERT(type == ARC_BUFC_DATA);
4493 zio_data_buf_free(buf, size);
4494 }
4495 }
4496
4497 /*
4498 * Free the arc data buffer.
4499 */
4500 static void
4501 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4502 {
4503 arc_state_t *state = hdr->b_l1hdr.b_state;
4504 arc_buf_contents_t type = arc_buf_type(hdr);
4505
4506 /* protected by hash lock, if in the hash table */
4507 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4508 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4509 ASSERT(state != arc_anon && state != arc_l2c_only);
4510
4511 (void) refcount_remove_many(&state->arcs_esize[type],
4512 size, tag);
4513 }
4514 (void) refcount_remove_many(&state->arcs_size, size, tag);
4515
4516 VERIFY3U(hdr->b_type, ==, type);
4517 if (type == ARC_BUFC_METADATA) {
4518 arc_space_return(size, ARC_SPACE_META);
4519 } else {
4520 ASSERT(type == ARC_BUFC_DATA);
4521 arc_space_return(size, ARC_SPACE_DATA);
4522 }
4523 }
4524
4525 /*
4526 * This routine is called whenever a buffer is accessed.
4527 * NOTE: the hash lock is dropped in this function.
4528 */
4529 static void
4530 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4531 {
4532 clock_t now;
4533
4534 ASSERT(MUTEX_HELD(hash_lock));
4535 ASSERT(HDR_HAS_L1HDR(hdr));
4536
4537 if (hdr->b_l1hdr.b_state == arc_anon) {
4538 /*
4539 * This buffer is not in the cache, and does not
4540 * appear in our "ghost" list. Add the new buffer
4541 * to the MRU state.
4542 */
4543
4544 ASSERT0(hdr->b_l1hdr.b_arc_access);
4545 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4546 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4547 arc_change_state(arc_mru, hdr, hash_lock);
4548
4549 } else if (hdr->b_l1hdr.b_state == arc_mru) {
4550 now = ddi_get_lbolt();
4551
4552 /*
4553 * If this buffer is here because of a prefetch, then either:
4554 * - clear the flag if this is a "referencing" read
4555 * (any subsequent access will bump this into the MFU state).
4556 * or
4557 * - move the buffer to the head of the list if this is
4558 * another prefetch (to make it less likely to be evicted).
4559 */
4560 if (HDR_PREFETCH(hdr)) {
4561 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4562 /* link protected by hash lock */
4563 ASSERT(multilist_link_active(
4564 &hdr->b_l1hdr.b_arc_node));
4565 } else {
4566 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4567 ARCSTAT_BUMP(arcstat_mru_hits);
4568 }
4569 hdr->b_l1hdr.b_arc_access = now;
4570 return;
4571 }
4572
4573 /*
4574 * This buffer has been "accessed" only once so far,
4575 * but it is still in the cache. Move it to the MFU
4576 * state.
4577 */
4578 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4579 /*
4580 * More than 125ms have passed since we
4581 * instantiated this buffer. Move it to the
4582 * most frequently used state.
4583 */
4584 hdr->b_l1hdr.b_arc_access = now;
4585 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4586 arc_change_state(arc_mfu, hdr, hash_lock);
4587 }
4588 ARCSTAT_BUMP(arcstat_mru_hits);
4589 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
4590 arc_state_t *new_state;
4591 /*
4592 * This buffer has been "accessed" recently, but
4593 * was evicted from the cache. Move it to the
4594 * MFU state.
4595 */
4596
4597 if (HDR_PREFETCH(hdr)) {
4598 new_state = arc_mru;
4599 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
4600 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4601 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4602 } else {
4603 new_state = arc_mfu;
4604 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4605 }
4606
4607 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4608 arc_change_state(new_state, hdr, hash_lock);
4609
4610 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
4611 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
4612 /*
4613 * This buffer has been accessed more than once and is
4614 * still in the cache. Keep it in the MFU state.
4615 *
4616 * NOTE: an add_reference() that occurred when we did
4617 * the arc_read() will have kicked this off the list.
4618 * If it was a prefetch, we will explicitly move it to
4619 * the head of the list now.
4620 */
4621 if ((HDR_PREFETCH(hdr)) != 0) {
4622 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4623 /* link protected by hash_lock */
4624 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4625 }
4626 ARCSTAT_BUMP(arcstat_mfu_hits);
4627 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4628 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4629 arc_state_t *new_state = arc_mfu;
4630 /*
4631 * This buffer has been accessed more than once but has
4632 * been evicted from the cache. Move it back to the
4633 * MFU state.
4634 */
4635
4636 if (HDR_PREFETCH(hdr)) {
4637 /*
4638 * This is a prefetch access...
4639 * move this block back to the MRU state.
4640 */
4641 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4642 new_state = arc_mru;
4643 }
4644
4645 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4646 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4647 arc_change_state(new_state, hdr, hash_lock);
4648
4649 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4650 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4651 /*
4652 * This buffer is on the 2nd Level ARC.
4653 */
4654
4655 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4656 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4657 arc_change_state(arc_mfu, hdr, hash_lock);
4658 } else {
4659 ASSERT(!"invalid arc state");
4660 }
4661 }
4662
4663 /* a generic arc_done_func_t which you can use */
4664 /* ARGSUSED */
4665 void
4666 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4667 {
4668 if (zio == NULL || zio->io_error == 0)
4669 bcopy(buf->b_data, arg, arc_buf_size(buf));
4670 arc_buf_destroy(buf, arg);
4671 }
4672
4673 /* a generic arc_done_func_t */
4674 void
4675 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
4676 {
4677 arc_buf_t **bufp = arg;
4678 if (zio && zio->io_error) {
4679 arc_buf_destroy(buf, arg);
4680 *bufp = NULL;
4681 } else {
4682 *bufp = buf;
4683 ASSERT(buf->b_data);
4684 }
4685 }
4686
4687 static void
4688 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
4689 {
4690 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
4691 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
4692 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
4693 } else {
4694 if (HDR_COMPRESSION_ENABLED(hdr)) {
4695 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
4696 BP_GET_COMPRESS(bp));
4697 }
4698 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
4699 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
4700 }
4701 }
4702
4703 static void
4704 arc_read_done(zio_t *zio)
4705 {
4706 arc_buf_hdr_t *hdr = zio->io_private;
4707 kmutex_t *hash_lock = NULL;
4708 arc_callback_t *callback_list;
4709 arc_callback_t *acb;
4710 boolean_t freeable = B_FALSE;
4711 boolean_t no_zio_error = (zio->io_error == 0);
4712
4713 /*
4714 * The hdr was inserted into hash-table and removed from lists
4715 * prior to starting I/O. We should find this header, since
4716 * it's in the hash table, and it should be legit since it's
4717 * not possible to evict it during the I/O. The only possible
4718 * reason for it not to be found is if we were freed during the
4719 * read.
4720 */
4721 if (HDR_IN_HASH_TABLE(hdr)) {
4722 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
4723 ASSERT3U(hdr->b_dva.dva_word[0], ==,
4724 BP_IDENTITY(zio->io_bp)->dva_word[0]);
4725 ASSERT3U(hdr->b_dva.dva_word[1], ==,
4726 BP_IDENTITY(zio->io_bp)->dva_word[1]);
4727
4728 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
4729 &hash_lock);
4730
4731 ASSERT((found == hdr &&
4732 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
4733 (found == hdr && HDR_L2_READING(hdr)));
4734 ASSERT3P(hash_lock, !=, NULL);
4735 }
4736
4737 if (no_zio_error) {
4738 /* byteswap if necessary */
4739 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
4740 if (BP_GET_LEVEL(zio->io_bp) > 0) {
4741 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
4742 } else {
4743 hdr->b_l1hdr.b_byteswap =
4744 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
4745 }
4746 } else {
4747 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
4748 }
4749 }
4750
4751 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
4752 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
4753 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
4754
4755 callback_list = hdr->b_l1hdr.b_acb;
4756 ASSERT3P(callback_list, !=, NULL);
4757
4758 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
4759 /*
4760 * Only call arc_access on anonymous buffers. This is because
4761 * if we've issued an I/O for an evicted buffer, we've already
4762 * called arc_access (to prevent any simultaneous readers from
4763 * getting confused).
4764 */
4765 arc_access(hdr, hash_lock);
4766 }
4767
4768 /*
4769 * If a read request has a callback (i.e. acb_done is not NULL), then we
4770 * make a buf containing the data according to the parameters which were
4771 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
4772 * aren't needlessly decompressing the data multiple times.
4773 */
4774 int callback_cnt = 0;
4775 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
4776 if (!acb->acb_done)
4777 continue;
4778
4779 /* This is a demand read since prefetches don't use callbacks */
4780 callback_cnt++;
4781
4782 int error = arc_buf_alloc_impl(hdr, acb->acb_private,
4783 acb->acb_compressed, no_zio_error, &acb->acb_buf);
4784 if (no_zio_error) {
4785 zio->io_error = error;
4786 }
4787 }
4788 hdr->b_l1hdr.b_acb = NULL;
4789 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
4790 if (callback_cnt == 0) {
4791 ASSERT(HDR_PREFETCH(hdr));
4792 ASSERT0(hdr->b_l1hdr.b_bufcnt);
4793 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
4794 }
4795
4796 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
4797 callback_list != NULL);
4798
4799 if (no_zio_error) {
4800 arc_hdr_verify(hdr, zio->io_bp);
4801 } else {
4802 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
4803 if (hdr->b_l1hdr.b_state != arc_anon)
4804 arc_change_state(arc_anon, hdr, hash_lock);
4805 if (HDR_IN_HASH_TABLE(hdr))
4806 buf_hash_remove(hdr);
4807 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4808 }
4809
4810 /*
4811 * Broadcast before we drop the hash_lock to avoid the possibility
4812 * that the hdr (and hence the cv) might be freed before we get to
4813 * the cv_broadcast().
4814 */
4815 cv_broadcast(&hdr->b_l1hdr.b_cv);
4816
4817 if (hash_lock != NULL) {
4818 mutex_exit(hash_lock);
4819 } else {
4820 /*
4821 * This block was freed while we waited for the read to
4822 * complete. It has been removed from the hash table and
4823 * moved to the anonymous state (so that it won't show up
4824 * in the cache).
4825 */
4826 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
4827 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4828 }
4829
4830 /* execute each callback and free its structure */
4831 while ((acb = callback_list) != NULL) {
4832 if (acb->acb_done)
4833 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
4834
4835 if (acb->acb_zio_dummy != NULL) {
4836 acb->acb_zio_dummy->io_error = zio->io_error;
4837 zio_nowait(acb->acb_zio_dummy);
4838 }
4839
4840 callback_list = acb->acb_next;
4841 kmem_free(acb, sizeof (arc_callback_t));
4842 }
4843
4844 if (freeable)
4845 arc_hdr_destroy(hdr);
4846 }
4847
4848 /*
4849 * "Read" the block at the specified DVA (in bp) via the
4850 * cache. If the block is found in the cache, invoke the provided
4851 * callback immediately and return. Note that the `zio' parameter
4852 * in the callback will be NULL in this case, since no IO was
4853 * required. If the block is not in the cache pass the read request
4854 * on to the spa with a substitute callback function, so that the
4855 * requested block will be added to the cache.
4856 *
4857 * If a read request arrives for a block that has a read in-progress,
4858 * either wait for the in-progress read to complete (and return the
4859 * results); or, if this is a read with a "done" func, add a record
4860 * to the read to invoke the "done" func when the read completes,
4861 * and return; or just return.
4862 *
4863 * arc_read_done() will invoke all the requested "done" functions
4864 * for readers of this block.
