1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24 * Copyright (c) 2011, 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 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
305 int zfs_arc_overflow_shift = 8;
306
307 /* shift of arc_c for calculating both min and max arc_p */
308 static int arc_p_min_shift = 4;
309
310 /* log2(fraction of arc to reclaim) */
311 static int arc_shrink_shift = 7;
312
313 /*
314 * log2(fraction of ARC which must be free to allow growing).
315 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
316 * when reading a new block into the ARC, we will evict an equal-sized block
317 * from the ARC.
318 *
319 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
320 * we will still not allow it to grow.
321 */
322 int arc_no_grow_shift = 5;
323
324
325 /*
326 * minimum lifespan of a prefetch block in clock ticks
327 * (initialized in arc_init())
328 */
329 static int arc_min_prefetch_lifespan;
330
331 /*
332 * If this percent of memory is free, don't throttle.
333 */
334 int arc_lotsfree_percent = 10;
335
336 static int arc_dead;
337
338 /*
339 * The arc has filled available memory and has now warmed up.
340 */
341 static boolean_t arc_warm;
342
343 /*
344 * log2 fraction of the zio arena to keep free.
345 */
346 int arc_zio_arena_free_shift = 2;
347
348 /*
349 * These tunables are for performance analysis.
350 */
351 uint64_t zfs_arc_max;
352 uint64_t zfs_arc_min;
353 uint64_t zfs_arc_meta_limit = 0;
354 uint64_t zfs_arc_meta_min = 0;
355 int zfs_arc_grow_retry = 0;
356 int zfs_arc_shrink_shift = 0;
357 int zfs_arc_p_min_shift = 0;
358 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
359
360 boolean_t zfs_compressed_arc_enabled = B_TRUE;
361
362 /*
363 * Note that buffers can be in one of 6 states:
364 * ARC_anon - anonymous (discussed below)
365 * ARC_mru - recently used, currently cached
366 * ARC_mru_ghost - recentely used, no longer in cache
367 * ARC_mfu - frequently used, currently cached
368 * ARC_mfu_ghost - frequently used, no longer in cache
369 * ARC_l2c_only - exists in L2ARC but not other states
370 * When there are no active references to the buffer, they are
371 * are linked onto a list in one of these arc states. These are
372 * the only buffers that can be evicted or deleted. Within each
373 * state there are multiple lists, one for meta-data and one for
374 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
375 * etc.) is tracked separately so that it can be managed more
376 * explicitly: favored over data, limited explicitly.
377 *
378 * Anonymous buffers are buffers that are not associated with
379 * a DVA. These are buffers that hold dirty block copies
380 * before they are written to stable storage. By definition,
381 * they are "ref'd" and are considered part of arc_mru
382 * that cannot be freed. Generally, they will aquire a DVA
383 * as they are written and migrate onto the arc_mru list.
384 *
385 * The ARC_l2c_only state is for buffers that are in the second
386 * level ARC but no longer in any of the ARC_m* lists. The second
387 * level ARC itself may also contain buffers that are in any of
388 * the ARC_m* states - meaning that a buffer can exist in two
389 * places. The reason for the ARC_l2c_only state is to keep the
390 * buffer header in the hash table, so that reads that hit the
391 * second level ARC benefit from these fast lookups.
392 */
393
394 typedef struct arc_state {
395 /*
396 * list of evictable buffers
397 */
398 multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
399 /*
400 * total amount of evictable data in this state
401 */
402 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
403 /*
404 * total amount of data in this state; this includes: evictable,
405 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
406 */
407 refcount_t arcs_size;
408 } arc_state_t;
409
410 /* The 6 states: */
411 static arc_state_t ARC_anon;
412 static arc_state_t ARC_mru;
413 static arc_state_t ARC_mru_ghost;
414 static arc_state_t ARC_mfu;
415 static arc_state_t ARC_mfu_ghost;
416 static arc_state_t ARC_l2c_only;
417
418 typedef struct arc_stats {
419 kstat_named_t arcstat_hits;
420 kstat_named_t arcstat_misses;
421 kstat_named_t arcstat_demand_data_hits;
422 kstat_named_t arcstat_demand_data_misses;
423 kstat_named_t arcstat_demand_metadata_hits;
424 kstat_named_t arcstat_demand_metadata_misses;
425 kstat_named_t arcstat_prefetch_data_hits;
426 kstat_named_t arcstat_prefetch_data_misses;
427 kstat_named_t arcstat_prefetch_metadata_hits;
428 kstat_named_t arcstat_prefetch_metadata_misses;
429 kstat_named_t arcstat_mru_hits;
430 kstat_named_t arcstat_mru_ghost_hits;
431 kstat_named_t arcstat_mfu_hits;
432 kstat_named_t arcstat_mfu_ghost_hits;
433 kstat_named_t arcstat_deleted;
434 /*
435 * Number of buffers that could not be evicted because the hash lock
436 * was held by another thread. The lock may not necessarily be held
437 * by something using the same buffer, since hash locks are shared
438 * by multiple buffers.
439 */
440 kstat_named_t arcstat_mutex_miss;
441 /*
442 * Number of buffers skipped because they have I/O in progress, are
443 * indrect prefetch buffers that have not lived long enough, or are
444 * not from the spa we're trying to evict from.
445 */
446 kstat_named_t arcstat_evict_skip;
447 /*
448 * Number of times arc_evict_state() was unable to evict enough
449 * buffers to reach it's target amount.
450 */
451 kstat_named_t arcstat_evict_not_enough;
452 kstat_named_t arcstat_evict_l2_cached;
453 kstat_named_t arcstat_evict_l2_eligible;
454 kstat_named_t arcstat_evict_l2_ineligible;
455 kstat_named_t arcstat_evict_l2_skip;
456 kstat_named_t arcstat_hash_elements;
457 kstat_named_t arcstat_hash_elements_max;
458 kstat_named_t arcstat_hash_collisions;
459 kstat_named_t arcstat_hash_chains;
460 kstat_named_t arcstat_hash_chain_max;
461 kstat_named_t arcstat_p;
462 kstat_named_t arcstat_c;
463 kstat_named_t arcstat_c_min;
464 kstat_named_t arcstat_c_max;
465 kstat_named_t arcstat_size;
466 /*
467 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
468 * Note that the compressed bytes may match the uncompressed bytes
469 * if the block is either not compressed or compressed arc is disabled.
470 */
471 kstat_named_t arcstat_compressed_size;
472 /*
473 * Uncompressed size of the data stored in b_pabd. If compressed
474 * arc is disabled then this value will be identical to the stat
475 * above.
476 */
477 kstat_named_t arcstat_uncompressed_size;
478 /*
479 * Number of bytes stored in all the arc_buf_t's. This is classified
480 * as "overhead" since this data is typically short-lived and will
481 * be evicted from the arc when it becomes unreferenced unless the
482 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
483 * values have been set (see comment in dbuf.c for more information).
484 */
485 kstat_named_t arcstat_overhead_size;
486 /*
487 * Number of bytes consumed by internal ARC structures necessary
488 * for tracking purposes; these structures are not actually
489 * backed by ARC buffers. This includes arc_buf_hdr_t structures
490 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
491 * caches), and arc_buf_t structures (allocated via arc_buf_t
492 * cache).
493 */
494 kstat_named_t arcstat_hdr_size;
495 /*
496 * Number of bytes consumed by ARC buffers of type equal to
497 * ARC_BUFC_DATA. This is generally consumed by buffers backing
498 * on disk user data (e.g. plain file contents).
499 */
500 kstat_named_t arcstat_data_size;
501 /*
502 * Number of bytes consumed by ARC buffers of type equal to
503 * ARC_BUFC_METADATA. This is generally consumed by buffers
504 * backing on disk data that is used for internal ZFS
505 * structures (e.g. ZAP, dnode, indirect blocks, etc).
506 */
507 kstat_named_t arcstat_metadata_size;
508 /*
509 * Number of bytes consumed by various buffers and structures
510 * not actually backed with ARC buffers. This includes bonus
511 * buffers (allocated directly via zio_buf_* functions),
512 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
513 * cache), and dnode_t structures (allocated via dnode_t cache).
514 */
515 kstat_named_t arcstat_other_size;
516 /*
517 * Total number of bytes consumed by ARC buffers residing in the
518 * arc_anon state. This includes *all* buffers in the arc_anon
519 * state; e.g. data, metadata, evictable, and unevictable buffers
520 * are all included in this value.
521 */
522 kstat_named_t arcstat_anon_size;
523 /*
524 * Number of bytes consumed by ARC buffers that meet the
525 * following criteria: backing buffers of type ARC_BUFC_DATA,
526 * residing in the arc_anon state, and are eligible for eviction
527 * (e.g. have no outstanding holds on the buffer).
528 */
529 kstat_named_t arcstat_anon_evictable_data;
530 /*
531 * Number of bytes consumed by ARC buffers that meet the
532 * following criteria: backing buffers of type ARC_BUFC_METADATA,
533 * residing in the arc_anon state, and are eligible for eviction
534 * (e.g. have no outstanding holds on the buffer).
535 */
536 kstat_named_t arcstat_anon_evictable_metadata;
537 /*
538 * Total number of bytes consumed by ARC buffers residing in the
539 * arc_mru state. This includes *all* buffers in the arc_mru
540 * state; e.g. data, metadata, evictable, and unevictable buffers
541 * are all included in this value.
542 */
543 kstat_named_t arcstat_mru_size;
544 /*
545 * Number of bytes consumed by ARC buffers that meet the
546 * following criteria: backing buffers of type ARC_BUFC_DATA,
547 * residing in the arc_mru state, and are eligible for eviction
548 * (e.g. have no outstanding holds on the buffer).
549 */
550 kstat_named_t arcstat_mru_evictable_data;
551 /*
552 * Number of bytes consumed by ARC buffers that meet the
553 * following criteria: backing buffers of type ARC_BUFC_METADATA,
554 * residing in the arc_mru state, and are eligible for eviction
555 * (e.g. have no outstanding holds on the buffer).
556 */
557 kstat_named_t arcstat_mru_evictable_metadata;
558 /*
559 * Total number of bytes that *would have been* consumed by ARC
560 * buffers in the arc_mru_ghost state. The key thing to note
561 * here, is the fact that this size doesn't actually indicate
562 * RAM consumption. The ghost lists only consist of headers and
563 * don't actually have ARC buffers linked off of these headers.
564 * Thus, *if* the headers had associated ARC buffers, these
565 * buffers *would have* consumed this number of bytes.
566 */
567 kstat_named_t arcstat_mru_ghost_size;
568 /*
569 * Number of bytes that *would have been* consumed by ARC
570 * buffers that are eligible for eviction, of type
571 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
572 */
573 kstat_named_t arcstat_mru_ghost_evictable_data;
574 /*
575 * Number of bytes that *would have been* consumed by ARC
576 * buffers that are eligible for eviction, of type
577 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
578 */
579 kstat_named_t arcstat_mru_ghost_evictable_metadata;
580 /*
581 * Total number of bytes consumed by ARC buffers residing in the
582 * arc_mfu state. This includes *all* buffers in the arc_mfu
583 * state; e.g. data, metadata, evictable, and unevictable buffers
584 * are all included in this value.
585 */
586 kstat_named_t arcstat_mfu_size;
587 /*
588 * Number of bytes consumed by ARC buffers that are eligible for
589 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
590 * state.
