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