4865 */
4866 int
4867 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
4868 void *private, zio_priority_t priority, int zio_flags,
4869 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
4870 {
4871 arc_buf_hdr_t *hdr = NULL;
4872 kmutex_t *hash_lock = NULL;
4873 zio_t *rzio;
4874 uint64_t guid = spa_load_guid(spa);
4875 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
4876
4877 ASSERT(!BP_IS_EMBEDDED(bp) ||
4878 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
4879
4880 top:
4881 if (!BP_IS_EMBEDDED(bp)) {
4882 /*
4883 * Embedded BP's have no DVA and require no I/O to "read".
4884 * Create an anonymous arc buf to back it.
4885 */
4886 hdr = buf_hash_find(guid, bp, &hash_lock);
4887 }
4888
4889 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
4890 arc_buf_t *buf = NULL;
4891 *arc_flags |= ARC_FLAG_CACHED;
4892
4893 if (HDR_IO_IN_PROGRESS(hdr)) {
4894
4895 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
4896 priority == ZIO_PRIORITY_SYNC_READ) {
4897 /*
4898 * This sync read must wait for an
4899 * in-progress async read (e.g. a predictive
4900 * prefetch). Async reads are queued
4901 * separately at the vdev_queue layer, so
4902 * this is a form of priority inversion.
4903 * Ideally, we would "inherit" the demand
4904 * i/o's priority by moving the i/o from
4905 * the async queue to the synchronous queue,
4906 * but there is currently no mechanism to do
4907 * so. Track this so that we can evaluate
4908 * the magnitude of this potential performance
4909 * problem.
4910 *
4911 * Note that if the prefetch i/o is already
4912 * active (has been issued to the device),
4913 * the prefetch improved performance, because
4914 * we issued it sooner than we would have
4915 * without the prefetch.
4916 */
4917 DTRACE_PROBE1(arc__sync__wait__for__async,
4918 arc_buf_hdr_t *, hdr);
4919 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
4920 }
4921 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4922 arc_hdr_clear_flags(hdr,
4923 ARC_FLAG_PREDICTIVE_PREFETCH);
4924 }
4925
4926 if (*arc_flags & ARC_FLAG_WAIT) {
4927 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4928 mutex_exit(hash_lock);
4929 goto top;
4930 }
4931 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4932
4933 if (done) {
4934 arc_callback_t *acb = NULL;
4935
4936 acb = kmem_zalloc(sizeof (arc_callback_t),
4937 KM_SLEEP);
4938 acb->acb_done = done;
4939 acb->acb_private = private;
4940 acb->acb_compressed = compressed_read;
4941 if (pio != NULL)
4942 acb->acb_zio_dummy = zio_null(pio,
4943 spa, NULL, NULL, NULL, zio_flags);
4944
4945 ASSERT3P(acb->acb_done, !=, NULL);
4946 acb->acb_next = hdr->b_l1hdr.b_acb;
4947 hdr->b_l1hdr.b_acb = acb;
4948 mutex_exit(hash_lock);
4949 return (0);
4950 }
4951 mutex_exit(hash_lock);
4952 return (0);
4953 }
4954
4955 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4956 hdr->b_l1hdr.b_state == arc_mfu);
4957
4958 if (done) {
4959 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4960 /*
4961 * This is a demand read which does not have to
4962 * wait for i/o because we did a predictive
4963 * prefetch i/o for it, which has completed.
4964 */
4965 DTRACE_PROBE1(
4966 arc__demand__hit__predictive__prefetch,
4967 arc_buf_hdr_t *, hdr);
4968 ARCSTAT_BUMP(
4969 arcstat_demand_hit_predictive_prefetch);
4970 arc_hdr_clear_flags(hdr,
4971 ARC_FLAG_PREDICTIVE_PREFETCH);
4972 }
4973 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
4974
4975 /* Get a buf with the desired data in it. */
4976 VERIFY0(arc_buf_alloc_impl(hdr, private,
4977 compressed_read, B_TRUE, &buf));
4978 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4979 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4980 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
4981 }
4982 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4983 arc_access(hdr, hash_lock);
4984 if (*arc_flags & ARC_FLAG_L2CACHE)
4985 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
4986 mutex_exit(hash_lock);
4987 ARCSTAT_BUMP(arcstat_hits);
4988 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4989 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4990 data, metadata, hits);
4991
4992 if (done)
4993 done(NULL, buf, private);
4994 } else {
4995 uint64_t lsize = BP_GET_LSIZE(bp);
4996 uint64_t psize = BP_GET_PSIZE(bp);
4997 arc_callback_t *acb;
4998 vdev_t *vd = NULL;
4999 uint64_t addr = 0;
5000 boolean_t devw = B_FALSE;
5001 uint64_t size;
5002
5003 if (hdr == NULL) {
5004 /* this block is not in the cache */
5005 arc_buf_hdr_t *exists = NULL;
5006 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5007 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5008 BP_GET_COMPRESS(bp), type);
5009
5010 if (!BP_IS_EMBEDDED(bp)) {
5011 hdr->b_dva = *BP_IDENTITY(bp);
5012 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5013 exists = buf_hash_insert(hdr, &hash_lock);
5014 }
5015 if (exists != NULL) {
5016 /* somebody beat us to the hash insert */
5017 mutex_exit(hash_lock);
5018 buf_discard_identity(hdr);
5019 arc_hdr_destroy(hdr);
5020 goto top; /* restart the IO request */
5021 }
5022 } else {
5023 /*
5024 * This block is in the ghost cache. If it was L2-only
5025 * (and thus didn't have an L1 hdr), we realloc the
5026 * header to add an L1 hdr.
5027 */
5028 if (!HDR_HAS_L1HDR(hdr)) {
5029 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5030 hdr_full_cache);
5031 }
5032 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5033 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5034 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5035 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5036 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5037 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5038
5039 /*
5040 * This is a delicate dance that we play here.
5041 * This hdr is in the ghost list so we access it
5042 * to move it out of the ghost list before we
5043 * initiate the read. If it's a prefetch then
5044 * it won't have a callback so we'll remove the
5045 * reference that arc_buf_alloc_impl() created. We
5046 * do this after we've called arc_access() to
5047 * avoid hitting an assert in remove_reference().
5048 */
5049 arc_access(hdr, hash_lock);
5050 arc_hdr_alloc_pabd(hdr);
5051 }
5052 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5053 size = arc_hdr_size(hdr);
5054
5055 /*
5056 * If compression is enabled on the hdr, then will do
5057 * RAW I/O and will store the compressed data in the hdr's
5058 * data block. Otherwise, the hdr's data block will contain
5059 * the uncompressed data.
5060 */
5061 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5062 zio_flags |= ZIO_FLAG_RAW;
5063 }
5064
5065 if (*arc_flags & ARC_FLAG_PREFETCH)
5066 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5067 if (*arc_flags & ARC_FLAG_L2CACHE)
5068 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5069 if (BP_GET_LEVEL(bp) > 0)
5070 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5071 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5072 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5073 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5074
5075 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5076 acb->acb_done = done;
5077 acb->acb_private = private;
5078 acb->acb_compressed = compressed_read;
5079
5080 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5081 hdr->b_l1hdr.b_acb = acb;
5082 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5083
5084 if (HDR_HAS_L2HDR(hdr) &&
5085 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5086 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5087 addr = hdr->b_l2hdr.b_daddr;
5088 /*
5089 * Lock out L2ARC device removal.
5090 */
5091 if (vdev_is_dead(vd) ||
5092 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5093 vd = NULL;
5094 }
5095
5096 if (priority == ZIO_PRIORITY_ASYNC_READ)
5097 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5098 else
5099 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5100
5101 if (hash_lock != NULL)
5102 mutex_exit(hash_lock);
5103
5104 /*
5105 * At this point, we have a level 1 cache miss. Try again in
5106 * L2ARC if possible.
5107 */
5108 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5109
5110 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5111 uint64_t, lsize, zbookmark_phys_t *, zb);
5112 ARCSTAT_BUMP(arcstat_misses);
5113 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5114 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5115 data, metadata, misses);
5116
5117 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5118 /*
5119 * Read from the L2ARC if the following are true:
5120 * 1. The L2ARC vdev was previously cached.
5121 * 2. This buffer still has L2ARC metadata.
5122 * 3. This buffer isn't currently writing to the L2ARC.
5123 * 4. The L2ARC entry wasn't evicted, which may
5124 * also have invalidated the vdev.
5125 * 5. This isn't prefetch and l2arc_noprefetch is set.
5126 */
5127 if (HDR_HAS_L2HDR(hdr) &&
5128 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5129 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5130 l2arc_read_callback_t *cb;
5131 abd_t *abd;
5132 uint64_t asize;
5133
5134 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5135 ARCSTAT_BUMP(arcstat_l2_hits);
5136
5137 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5138 KM_SLEEP);
5139 cb->l2rcb_hdr = hdr;
5140 cb->l2rcb_bp = *bp;
5141 cb->l2rcb_zb = *zb;
5142 cb->l2rcb_flags = zio_flags;
5143
5144 asize = vdev_psize_to_asize(vd, size);
5145 if (asize != size) {
5146 abd = abd_alloc_for_io(asize,
5147 HDR_ISTYPE_METADATA(hdr));
5148 cb->l2rcb_abd = abd;
5149 } else {
5150 abd = hdr->b_l1hdr.b_pabd;
5151 }
5152
5153 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5154 addr + asize <= vd->vdev_psize -
5155 VDEV_LABEL_END_SIZE);
5156
5157 /*
5158 * l2arc read. The SCL_L2ARC lock will be
5159 * released by l2arc_read_done().
5160 * Issue a null zio if the underlying buffer
5161 * was squashed to zero size by compression.
5162 */
5163 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5164 ZIO_COMPRESS_EMPTY);
5165 rzio = zio_read_phys(pio, vd, addr,
5166 asize, abd,
5167 ZIO_CHECKSUM_OFF,
5168 l2arc_read_done, cb, priority,
5169 zio_flags | ZIO_FLAG_DONT_CACHE |
5170 ZIO_FLAG_CANFAIL |
5171 ZIO_FLAG_DONT_PROPAGATE |
5172 ZIO_FLAG_DONT_RETRY, B_FALSE);
5173 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5174 zio_t *, rzio);
5175 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5176
5177 if (*arc_flags & ARC_FLAG_NOWAIT) {
5178 zio_nowait(rzio);
5179 return (0);
5180 }
5181
5182 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5183 if (zio_wait(rzio) == 0)
5184 return (0);
5185
5186 /* l2arc read error; goto zio_read() */
5187 } else {
5188 DTRACE_PROBE1(l2arc__miss,
5189 arc_buf_hdr_t *, hdr);
5190 ARCSTAT_BUMP(arcstat_l2_misses);
5191 if (HDR_L2_WRITING(hdr))
5192 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5193 spa_config_exit(spa, SCL_L2ARC, vd);
5194 }
5195 } else {
5196 if (vd != NULL)
5197 spa_config_exit(spa, SCL_L2ARC, vd);
5198 if (l2arc_ndev != 0) {
5199 DTRACE_PROBE1(l2arc__miss,
5200 arc_buf_hdr_t *, hdr);
5201 ARCSTAT_BUMP(arcstat_l2_misses);
5202 }
5203 }
5204
5205 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
5206 arc_read_done, hdr, priority, zio_flags, zb);
5207
5208 if (*arc_flags & ARC_FLAG_WAIT)
5209 return (zio_wait(rzio));
5210
5211 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5212 zio_nowait(rzio);
5213 }
5214 return (0);
5215 }
5216
5217 /*
5218 * Notify the arc that a block was freed, and thus will never be used again.