591 */
592 kstat_named_t arcstat_mfu_evictable_data;
593 /*
594 * Number of bytes consumed by ARC buffers that are eligible for
595 * eviction, of type ARC_BUFC_METADATA, and reside in the
596 * arc_mfu state.
597 */
598 kstat_named_t arcstat_mfu_evictable_metadata;
599 /*
600 * Total number of bytes that *would have been* consumed by ARC
601 * buffers in the arc_mfu_ghost state. See the comment above
602 * arcstat_mru_ghost_size for more details.
603 */
604 kstat_named_t arcstat_mfu_ghost_size;
605 /*
606 * Number of bytes that *would have been* consumed by ARC
607 * buffers that are eligible for eviction, of type
608 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
609 */
610 kstat_named_t arcstat_mfu_ghost_evictable_data;
611 /*
612 * Number of bytes that *would have been* consumed by ARC
613 * buffers that are eligible for eviction, of type
614 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
615 */
616 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
617 kstat_named_t arcstat_l2_hits;
618 kstat_named_t arcstat_l2_misses;
619 kstat_named_t arcstat_l2_feeds;
620 kstat_named_t arcstat_l2_rw_clash;
621 kstat_named_t arcstat_l2_read_bytes;
622 kstat_named_t arcstat_l2_write_bytes;
623 kstat_named_t arcstat_l2_writes_sent;
624 kstat_named_t arcstat_l2_writes_done;
625 kstat_named_t arcstat_l2_writes_error;
626 kstat_named_t arcstat_l2_writes_lock_retry;
627 kstat_named_t arcstat_l2_evict_lock_retry;
628 kstat_named_t arcstat_l2_evict_reading;
629 kstat_named_t arcstat_l2_evict_l1cached;
630 kstat_named_t arcstat_l2_free_on_write;
631 kstat_named_t arcstat_l2_abort_lowmem;
632 kstat_named_t arcstat_l2_cksum_bad;
633 kstat_named_t arcstat_l2_io_error;
634 kstat_named_t arcstat_l2_lsize;
635 kstat_named_t arcstat_l2_psize;
636 kstat_named_t arcstat_l2_hdr_size;
637 kstat_named_t arcstat_memory_throttle_count;
638 kstat_named_t arcstat_meta_used;
639 kstat_named_t arcstat_meta_limit;
640 kstat_named_t arcstat_meta_max;
641 kstat_named_t arcstat_meta_min;
642 kstat_named_t arcstat_sync_wait_for_async;
643 kstat_named_t arcstat_demand_hit_predictive_prefetch;
644 } arc_stats_t;
645
646 static arc_stats_t arc_stats = {
647 { "hits", KSTAT_DATA_UINT64 },
648 { "misses", KSTAT_DATA_UINT64 },
649 { "demand_data_hits", KSTAT_DATA_UINT64 },
650 { "demand_data_misses", KSTAT_DATA_UINT64 },
651 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
652 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
653 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
654 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
655 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
656 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
657 { "mru_hits", KSTAT_DATA_UINT64 },
658 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
659 { "mfu_hits", KSTAT_DATA_UINT64 },
660 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
661 { "deleted", KSTAT_DATA_UINT64 },
662 { "mutex_miss", KSTAT_DATA_UINT64 },
663 { "evict_skip", KSTAT_DATA_UINT64 },
664 { "evict_not_enough", KSTAT_DATA_UINT64 },
665 { "evict_l2_cached", KSTAT_DATA_UINT64 },
666 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
667 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
668 { "evict_l2_skip", KSTAT_DATA_UINT64 },
669 { "hash_elements", KSTAT_DATA_UINT64 },
670 { "hash_elements_max", KSTAT_DATA_UINT64 },
671 { "hash_collisions", KSTAT_DATA_UINT64 },
672 { "hash_chains", KSTAT_DATA_UINT64 },
673 { "hash_chain_max", KSTAT_DATA_UINT64 },
674 { "p", KSTAT_DATA_UINT64 },
675 { "c", KSTAT_DATA_UINT64 },
676 { "c_min", KSTAT_DATA_UINT64 },
677 { "c_max", KSTAT_DATA_UINT64 },
678 { "size", KSTAT_DATA_UINT64 },
679 { "compressed_size", KSTAT_DATA_UINT64 },
680 { "uncompressed_size", KSTAT_DATA_UINT64 },
681 { "overhead_size", KSTAT_DATA_UINT64 },
682 { "hdr_size", KSTAT_DATA_UINT64 },
683 { "data_size", KSTAT_DATA_UINT64 },
684 { "metadata_size", KSTAT_DATA_UINT64 },
685 { "other_size", KSTAT_DATA_UINT64 },
686 { "anon_size", KSTAT_DATA_UINT64 },
687 { "anon_evictable_data", KSTAT_DATA_UINT64 },
688 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
689 { "mru_size", KSTAT_DATA_UINT64 },
690 { "mru_evictable_data", KSTAT_DATA_UINT64 },
691 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
692 { "mru_ghost_size", KSTAT_DATA_UINT64 },
693 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
694 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
695 { "mfu_size", KSTAT_DATA_UINT64 },
696 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
697 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
698 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
699 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
700 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
701 { "l2_hits", KSTAT_DATA_UINT64 },
702 { "l2_misses", KSTAT_DATA_UINT64 },
703 { "l2_feeds", KSTAT_DATA_UINT64 },
704 { "l2_rw_clash", KSTAT_DATA_UINT64 },
705 { "l2_read_bytes", KSTAT_DATA_UINT64 },
706 { "l2_write_bytes", KSTAT_DATA_UINT64 },
707 { "l2_writes_sent", KSTAT_DATA_UINT64 },
708 { "l2_writes_done", KSTAT_DATA_UINT64 },
709 { "l2_writes_error", KSTAT_DATA_UINT64 },
710 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
711 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
712 { "l2_evict_reading", KSTAT_DATA_UINT64 },
713 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
714 { "l2_free_on_write", KSTAT_DATA_UINT64 },
715 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
716 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
717 { "l2_io_error", KSTAT_DATA_UINT64 },
718 { "l2_size", KSTAT_DATA_UINT64 },
719 { "l2_asize", KSTAT_DATA_UINT64 },
720 { "l2_hdr_size", KSTAT_DATA_UINT64 },
721 { "memory_throttle_count", KSTAT_DATA_UINT64 },
722 { "arc_meta_used", KSTAT_DATA_UINT64 },
723 { "arc_meta_limit", KSTAT_DATA_UINT64 },
724 { "arc_meta_max", KSTAT_DATA_UINT64 },
725 { "arc_meta_min", KSTAT_DATA_UINT64 },
726 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
727 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
728 };
729
730 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
731
732 #define ARCSTAT_INCR(stat, val) \
733 atomic_add_64(&arc_stats.stat.value.ui64, (val))
734
735 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
736 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
737
738 #define ARCSTAT_MAX(stat, val) { \
739 uint64_t m; \
740 while ((val) > (m = arc_stats.stat.value.ui64) && \
741 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
742 continue; \
743 }
744
745 #define ARCSTAT_MAXSTAT(stat) \
746 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
747
748 /*
749 * We define a macro to allow ARC hits/misses to be easily broken down by
750 * two separate conditions, giving a total of four different subtypes for
751 * each of hits and misses (so eight statistics total).
752 */
753 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
754 if (cond1) { \
755 if (cond2) { \
756 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
757 } else { \
758 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
759 } \
760 } else { \
761 if (cond2) { \
762 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
763 } else { \
764 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
765 } \
766 }
767
768 kstat_t *arc_ksp;
769 static arc_state_t *arc_anon;
770 static arc_state_t *arc_mru;
771 static arc_state_t *arc_mru_ghost;
772 static arc_state_t *arc_mfu;
773 static arc_state_t *arc_mfu_ghost;
774 static arc_state_t *arc_l2c_only;
775
776 /*
777 * There are several ARC variables that are critical to export as kstats --
778 * but we don't want to have to grovel around in the kstat whenever we wish to
779 * manipulate them. For these variables, we therefore define them to be in
780 * terms of the statistic variable. This assures that we are not introducing
781 * the possibility of inconsistency by having shadow copies of the variables,
782 * while still allowing the code to be readable.
783 */
784 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
785 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
786 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
787 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
788 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
789 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
790 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
791 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
792 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
793
794 /* compressed size of entire arc */
795 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
796 /* uncompressed size of entire arc */
797 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
798 /* number of bytes in the arc from arc_buf_t's */
799 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
800
801 static int arc_no_grow; /* Don't try to grow cache size */
802 static uint64_t arc_tempreserve;
803 static uint64_t arc_loaned_bytes;
804
805 typedef struct arc_callback arc_callback_t;
806
807 struct arc_callback {
808 void *acb_private;
809 arc_done_func_t *acb_done;
810 arc_buf_t *acb_buf;
811 boolean_t acb_compressed;
812 zio_t *acb_zio_dummy;
813 arc_callback_t *acb_next;
814 };
815
816 typedef struct arc_write_callback arc_write_callback_t;
817
818 struct arc_write_callback {
819 void *awcb_private;
820 arc_done_func_t *awcb_ready;
821 arc_done_func_t *awcb_children_ready;
822 arc_done_func_t *awcb_physdone;
823 arc_done_func_t *awcb_done;
824 arc_buf_t *awcb_buf;
825 };
826
827 /*
828 * ARC buffers are separated into multiple structs as a memory saving measure:
829 * - Common fields struct, always defined, and embedded within it:
830 * - L2-only fields, always allocated but undefined when not in L2ARC
831 * - L1-only fields, only allocated when in L1ARC
832 *
833 * Buffer in L1 Buffer only in L2
834 * +------------------------+ +------------------------+
835 * | arc_buf_hdr_t | | arc_buf_hdr_t |
836 * | | | |
837 * | | | |
838 * | | | |
839 * +------------------------+ +------------------------+
840 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
841 * | (undefined if L1-only) | | |
842 * +------------------------+ +------------------------+
843 * | l1arc_buf_hdr_t |
844 * | |
845 * | |
846 * | |
847 * | |
848 * +------------------------+
849 *
850 * Because it's possible for the L2ARC to become extremely large, we can wind
851 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
852 * is minimized by only allocating the fields necessary for an L1-cached buffer
853 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
854 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
855 * words in pointers. arc_hdr_realloc() is used to switch a header between
856 * these two allocation states.
857 */
858 typedef struct l1arc_buf_hdr {
859 kmutex_t b_freeze_lock;
860 zio_cksum_t *b_freeze_cksum;
861 #ifdef ZFS_DEBUG
862 /*
863 * Used for debugging with kmem_flags - by allocating and freeing
864 * b_thawed when the buffer is thawed, we get a record of the stack
865 * trace that thawed it.
866 */
867 void *b_thawed;
868 #endif
869
870 arc_buf_t *b_buf;
871 uint32_t b_bufcnt;
872 /* for waiting on writes to complete */
873 kcondvar_t b_cv;
874 uint8_t b_byteswap;
875
876 /* protected by arc state mutex */
877 arc_state_t *b_state;
878 multilist_node_t b_arc_node;
879
880 /* updated atomically */
881 clock_t b_arc_access;
882
883 /* self protecting */
884 refcount_t b_refcnt;
885
886 arc_callback_t *b_acb;
887 abd_t *b_pabd;
888 } l1arc_buf_hdr_t;
889
890 typedef struct l2arc_dev l2arc_dev_t;
891
892 typedef struct l2arc_buf_hdr {
893 /* protected by arc_buf_hdr mutex */
894 l2arc_dev_t *b_dev; /* L2ARC device */
895 uint64_t b_daddr; /* disk address, offset byte */
896
897 list_node_t b_l2node;
898 } l2arc_buf_hdr_t;
899
900 struct arc_buf_hdr {
901 /* protected by hash lock */
902 dva_t b_dva;
903 uint64_t b_birth;
904
905 arc_buf_contents_t b_type;
906 arc_buf_hdr_t *b_hash_next;
907 arc_flags_t b_flags;
908
909 /*
910 * This field stores the size of the data buffer after
911 * compression, and is set in the arc's zio completion handlers.