5219 */
5220 void
5221 arc_freed(spa_t *spa, const blkptr_t *bp)
5222 {
5223 arc_buf_hdr_t *hdr;
5224 kmutex_t *hash_lock;
5225 uint64_t guid = spa_load_guid(spa);
5226
5227 ASSERT(!BP_IS_EMBEDDED(bp));
5228
5229 hdr = buf_hash_find(guid, bp, &hash_lock);
5230 if (hdr == NULL)
5231 return;
5232
5233 /*
5234 * We might be trying to free a block that is still doing I/O
5235 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5236 * dmu_sync-ed block). If this block is being prefetched, then it
5237 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5238 * until the I/O completes. A block may also have a reference if it is
5239 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5240 * have written the new block to its final resting place on disk but
5241 * without the dedup flag set. This would have left the hdr in the MRU
5242 * state and discoverable. When the txg finally syncs it detects that
5243 * the block was overridden in open context and issues an override I/O.
5244 * Since this is a dedup block, the override I/O will determine if the
5245 * block is already in the DDT. If so, then it will replace the io_bp
5246 * with the bp from the DDT and allow the I/O to finish. When the I/O
5247 * reaches the done callback, dbuf_write_override_done, it will
5248 * check to see if the io_bp and io_bp_override are identical.
5249 * If they are not, then it indicates that the bp was replaced with
5250 * the bp in the DDT and the override bp is freed. This allows
5251 * us to arrive here with a reference on a block that is being
5252 * freed. So if we have an I/O in progress, or a reference to
5253 * this hdr, then we don't destroy the hdr.
5254 */
5255 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5256 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5257 arc_change_state(arc_anon, hdr, hash_lock);
5258 arc_hdr_destroy(hdr);
5259 mutex_exit(hash_lock);
5260 } else {
5261 mutex_exit(hash_lock);
5262 }
5263
5264 }
5265
5266 /*
5267 * Release this buffer from the cache, making it an anonymous buffer. This
5268 * must be done after a read and prior to modifying the buffer contents.
5269 * If the buffer has more than one reference, we must make
5270 * a new hdr for the buffer.
5271 */
5272 void
5273 arc_release(arc_buf_t *buf, void *tag)
5274 {
5275 arc_buf_hdr_t *hdr = buf->b_hdr;
5276
5277 /*
5278 * It would be nice to assert that if it's DMU metadata (level >
5279 * 0 || it's the dnode file), then it must be syncing context.
5280 * But we don't know that information at this level.
5281 */
5282
5283 mutex_enter(&buf->b_evict_lock);
5284
5285 ASSERT(HDR_HAS_L1HDR(hdr));
5286
5287 /*
5288 * We don't grab the hash lock prior to this check, because if
5289 * the buffer's header is in the arc_anon state, it won't be
5290 * linked into the hash table.
5291 */
5292 if (hdr->b_l1hdr.b_state == arc_anon) {
5293 mutex_exit(&buf->b_evict_lock);
5294 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5295 ASSERT(!HDR_IN_HASH_TABLE(hdr));
5296 ASSERT(!HDR_HAS_L2HDR(hdr));
5297 ASSERT(HDR_EMPTY(hdr));
5298
5299 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5300 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5301 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5302
5303 hdr->b_l1hdr.b_arc_access = 0;
5304
5305 /*
5306 * If the buf is being overridden then it may already
5307 * have a hdr that is not empty.
5308 */
5309 buf_discard_identity(hdr);
5310 arc_buf_thaw(buf);
5311
5312 return;
5313 }
5314
5315 kmutex_t *hash_lock = HDR_LOCK(hdr);
5316 mutex_enter(hash_lock);
5317
5318 /*
5319 * This assignment is only valid as long as the hash_lock is
5320 * held, we must be careful not to reference state or the
5321 * b_state field after dropping the lock.
5322 */
5323 arc_state_t *state = hdr->b_l1hdr.b_state;
5324 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5325 ASSERT3P(state, !=, arc_anon);
5326
5327 /* this buffer is not on any list */
5328 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
5329
5330 if (HDR_HAS_L2HDR(hdr)) {
5331 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5332
5333 /*
5334 * We have to recheck this conditional again now that
5335 * we're holding the l2ad_mtx to prevent a race with
5336 * another thread which might be concurrently calling
5337 * l2arc_evict(). In that case, l2arc_evict() might have
5338 * destroyed the header's L2 portion as we were waiting
5339 * to acquire the l2ad_mtx.
5340 */
5341 if (HDR_HAS_L2HDR(hdr))
5342 arc_hdr_l2hdr_destroy(hdr);
5343
5344 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5345 }
5346
5347 /*
5348 * Do we have more than one buf?
5349 */
5350 if (hdr->b_l1hdr.b_bufcnt > 1) {
5351 arc_buf_hdr_t *nhdr;
5352 uint64_t spa = hdr->b_spa;
5353 uint64_t psize = HDR_GET_PSIZE(hdr);
5354 uint64_t lsize = HDR_GET_LSIZE(hdr);
5355 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5356 arc_buf_contents_t type = arc_buf_type(hdr);
5357 VERIFY3U(hdr->b_type, ==, type);
5358
5359 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5360 (void) remove_reference(hdr, hash_lock, tag);
5361
5362 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
5363 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5364 ASSERT(ARC_BUF_LAST(buf));
5365 }
5366
5367 /*
5368 * Pull the data off of this hdr and attach it to
5369 * a new anonymous hdr. Also find the last buffer
5370 * in the hdr's buffer list.
5371 */
5372 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
5373 ASSERT3P(lastbuf, !=, NULL);
5374
5375 /*
5376 * If the current arc_buf_t and the hdr are sharing their data
5377 * buffer, then we must stop sharing that block.
5378 */
5379 if (arc_buf_is_shared(buf)) {
5380 VERIFY(!arc_buf_is_shared(lastbuf));
5381
5382 /*
5383 * First, sever the block sharing relationship between
5384 * buf and the arc_buf_hdr_t.
5385 */
5386 arc_unshare_buf(hdr, buf);
5387
5388 /*
5389 * Now we need to recreate the hdr's b_pabd. Since we
5390 * have lastbuf handy, we try to share with it, but if
5391 * we can't then we allocate a new b_pabd and copy the
5392 * data from buf into it.
5393 */
5394 if (arc_can_share(hdr, lastbuf)) {
5395 arc_share_buf(hdr, lastbuf);
5396 } else {
5397 arc_hdr_alloc_pabd(hdr);
5398 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
5399 buf->b_data, psize);
5400 }
5401 VERIFY3P(lastbuf->b_data, !=, NULL);
5402 } else if (HDR_SHARED_DATA(hdr)) {
5403 /*
5404 * Uncompressed shared buffers are always at the end
5405 * of the list. Compressed buffers don't have the
5406 * same requirements. This makes it hard to
5407 * simply assert that the lastbuf is shared so
5408 * we rely on the hdr's compression flags to determine
5409 * if we have a compressed, shared buffer.
5410 */
5411 ASSERT(arc_buf_is_shared(lastbuf) ||
5412 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
5413 ASSERT(!ARC_BUF_SHARED(buf));
5414 }
5415 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5416 ASSERT3P(state, !=, arc_l2c_only);
5417
5418 (void) refcount_remove_many(&state->arcs_size,
5419 arc_buf_size(buf), buf);
5420
5421 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5422 ASSERT3P(state, !=, arc_l2c_only);
5423 (void) refcount_remove_many(&state->arcs_esize[type],
5424 arc_buf_size(buf), buf);
5425 }
5426
5427 hdr->b_l1hdr.b_bufcnt -= 1;
5428 arc_cksum_verify(buf);
5429 arc_buf_unwatch(buf);
5430
5431 mutex_exit(hash_lock);
5432
5433 /*
5434 * Allocate a new hdr. The new hdr will contain a b_pabd
5435 * buffer which will be freed in arc_write().
5436 */
5437 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5438 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5439 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5440 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5441 VERIFY3U(nhdr->b_type, ==, type);
5442 ASSERT(!HDR_SHARED_DATA(nhdr));
5443
5444 nhdr->b_l1hdr.b_buf = buf;
5445 nhdr->b_l1hdr.b_bufcnt = 1;
5446 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5447 buf->b_hdr = nhdr;
5448
5449 mutex_exit(&buf->b_evict_lock);
5450 (void) refcount_add_many(&arc_anon->arcs_size,
5451 arc_buf_size(buf), buf);
5452 } else {
5453 mutex_exit(&buf->b_evict_lock);
5454 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5455 /* protected by hash lock, or hdr is on arc_anon */
5456 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5457 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5458 arc_change_state(arc_anon, hdr, hash_lock);
5459 hdr->b_l1hdr.b_arc_access = 0;
5460 mutex_exit(hash_lock);
5461
5462 buf_discard_identity(hdr);
5463 arc_buf_thaw(buf);
5464 }
5465 }
5466
5467 int
5468 arc_released(arc_buf_t *buf)
5469 {
5470 int released;
5471
5472 mutex_enter(&buf->b_evict_lock);
5473 released = (buf->b_data != NULL &&
5474 buf->b_hdr->b_l1hdr.b_state == arc_anon);
5475 mutex_exit(&buf->b_evict_lock);
5476 return (released);
5477 }
5478
5479 #ifdef ZFS_DEBUG
5480 int
5481 arc_referenced(arc_buf_t *buf)
5482 {
5483 int referenced;
5484
5485 mutex_enter(&buf->b_evict_lock);
5486 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5487 mutex_exit(&buf->b_evict_lock);
5488 return (referenced);
5489 }
5490 #endif
5491
5492 static void
5493 arc_write_ready(zio_t *zio)
5494 {
5495 arc_write_callback_t *callback = zio->io_private;
5496 arc_buf_t *buf = callback->awcb_buf;
5497 arc_buf_hdr_t *hdr = buf->b_hdr;
5498 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
5499
5500 ASSERT(HDR_HAS_L1HDR(hdr));
5501 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5502 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
5503
5504 /*
5505 * If we're reexecuting this zio because the pool suspended, then
5506 * cleanup any state that was previously set the first time the
5507 * callback was invoked.
5508 */
5509 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
5510 arc_cksum_free(hdr);
5511 arc_buf_unwatch(buf);
5512 if (hdr->b_l1hdr.b_pabd != NULL) {
5513 if (arc_buf_is_shared(buf)) {
5514 arc_unshare_buf(hdr, buf);
5515 } else {
5516 arc_hdr_free_pabd(hdr);
5517 }
5518 }
5519 }
5520 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5521 ASSERT(!HDR_SHARED_DATA(hdr));
5522 ASSERT(!arc_buf_is_shared(buf));
5523
5524 callback->awcb_ready(zio, buf, callback->awcb_private);
5525
5526 if (HDR_IO_IN_PROGRESS(hdr))
5527 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
5528
5529 arc_cksum_compute(buf);
5530 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5531
5532 enum zio_compress compress;
5533 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5534 compress = ZIO_COMPRESS_OFF;
5535 } else {
5536 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
5537 compress = BP_GET_COMPRESS(zio->io_bp);
5538 }
5539 HDR_SET_PSIZE(hdr, psize);
5540 arc_hdr_set_compress(hdr, compress);
5541
5542
5543 /*
5544 * Fill the hdr with data. If the hdr is compressed, the data we want
5545 * is available from the zio, otherwise we can take it from the buf.
5546 *
5547 * We might be able to share the buf's data with the hdr here. However,
5548 * doing so would cause the ARC to be full of linear ABDs if we write a
5549 * lot of shareable data. As a compromise, we check whether scattered
5550 * ABDs are allowed, and assume that if they are then the user wants
5551 * the ARC to be primarily filled with them regardless of the data being
5552 * written. Therefore, if they're allowed then we allocate one and copy
5553 * the data into it; otherwise, we share the data directly if we can.
5554 */
5555 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
5556 arc_hdr_alloc_pabd(hdr);
5557
5558 /*
5559 * Ideally, we would always copy the io_abd into b_pabd, but the
5560 * user may have disabled compressed ARC, thus we must check the
5561 * hdr's compression setting rather than the io_bp's.