912 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
913 *
914 * While the block pointers can store up to 32MB in their psize
915 * field, we can only store up to 32MB minus 512B. This is due
916 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
917 * a field of zeros represents 512B in the bp). We can't use a
918 * bias of 1 since we need to reserve a psize of zero, here, to
919 * represent holes and embedded blocks.
920 *
921 * This isn't a problem in practice, since the maximum size of a
922 * buffer is limited to 16MB, so we never need to store 32MB in
923 * this field. Even in the upstream illumos code base, the
924 * maximum size of a buffer is limited to 16MB.
925 */
926 uint16_t b_psize;
927
928 /*
929 * This field stores the size of the data buffer before
930 * compression, and cannot change once set. It is in units
931 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes)
932 */
933 uint16_t b_lsize; /* immutable */
934 uint64_t b_spa; /* immutable */
935
936 /* L2ARC fields. Undefined when not in L2ARC. */
937 l2arc_buf_hdr_t b_l2hdr;
938 /* L1ARC fields. Undefined when in l2arc_only state */
939 l1arc_buf_hdr_t b_l1hdr;
940 };
941
942 #define GHOST_STATE(state) \
943 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
944 (state) == arc_l2c_only)
945
946 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
947 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
948 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
949 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
950 #define HDR_COMPRESSION_ENABLED(hdr) \
951 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
952
953 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
954 #define HDR_L2_READING(hdr) \
955 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
956 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
957 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
958 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
959 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
960 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
961
962 #define HDR_ISTYPE_METADATA(hdr) \
963 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
964 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
965
966 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
967 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
968
969 /* For storing compression mode in b_flags */
970 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
971
972 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
973 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
974 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
975 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
976
977 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
978 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
979 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
980
981 /*
982 * Other sizes
983 */
984
985 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
986 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
987
988 /*
989 * Hash table routines
990 */
991
992 #define HT_LOCK_PAD 64
993
994 struct ht_lock {
995 kmutex_t ht_lock;
996 #ifdef _KERNEL
997 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
998 #endif
999 };
1000
1001 #define BUF_LOCKS 256
1002 typedef struct buf_hash_table {
1003 uint64_t ht_mask;
1004 arc_buf_hdr_t **ht_table;
1005 struct ht_lock ht_locks[BUF_LOCKS];
1006 } buf_hash_table_t;
1007
1008 static buf_hash_table_t buf_hash_table;
1009
1010 #define BUF_HASH_INDEX(spa, dva, birth) \
1011 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1012 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1013 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1014 #define HDR_LOCK(hdr) \
1015 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1016
1017 uint64_t zfs_crc64_table[256];
1018
1019 /*
1020 * Level 2 ARC
1021 */
1022
1023 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1024 #define L2ARC_HEADROOM 2 /* num of writes */
1025 /*
1026 * If we discover during ARC scan any buffers to be compressed, we boost
1027 * our headroom for the next scanning cycle by this percentage multiple.
1028 */
1029 #define L2ARC_HEADROOM_BOOST 200
1030 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1031 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1032
1033 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1034 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1035
1036 /* L2ARC Performance Tunables */
1037 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1038 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1039 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1040 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1041 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1042 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1043 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1044 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1045 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1046
1047 /*
1048 * L2ARC Internals
1049 */
1050 struct l2arc_dev {
1051 vdev_t *l2ad_vdev; /* vdev */
1052 spa_t *l2ad_spa; /* spa */
1053 uint64_t l2ad_hand; /* next write location */
1054 uint64_t l2ad_start; /* first addr on device */
1055 uint64_t l2ad_end; /* last addr on device */
1056 boolean_t l2ad_first; /* first sweep through */
1057 boolean_t l2ad_writing; /* currently writing */
1058 kmutex_t l2ad_mtx; /* lock for buffer list */
1059 list_t l2ad_buflist; /* buffer list */
1060 list_node_t l2ad_node; /* device list node */
1061 refcount_t l2ad_alloc; /* allocated bytes */
1062 };
1063
1064 static list_t L2ARC_dev_list; /* device list */
1065 static list_t *l2arc_dev_list; /* device list pointer */
1066 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1067 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1068 static list_t L2ARC_free_on_write; /* free after write buf list */
1069 static list_t *l2arc_free_on_write; /* free after write list ptr */
1070 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1071 static uint64_t l2arc_ndev; /* number of devices */
1072
1073 typedef struct l2arc_read_callback {
1074 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1075 blkptr_t l2rcb_bp; /* original blkptr */
1076 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1077 int l2rcb_flags; /* original flags */
1078 abd_t *l2rcb_abd; /* temporary buffer */
1079 } l2arc_read_callback_t;
1080
1081 typedef struct l2arc_write_callback {
1082 l2arc_dev_t *l2wcb_dev; /* device info */
1083 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1084 } l2arc_write_callback_t;
1085
1086 typedef struct l2arc_data_free {
1087 /* protected by l2arc_free_on_write_mtx */
1088 abd_t *l2df_abd;
1089 size_t l2df_size;
1090 arc_buf_contents_t l2df_type;
1091 list_node_t l2df_list_node;
1092 } l2arc_data_free_t;
1093
1094 static kmutex_t l2arc_feed_thr_lock;
1095 static kcondvar_t l2arc_feed_thr_cv;
1096 static uint8_t l2arc_thread_exit;
1097
1098 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1099 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1100 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1101 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1102 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1103 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1104 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1105 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1106 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1107 static boolean_t arc_is_overflowing();
1108 static void arc_buf_watch(arc_buf_t *);
1109
1110 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1111 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1112 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1113 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1114
1115 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1116 static void l2arc_read_done(zio_t *);
1117
1118 static uint64_t
1119 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1120 {
1121 uint8_t *vdva = (uint8_t *)dva;
1122 uint64_t crc = -1ULL;
1123 int i;
1124
1125 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1126
1127 for (i = 0; i < sizeof (dva_t); i++)
1128 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1129
1130 crc ^= (spa>>8) ^ birth;
1131
1132 return (crc);
1133 }
1134
1135 #define HDR_EMPTY(hdr) \
1136 ((hdr)->b_dva.dva_word[0] == 0 && \
1137 (hdr)->b_dva.dva_word[1] == 0)
1138
1139 #define HDR_EQUAL(spa, dva, birth, hdr) \
1140 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1141 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1142 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1143
1144 static void
1145 buf_discard_identity(arc_buf_hdr_t *hdr)
1146 {
1147 hdr->b_dva.dva_word[0] = 0;
1148 hdr->b_dva.dva_word[1] = 0;
1149 hdr->b_birth = 0;
1150 }
1151
1152 static arc_buf_hdr_t *
1153 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1154 {
1155 const dva_t *dva = BP_IDENTITY(bp);
1156 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1157 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1158 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1159 arc_buf_hdr_t *hdr;
1160
1161 mutex_enter(hash_lock);
1162 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1163 hdr = hdr->b_hash_next) {
1164 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1165 *lockp = hash_lock;
1166 return (hdr);
1167 }
1168 }
1169 mutex_exit(hash_lock);
1170 *lockp = NULL;
1171 return (NULL);
1172 }
1173
1174 /*
1175 * Insert an entry into the hash table. If there is already an element
1176 * equal to elem in the hash table, then the already existing element
1177 * will be returned and the new element will not be inserted.
1178 * Otherwise returns NULL.
1179 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1180 */
1181 static arc_buf_hdr_t *
1182 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1183 {
1184 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1185 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1186 arc_buf_hdr_t *fhdr;
1187 uint32_t i;
1188
1189 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1190 ASSERT(hdr->b_birth != 0);
1191 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1192
1193 if (lockp != NULL) {
1194 *lockp = hash_lock;
1195 mutex_enter(hash_lock);
1196 } else {
1197 ASSERT(MUTEX_HELD(hash_lock));
1198 }
1199
1200 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1201 fhdr = fhdr->b_hash_next, i++) {
1202 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1203 return (fhdr);
1204 }
1205
1206 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1207 buf_hash_table.ht_table[idx] = hdr;
1208 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1209
1210 /* collect some hash table performance data */
1211 if (i > 0) {
1212 ARCSTAT_BUMP(arcstat_hash_collisions);
1213 if (i == 1)
1214 ARCSTAT_BUMP(arcstat_hash_chains);
1215
1216 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1217 }
1218
1219 ARCSTAT_BUMP(arcstat_hash_elements);
1220 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1221
1222 return (NULL);
1223 }
1224
1225 static void
1226 buf_hash_remove(arc_buf_hdr_t *hdr)
1227 {
1228 arc_buf_hdr_t *fhdr, **hdrp;
1229 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1230
1231 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1232 ASSERT(HDR_IN_HASH_TABLE(hdr));
1233
1234 hdrp = &buf_hash_table.ht_table[idx];
1235 while ((fhdr = *hdrp) != hdr) {
1236 ASSERT3P(fhdr, !=, NULL);
1237 hdrp = &fhdr->b_hash_next;
1238 }
1239 *hdrp = hdr->b_hash_next;
1240 hdr->b_hash_next = NULL;
1241 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1242
1243 /* collect some hash table performance data */
1244 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1245
1246 if (buf_hash_table.ht_table[idx] &&
1247 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1248 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1249 }
1250
1251 /*
1252 * Global data structures and functions for the buf kmem cache.
1253 */
1254 static kmem_cache_t *hdr_full_cache;
1255 static kmem_cache_t *hdr_l2only_cache;
1256 static kmem_cache_t *buf_cache;
1257
1258 static void
1259 buf_fini(void)
1260 {
1261 int i;
1262
1263 kmem_free(buf_hash_table.ht_table,
1264 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1265 for (i = 0; i < BUF_LOCKS; i++)
1266 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1267 kmem_cache_destroy(hdr_full_cache);
1268 kmem_cache_destroy(hdr_l2only_cache);
1269 kmem_cache_destroy(buf_cache);
1270 }
1271
1272 /*
1273 * Constructor callback - called when the cache is empty
1274 * and a new buf is requested.
1275 */
1276 /* ARGSUSED */
1277 static int
1278 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1279 {
1280 arc_buf_hdr_t *hdr = vbuf;
1281
1282 bzero(hdr, HDR_FULL_SIZE);
1283 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1284 refcount_create(&hdr->b_l1hdr.b_refcnt);
1285 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1286 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1287 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1288
1289 return (0);
1290 }
1291
1292 /* ARGSUSED */
1293 static int
1294 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1295 {
1296 arc_buf_hdr_t *hdr = vbuf;
1297
1298 bzero(hdr, HDR_L2ONLY_SIZE);
1299 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1300
1301 return (0);
1302 }
1303
1304 /* ARGSUSED */
1305 static int
1306 buf_cons(void *vbuf, void *unused, int kmflag)
1307 {
1308 arc_buf_t *buf = vbuf;
1309
1310 bzero(buf, sizeof (arc_buf_t));
1311 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1312 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1313
1314 return (0);
1315 }
1316
1317 /*
1318 * Destructor callback - called when a cached buf is
1319 * no longer required.