5562 */
5563 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5564 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
5565 ZIO_COMPRESS_OFF);
5566 ASSERT3U(psize, >, 0);
5567
5568 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
5569 } else {
5570 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
5571
5572 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
5573 arc_buf_size(buf));
5574 }
5575 } else {
5576 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
5577 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
5578 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5579
5580 arc_share_buf(hdr, buf);
5581 }
5582
5583 arc_hdr_verify(hdr, zio->io_bp);
5584 }
5585
5586 static void
5587 arc_write_children_ready(zio_t *zio)
5588 {
5589 arc_write_callback_t *callback = zio->io_private;
5590 arc_buf_t *buf = callback->awcb_buf;
5591
5592 callback->awcb_children_ready(zio, buf, callback->awcb_private);
5593 }
5594
5595 /*
5596 * The SPA calls this callback for each physical write that happens on behalf
5597 * of a logical write. See the comment in dbuf_write_physdone() for details.
5598 */
5599 static void
5600 arc_write_physdone(zio_t *zio)
5601 {
5602 arc_write_callback_t *cb = zio->io_private;
5603 if (cb->awcb_physdone != NULL)
5604 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
5605 }
5606
5607 static void
5608 arc_write_done(zio_t *zio)
5609 {
5610 arc_write_callback_t *callback = zio->io_private;
5611 arc_buf_t *buf = callback->awcb_buf;
5612 arc_buf_hdr_t *hdr = buf->b_hdr;
5613
5614 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5615
5616 if (zio->io_error == 0) {
5617 arc_hdr_verify(hdr, zio->io_bp);
5618
5619 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5620 buf_discard_identity(hdr);
5621 } else {
5622 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
5623 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
5624 }
5625 } else {
5626 ASSERT(HDR_EMPTY(hdr));
5627 }
5628
5629 /*
5630 * If the block to be written was all-zero or compressed enough to be
5631 * embedded in the BP, no write was performed so there will be no
5632 * dva/birth/checksum. The buffer must therefore remain anonymous
5633 * (and uncached).
5634 */
5635 if (!HDR_EMPTY(hdr)) {
5636 arc_buf_hdr_t *exists;
5637 kmutex_t *hash_lock;
5638
5639 ASSERT3U(zio->io_error, ==, 0);
5640
5641 arc_cksum_verify(buf);
5642
5643 exists = buf_hash_insert(hdr, &hash_lock);
5644 if (exists != NULL) {
5645 /*
5646 * This can only happen if we overwrite for
5647 * sync-to-convergence, because we remove
5648 * buffers from the hash table when we arc_free().
5649 */
5650 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
5651 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5652 panic("bad overwrite, hdr=%p exists=%p",
5653 (void *)hdr, (void *)exists);
5654 ASSERT(refcount_is_zero(
5655 &exists->b_l1hdr.b_refcnt));
5656 arc_change_state(arc_anon, exists, hash_lock);
5657 mutex_exit(hash_lock);
5658 arc_hdr_destroy(exists);
5659 exists = buf_hash_insert(hdr, &hash_lock);
5660 ASSERT3P(exists, ==, NULL);
5661 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
5662 /* nopwrite */
5663 ASSERT(zio->io_prop.zp_nopwrite);
5664 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5665 panic("bad nopwrite, hdr=%p exists=%p",
5666 (void *)hdr, (void *)exists);
5667 } else {
5668 /* Dedup */
5669 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
5670 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
5671 ASSERT(BP_GET_DEDUP(zio->io_bp));
5672 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
5673 }
5674 }
5675 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5676 /* if it's not anon, we are doing a scrub */
5677 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
5678 arc_access(hdr, hash_lock);
5679 mutex_exit(hash_lock);
5680 } else {
5681 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5682 }
5683
5684 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5685 callback->awcb_done(zio, buf, callback->awcb_private);
5686
5687 abd_put(zio->io_abd);
5688 kmem_free(callback, sizeof (arc_write_callback_t));
5689 }
5690
5691 zio_t *
5692 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
5693 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
5694 arc_done_func_t *children_ready, arc_done_func_t *physdone,
5695 arc_done_func_t *done, void *private, zio_priority_t priority,
5696 int zio_flags, const zbookmark_phys_t *zb)
5697 {
5698 arc_buf_hdr_t *hdr = buf->b_hdr;
5699 arc_write_callback_t *callback;
5700 zio_t *zio;
5701 zio_prop_t localprop = *zp;
5702
5703 ASSERT3P(ready, !=, NULL);
5704 ASSERT3P(done, !=, NULL);
5705 ASSERT(!HDR_IO_ERROR(hdr));
5706 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5707 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5708 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
5709 if (l2arc)
5710 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5711 if (ARC_BUF_COMPRESSED(buf)) {
5712 /*
5713 * We're writing a pre-compressed buffer. Make the
5714 * compression algorithm requested by the zio_prop_t match
5715 * the pre-compressed buffer's compression algorithm.
5716 */
5717 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
5718
5719 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
5720 zio_flags |= ZIO_FLAG_RAW;
5721 }
5722 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
5723 callback->awcb_ready = ready;
5724 callback->awcb_children_ready = children_ready;
5725 callback->awcb_physdone = physdone;
5726 callback->awcb_done = done;
5727 callback->awcb_private = private;
5728 callback->awcb_buf = buf;
5729
5730 /*
5731 * The hdr's b_pabd is now stale, free it now. A new data block
5732 * will be allocated when the zio pipeline calls arc_write_ready().
5733 */
5734 if (hdr->b_l1hdr.b_pabd != NULL) {
5735 /*
5736 * If the buf is currently sharing the data block with
5737 * the hdr then we need to break that relationship here.
5738 * The hdr will remain with a NULL data pointer and the
5739 * buf will take sole ownership of the block.
5740 */
5741 if (arc_buf_is_shared(buf)) {
5742 arc_unshare_buf(hdr, buf);
5743 } else {
5744 arc_hdr_free_pabd(hdr);
5745 }
5746 VERIFY3P(buf->b_data, !=, NULL);
5747 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
5748 }
5749 ASSERT(!arc_buf_is_shared(buf));
5750 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5751
5752 zio = zio_write(pio, spa, txg, bp,
5753 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
5754 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
5755 (children_ready != NULL) ? arc_write_children_ready : NULL,
5756 arc_write_physdone, arc_write_done, callback,
5757 priority, zio_flags, zb);
5758
5759 return (zio);
5760 }
5761
5762 static int
5763 arc_memory_throttle(uint64_t reserve, uint64_t txg)
5764 {
5765 #ifdef _KERNEL
5766 uint64_t available_memory = ptob(freemem);
5767 static uint64_t page_load = 0;
5768 static uint64_t last_txg = 0;
5769
5770 #if defined(__i386)
5771 available_memory =
5772 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
5773 #endif
5774
5775 if (freemem > physmem * arc_lotsfree_percent / 100)
5776 return (0);
5777
5778 if (txg > last_txg) {
5779 last_txg = txg;
5780 page_load = 0;
5781 }
5782 /*
5783 * If we are in pageout, we know that memory is already tight,
5784 * the arc is already going to be evicting, so we just want to
5785 * continue to let page writes occur as quickly as possible.
5786 */
5787 if (curproc == proc_pageout) {
5788 if (page_load > MAX(ptob(minfree), available_memory) / 4)
5789 return (SET_ERROR(ERESTART));
5790 /* Note: reserve is inflated, so we deflate */
5791 page_load += reserve / 8;
5792 return (0);
5793 } else if (page_load > 0 && arc_reclaim_needed()) {
5794 /* memory is low, delay before restarting */
5795 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
5796 return (SET_ERROR(EAGAIN));
5797 }
5798 page_load = 0;
5799 #endif
5800 return (0);
5801 }
5802
5803 void
5804 arc_tempreserve_clear(uint64_t reserve)
5805 {
5806 atomic_add_64(&arc_tempreserve, -reserve);
5807 ASSERT((int64_t)arc_tempreserve >= 0);
5808 }
5809
5810 int
5811 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
5812 {
5813 int error;
5814 uint64_t anon_size;
5815
5816 if (reserve > arc_c/4 && !arc_no_grow)
5817 arc_c = MIN(arc_c_max, reserve * 4);
5818 if (reserve > arc_c)
5819 return (SET_ERROR(ENOMEM));
5820
5821 /*
5822 * Don't count loaned bufs as in flight dirty data to prevent long
5823 * network delays from blocking transactions that are ready to be
5824 * assigned to a txg.
5825 */
5826
5827 /* assert that it has not wrapped around */
5828 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
5829
5830 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
5831 arc_loaned_bytes), 0);
5832
5833 /*
5834 * Writes will, almost always, require additional memory allocations
5835 * in order to compress/encrypt/etc the data. We therefore need to
5836 * make sure that there is sufficient available memory for this.
5837 */
5838 error = arc_memory_throttle(reserve, txg);
5839 if (error != 0)
5840 return (error);
5841
5842 /*
5843 * Throttle writes when the amount of dirty data in the cache
5844 * gets too large. We try to keep the cache less than half full
5845 * of dirty blocks so that our sync times don't grow too large.
5846 * Note: if two requests come in concurrently, we might let them
5847 * both succeed, when one of them should fail. Not a huge deal.
5848 */
5849
5850 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
5851 anon_size > arc_c / 4) {
5852 uint64_t meta_esize =
5853 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
5854 uint64_t data_esize =
5855 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
5856 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5857 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5858 arc_tempreserve >> 10, meta_esize >> 10,
5859 data_esize >> 10, reserve >> 10, arc_c >> 10);
5860 return (SET_ERROR(ERESTART));
5861 }
5862 atomic_add_64(&arc_tempreserve, reserve);
5863 return (0);
5864 }
5865
5866 static void
5867 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
5868 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
5869 {
5870 size->value.ui64 = refcount_count(&state->arcs_size);
5871 evict_data->value.ui64 =
5872 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
5873 evict_metadata->value.ui64 =
5874 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
5875 }
5876
5877 static int
5878 arc_kstat_update(kstat_t *ksp, int rw)
5879 {
5880 arc_stats_t *as = ksp->ks_data;
5881
5882 if (rw == KSTAT_WRITE) {
5883 return (EACCES);
5884 } else {
5885 arc_kstat_update_state(arc_anon,
5886 &as->arcstat_anon_size,
5887 &as->arcstat_anon_evictable_data,
5888 &as->arcstat_anon_evictable_metadata);
5889 arc_kstat_update_state(arc_mru,
5890 &as->arcstat_mru_size,
5891 &as->arcstat_mru_evictable_data,
5892 &as->arcstat_mru_evictable_metadata);
5893 arc_kstat_update_state(arc_mru_ghost,
5894 &as->arcstat_mru_ghost_size,
5895 &as->arcstat_mru_ghost_evictable_data,
5896 &as->arcstat_mru_ghost_evictable_metadata);
5897 arc_kstat_update_state(arc_mfu,
5898 &as->arcstat_mfu_size,
5899 &as->arcstat_mfu_evictable_data,
5900 &as->arcstat_mfu_evictable_metadata);
5901 arc_kstat_update_state(arc_mfu_ghost,
5902 &as->arcstat_mfu_ghost_size,
5903 &as->arcstat_mfu_ghost_evictable_data,
5904 &as->arcstat_mfu_ghost_evictable_metadata);
5905
5906 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
5907 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
5908 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
5909 ARCSTAT(arcstat_metadata_size) =
5910 aggsum_value(&astat_metadata_size);
5911 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
5912 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size);
5913 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
5914 }
5915
5916 return (0);
5917 }
5918
5919 /*
5920 * This function *must* return indices evenly distributed between all
5921 * sublists of the multilist. This is needed due to how the ARC eviction
5922 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5923 * distributed between all sublists and uses this assumption when
5924 * deciding which sublist to evict from and how much to evict from it.