1320 */
1321 /* ARGSUSED */
1322 static void
1323 hdr_full_dest(void *vbuf, void *unused)
1324 {
1325 arc_buf_hdr_t *hdr = vbuf;
1326
1327 ASSERT(HDR_EMPTY(hdr));
1328 cv_destroy(&hdr->b_l1hdr.b_cv);
1329 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1330 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1331 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1332 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1333 }
1334
1335 /* ARGSUSED */
1336 static void
1337 hdr_l2only_dest(void *vbuf, void *unused)
1338 {
1339 arc_buf_hdr_t *hdr = vbuf;
1340
1341 ASSERT(HDR_EMPTY(hdr));
1342 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1343 }
1344
1345 /* ARGSUSED */
1346 static void
1347 buf_dest(void *vbuf, void *unused)
1348 {
1349 arc_buf_t *buf = vbuf;
1350
1351 mutex_destroy(&buf->b_evict_lock);
1352 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1353 }
1354
1355 /*
1356 * Reclaim callback -- invoked when memory is low.
1357 */
1358 /* ARGSUSED */
1359 static void
1360 hdr_recl(void *unused)
1361 {
1362 dprintf("hdr_recl called\n");
1363 /*
1364 * umem calls the reclaim func when we destroy the buf cache,
1365 * which is after we do arc_fini().
1366 */
1367 if (!arc_dead)
1368 cv_signal(&arc_reclaim_thread_cv);
1369 }
1370
1371 static void
1372 buf_init(void)
1373 {
1374 uint64_t *ct;
1375 uint64_t hsize = 1ULL << 12;
1376 int i, j;
1377
1378 /*
1379 * The hash table is big enough to fill all of physical memory
1380 * with an average block size of zfs_arc_average_blocksize (default 8K).
1381 * By default, the table will take up
1382 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1383 */
1384 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1385 hsize <<= 1;
1386 retry:
1387 buf_hash_table.ht_mask = hsize - 1;
1388 buf_hash_table.ht_table =
1389 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1390 if (buf_hash_table.ht_table == NULL) {
1391 ASSERT(hsize > (1ULL << 8));
1392 hsize >>= 1;
1393 goto retry;
1394 }
1395
1396 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1397 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1398 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1399 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1400 NULL, NULL, 0);
1401 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1402 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1403
1404 for (i = 0; i < 256; i++)
1405 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1406 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1407
1408 for (i = 0; i < BUF_LOCKS; i++) {
1409 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1410 NULL, MUTEX_DEFAULT, NULL);
1411 }
1412 }
1413
1414 /*
1415 * This is the size that the buf occupies in memory. If the buf is compressed,
1416 * it will correspond to the compressed size. You should use this method of
1417 * getting the buf size unless you explicitly need the logical size.
1418 */
1419 int32_t
1420 arc_buf_size(arc_buf_t *buf)
1421 {
1422 return (ARC_BUF_COMPRESSED(buf) ?
1423 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1424 }
1425
1426 int32_t
1427 arc_buf_lsize(arc_buf_t *buf)
1428 {
1429 return (HDR_GET_LSIZE(buf->b_hdr));
1430 }
1431
1432 enum zio_compress
1433 arc_get_compression(arc_buf_t *buf)
1434 {
1435 return (ARC_BUF_COMPRESSED(buf) ?
1436 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1437 }
1438
1439 #define ARC_MINTIME (hz>>4) /* 62 ms */
1440
1441 static inline boolean_t
1442 arc_buf_is_shared(arc_buf_t *buf)
1443 {
1444 boolean_t shared = (buf->b_data != NULL &&
1445 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1446 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1447 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1448 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1449 IMPLY(shared, ARC_BUF_SHARED(buf));
1450 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1451
1452 /*
1453 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1454 * already being shared" requirement prevents us from doing that.
1455 */
1456
1457 return (shared);
1458 }
1459
1460 /*
1461 * Free the checksum associated with this header. If there is no checksum, this
1462 * is a no-op.
1463 */
1464 static inline void
1465 arc_cksum_free(arc_buf_hdr_t *hdr)
1466 {
1467 ASSERT(HDR_HAS_L1HDR(hdr));
1468 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1469 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1470 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1471 hdr->b_l1hdr.b_freeze_cksum = NULL;
1472 }
1473 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1474 }
1475
1476 /*
1477 * Return true iff at least one of the bufs on hdr is not compressed.
1478 */
1479 static boolean_t
1480 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1481 {
1482 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1483 if (!ARC_BUF_COMPRESSED(b)) {
1484 return (B_TRUE);
1485 }
1486 }
1487 return (B_FALSE);
1488 }
1489
1490 /*
1491 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1492 * matches the checksum that is stored in the hdr. If there is no checksum,
1493 * or if the buf is compressed, this is a no-op.
1494 */
1495 static void
1496 arc_cksum_verify(arc_buf_t *buf)
1497 {
1498 arc_buf_hdr_t *hdr = buf->b_hdr;
1499 zio_cksum_t zc;
1500
1501 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1502 return;
1503
1504 if (ARC_BUF_COMPRESSED(buf)) {
1505 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1506 arc_hdr_has_uncompressed_buf(hdr));
1507 return;
1508 }
1509
1510 ASSERT(HDR_HAS_L1HDR(hdr));
1511
1512 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1513 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1514 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1515 return;
1516 }
1517
1518 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1519 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1520 panic("buffer modified while frozen!");
1521 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1522 }
1523
1524 static boolean_t
1525 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1526 {
1527 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1528 boolean_t valid_cksum;
1529
1530 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1531 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1532
1533 /*
1534 * We rely on the blkptr's checksum to determine if the block
1535 * is valid or not. When compressed arc is enabled, the l2arc
1536 * writes the block to the l2arc just as it appears in the pool.
1537 * This allows us to use the blkptr's checksum to validate the
1538 * data that we just read off of the l2arc without having to store
1539 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1540 * arc is disabled, then the data written to the l2arc is always
1541 * uncompressed and won't match the block as it exists in the main
1542 * pool. When this is the case, we must first compress it if it is
1543 * compressed on the main pool before we can validate the checksum.
1544 */
1545 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1546 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1547 uint64_t lsize = HDR_GET_LSIZE(hdr);
1548 uint64_t csize;
1549
1550 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1551 csize = zio_compress_data(compress, zio->io_abd,
1552 abd_to_buf(cdata), lsize);
1553
1554 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1555 if (csize < HDR_GET_PSIZE(hdr)) {
1556 /*
1557 * Compressed blocks are always a multiple of the
1558 * smallest ashift in the pool. Ideally, we would
1559 * like to round up the csize to the next
1560 * spa_min_ashift but that value may have changed
1561 * since the block was last written. Instead,
1562 * we rely on the fact that the hdr's psize
1563 * was set to the psize of the block when it was
1564 * last written. We set the csize to that value
1565 * and zero out any part that should not contain
1566 * data.
1567 */
1568 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
1569 csize = HDR_GET_PSIZE(hdr);
1570 }
1571 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
1572 }
1573
1574 /*
1575 * Block pointers always store the checksum for the logical data.
1576 * If the block pointer has the gang bit set, then the checksum
1577 * it represents is for the reconstituted data and not for an
1578 * individual gang member. The zio pipeline, however, must be able to
1579 * determine the checksum of each of the gang constituents so it
1580 * treats the checksum comparison differently than what we need
1581 * for l2arc blocks. This prevents us from using the
1582 * zio_checksum_error() interface directly. Instead we must call the
1583 * zio_checksum_error_impl() so that we can ensure the checksum is
1584 * generated using the correct checksum algorithm and accounts for the
1585 * logical I/O size and not just a gang fragment.
1586 */
1587 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1588 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1589 zio->io_offset, NULL) == 0);
1590 zio_pop_transforms(zio);
1591 return (valid_cksum);
1592 }
1593
1594 /*
1595 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1596 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1597 * isn't modified later on. If buf is compressed or there is already a checksum
1598 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1599 */
1600 static void
1601 arc_cksum_compute(arc_buf_t *buf)
1602 {
1603 arc_buf_hdr_t *hdr = buf->b_hdr;
1604
1605 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1606 return;
1607
1608 ASSERT(HDR_HAS_L1HDR(hdr));
1609
1610 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1611 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1612 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
1613 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1614 return;
1615 } else if (ARC_BUF_COMPRESSED(buf)) {
1616 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1617 return;
1618 }
1619
1620 ASSERT(!ARC_BUF_COMPRESSED(buf));
1621 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1622 KM_SLEEP);
1623 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1624 hdr->b_l1hdr.b_freeze_cksum);
1625 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1626 arc_buf_watch(buf);
1627 }
1628
1629 #ifndef _KERNEL
1630 typedef struct procctl {
1631 long cmd;
1632 prwatch_t prwatch;
1633 } procctl_t;
1634 #endif
1635
1636 /* ARGSUSED */
1637 static void
1638 arc_buf_unwatch(arc_buf_t *buf)
1639 {
1640 #ifndef _KERNEL
1641 if (arc_watch) {
1642 int result;
1643 procctl_t ctl;
1644 ctl.cmd = PCWATCH;
1645 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1646 ctl.prwatch.pr_size = 0;
1647 ctl.prwatch.pr_wflags = 0;
1648 result = write(arc_procfd, &ctl, sizeof (ctl));
1649 ASSERT3U(result, ==, sizeof (ctl));
1650 }
1651 #endif
1652 }
1653
1654 /* ARGSUSED */
1655 static void
1656 arc_buf_watch(arc_buf_t *buf)
1657 {
1658 #ifndef _KERNEL
1659 if (arc_watch) {
1660 int result;
1661 procctl_t ctl;
1662 ctl.cmd = PCWATCH;
1663 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1664 ctl.prwatch.pr_size = arc_buf_size(buf);
1665 ctl.prwatch.pr_wflags = WA_WRITE;
1666 result = write(arc_procfd, &ctl, sizeof (ctl));
1667 ASSERT3U(result, ==, sizeof (ctl));
1668 }
1669 #endif
1670 }
1671
1672 static arc_buf_contents_t
1673 arc_buf_type(arc_buf_hdr_t *hdr)
1674 {
1675 arc_buf_contents_t type;
1676 if (HDR_ISTYPE_METADATA(hdr)) {
1677 type = ARC_BUFC_METADATA;
1678 } else {
1679 type = ARC_BUFC_DATA;
1680 }
1681 VERIFY3U(hdr->b_type, ==, type);
1682 return (type);
1683 }
1684
1685 boolean_t
1686 arc_is_metadata(arc_buf_t *buf)
1687 {
1688 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1689 }
1690
1691 static uint32_t
1692 arc_bufc_to_flags(arc_buf_contents_t type)
1693 {
1694 switch (type) {
1695 case ARC_BUFC_DATA:
1696 /* metadata field is 0 if buffer contains normal data */
1697 return (0);
1698 case ARC_BUFC_METADATA:
1699 return (ARC_FLAG_BUFC_METADATA);
1700 default:
1701 break;
1702 }
1703 panic("undefined ARC buffer type!");
1704 return ((uint32_t)-1);
1705 }
1706
1707 void
1708 arc_buf_thaw(arc_buf_t *buf)
1709 {
1710 arc_buf_hdr_t *hdr = buf->b_hdr;
1711
1712 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1713 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1714
1715 arc_cksum_verify(buf);
1716
1717 /*
1718 * Compressed buffers do not manipulate the b_freeze_cksum or
1719 * allocate b_thawed.