5925 */
5926 unsigned int
5927 arc_state_multilist_index_func(multilist_t *ml, void *obj)
5928 {
5929 arc_buf_hdr_t *hdr = obj;
5930
5931 /*
5932 * We rely on b_dva to generate evenly distributed index
5933 * numbers using buf_hash below. So, as an added precaution,
5934 * let's make sure we never add empty buffers to the arc lists.
5935 */
5936 ASSERT(!HDR_EMPTY(hdr));
5937
5938 /*
5939 * The assumption here, is the hash value for a given
5940 * arc_buf_hdr_t will remain constant throughout it's lifetime
5941 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
5942 * Thus, we don't need to store the header's sublist index
5943 * on insertion, as this index can be recalculated on removal.
5944 *
5945 * Also, the low order bits of the hash value are thought to be
5946 * distributed evenly. Otherwise, in the case that the multilist
5947 * has a power of two number of sublists, each sublists' usage
5948 * would not be evenly distributed.
5949 */
5950 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
5951 multilist_get_num_sublists(ml));
5952 }
5953
5954 static void
5955 arc_state_init(void)
5956 {
5957 arc_anon = &ARC_anon;
5958 arc_mru = &ARC_mru;
5959 arc_mru_ghost = &ARC_mru_ghost;
5960 arc_mfu = &ARC_mfu;
5961 arc_mfu_ghost = &ARC_mfu_ghost;
5962 arc_l2c_only = &ARC_l2c_only;
5963
5964 arc_mru->arcs_list[ARC_BUFC_METADATA] =
5965 multilist_create(sizeof (arc_buf_hdr_t),
5966 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5967 arc_state_multilist_index_func);
5968 arc_mru->arcs_list[ARC_BUFC_DATA] =
5969 multilist_create(sizeof (arc_buf_hdr_t),
5970 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5971 arc_state_multilist_index_func);
5972 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
5973 multilist_create(sizeof (arc_buf_hdr_t),
5974 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5975 arc_state_multilist_index_func);
5976 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
5977 multilist_create(sizeof (arc_buf_hdr_t),
5978 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5979 arc_state_multilist_index_func);
5980 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
5981 multilist_create(sizeof (arc_buf_hdr_t),
5982 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5983 arc_state_multilist_index_func);
5984 arc_mfu->arcs_list[ARC_BUFC_DATA] =
5985 multilist_create(sizeof (arc_buf_hdr_t),
5986 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5987 arc_state_multilist_index_func);
5988 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
5989 multilist_create(sizeof (arc_buf_hdr_t),
5990 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5991 arc_state_multilist_index_func);
5992 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
5993 multilist_create(sizeof (arc_buf_hdr_t),
5994 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5995 arc_state_multilist_index_func);
5996 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
5997 multilist_create(sizeof (arc_buf_hdr_t),
5998 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5999 arc_state_multilist_index_func);
6000 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6001 multilist_create(sizeof (arc_buf_hdr_t),
6002 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6003 arc_state_multilist_index_func);
6004
6005 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6006 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6007 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6008 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6009 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6010 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6011 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6012 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6013 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6014 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6015 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6016 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6017
6018 refcount_create(&arc_anon->arcs_size);
6019 refcount_create(&arc_mru->arcs_size);
6020 refcount_create(&arc_mru_ghost->arcs_size);
6021 refcount_create(&arc_mfu->arcs_size);
6022 refcount_create(&arc_mfu_ghost->arcs_size);
6023 refcount_create(&arc_l2c_only->arcs_size);
6024
6025 aggsum_init(&arc_meta_used, 0);
6026 aggsum_init(&arc_size, 0);
6027 aggsum_init(&astat_data_size, 0);
6028 aggsum_init(&astat_metadata_size, 0);
6029 aggsum_init(&astat_hdr_size, 0);
6030 aggsum_init(&astat_other_size, 0);
6031 aggsum_init(&astat_l2_hdr_size, 0);
6032 }
6033
6034 static void
6035 arc_state_fini(void)
6036 {
6037 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6038 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6039 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6040 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6041 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6042 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6043 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6044 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6045 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6046 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6047 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6048 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6049
6050 refcount_destroy(&arc_anon->arcs_size);
6051 refcount_destroy(&arc_mru->arcs_size);
6052 refcount_destroy(&arc_mru_ghost->arcs_size);
6053 refcount_destroy(&arc_mfu->arcs_size);
6054 refcount_destroy(&arc_mfu_ghost->arcs_size);
6055 refcount_destroy(&arc_l2c_only->arcs_size);
6056
6057 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6058 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6059 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6060 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6061 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
6062 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6063 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
6064 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6065 }
6066
6067 uint64_t
6068 arc_max_bytes(void)
6069 {
6070 return (arc_c_max);
6071 }
6072
6073 void
6074 arc_init(void)
6075 {
6076 /*
6077 * allmem is "all memory that we could possibly use".
6078 */
6079 #ifdef _KERNEL
6080 uint64_t allmem = ptob(physmem - swapfs_minfree);
6081 #else
6082 uint64_t allmem = (physmem * PAGESIZE) / 2;
6083 #endif
6084
6085 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
6086 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
6087 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
6088
6089 /* Convert seconds to clock ticks */
6090 arc_min_prefetch_lifespan = 1 * hz;
6091
6092 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
6093 arc_c_min = MAX(allmem / 32, 64 << 20);
6094 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
6095 if (allmem >= 1 << 30)
6096 arc_c_max = allmem - (1 << 30);
6097 else
6098 arc_c_max = arc_c_min;
6099 arc_c_max = MAX(allmem * 3 / 4, arc_c_max);
6100
6101 /*
6102 * In userland, there's only the memory pressure that we artificially
6103 * create (see arc_available_memory()). Don't let arc_c get too
6104 * small, because it can cause transactions to be larger than
6105 * arc_c, causing arc_tempreserve_space() to fail.
6106 */
6107 #ifndef _KERNEL
6108 arc_c_min = arc_c_max / 2;
6109 #endif
6110
6111 /*
6112 * Allow the tunables to override our calculations if they are
6113 * reasonable (ie. over 64MB)
6114 */
6115 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) {
6116 arc_c_max = zfs_arc_max;
6117 arc_c_min = MIN(arc_c_min, arc_c_max);
6118 }
6119 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
6120 arc_c_min = zfs_arc_min;
6121
6122 arc_c = arc_c_max;
6123 arc_p = (arc_c >> 1);
6124
6125 /* limit meta-data to 1/4 of the arc capacity */
6126 arc_meta_limit = arc_c_max / 4;
6127
6128 #ifdef _KERNEL
6129 /*
6130 * Metadata is stored in the kernel's heap. Don't let us
6131 * use more than half the heap for the ARC.
6132 */
6133 arc_meta_limit = MIN(arc_meta_limit,
6134 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
6135 #endif
6136
6137 /* Allow the tunable to override if it is reasonable */
6138 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6139 arc_meta_limit = zfs_arc_meta_limit;
6140
6141 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6142 arc_c_min = arc_meta_limit / 2;
6143
6144 if (zfs_arc_meta_min > 0) {
6145 arc_meta_min = zfs_arc_meta_min;
6146 } else {
6147 arc_meta_min = arc_c_min / 2;
6148 }
6149
6150 if (zfs_arc_grow_retry > 0)
6151 arc_grow_retry = zfs_arc_grow_retry;
6152
6153 if (zfs_arc_shrink_shift > 0)
6154 arc_shrink_shift = zfs_arc_shrink_shift;
6155
6156 /*
6157 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6158 */
6159 if (arc_no_grow_shift >= arc_shrink_shift)
6160 arc_no_grow_shift = arc_shrink_shift - 1;
6161
6162 if (zfs_arc_p_min_shift > 0)
6163 arc_p_min_shift = zfs_arc_p_min_shift;
6164
6165 /* if kmem_flags are set, lets try to use less memory */
6166 if (kmem_debugging())
6167 arc_c = arc_c / 2;
6168 if (arc_c < arc_c_min)
6169 arc_c = arc_c_min;
6170
6171 arc_state_init();
6172 buf_init();
6173
6174 arc_reclaim_thread_exit = B_FALSE;
6175
6176 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6177 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6178
6179 if (arc_ksp != NULL) {
6180 arc_ksp->ks_data = &arc_stats;
6181 arc_ksp->ks_update = arc_kstat_update;
6182 kstat_install(arc_ksp);
6183 }
6184
6185 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
6186 TS_RUN, minclsyspri);
6187
6188 arc_dead = B_FALSE;
6189 arc_warm = B_FALSE;
6190
6191 /*
6192 * Calculate maximum amount of dirty data per pool.
6193 *
6194 * If it has been set by /etc/system, take that.
6195 * Otherwise, use a percentage of physical memory defined by
6196 * zfs_dirty_data_max_percent (default 10%) with a cap at
6197 * zfs_dirty_data_max_max (default 4GB).
6198 */
6199 if (zfs_dirty_data_max == 0) {
6200 zfs_dirty_data_max = physmem * PAGESIZE *
6201 zfs_dirty_data_max_percent / 100;
6202 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6203 zfs_dirty_data_max_max);
6204 }
6205 }
6206
6207 void
6208 arc_fini(void)
6209 {
6210 mutex_enter(&arc_reclaim_lock);
6211 arc_reclaim_thread_exit = B_TRUE;
6212 /*
6213 * The reclaim thread will set arc_reclaim_thread_exit back to
6214 * B_FALSE when it is finished exiting; we're waiting for that.
6215 */
6216 while (arc_reclaim_thread_exit) {
6217 cv_signal(&arc_reclaim_thread_cv);
6218 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6219 }
6220 mutex_exit(&arc_reclaim_lock);
6221
6222 /* Use B_TRUE to ensure *all* buffers are evicted */
6223 arc_flush(NULL, B_TRUE);
6224
6225 arc_dead = B_TRUE;
6226
6227 if (arc_ksp != NULL) {
6228 kstat_delete(arc_ksp);
6229 arc_ksp = NULL;
6230 }
6231
6232 mutex_destroy(&arc_reclaim_lock);
6233 cv_destroy(&arc_reclaim_thread_cv);
6234 cv_destroy(&arc_reclaim_waiters_cv);
6235
6236 arc_state_fini();
6237 buf_fini();
6238
6239 ASSERT0(arc_loaned_bytes);
6240 }
6241
6242 /*
6243 * Level 2 ARC
6244 *
6245 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6246 * It uses dedicated storage devices to hold cached data, which are populated
6247 * using large infrequent writes. The main role of this cache is to boost
6248 * the performance of random read workloads. The intended L2ARC devices
6249 * include short-stroked disks, solid state disks, and other media with
6250 * substantially faster read latency than disk.
6251 *
6252 * +-----------------------+
6253 * | ARC |
6254 * +-----------------------+
6255 * | ^ ^
6256 * | | |
6257 * l2arc_feed_thread() arc_read()
6258 * | | |
6259 * | l2arc read |
6260 * V | |
6261 * +---------------+ |
6262 * | L2ARC | |
6263 * +---------------+ |
6264 * | ^ |
6265 * l2arc_write() | |
6266 * | | |
6267 * V | |
6268 * +-------+ +-------+
6269 * | vdev | | vdev |
6270 * | cache | | cache |
6271 * +-------+ +-------+
6272 * +=========+ .-----.
6273 * : L2ARC : |-_____-|
6274 * : devices : | Disks |
6275 * +=========+ `-_____-'
6276 *
6277 * Read requests are satisfied from the following sources, in order:
6278 *
6279 * 1) ARC
6280 * 2) vdev cache of L2ARC devices
6281 * 3) L2ARC devices
6282 * 4) vdev cache of disks
6283 * 5) disks
6284 *
6285 * Some L2ARC device types exhibit extremely slow write performance.