1720 */
1721 if (ARC_BUF_COMPRESSED(buf)) {
1722 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1723 arc_hdr_has_uncompressed_buf(hdr));
1724 return;
1725 }
1726
1727 ASSERT(HDR_HAS_L1HDR(hdr));
1728 arc_cksum_free(hdr);
1729
1730 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1731 #ifdef ZFS_DEBUG
1732 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1733 if (hdr->b_l1hdr.b_thawed != NULL)
1734 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1735 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1736 }
1737 #endif
1738
1739 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1740
1741 arc_buf_unwatch(buf);
1742 }
1743
1744 void
1745 arc_buf_freeze(arc_buf_t *buf)
1746 {
1747 arc_buf_hdr_t *hdr = buf->b_hdr;
1748 kmutex_t *hash_lock;
1749
1750 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1751 return;
1752
1753 if (ARC_BUF_COMPRESSED(buf)) {
1754 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1755 arc_hdr_has_uncompressed_buf(hdr));
1756 return;
1757 }
1758
1759 hash_lock = HDR_LOCK(hdr);
1760 mutex_enter(hash_lock);
1761
1762 ASSERT(HDR_HAS_L1HDR(hdr));
1763 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
1764 hdr->b_l1hdr.b_state == arc_anon);
1765 arc_cksum_compute(buf);
1766 mutex_exit(hash_lock);
1767 }
1768
1769 /*
1770 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1771 * the following functions should be used to ensure that the flags are
1772 * updated in a thread-safe way. When manipulating the flags either
1773 * the hash_lock must be held or the hdr must be undiscoverable. This
1774 * ensures that we're not racing with any other threads when updating
1775 * the flags.
1776 */
1777 static inline void
1778 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1779 {
1780 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1781 hdr->b_flags |= flags;
1782 }
1783
1784 static inline void
1785 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1786 {
1787 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1788 hdr->b_flags &= ~flags;
1789 }
1790
1791 /*
1792 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1793 * done in a special way since we have to clear and set bits
1794 * at the same time. Consumers that wish to set the compression bits
1795 * must use this function to ensure that the flags are updated in
1796 * thread-safe manner.
1797 */
1798 static void
1799 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1800 {
1801 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1802
1803 /*
1804 * Holes and embedded blocks will always have a psize = 0 so
1805 * we ignore the compression of the blkptr and set the
1806 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
1807 * Holes and embedded blocks remain anonymous so we don't
1808 * want to uncompress them. Mark them as uncompressed.
1809 */
1810 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1811 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1812 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF);
1813 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1814 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1815 } else {
1816 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1817 HDR_SET_COMPRESS(hdr, cmp);
1818 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1819 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1820 }
1821 }
1822
1823 /*
1824 * Looks for another buf on the same hdr which has the data decompressed, copies
1825 * from it, and returns true. If no such buf exists, returns false.
1826 */
1827 static boolean_t
1828 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1829 {
1830 arc_buf_hdr_t *hdr = buf->b_hdr;
1831 boolean_t copied = B_FALSE;
1832
1833 ASSERT(HDR_HAS_L1HDR(hdr));
1834 ASSERT3P(buf->b_data, !=, NULL);
1835 ASSERT(!ARC_BUF_COMPRESSED(buf));
1836
1837 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1838 from = from->b_next) {
1839 /* can't use our own data buffer */
1840 if (from == buf) {
1841 continue;
1842 }
1843
1844 if (!ARC_BUF_COMPRESSED(from)) {
1845 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1846 copied = B_TRUE;
1847 break;
1848 }
1849 }
1850
1851 /*
1852 * There were no decompressed bufs, so there should not be a
1853 * checksum on the hdr either.
1854 */
1855 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1856
1857 return (copied);
1858 }
1859
1860 /*
1861 * Given a buf that has a data buffer attached to it, this function will
1862 * efficiently fill the buf with data of the specified compression setting from
1863 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1864 * are already sharing a data buf, no copy is performed.
1865 *
1866 * If the buf is marked as compressed but uncompressed data was requested, this
1867 * will allocate a new data buffer for the buf, remove that flag, and fill the
1868 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1869 * uncompressed data, and (since we haven't added support for it yet) if you
1870 * want compressed data your buf must already be marked as compressed and have
1871 * the correct-sized data buffer.
1872 */
1873 static int
1874 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
1875 {
1876 arc_buf_hdr_t *hdr = buf->b_hdr;
1877 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
1878 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
1879
1880 ASSERT3P(buf->b_data, !=, NULL);
1881 IMPLY(compressed, hdr_compressed);
1882 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
1883
1884 if (hdr_compressed == compressed) {
1885 if (!arc_buf_is_shared(buf)) {
1886 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
1887 arc_buf_size(buf));
1888 }
1889 } else {
1890 ASSERT(hdr_compressed);
1891 ASSERT(!compressed);
1892 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
1893
1894 /*
1895 * If the buf is sharing its data with the hdr, unlink it and
1896 * allocate a new data buffer for the buf.
1897 */
1898 if (arc_buf_is_shared(buf)) {
1899 ASSERT(ARC_BUF_COMPRESSED(buf));
1900
1901 /* We need to give the buf it's own b_data */
1902 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
1903 buf->b_data =
1904 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
1905 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
1906
1907 /* Previously overhead was 0; just add new overhead */
1908 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
1909 } else if (ARC_BUF_COMPRESSED(buf)) {
1910 /* We need to reallocate the buf's b_data */
1911 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
1912 buf);
1913 buf->b_data =
1914 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
1915
1916 /* We increased the size of b_data; update overhead */
1917 ARCSTAT_INCR(arcstat_overhead_size,
1918 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
1919 }
1920
1921 /*
1922 * Regardless of the buf's previous compression settings, it
1923 * should not be compressed at the end of this function.
1924 */
1925 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1926
1927 /*
1928 * Try copying the data from another buf which already has a
1929 * decompressed version. If that's not possible, it's time to
1930 * bite the bullet and decompress the data from the hdr.
1931 */
1932 if (arc_buf_try_copy_decompressed_data(buf)) {
1933 /* Skip byteswapping and checksumming (already done) */
1934 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
1935 return (0);
1936 } else {
1937 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1938 hdr->b_l1hdr.b_pabd, buf->b_data,
1939 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
1940
1941 /*
1942 * Absent hardware errors or software bugs, this should
1943 * be impossible, but log it anyway so we can debug it.
1944 */
1945 if (error != 0) {
1946 zfs_dbgmsg(
1947 "hdr %p, compress %d, psize %d, lsize %d",
1948 hdr, HDR_GET_COMPRESS(hdr),
1949 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
1950 return (SET_ERROR(EIO));
1951 }
1952 }
1953 }
1954
1955 /* Byteswap the buf's data if necessary */
1956 if (bswap != DMU_BSWAP_NUMFUNCS) {
1957 ASSERT(!HDR_SHARED_DATA(hdr));
1958 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
1959 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
1960 }
1961
1962 /* Compute the hdr's checksum if necessary */
1963 arc_cksum_compute(buf);
1964
1965 return (0);
1966 }
1967
1968 int
1969 arc_decompress(arc_buf_t *buf)
1970 {
1971 return (arc_buf_fill(buf, B_FALSE));
1972 }
1973
1974 /*
1975 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1976 */
1977 static uint64_t
1978 arc_hdr_size(arc_buf_hdr_t *hdr)
1979 {
1980 uint64_t size;
1981
1982 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1983 HDR_GET_PSIZE(hdr) > 0) {
1984 size = HDR_GET_PSIZE(hdr);
1985 } else {
1986 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1987 size = HDR_GET_LSIZE(hdr);
1988 }
1989 return (size);
1990 }
1991
1992 /*
1993 * Increment the amount of evictable space in the arc_state_t's refcount.
1994 * We account for the space used by the hdr and the arc buf individually
1995 * so that we can add and remove them from the refcount individually.
1996 */
1997 static void
1998 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
1999 {
2000 arc_buf_contents_t type = arc_buf_type(hdr);
2001
2002 ASSERT(HDR_HAS_L1HDR(hdr));
2003
2004 if (GHOST_STATE(state)) {
2005 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2006 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2007 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2008 (void) refcount_add_many(&state->arcs_esize[type],
2009 HDR_GET_LSIZE(hdr), hdr);
2010 return;
2011 }
2012
2013 ASSERT(!GHOST_STATE(state));
2014 if (hdr->b_l1hdr.b_pabd != NULL) {
2015 (void) refcount_add_many(&state->arcs_esize[type],
2016 arc_hdr_size(hdr), hdr);
2017 }
2018 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2019 buf = buf->b_next) {
2020 if (arc_buf_is_shared(buf))
2021 continue;
2022 (void) refcount_add_many(&state->arcs_esize[type],
2023 arc_buf_size(buf), buf);
2024 }
2025 }
2026
2027 /*
2028 * Decrement the amount of evictable space in the arc_state_t's refcount.
2029 * We account for the space used by the hdr and the arc buf individually
2030 * so that we can add and remove them from the refcount individually.
2031 */
2032 static void
2033 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2034 {
2035 arc_buf_contents_t type = arc_buf_type(hdr);
2036
2037 ASSERT(HDR_HAS_L1HDR(hdr));
2038
2039 if (GHOST_STATE(state)) {
2040 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2041 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2042 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2043 (void) refcount_remove_many(&state->arcs_esize[type],
2044 HDR_GET_LSIZE(hdr), hdr);
2045 return;
2046 }
2047
2048 ASSERT(!GHOST_STATE(state));
2049 if (hdr->b_l1hdr.b_pabd != NULL) {
2050 (void) refcount_remove_many(&state->arcs_esize[type],
2051 arc_hdr_size(hdr), hdr);
2052 }
2053 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2054 buf = buf->b_next) {
2055 if (arc_buf_is_shared(buf))
2056 continue;
2057 (void) refcount_remove_many(&state->arcs_esize[type],
2058 arc_buf_size(buf), buf);
2059 }
2060 }
2061
2062 /*
2063 * Add a reference to this hdr indicating that someone is actively
2064 * referencing that memory. When the refcount transitions from 0 to 1,
2065 * we remove it from the respective arc_state_t list to indicate that
2066 * it is not evictable.
2067 */
2068 static void
2069 add_reference(arc_buf_hdr_t *hdr, void *tag)
2070 {
2071 ASSERT(HDR_HAS_L1HDR(hdr));
2072 if (!MUTEX_HELD(HDR_LOCK(hdr))) {
2073 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2074 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2075 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2076 }
2077
2078 arc_state_t *state = hdr->b_l1hdr.b_state;
2079
2080 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2081 (state != arc_anon)) {
2082 /* We don't use the L2-only state list. */
2083 if (state != arc_l2c_only) {
2084 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2085 hdr);
2086 arc_evictable_space_decrement(hdr, state);
2087 }
2088 /* remove the prefetch flag if we get a reference */
2089 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2090 }
2091 }
2092
2093 /*
2094 * Remove a reference from this hdr. When the reference transitions from
2095 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2096 * list making it eligible for eviction.