6286 * To accommodate for this there are some significant differences between
6287 * the L2ARC and traditional cache design:
6288 *
6289 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6290 * the ARC behave as usual, freeing buffers and placing headers on ghost
6291 * lists. The ARC does not send buffers to the L2ARC during eviction as
6292 * this would add inflated write latencies for all ARC memory pressure.
6293 *
6294 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6295 * It does this by periodically scanning buffers from the eviction-end of
6296 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6297 * not already there. It scans until a headroom of buffers is satisfied,
6298 * which itself is a buffer for ARC eviction. If a compressible buffer is
6299 * found during scanning and selected for writing to an L2ARC device, we
6300 * temporarily boost scanning headroom during the next scan cycle to make
6301 * sure we adapt to compression effects (which might significantly reduce
6302 * the data volume we write to L2ARC). The thread that does this is
6303 * l2arc_feed_thread(), illustrated below; example sizes are included to
6304 * provide a better sense of ratio than this diagram:
6305 *
6306 * head --> tail
6307 * +---------------------+----------+
6308 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6309 * +---------------------+----------+ | o L2ARC eligible
6310 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6311 * +---------------------+----------+ |
6312 * 15.9 Gbytes ^ 32 Mbytes |
6313 * headroom |
6314 * l2arc_feed_thread()
6315 * |
6316 * l2arc write hand <--[oooo]--'
6317 * | 8 Mbyte
6318 * | write max
6319 * V
6320 * +==============================+
6321 * L2ARC dev |####|#|###|###| |####| ... |
6322 * +==============================+
6323 * 32 Gbytes
6324 *
6325 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6326 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6327 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6328 * safe to say that this is an uncommon case, since buffers at the end of
6329 * the ARC lists have moved there due to inactivity.
6330 *
6331 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6332 * then the L2ARC simply misses copying some buffers. This serves as a
6333 * pressure valve to prevent heavy read workloads from both stalling the ARC
6334 * with waits and clogging the L2ARC with writes. This also helps prevent
6335 * the potential for the L2ARC to churn if it attempts to cache content too
6336 * quickly, such as during backups of the entire pool.
6337 *
6338 * 5. After system boot and before the ARC has filled main memory, there are
6339 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6340 * lists can remain mostly static. Instead of searching from tail of these
6341 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6342 * for eligible buffers, greatly increasing its chance of finding them.
6343 *
6344 * The L2ARC device write speed is also boosted during this time so that
6345 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6346 * there are no L2ARC reads, and no fear of degrading read performance
6347 * through increased writes.
6348 *
6349 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6350 * the vdev queue can aggregate them into larger and fewer writes. Each
6351 * device is written to in a rotor fashion, sweeping writes through
6352 * available space then repeating.
6353 *
6354 * 7. The L2ARC does not store dirty content. It never needs to flush
6355 * write buffers back to disk based storage.
6356 *
6357 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6358 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6359 *
6360 * The performance of the L2ARC can be tweaked by a number of tunables, which
6361 * may be necessary for different workloads:
6362 *
6363 * l2arc_write_max max write bytes per interval
6364 * l2arc_write_boost extra write bytes during device warmup
6365 * l2arc_noprefetch skip caching prefetched buffers
6366 * l2arc_headroom number of max device writes to precache
6367 * l2arc_headroom_boost when we find compressed buffers during ARC
6368 * scanning, we multiply headroom by this
6369 * percentage factor for the next scan cycle,
6370 * since more compressed buffers are likely to
6371 * be present
6372 * l2arc_feed_secs seconds between L2ARC writing
6373 *
6374 * Tunables may be removed or added as future performance improvements are
6375 * integrated, and also may become zpool properties.
6376 *
6377 * There are three key functions that control how the L2ARC warms up:
6378 *
6379 * l2arc_write_eligible() check if a buffer is eligible to cache
6380 * l2arc_write_size() calculate how much to write
6381 * l2arc_write_interval() calculate sleep delay between writes
6382 *
6383 * These three functions determine what to write, how much, and how quickly
6384 * to send writes.
6385 */
6386
6387 static boolean_t
6388 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
6389 {
6390 /*
6391 * A buffer is *not* eligible for the L2ARC if it:
6392 * 1. belongs to a different spa.
6393 * 2. is already cached on the L2ARC.
6394 * 3. has an I/O in progress (it may be an incomplete read).
6395 * 4. is flagged not eligible (zfs property).
6396 */
6397 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
6398 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
6399 return (B_FALSE);
6400
6401 return (B_TRUE);
6402 }
6403
6404 static uint64_t
6405 l2arc_write_size(void)
6406 {
6407 uint64_t size;
6408
6409 /*
6410 * Make sure our globals have meaningful values in case the user
6411 * altered them.
6412 */
6413 size = l2arc_write_max;
6414 if (size == 0) {
6415 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
6416 "be greater than zero, resetting it to the default (%d)",
6417 L2ARC_WRITE_SIZE);
6418 size = l2arc_write_max = L2ARC_WRITE_SIZE;
6419 }
6420
6421 if (arc_warm == B_FALSE)
6422 size += l2arc_write_boost;
6423
6424 return (size);
6425
6426 }
6427
6428 static clock_t
6429 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
6430 {
6431 clock_t interval, next, now;
6432
6433 /*
6434 * If the ARC lists are busy, increase our write rate; if the
6435 * lists are stale, idle back. This is achieved by checking
6436 * how much we previously wrote - if it was more than half of
6437 * what we wanted, schedule the next write much sooner.
6438 */
6439 if (l2arc_feed_again && wrote > (wanted / 2))
6440 interval = (hz * l2arc_feed_min_ms) / 1000;
6441 else
6442 interval = hz * l2arc_feed_secs;
6443
6444 now = ddi_get_lbolt();
6445 next = MAX(now, MIN(now + interval, began + interval));
6446
6447 return (next);
6448 }
6449
6450 /*
6451 * Cycle through L2ARC devices. This is how L2ARC load balances.
6452 * If a device is returned, this also returns holding the spa config lock.
6453 */
6454 static l2arc_dev_t *
6455 l2arc_dev_get_next(void)
6456 {
6457 l2arc_dev_t *first, *next = NULL;
6458
6459 /*
6460 * Lock out the removal of spas (spa_namespace_lock), then removal
6461 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6462 * both locks will be dropped and a spa config lock held instead.
6463 */
6464 mutex_enter(&spa_namespace_lock);
6465 mutex_enter(&l2arc_dev_mtx);
6466
6467 /* if there are no vdevs, there is nothing to do */
6468 if (l2arc_ndev == 0)
6469 goto out;
6470
6471 first = NULL;
6472 next = l2arc_dev_last;
6473 do {
6474 /* loop around the list looking for a non-faulted vdev */
6475 if (next == NULL) {
6476 next = list_head(l2arc_dev_list);
6477 } else {
6478 next = list_next(l2arc_dev_list, next);
6479 if (next == NULL)
6480 next = list_head(l2arc_dev_list);
6481 }
6482
6483 /* if we have come back to the start, bail out */
6484 if (first == NULL)
6485 first = next;
6486 else if (next == first)
6487 break;
6488
6489 } while (vdev_is_dead(next->l2ad_vdev));
6490
6491 /* if we were unable to find any usable vdevs, return NULL */
6492 if (vdev_is_dead(next->l2ad_vdev))
6493 next = NULL;
6494
6495 l2arc_dev_last = next;
6496
6497 out:
6498 mutex_exit(&l2arc_dev_mtx);
6499
6500 /*
6501 * Grab the config lock to prevent the 'next' device from being
6502 * removed while we are writing to it.
6503 */
6504 if (next != NULL)
6505 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
6506 mutex_exit(&spa_namespace_lock);
6507
6508 return (next);
6509 }
6510
6511 /*
6512 * Free buffers that were tagged for destruction.
6513 */
6514 static void
6515 l2arc_do_free_on_write()
6516 {
6517 list_t *buflist;
6518 l2arc_data_free_t *df, *df_prev;
6519
6520 mutex_enter(&l2arc_free_on_write_mtx);
6521 buflist = l2arc_free_on_write;
6522
6523 for (df = list_tail(buflist); df; df = df_prev) {
6524 df_prev = list_prev(buflist, df);
6525 ASSERT3P(df->l2df_abd, !=, NULL);
6526 abd_free(df->l2df_abd);
6527 list_remove(buflist, df);
6528 kmem_free(df, sizeof (l2arc_data_free_t));
6529 }
6530
6531 mutex_exit(&l2arc_free_on_write_mtx);
6532 }
6533
6534 /*
6535 * A write to a cache device has completed. Update all headers to allow
6536 * reads from these buffers to begin.
6537 */
6538 static void
6539 l2arc_write_done(zio_t *zio)
6540 {
6541 l2arc_write_callback_t *cb;
6542 l2arc_dev_t *dev;
6543 list_t *buflist;
6544 arc_buf_hdr_t *head, *hdr, *hdr_prev;
6545 kmutex_t *hash_lock;
6546 int64_t bytes_dropped = 0;
6547
6548 cb = zio->io_private;
6549 ASSERT3P(cb, !=, NULL);
6550 dev = cb->l2wcb_dev;
6551 ASSERT3P(dev, !=, NULL);
6552 head = cb->l2wcb_head;
6553 ASSERT3P(head, !=, NULL);
6554 buflist = &dev->l2ad_buflist;
6555 ASSERT3P(buflist, !=, NULL);
6556 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
6557 l2arc_write_callback_t *, cb);
6558
6559 if (zio->io_error != 0)
6560 ARCSTAT_BUMP(arcstat_l2_writes_error);
6561
6562 /*
6563 * All writes completed, or an error was hit.
6564 */
6565 top:
6566 mutex_enter(&dev->l2ad_mtx);
6567 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
6568 hdr_prev = list_prev(buflist, hdr);
6569
6570 hash_lock = HDR_LOCK(hdr);
6571
6572 /*
6573 * We cannot use mutex_enter or else we can deadlock
6574 * with l2arc_write_buffers (due to swapping the order
6575 * the hash lock and l2ad_mtx are taken).
6576 */
6577 if (!mutex_tryenter(hash_lock)) {
6578 /*
6579 * Missed the hash lock. We must retry so we
6580 * don't leave the ARC_FLAG_L2_WRITING bit set.
6581 */
6582 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
6583
6584 /*
6585 * We don't want to rescan the headers we've
6586 * already marked as having been written out, so
6587 * we reinsert the head node so we can pick up
6588 * where we left off.
6589 */
6590 list_remove(buflist, head);
6591 list_insert_after(buflist, hdr, head);
6592
6593 mutex_exit(&dev->l2ad_mtx);
6594
6595 /*
6596 * We wait for the hash lock to become available
6597 * to try and prevent busy waiting, and increase
6598 * the chance we'll be able to acquire the lock
6599 * the next time around.
6600 */
6601 mutex_enter(hash_lock);
6602 mutex_exit(hash_lock);
6603 goto top;
6604 }
6605
6606 /*
6607 * We could not have been moved into the arc_l2c_only
6608 * state while in-flight due to our ARC_FLAG_L2_WRITING
6609 * bit being set. Let's just ensure that's being enforced.
6610 */
6611 ASSERT(HDR_HAS_L1HDR(hdr));
6612
6613 if (zio->io_error != 0) {
6614 /*
6615 * Error - drop L2ARC entry.
6616 */
6617 list_remove(buflist, hdr);
6618 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
6619
6620 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
6621 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
6622
6623 bytes_dropped += arc_hdr_size(hdr);
6624 (void) refcount_remove_many(&dev->l2ad_alloc,
6625 arc_hdr_size(hdr), hdr);
6626 }
6627
6628 /*
6629 * Allow ARC to begin reads and ghost list evictions to
6630 * this L2ARC entry.