2097 */
2098 static int
2099 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2100 {
2101 int cnt;
2102 arc_state_t *state = hdr->b_l1hdr.b_state;
2103
2104 ASSERT(HDR_HAS_L1HDR(hdr));
2105 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2106 ASSERT(!GHOST_STATE(state));
2107
2108 /*
2109 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2110 * check to prevent usage of the arc_l2c_only list.
2111 */
2112 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2113 (state != arc_anon)) {
2114 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2115 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2116 arc_evictable_space_increment(hdr, state);
2117 }
2118 return (cnt);
2119 }
2120
2121 /*
2122 * Move the supplied buffer to the indicated state. The hash lock
2123 * for the buffer must be held by the caller.
2124 */
2125 static void
2126 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2127 kmutex_t *hash_lock)
2128 {
2129 arc_state_t *old_state;
2130 int64_t refcnt;
2131 uint32_t bufcnt;
2132 boolean_t update_old, update_new;
2133 arc_buf_contents_t buftype = arc_buf_type(hdr);
2134
2135 /*
2136 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2137 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2138 * L1 hdr doesn't always exist when we change state to arc_anon before
2139 * destroying a header, in which case reallocating to add the L1 hdr is
2140 * pointless.
2141 */
2142 if (HDR_HAS_L1HDR(hdr)) {
2143 old_state = hdr->b_l1hdr.b_state;
2144 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
2145 bufcnt = hdr->b_l1hdr.b_bufcnt;
2146 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL);
2147 } else {
2148 old_state = arc_l2c_only;
2149 refcnt = 0;
2150 bufcnt = 0;
2151 update_old = B_FALSE;
2152 }
2153 update_new = update_old;
2154
2155 ASSERT(MUTEX_HELD(hash_lock));
2156 ASSERT3P(new_state, !=, old_state);
2157 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2158 ASSERT(old_state != arc_anon || bufcnt <= 1);
2159
2160 /*
2161 * If this buffer is evictable, transfer it from the
2162 * old state list to the new state list.
2163 */
2164 if (refcnt == 0) {
2165 if (old_state != arc_anon && old_state != arc_l2c_only) {
2166 ASSERT(HDR_HAS_L1HDR(hdr));
2167 multilist_remove(old_state->arcs_list[buftype], hdr);
2168
2169 if (GHOST_STATE(old_state)) {
2170 ASSERT0(bufcnt);
2171 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2172 update_old = B_TRUE;
2173 }
2174 arc_evictable_space_decrement(hdr, old_state);
2175 }
2176 if (new_state != arc_anon && new_state != arc_l2c_only) {
2177
2178 /*
2179 * An L1 header always exists here, since if we're
2180 * moving to some L1-cached state (i.e. not l2c_only or
2181 * anonymous), we realloc the header to add an L1hdr
2182 * beforehand.
2183 */
2184 ASSERT(HDR_HAS_L1HDR(hdr));
2185 multilist_insert(new_state->arcs_list[buftype], hdr);
2186
2187 if (GHOST_STATE(new_state)) {
2188 ASSERT0(bufcnt);
2189 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2190 update_new = B_TRUE;
2191 }
2192 arc_evictable_space_increment(hdr, new_state);
2193 }
2194 }
2195
2196 ASSERT(!HDR_EMPTY(hdr));
2197 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2198 buf_hash_remove(hdr);
2199
2200 /* adjust state sizes (ignore arc_l2c_only) */
2201
2202 if (update_new && new_state != arc_l2c_only) {
2203 ASSERT(HDR_HAS_L1HDR(hdr));
2204 if (GHOST_STATE(new_state)) {
2205 ASSERT0(bufcnt);
2206
2207 /*
2208 * When moving a header to a ghost state, we first
2209 * remove all arc buffers. Thus, we'll have a
2210 * bufcnt of zero, and no arc buffer to use for
2211 * the reference. As a result, we use the arc
2212 * header pointer for the reference.
2213 */
2214 (void) refcount_add_many(&new_state->arcs_size,
2215 HDR_GET_LSIZE(hdr), hdr);
2216 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2217 } else {
2218 uint32_t buffers = 0;
2219
2220 /*
2221 * Each individual buffer holds a unique reference,
2222 * thus we must remove each of these references one
2223 * at a time.
2224 */
2225 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2226 buf = buf->b_next) {
2227 ASSERT3U(bufcnt, !=, 0);
2228 buffers++;
2229
2230 /*
2231 * When the arc_buf_t is sharing the data
2232 * block with the hdr, the owner of the
2233 * reference belongs to the hdr. Only
2234 * add to the refcount if the arc_buf_t is
2235 * not shared.
2236 */
2237 if (arc_buf_is_shared(buf))
2238 continue;
2239
2240 (void) refcount_add_many(&new_state->arcs_size,
2241 arc_buf_size(buf), buf);
2242 }
2243 ASSERT3U(bufcnt, ==, buffers);
2244
2245 if (hdr->b_l1hdr.b_pabd != NULL) {
2246 (void) refcount_add_many(&new_state->arcs_size,
2247 arc_hdr_size(hdr), hdr);
2248 } else {
2249 ASSERT(GHOST_STATE(old_state));
2250 }
2251 }
2252 }
2253
2254 if (update_old && old_state != arc_l2c_only) {
2255 ASSERT(HDR_HAS_L1HDR(hdr));
2256 if (GHOST_STATE(old_state)) {
2257 ASSERT0(bufcnt);
2258 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2259
2260 /*
2261 * When moving a header off of a ghost state,
2262 * the header will not contain any arc buffers.
2263 * We use the arc header pointer for the reference
2264 * which is exactly what we did when we put the
2265 * header on the ghost state.
2266 */
2267
2268 (void) refcount_remove_many(&old_state->arcs_size,
2269 HDR_GET_LSIZE(hdr), hdr);
2270 } else {
2271 uint32_t buffers = 0;
2272
2273 /*
2274 * Each individual buffer holds a unique reference,
2275 * thus we must remove each of these references one
2276 * at a time.
2277 */
2278 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2279 buf = buf->b_next) {
2280 ASSERT3U(bufcnt, !=, 0);
2281 buffers++;
2282
2283 /*
2284 * When the arc_buf_t is sharing the data
2285 * block with the hdr, the owner of the
2286 * reference belongs to the hdr. Only
2287 * add to the refcount if the arc_buf_t is
2288 * not shared.
2289 */
2290 if (arc_buf_is_shared(buf))
2291 continue;
2292
2293 (void) refcount_remove_many(
2294 &old_state->arcs_size, arc_buf_size(buf),
2295 buf);
2296 }
2297 ASSERT3U(bufcnt, ==, buffers);
2298 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2299 (void) refcount_remove_many(
2300 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2301 }
2302 }
2303
2304 if (HDR_HAS_L1HDR(hdr))
2305 hdr->b_l1hdr.b_state = new_state;
2306
2307 /*
2308 * L2 headers should never be on the L2 state list since they don't
2309 * have L1 headers allocated.
2310 */
2311 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2312 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2313 }
2314
2315 void
2316 arc_space_consume(uint64_t space, arc_space_type_t type)
2317 {
2318 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2319
2320 switch (type) {
2321 case ARC_SPACE_DATA:
2322 ARCSTAT_INCR(arcstat_data_size, space);
2323 break;
2324 case ARC_SPACE_META:
2325 ARCSTAT_INCR(arcstat_metadata_size, space);
2326 break;
2327 case ARC_SPACE_OTHER:
2328 ARCSTAT_INCR(arcstat_other_size, space);
2329 break;
2330 case ARC_SPACE_HDRS:
2331 ARCSTAT_INCR(arcstat_hdr_size, space);
2332 break;
2333 case ARC_SPACE_L2HDRS:
2334 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
2335 break;
2336 }
2337
2338 if (type != ARC_SPACE_DATA)
2339 ARCSTAT_INCR(arcstat_meta_used, space);
2340
2341 atomic_add_64(&arc_size, space);
2342 }
2343
2344 void
2345 arc_space_return(uint64_t space, arc_space_type_t type)
2346 {
2347 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2348
2349 switch (type) {
2350 case ARC_SPACE_DATA:
2351 ARCSTAT_INCR(arcstat_data_size, -space);
2352 break;
2353 case ARC_SPACE_META:
2354 ARCSTAT_INCR(arcstat_metadata_size, -space);
2355 break;
2356 case ARC_SPACE_OTHER:
2357 ARCSTAT_INCR(arcstat_other_size, -space);
2358 break;
2359 case ARC_SPACE_HDRS:
2360 ARCSTAT_INCR(arcstat_hdr_size, -space);
2361 break;
2362 case ARC_SPACE_L2HDRS:
2363 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
2364 break;
2365 }
2366
2367 if (type != ARC_SPACE_DATA) {
2368 ASSERT(arc_meta_used >= space);
2369 if (arc_meta_max < arc_meta_used)
2370 arc_meta_max = arc_meta_used;
2371 ARCSTAT_INCR(arcstat_meta_used, -space);
2372 }
2373
2374 ASSERT(arc_size >= space);
2375 atomic_add_64(&arc_size, -space);
2376 }
2377
2378 /*
2379 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2380 * with the hdr's b_pabd.
2381 */
2382 static boolean_t
2383 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2384 {
2385 /*
2386 * The criteria for sharing a hdr's data are:
2387 * 1. the hdr's compression matches the buf's compression
2388 * 2. the hdr doesn't need to be byteswapped
2389 * 3. the hdr isn't already being shared
2390 * 4. the buf is either compressed or it is the last buf in the hdr list
2391 *
2392 * Criterion #4 maintains the invariant that shared uncompressed
2393 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2394 * might ask, "if a compressed buf is allocated first, won't that be the
2395 * last thing in the list?", but in that case it's impossible to create
2396 * a shared uncompressed buf anyway (because the hdr must be compressed
2397 * to have the compressed buf). You might also think that #3 is
2398 * sufficient to make this guarantee, however it's possible
2399 * (specifically in the rare L2ARC write race mentioned in
2400 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2401 * is sharable, but wasn't at the time of its allocation. Rather than
2402 * allow a new shared uncompressed buf to be created and then shuffle
2403 * the list around to make it the last element, this simply disallows
2404 * sharing if the new buf isn't the first to be added.
2405 */
2406 ASSERT3P(buf->b_hdr, ==, hdr);
2407 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF;
2408 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2409 return (buf_compressed == hdr_compressed &&
2410 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2411 !HDR_SHARED_DATA(hdr) &&
2412 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2413 }
2414
2415 /*
2416 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2417 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2418 * copy was made successfully, or an error code otherwise.
2419 */
2420 static int
2421 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2422 boolean_t fill, arc_buf_t **ret)
2423 {
2424 arc_buf_t *buf;
2425
2426 ASSERT(HDR_HAS_L1HDR(hdr));
2427 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2428 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2429 hdr->b_type == ARC_BUFC_METADATA);
2430 ASSERT3P(ret, !=, NULL);
2431 ASSERT3P(*ret, ==, NULL);
2432
2433 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2434 buf->b_hdr = hdr;
2435 buf->b_data = NULL;
2436 buf->b_next = hdr->b_l1hdr.b_buf;
2437 buf->b_flags = 0;
2438
2439 add_reference(hdr, tag);
2440
2441 /*
2442 * We're about to change the hdr's b_flags. We must either
2443 * hold the hash_lock or be undiscoverable.
2444 */
2445 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2446
2447 /*
2448 * Only honor requests for compressed bufs if the hdr is actually
2449 * compressed.
2450 */
2451 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF)
2452 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2453
2454 /*
2455 * If the hdr's data can be shared then we share the data buffer and
2456 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2457 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2458 * buffer to store the buf's data.
2459 *
2460 * There are two additional restrictions here because we're sharing
2461 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2462 * actively involved in an L2ARC write, because if this buf is used by
2463 * an arc_write() then the hdr's data buffer will be released when the
2464 * write completes, even though the L2ARC write might still be using it.
2465 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2466 * need to be ABD-aware.
2467 */
2468 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2469 abd_is_linear(hdr->b_l1hdr.b_pabd);
2470
2471 /* Set up b_data and sharing */
2472 if (can_share) {
2473 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2474 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2475 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2476 } else {
2477 buf->b_data =
2478 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2479 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2480 }
2481 VERIFY3P(buf->b_data, !=, NULL);
2482
2483 hdr->b_l1hdr.b_buf = buf;
2484 hdr->b_l1hdr.b_bufcnt += 1;
2485
2486 /*
2487 * If the user wants the data from the hdr, we need to either copy or
2488 * decompress the data.
2489 */
2490 if (fill) {
2491 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
2492 }
2493
2494 return (0);
2495 }
2496
2497 static char *arc_onloan_tag = "onloan";
2498
2499 static inline void
2500 arc_loaned_bytes_update(int64_t delta)
2501 {
2502 atomic_add_64(&arc_loaned_bytes, delta);
2503
2504 /* assert that it did not wrap around */
2505 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2506 }
2507
2508 /*
2509 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2510 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2511 * buffers must be returned to the arc before they can be used by the DMU or
2512 * freed.
2513 */
2514 arc_buf_t *
2515 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2516 {
2517 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2518 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2519
2520 arc_loaned_bytes_update(size);
2521
2522 return (buf);
2523 }
2524
2525 arc_buf_t *
2526 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2527 enum zio_compress compression_type)
2528 {
2529 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2530 psize, lsize, compression_type);
2531
2532 arc_loaned_bytes_update(psize);
2533
2534 return (buf);
2535 }
2536
2537
2538 /*
2539 * Return a loaned arc buffer to the arc.
2540 */
2541 void
2542 arc_return_buf(arc_buf_t *buf, void *tag)
2543 {
2544 arc_buf_hdr_t *hdr = buf->b_hdr;
2545
2546 ASSERT3P(buf->b_data, !=, NULL);
2547 ASSERT(HDR_HAS_L1HDR(hdr));
2548 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2549 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2550
2551 arc_loaned_bytes_update(-arc_buf_size(buf));
2552 }
2553
2554 /* Detach an arc_buf from a dbuf (tag) */
2555 void
2556 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2557 {
2558 arc_buf_hdr_t *hdr = buf->b_hdr;
2559
2560 ASSERT3P(buf->b_data, !=, NULL);
2561 ASSERT(HDR_HAS_L1HDR(hdr));
2562 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2563 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2564
2565 arc_loaned_bytes_update(arc_buf_size(buf));
2566 }
2567
2568 static void
2569 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2570 {
2571 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2572
2573 df->l2df_abd = abd;
2574 df->l2df_size = size;
2575 df->l2df_type = type;
2576 mutex_enter(&l2arc_free_on_write_mtx);
2577 list_insert_head(l2arc_free_on_write, df);
2578 mutex_exit(&l2arc_free_on_write_mtx);
2579 }
2580
2581 static void
2582 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2583 {
2584 arc_state_t *state = hdr->b_l1hdr.b_state;
2585 arc_buf_contents_t type = arc_buf_type(hdr);
2586 uint64_t size = arc_hdr_size(hdr);
2587
2588 /* protected by hash lock, if in the hash table */
2589 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2590 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2591 ASSERT(state != arc_anon && state != arc_l2c_only);
2592
2593 (void) refcount_remove_many(&state->arcs_esize[type],
2594 size, hdr);
2595 }
2596 (void) refcount_remove_many(&state->arcs_size, size, hdr);
2597 if (type == ARC_BUFC_METADATA) {
2598 arc_space_return(size, ARC_SPACE_META);
2599 } else {
2600 ASSERT(type == ARC_BUFC_DATA);
2601 arc_space_return(size, ARC_SPACE_DATA);
2602 }
2603
2604 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2605 }
2606
2607 /*
2608 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2609 * data buffer, we transfer the refcount ownership to the hdr and update
2610 * the appropriate kstats.
2611 */
2612 static void
2613 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2614 {
2615 arc_state_t *state = hdr->b_l1hdr.b_state;
2616
2617 ASSERT(arc_can_share(hdr, buf));
2618 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2619 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2620
2621 /*
2622 * Start sharing the data buffer. We transfer the
2623 * refcount ownership to the hdr since it always owns
2624 * the refcount whenever an arc_buf_t is shared.
2625 */
2626 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2627 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2628 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2629 HDR_ISTYPE_METADATA(hdr));
2630 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2631 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2632
2633 /*
2634 * Since we've transferred ownership to the hdr we need
2635 * to increment its compressed and uncompressed kstats and
2636 * decrement the overhead size.
2637 */
2638 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2639 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2640 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2641 }
2642
2643 static void
2644 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2645 {
2646 arc_state_t *state = hdr->b_l1hdr.b_state;
2647
2648 ASSERT(arc_buf_is_shared(buf));
2649 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2650 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2651
2652 /*
2653 * We are no longer sharing this buffer so we need
2654 * to transfer its ownership to the rightful owner.
2655 */
2656 refcount_transfer_ownership(&state->arcs_size, hdr, buf);
2657 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2658 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
2659 abd_put(hdr->b_l1hdr.b_pabd);
2660 hdr->b_l1hdr.b_pabd = NULL;
2661 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2662
2663 /*
2664 * Since the buffer is no longer shared between
2665 * the arc buf and the hdr, count it as overhead.
2666 */
2667 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2668 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2669 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2670 }
2671
2672 /*
2673 * Remove an arc_buf_t from the hdr's buf list and return the last
2674 * arc_buf_t on the list. If no buffers remain on the list then return
2675 * NULL.
2676 */
2677 static arc_buf_t *
2678 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2679 {
2680 ASSERT(HDR_HAS_L1HDR(hdr));
2681 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2682
2683 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
2684 arc_buf_t *lastbuf = NULL;
2685
2686 /*
2687 * Remove the buf from the hdr list and locate the last
2688 * remaining buffer on the list.
2689 */
2690 while (*bufp != NULL) {
2691 if (*bufp == buf)
2692 *bufp = buf->b_next;
2693
2694 /*
2695 * If we've removed a buffer in the middle of
2696 * the list then update the lastbuf and update
2697 * bufp.
2698 */
2699 if (*bufp != NULL) {
2700 lastbuf = *bufp;
2701 bufp = &(*bufp)->b_next;
2702 }
2703 }
2704 buf->b_next = NULL;
2705 ASSERT3P(lastbuf, !=, buf);
2706 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
2707 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
2708 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
2709
2710 return (lastbuf);
2711 }
2712
2713 /*
2714 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
2715 * list and free it.
2716 */
2717 static void
2718 arc_buf_destroy_impl(arc_buf_t *buf)
2719 {
2720 arc_buf_hdr_t *hdr = buf->b_hdr;
2721
2722 /*
2723 * Free up the data associated with the buf but only if we're not
2724 * sharing this with the hdr. If we are sharing it with the hdr, the
2725 * hdr is responsible for doing the free.
2726 */
2727 if (buf->b_data != NULL) {
2728 /*
2729 * We're about to change the hdr's b_flags. We must either
2730 * hold the hash_lock or be undiscoverable.
2731 */
2732 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2733
2734 arc_cksum_verify(buf);
2735 arc_buf_unwatch(buf);
2736
2737 if (arc_buf_is_shared(buf)) {
2738 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2739 } else {
2740 uint64_t size = arc_buf_size(buf);
2741 arc_free_data_buf(hdr, buf->b_data, size, buf);
2742 ARCSTAT_INCR(arcstat_overhead_size, -size);
2743 }
2744 buf->b_data = NULL;
2745
2746 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
2747 hdr->b_l1hdr.b_bufcnt -= 1;
2748 }
2749
2750 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
2751
2752 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
2753 /*
2754 * If the current arc_buf_t is sharing its data buffer with the
2755 * hdr, then reassign the hdr's b_pabd to share it with the new
2756 * buffer at the end of the list. The shared buffer is always
2757 * the last one on the hdr's buffer list.
2758 *
2759 * There is an equivalent case for compressed bufs, but since
2760 * they aren't guaranteed to be the last buf in the list and
2761 * that is an exceedingly rare case, we just allow that space be
2762 * wasted temporarily.
2763 */
2764 if (lastbuf != NULL) {
2765 /* Only one buf can be shared at once */
2766 VERIFY(!arc_buf_is_shared(lastbuf));
2767 /* hdr is uncompressed so can't have compressed buf */
2768 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
2769
2770 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2771 arc_hdr_free_pabd(hdr);
2772
2773 /*
2774 * We must setup a new shared block between the
2775 * last buffer and the hdr. The data would have
2776 * been allocated by the arc buf so we need to transfer
2777 * ownership to the hdr since it's now being shared.
2778 */
2779 arc_share_buf(hdr, lastbuf);
2780 }
2781 } else if (HDR_SHARED_DATA(hdr)) {
2782 /*
2783 * Uncompressed shared buffers are always at the end
2784 * of the list. Compressed buffers don't have the
2785 * same requirements. This makes it hard to
2786 * simply assert that the lastbuf is shared so
2787 * we rely on the hdr's compression flags to determine
2788 * if we have a compressed, shared buffer.
2789 */
2790 ASSERT3P(lastbuf, !=, NULL);
2791 ASSERT(arc_buf_is_shared(lastbuf) ||
2792 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
2793 }
2794
2795 /*
2796 * Free the checksum if we're removing the last uncompressed buf from
2797 * this hdr.
2798 */
2799 if (!arc_hdr_has_uncompressed_buf(hdr)) {
2800 arc_cksum_free(hdr);
2801 }
2802
2803 /* clean up the buf */
2804 buf->b_hdr = NULL;
2805 kmem_cache_free(buf_cache, buf);
2806 }
2807
2808 static void
2809 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
2810 {
2811 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2812 ASSERT(HDR_HAS_L1HDR(hdr));
2813 ASSERT(!HDR_SHARED_DATA(hdr));
2814
2815 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2816 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
2817 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2818 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2819
2820 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2821 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2822 }
2823
2824 static void
2825 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
2826 {
2827 ASSERT(HDR_HAS_L1HDR(hdr));
2828 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2829
2830 /*
2831 * If the hdr is currently being written to the l2arc then
2832 * we defer freeing the data by adding it to the l2arc_free_on_write
2833 * list. The l2arc will free the data once it's finished
2834 * writing it to the l2arc device.