6631 */
6632 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
6633
6634 mutex_exit(hash_lock);
6635 }
6636
6637 atomic_inc_64(&l2arc_writes_done);
6638 list_remove(buflist, head);
6639 ASSERT(!HDR_HAS_L1HDR(head));
6640 kmem_cache_free(hdr_l2only_cache, head);
6641 mutex_exit(&dev->l2ad_mtx);
6642
6643 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
6644
6645 l2arc_do_free_on_write();
6646
6647 kmem_free(cb, sizeof (l2arc_write_callback_t));
6648 }
6649
6650 /*
6651 * A read to a cache device completed. Validate buffer contents before
6652 * handing over to the regular ARC routines.
6653 */
6654 static void
6655 l2arc_read_done(zio_t *zio)
6656 {
6657 l2arc_read_callback_t *cb;
6658 arc_buf_hdr_t *hdr;
6659 kmutex_t *hash_lock;
6660 boolean_t valid_cksum;
6661
6662 ASSERT3P(zio->io_vd, !=, NULL);
6663 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
6664
6665 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
6666
6667 cb = zio->io_private;
6668 ASSERT3P(cb, !=, NULL);
6669 hdr = cb->l2rcb_hdr;
6670 ASSERT3P(hdr, !=, NULL);
6671
6672 hash_lock = HDR_LOCK(hdr);
6673 mutex_enter(hash_lock);
6674 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6675
6676 /*
6677 * If the data was read into a temporary buffer,
6678 * move it and free the buffer.
6679 */
6680 if (cb->l2rcb_abd != NULL) {
6681 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
6682 if (zio->io_error == 0) {
6683 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
6684 arc_hdr_size(hdr));
6685 }
6686
6687 /*
6688 * The following must be done regardless of whether
6689 * there was an error:
6690 * - free the temporary buffer
6691 * - point zio to the real ARC buffer
6692 * - set zio size accordingly
6693 * These are required because zio is either re-used for
6694 * an I/O of the block in the case of the error
6695 * or the zio is passed to arc_read_done() and it
6696 * needs real data.
6697 */
6698 abd_free(cb->l2rcb_abd);
6699 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
6700 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
6701 }
6702
6703 ASSERT3P(zio->io_abd, !=, NULL);
6704
6705 /*
6706 * Check this survived the L2ARC journey.
6707 */
6708 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
6709 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
6710 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
6711
6712 valid_cksum = arc_cksum_is_equal(hdr, zio);
6713 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
6714 mutex_exit(hash_lock);
6715 zio->io_private = hdr;
6716 arc_read_done(zio);
6717 } else {
6718 mutex_exit(hash_lock);
6719 /*
6720 * Buffer didn't survive caching. Increment stats and
6721 * reissue to the original storage device.
6722 */
6723 if (zio->io_error != 0) {
6724 ARCSTAT_BUMP(arcstat_l2_io_error);
6725 } else {
6726 zio->io_error = SET_ERROR(EIO);
6727 }
6728 if (!valid_cksum)
6729 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
6730
6731 /*
6732 * If there's no waiter, issue an async i/o to the primary
6733 * storage now. If there *is* a waiter, the caller must
6734 * issue the i/o in a context where it's OK to block.
6735 */
6736 if (zio->io_waiter == NULL) {
6737 zio_t *pio = zio_unique_parent(zio);
6738
6739 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
6740
6741 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
6742 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
6743 hdr, zio->io_priority, cb->l2rcb_flags,
6744 &cb->l2rcb_zb));
6745 }
6746 }
6747
6748 kmem_free(cb, sizeof (l2arc_read_callback_t));
6749 }
6750
6751 /*
6752 * This is the list priority from which the L2ARC will search for pages to
6753 * cache. This is used within loops (0..3) to cycle through lists in the
6754 * desired order. This order can have a significant effect on cache
6755 * performance.
6756 *
6757 * Currently the metadata lists are hit first, MFU then MRU, followed by
6758 * the data lists. This function returns a locked list, and also returns
6759 * the lock pointer.
6760 */
6761 static multilist_sublist_t *
6762 l2arc_sublist_lock(int list_num)
6763 {
6764 multilist_t *ml = NULL;
6765 unsigned int idx;
6766
6767 ASSERT(list_num >= 0 && list_num <= 3);
6768
6769 switch (list_num) {
6770 case 0:
6771 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
6772 break;
6773 case 1:
6774 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
6775 break;
6776 case 2:
6777 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
6778 break;
6779 case 3:
6780 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
6781 break;
6782 }
6783
6784 /*
6785 * Return a randomly-selected sublist. This is acceptable
6786 * because the caller feeds only a little bit of data for each
6787 * call (8MB). Subsequent calls will result in different
6788 * sublists being selected.
6789 */
6790 idx = multilist_get_random_index(ml);
6791 return (multilist_sublist_lock(ml, idx));
6792 }
6793
6794 /*
6795 * Evict buffers from the device write hand to the distance specified in
6796 * bytes. This distance may span populated buffers, it may span nothing.
6797 * This is clearing a region on the L2ARC device ready for writing.
6798 * If the 'all' boolean is set, every buffer is evicted.
6799 */
6800 static void
6801 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
6802 {
6803 list_t *buflist;
6804 arc_buf_hdr_t *hdr, *hdr_prev;
6805 kmutex_t *hash_lock;
6806 uint64_t taddr;
6807
6808 buflist = &dev->l2ad_buflist;
6809
6810 if (!all && dev->l2ad_first) {
6811 /*
6812 * This is the first sweep through the device. There is
6813 * nothing to evict.
6814 */
6815 return;
6816 }
6817
6818 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
6819 /*
6820 * When nearing the end of the device, evict to the end
6821 * before the device write hand jumps to the start.
6822 */
6823 taddr = dev->l2ad_end;
6824 } else {
6825 taddr = dev->l2ad_hand + distance;
6826 }
6827 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
6828 uint64_t, taddr, boolean_t, all);
6829
6830 top:
6831 mutex_enter(&dev->l2ad_mtx);
6832 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
6833 hdr_prev = list_prev(buflist, hdr);
6834
6835 hash_lock = HDR_LOCK(hdr);
6836
6837 /*
6838 * We cannot use mutex_enter or else we can deadlock
6839 * with l2arc_write_buffers (due to swapping the order
6840 * the hash lock and l2ad_mtx are taken).
6841 */
6842 if (!mutex_tryenter(hash_lock)) {
6843 /*
6844 * Missed the hash lock. Retry.
6845 */
6846 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
6847 mutex_exit(&dev->l2ad_mtx);
6848 mutex_enter(hash_lock);
6849 mutex_exit(hash_lock);
6850 goto top;
6851 }
6852
6853 /*
6854 * A header can't be on this list if it doesn't have L2 header.
6855 */
6856 ASSERT(HDR_HAS_L2HDR(hdr));
6857
6858 /* Ensure this header has finished being written. */
6859 ASSERT(!HDR_L2_WRITING(hdr));
6860 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
6861
6862 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
6863 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
6864 /*
6865 * We've evicted to the target address,
6866 * or the end of the device.
6867 */
6868 mutex_exit(hash_lock);
6869 break;
6870 }
6871
6872 if (!HDR_HAS_L1HDR(hdr)) {
6873 ASSERT(!HDR_L2_READING(hdr));
6874 /*
6875 * This doesn't exist in the ARC. Destroy.
6876 * arc_hdr_destroy() will call list_remove()
6877 * and decrement arcstat_l2_lsize.
6878 */
6879 arc_change_state(arc_anon, hdr, hash_lock);
6880 arc_hdr_destroy(hdr);
6881 } else {
6882 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
6883 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
6884 /*
6885 * Invalidate issued or about to be issued
6886 * reads, since we may be about to write
6887 * over this location.
6888 */
6889 if (HDR_L2_READING(hdr)) {
6890 ARCSTAT_BUMP(arcstat_l2_evict_reading);
6891 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
6892 }
6893
6894 arc_hdr_l2hdr_destroy(hdr);
6895 }
6896 mutex_exit(hash_lock);
6897 }
6898 mutex_exit(&dev->l2ad_mtx);
6899 }
6900
6901 /*
6902 * Find and write ARC buffers to the L2ARC device.
6903 *
6904 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
6905 * for reading until they have completed writing.
6906 * The headroom_boost is an in-out parameter used to maintain headroom boost
6907 * state between calls to this function.
6908 *
6909 * Returns the number of bytes actually written (which may be smaller than
6910 * the delta by which the device hand has changed due to alignment).
6911 */
6912 static uint64_t
6913 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
6914 {
6915 arc_buf_hdr_t *hdr, *hdr_prev, *head;
6916 uint64_t write_asize, write_psize, write_lsize, headroom;
6917 boolean_t full;
6918 l2arc_write_callback_t *cb;
6919 zio_t *pio, *wzio;
6920 uint64_t guid = spa_load_guid(spa);
6921
6922 ASSERT3P(dev->l2ad_vdev, !=, NULL);
6923
6924 pio = NULL;
6925 write_lsize = write_asize = write_psize = 0;
6926 full = B_FALSE;
6927 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
6928 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
6929
6930 /*
6931 * Copy buffers for L2ARC writing.
6932 */
6933 for (int try = 0; try <= 3; try++) {
6934 multilist_sublist_t *mls = l2arc_sublist_lock(try);
6935 uint64_t passed_sz = 0;
6936
6937 /*
6938 * L2ARC fast warmup.
6939 *
6940 * Until the ARC is warm and starts to evict, read from the
6941 * head of the ARC lists rather than the tail.
6942 */
6943 if (arc_warm == B_FALSE)
6944 hdr = multilist_sublist_head(mls);
6945 else
6946 hdr = multilist_sublist_tail(mls);
6947
6948 headroom = target_sz * l2arc_headroom;
6949 if (zfs_compressed_arc_enabled)
6950 headroom = (headroom * l2arc_headroom_boost) / 100;
6951
6952 for (; hdr; hdr = hdr_prev) {
6953 kmutex_t *hash_lock;
6954
6955 if (arc_warm == B_FALSE)
6956 hdr_prev = multilist_sublist_next(mls, hdr);
6957 else
6958 hdr_prev = multilist_sublist_prev(mls, hdr);
6959
6960 hash_lock = HDR_LOCK(hdr);
6961 if (!mutex_tryenter(hash_lock)) {
6962 /*
6963 * Skip this buffer rather than waiting.
6964 */
6965 continue;
6966 }
6967
6968 passed_sz += HDR_GET_LSIZE(hdr);
6969 if (passed_sz > headroom) {
6970 /*
6971 * Searched too far.
6972 */
6973 mutex_exit(hash_lock);
6974 break;
6975 }
6976
6977 if (!l2arc_write_eligible(guid, hdr)) {
6978 mutex_exit(hash_lock);
6979 continue;
6980 }
6981
6982 /*
6983 * We rely on the L1 portion of the header below, so
6984 * it's invalid for this header to have been evicted out
6985 * of the ghost cache, prior to being written out. The
6986 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6987 */
6988 ASSERT(HDR_HAS_L1HDR(hdr));
6989
6990 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
6991 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6992 ASSERT3U(arc_hdr_size(hdr), >, 0);
6993 uint64_t psize = arc_hdr_size(hdr);
6994 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
6995 psize);
6996
6997 if ((write_asize + asize) > target_sz) {
6998 full = B_TRUE;
6999 mutex_exit(hash_lock);
7000 break;
7001 }
7002
7003 if (pio == NULL) {
7004 /*
7005 * Insert a dummy header on the buflist so
7006 * l2arc_write_done() can find where the
7007 * write buffers begin without searching.
7008 */
7009 mutex_enter(&dev->l2ad_mtx);
7010 list_insert_head(&dev->l2ad_buflist, head);
7011 mutex_exit(&dev->l2ad_mtx);
7012
7013 cb = kmem_alloc(
7014 sizeof (l2arc_write_callback_t), KM_SLEEP);
7015 cb->l2wcb_dev = dev;
7016 cb->l2wcb_head = head;
7017 pio = zio_root(spa, l2arc_write_done, cb,
7018 ZIO_FLAG_CANFAIL);
7019 }
7020
7021 hdr->b_l2hdr.b_dev = dev;
7022 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
7023 arc_hdr_set_flags(hdr,
7024 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
7025
7026 mutex_enter(&dev->l2ad_mtx);
7027 list_insert_head(&dev->l2ad_buflist, hdr);
7028 mutex_exit(&dev->l2ad_mtx);
7029
7030 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
7031
7032 /*
7033 * Normally the L2ARC can use the hdr's data, but if
7034 * we're sharing data between the hdr and one of its
7035 * bufs, L2ARC needs its own copy of the data so that
7036 * the ZIO below can't race with the buf consumer.
7037 * Another case where we need to create a copy of the
7038 * data is when the buffer size is not device-aligned
7039 * and we need to pad the block to make it such.
7040 * That also keeps the clock hand suitably aligned.
7041 *
7042 * To ensure that the copy will be available for the
7043 * lifetime of the ZIO and be cleaned up afterwards, we
7044 * add it to the l2arc_free_on_write queue.
7045 */
7046 abd_t *to_write;
7047 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
7048 to_write = hdr->b_l1hdr.b_pabd;
7049 } else {
7050 to_write = abd_alloc_for_io(asize,
7051 HDR_ISTYPE_METADATA(hdr));
7052 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
7053 if (asize != psize) {
7054 abd_zero_off(to_write, psize,
7055 asize - psize);
7056 }
7057 l2arc_free_abd_on_write(to_write, asize,
7058 arc_buf_type(hdr));
7059 }
7060 wzio = zio_write_phys(pio, dev->l2ad_vdev,
7061 hdr->b_l2hdr.b_daddr, asize, to_write,
7062 ZIO_CHECKSUM_OFF, NULL, hdr,
7063 ZIO_PRIORITY_ASYNC_WRITE,
7064 ZIO_FLAG_CANFAIL, B_FALSE);
7065
7066 write_lsize += HDR_GET_LSIZE(hdr);
7067 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
7068 zio_t *, wzio);
7069
7070 write_psize += psize;
7071 write_asize += asize;
7072 dev->l2ad_hand += asize;
7073
7074 mutex_exit(hash_lock);
7075
7076 (void) zio_nowait(wzio);
7077 }
7078
7079 multilist_sublist_unlock(mls);
7080
7081 if (full == B_TRUE)
7082 break;
7083 }
7084
7085 /* No buffers selected for writing? */
7086 if (pio == NULL) {
7087 ASSERT0(write_lsize);
7088 ASSERT(!HDR_HAS_L1HDR(head));
7089 kmem_cache_free(hdr_l2only_cache, head);
7090 return (0);
7091 }
7092
7093 ASSERT3U(write_asize, <=, target_sz);
7094 ARCSTAT_BUMP(arcstat_l2_writes_sent);
7095 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
7096 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
7097 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
7098 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
7099
7100 /*
7101 * Bump device hand to the device start if it is approaching the end.
7102 * l2arc_evict() will already have evicted ahead for this case.
7103 */
7104 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7105 dev->l2ad_hand = dev->l2ad_start;
7106 dev->l2ad_first = B_FALSE;
7107 }
7108
7109 dev->l2ad_writing = B_TRUE;
7110 (void) zio_wait(pio);
7111 dev->l2ad_writing = B_FALSE;
7112
7113 return (write_asize);
7114 }
7115
7116 /*
7117 * This thread feeds the L2ARC at regular intervals. This is the beating
7118 * heart of the L2ARC.
7119 */
7120 /* ARGSUSED */
7121 static void
7122 l2arc_feed_thread(void *unused)
7123 {
7124 callb_cpr_t cpr;
7125 l2arc_dev_t *dev;
7126 spa_t *spa;
7127 uint64_t size, wrote;
7128 clock_t begin, next = ddi_get_lbolt();
7129
7130 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7131
7132 mutex_enter(&l2arc_feed_thr_lock);
7133
7134 while (l2arc_thread_exit == 0) {
7135 CALLB_CPR_SAFE_BEGIN(&cpr);
7136 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7137 next);
7138 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7139 next = ddi_get_lbolt() + hz;
7140
7141 /*
7142 * Quick check for L2ARC devices.
7143 */
7144 mutex_enter(&l2arc_dev_mtx);
7145 if (l2arc_ndev == 0) {
7146 mutex_exit(&l2arc_dev_mtx);
7147 continue;
7148 }
7149 mutex_exit(&l2arc_dev_mtx);
7150 begin = ddi_get_lbolt();
7151
7152 /*
7153 * This selects the next l2arc device to write to, and in
7154 * doing so the next spa to feed from: dev->l2ad_spa. This
7155 * will return NULL if there are now no l2arc devices or if
7156 * they are all faulted.
7157 *
7158 * If a device is returned, its spa's config lock is also
7159 * held to prevent device removal. l2arc_dev_get_next()
7160 * will grab and release l2arc_dev_mtx.
7161 */
7162 if ((dev = l2arc_dev_get_next()) == NULL)
7163 continue;
7164
7165 spa = dev->l2ad_spa;
7166 ASSERT3P(spa, !=, NULL);
7167
7168 /*
7169 * If the pool is read-only then force the feed thread to
7170 * sleep a little longer.
7171 */
7172 if (!spa_writeable(spa)) {
7173 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7174 spa_config_exit(spa, SCL_L2ARC, dev);
7175 continue;
7176 }
7177
7178 /*
7179 * Avoid contributing to memory pressure.
7180 */
7181 if (arc_reclaim_needed()) {
7182 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7183 spa_config_exit(spa, SCL_L2ARC, dev);
7184 continue;
7185 }
7186
7187 ARCSTAT_BUMP(arcstat_l2_feeds);
7188
7189 size = l2arc_write_size();
7190
7191 /*
7192 * Evict L2ARC buffers that will be overwritten.
7193 */
7194 l2arc_evict(dev, size, B_FALSE);
7195
7196 /*
7197 * Write ARC buffers.
7198 */
7199 wrote = l2arc_write_buffers(spa, dev, size);
7200
7201 /*
7202 * Calculate interval between writes.
7203 */
7204 next = l2arc_write_interval(begin, size, wrote);
7205 spa_config_exit(spa, SCL_L2ARC, dev);
7206 }
7207
7208 l2arc_thread_exit = 0;
7209 cv_broadcast(&l2arc_feed_thr_cv);
7210 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7211 thread_exit();
7212 }
7213
7214 boolean_t
7215 l2arc_vdev_present(vdev_t *vd)
7216 {
7217 l2arc_dev_t *dev;
7218
7219 mutex_enter(&l2arc_dev_mtx);
7220 for (dev = list_head(l2arc_dev_list); dev != NULL;
7221 dev = list_next(l2arc_dev_list, dev)) {
7222 if (dev->l2ad_vdev == vd)
7223 break;
7224 }
7225 mutex_exit(&l2arc_dev_mtx);
7226
7227 return (dev != NULL);
7228 }
7229
7230 /*
7231 * Add a vdev for use by the L2ARC. By this point the spa has already
7232 * validated the vdev and opened it.
7233 */
7234 void
7235 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7236 {
7237 l2arc_dev_t *adddev;
7238
7239 ASSERT(!l2arc_vdev_present(vd));
7240
7241 /*
7242 * Create a new l2arc device entry.
7243 */
7244 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7245 adddev->l2ad_spa = spa;
7246 adddev->l2ad_vdev = vd;
7247 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7248 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7249 adddev->l2ad_hand = adddev->l2ad_start;
7250 adddev->l2ad_first = B_TRUE;
7251 adddev->l2ad_writing = B_FALSE;
7252
7253 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7254 /*
7255 * This is a list of all ARC buffers that are still valid on the
7256 * device.
7257 */
7258 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7259 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7260
7261 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7262 refcount_create(&adddev->l2ad_alloc);
7263
7264 /*
7265 * Add device to global list
7266 */
7267 mutex_enter(&l2arc_dev_mtx);
7268 list_insert_head(l2arc_dev_list, adddev);
7269 atomic_inc_64(&l2arc_ndev);
7270 mutex_exit(&l2arc_dev_mtx);
7271 }
7272
7273 /*
7274 * Remove a vdev from the L2ARC.
7275 */
7276 void
7277 l2arc_remove_vdev(vdev_t *vd)
7278 {
7279 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7280
7281 /*
7282 * Find the device by vdev
7283 */
7284 mutex_enter(&l2arc_dev_mtx);
7285 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7286 nextdev = list_next(l2arc_dev_list, dev);
7287 if (vd == dev->l2ad_vdev) {
7288 remdev = dev;
7289 break;
7290 }
7291 }
7292 ASSERT3P(remdev, !=, NULL);
7293
7294 /*
7295 * Remove device from global list
7296 */
7297 list_remove(l2arc_dev_list, remdev);
7298 l2arc_dev_last = NULL; /* may have been invalidated */
7299 atomic_dec_64(&l2arc_ndev);
7300 mutex_exit(&l2arc_dev_mtx);
7301
7302 /*
7303 * Clear all buflists and ARC references. L2ARC device flush.
7304 */
7305 l2arc_evict(remdev, 0, B_TRUE);
7306 list_destroy(&remdev->l2ad_buflist);
7307 mutex_destroy(&remdev->l2ad_mtx);
7308 refcount_destroy(&remdev->l2ad_alloc);
7309 kmem_free(remdev, sizeof (l2arc_dev_t));
7310 }
7311
7312 void
7313 l2arc_init(void)
7314 {
7315 l2arc_thread_exit = 0;
7316 l2arc_ndev = 0;
7317 l2arc_writes_sent = 0;
7318 l2arc_writes_done = 0;
7319
7320 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7321 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7322 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7323 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7324
7325 l2arc_dev_list = &L2ARC_dev_list;
7326 l2arc_free_on_write = &L2ARC_free_on_write;
7327 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7328 offsetof(l2arc_dev_t, l2ad_node));
7329 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7330 offsetof(l2arc_data_free_t, l2df_list_node));
7331 }
7332
7333 void
7334 l2arc_fini(void)
7335 {
7336 /*
7337 * This is called from dmu_fini(), which is called from spa_fini();
7338 * Because of this, we can assume that all l2arc devices have
7339 * already been removed when the pools themselves were removed.
7340 */
7341
7342 l2arc_do_free_on_write();
7343
7344 mutex_destroy(&l2arc_feed_thr_lock);
7345 cv_destroy(&l2arc_feed_thr_cv);
7346 mutex_destroy(&l2arc_dev_mtx);
7347 mutex_destroy(&l2arc_free_on_write_mtx);
7348
7349 list_destroy(l2arc_dev_list);
7350 list_destroy(l2arc_free_on_write);
7351 }
7352
7353 void
7354 l2arc_start(void)
7355 {
7356 if (!(spa_mode_global & FWRITE))
7357 return;
7358
7359 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7360 TS_RUN, minclsyspri);
7361 }
7362
7363 void
7364 l2arc_stop(void)
7365 {
7366 if (!(spa_mode_global & FWRITE))
7367 return;
7368
7369 mutex_enter(&l2arc_feed_thr_lock);
7370 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7371 l2arc_thread_exit = 1;
7372 while (l2arc_thread_exit != 0)
7373 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7374 mutex_exit(&l2arc_feed_thr_lock);
7375 }