2835 */
2836 if (HDR_L2_WRITING(hdr)) {
2837 arc_hdr_free_on_write(hdr);
2838 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2839 } else {
2840 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
2841 arc_hdr_size(hdr), hdr);
2842 }
2843 hdr->b_l1hdr.b_pabd = NULL;
2844 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2845
2846 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2847 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2848 }
2849
2850 static arc_buf_hdr_t *
2851 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
2852 enum zio_compress compression_type, arc_buf_contents_t type)
2853 {
2854 arc_buf_hdr_t *hdr;
2855
2856 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
2857
2858 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
2859 ASSERT(HDR_EMPTY(hdr));
2860 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2861 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
2862 HDR_SET_PSIZE(hdr, psize);
2863 HDR_SET_LSIZE(hdr, lsize);
2864 hdr->b_spa = spa;
2865 hdr->b_type = type;
2866 hdr->b_flags = 0;
2867 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
2868 arc_hdr_set_compress(hdr, compression_type);
2869
2870 hdr->b_l1hdr.b_state = arc_anon;
2871 hdr->b_l1hdr.b_arc_access = 0;
2872 hdr->b_l1hdr.b_bufcnt = 0;
2873 hdr->b_l1hdr.b_buf = NULL;
2874
2875 /*
2876 * Allocate the hdr's buffer. This will contain either
2877 * the compressed or uncompressed data depending on the block
2878 * it references and compressed arc enablement.
2879 */
2880 arc_hdr_alloc_pabd(hdr);
2881 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2882
2883 return (hdr);
2884 }
2885
2886 /*
2887 * Transition between the two allocation states for the arc_buf_hdr struct.
2888 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
2889 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
2890 * version is used when a cache buffer is only in the L2ARC in order to reduce
2891 * memory usage.
2892 */
2893 static arc_buf_hdr_t *
2894 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
2895 {
2896 ASSERT(HDR_HAS_L2HDR(hdr));
2897
2898 arc_buf_hdr_t *nhdr;
2899 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2900
2901 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
2902 (old == hdr_l2only_cache && new == hdr_full_cache));
2903
2904 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
2905
2906 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
2907 buf_hash_remove(hdr);
2908
2909 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
2910
2911 if (new == hdr_full_cache) {
2912 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
2913 /*
2914 * arc_access and arc_change_state need to be aware that a
2915 * header has just come out of L2ARC, so we set its state to
2916 * l2c_only even though it's about to change.
2917 */
2918 nhdr->b_l1hdr.b_state = arc_l2c_only;
2919
2920 /* Verify previous threads set to NULL before freeing */
2921 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
2922 } else {
2923 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2924 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2925 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2926
2927 /*
2928 * If we've reached here, We must have been called from
2929 * arc_evict_hdr(), as such we should have already been
2930 * removed from any ghost list we were previously on
2931 * (which protects us from racing with arc_evict_state),
2932 * thus no locking is needed during this check.
2933 */
2934 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2935
2936 /*
2937 * A buffer must not be moved into the arc_l2c_only
2938 * state if it's not finished being written out to the
2939 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
2940 * might try to be accessed, even though it was removed.
2941 */
2942 VERIFY(!HDR_L2_WRITING(hdr));
2943 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2944
2945 #ifdef ZFS_DEBUG
2946 if (hdr->b_l1hdr.b_thawed != NULL) {
2947 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2948 hdr->b_l1hdr.b_thawed = NULL;
2949 }
2950 #endif
2951
2952 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
2953 }
2954 /*
2955 * The header has been reallocated so we need to re-insert it into any
2956 * lists it was on.
2957 */
2958 (void) buf_hash_insert(nhdr, NULL);
2959
2960 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
2961
2962 mutex_enter(&dev->l2ad_mtx);
2963
2964 /*
2965 * We must place the realloc'ed header back into the list at
2966 * the same spot. Otherwise, if it's placed earlier in the list,
2967 * l2arc_write_buffers() could find it during the function's
2968 * write phase, and try to write it out to the l2arc.
2969 */
2970 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
2971 list_remove(&dev->l2ad_buflist, hdr);
2972
2973 mutex_exit(&dev->l2ad_mtx);
2974
2975 /*
2976 * Since we're using the pointer address as the tag when
2977 * incrementing and decrementing the l2ad_alloc refcount, we
2978 * must remove the old pointer (that we're about to destroy) and
2979 * add the new pointer to the refcount. Otherwise we'd remove
2980 * the wrong pointer address when calling arc_hdr_destroy() later.
2981 */
2982
2983 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
2984 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr);
2985
2986 buf_discard_identity(hdr);
2987 kmem_cache_free(old, hdr);
2988
2989 return (nhdr);
2990 }
2991
2992 /*
2993 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
2994 * The buf is returned thawed since we expect the consumer to modify it.
2995 */
2996 arc_buf_t *
2997 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
2998 {
2999 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3000 ZIO_COMPRESS_OFF, type);
3001 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3002
3003 arc_buf_t *buf = NULL;
3004 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf));
3005 arc_buf_thaw(buf);
3006
3007 return (buf);
3008 }
3009
3010 /*
3011 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3012 * for bufs containing metadata.
3013 */
3014 arc_buf_t *
3015 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3016 enum zio_compress compression_type)
3017 {
3018 ASSERT3U(lsize, >, 0);
3019 ASSERT3U(lsize, >=, psize);
3020 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3021 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3022
3023 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3024 compression_type, ARC_BUFC_DATA);
3025 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3026
3027 arc_buf_t *buf = NULL;
3028 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3029 arc_buf_thaw(buf);
3030 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3031
3032 if (!arc_buf_is_shared(buf)) {
3033 /*
3034 * To ensure that the hdr has the correct data in it if we call
3035 * arc_decompress() on this buf before it's been written to
3036 * disk, it's easiest if we just set up sharing between the
3037 * buf and the hdr.
3038 */
3039 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3040 arc_hdr_free_pabd(hdr);
3041 arc_share_buf(hdr, buf);
3042 }
3043
3044 return (buf);
3045 }
3046
3047 static void
3048 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3049 {
3050 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3051 l2arc_dev_t *dev = l2hdr->b_dev;
3052 uint64_t psize = arc_hdr_size(hdr);
3053
3054 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3055 ASSERT(HDR_HAS_L2HDR(hdr));
3056
3057 list_remove(&dev->l2ad_buflist, hdr);
3058
3059 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3060 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3061
3062 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3063
3064 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3065 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3066 }
3067
3068 static void
3069 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3070 {
3071 if (HDR_HAS_L1HDR(hdr)) {
3072 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3073 hdr->b_l1hdr.b_bufcnt > 0);
3074 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3075 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3076 }
3077 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3078 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3079
3080 if (!HDR_EMPTY(hdr))
3081 buf_discard_identity(hdr);
3082
3083 if (HDR_HAS_L2HDR(hdr)) {
3084 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3085 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3086
3087 if (!buflist_held)
3088 mutex_enter(&dev->l2ad_mtx);
3089
3090 /*
3091 * Even though we checked this conditional above, we
3092 * need to check this again now that we have the
3093 * l2ad_mtx. This is because we could be racing with
3094 * another thread calling l2arc_evict() which might have
3095 * destroyed this header's L2 portion as we were waiting
3096 * to acquire the l2ad_mtx. If that happens, we don't
3097 * want to re-destroy the header's L2 portion.
3098 */
3099 if (HDR_HAS_L2HDR(hdr))
3100 arc_hdr_l2hdr_destroy(hdr);
3101
3102 if (!buflist_held)
3103 mutex_exit(&dev->l2ad_mtx);
3104 }
3105
3106 if (HDR_HAS_L1HDR(hdr)) {
3107 arc_cksum_free(hdr);
3108
3109 while (hdr->b_l1hdr.b_buf != NULL)
3110 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3111
3112 #ifdef ZFS_DEBUG
3113 if (hdr->b_l1hdr.b_thawed != NULL) {
3114 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3115 hdr->b_l1hdr.b_thawed = NULL;
3116 }
3117 #endif
3118
3119 if (hdr->b_l1hdr.b_pabd != NULL) {
3120 arc_hdr_free_pabd(hdr);
3121 }
3122 }
3123
3124 ASSERT3P(hdr->b_hash_next, ==, NULL);
3125 if (HDR_HAS_L1HDR(hdr)) {
3126 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3127 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3128 kmem_cache_free(hdr_full_cache, hdr);
3129 } else {
3130 kmem_cache_free(hdr_l2only_cache, hdr);
3131 }
3132 }
3133
3134 void
3135 arc_buf_destroy(arc_buf_t *buf, void* tag)
3136 {
3137 arc_buf_hdr_t *hdr = buf->b_hdr;
3138 kmutex_t *hash_lock = HDR_LOCK(hdr);
3139
3140 if (hdr->b_l1hdr.b_state == arc_anon) {
3141 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3142 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3143 VERIFY0(remove_reference(hdr, NULL, tag));
3144 arc_hdr_destroy(hdr);
3145 return;
3146 }
3147
3148 mutex_enter(hash_lock);
3149 ASSERT3P(hdr, ==, buf->b_hdr);
3150 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3151 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3152 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3153 ASSERT3P(buf->b_data, !=, NULL);
3154
3155 (void) remove_reference(hdr, hash_lock, tag);
3156 arc_buf_destroy_impl(buf);
3157 mutex_exit(hash_lock);
3158 }
3159
3160 /*
3161 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3162 * state of the header is dependent on it's state prior to entering this
3163 * function. The following transitions are possible:
3164 *
3165 * - arc_mru -> arc_mru_ghost
3166 * - arc_mfu -> arc_mfu_ghost
3167 * - arc_mru_ghost -> arc_l2c_only
3168 * - arc_mru_ghost -> deleted
3169 * - arc_mfu_ghost -> arc_l2c_only
3170 * - arc_mfu_ghost -> deleted
3171 */
3172 static int64_t
3173 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3174 {
3175 arc_state_t *evicted_state, *state;
3176 int64_t bytes_evicted = 0;
3177
3178 ASSERT(MUTEX_HELD(hash_lock));
3179 ASSERT(HDR_HAS_L1HDR(hdr));
3180
3181 state = hdr->b_l1hdr.b_state;
3182 if (GHOST_STATE(state)) {
3183 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3184 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3185
3186 /*
3187 * l2arc_write_buffers() relies on a header's L1 portion
3188 * (i.e. its b_pabd field) during it's write phase.
3189 * Thus, we cannot push a header onto the arc_l2c_only
3190 * state (removing it's L1 piece) until the header is
3191 * done being written to the l2arc.
3192 */
3193 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3194 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3195 return (bytes_evicted);
3196 }
3197
3198 ARCSTAT_BUMP(arcstat_deleted);
3199 bytes_evicted += HDR_GET_LSIZE(hdr);
3200
3201 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3202
3203 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3204 if (HDR_HAS_L2HDR(hdr)) {
3205 /*
3206 * This buffer is cached on the 2nd Level ARC;
3207 * don't destroy the header.
3208 */
3209 arc_change_state(arc_l2c_only, hdr, hash_lock);
3210 /*
3211 * dropping from L1+L2 cached to L2-only,
3212 * realloc to remove the L1 header.
3213 */
3214 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3215 hdr_l2only_cache);
3216 } else {
3217 arc_change_state(arc_anon, hdr, hash_lock);
3218 arc_hdr_destroy(hdr);
3219 }
3220 return (bytes_evicted);
3221 }
3222
3223 ASSERT(state == arc_mru || state == arc_mfu);
3224 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3225
3226 /* prefetch buffers have a minimum lifespan */
3227 if (HDR_IO_IN_PROGRESS(hdr) ||
3228 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3229 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3230 arc_min_prefetch_lifespan)) {
3231 ARCSTAT_BUMP(arcstat_evict_skip);
3232 return (bytes_evicted);
3233 }
3234
3235 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3236 while (hdr->b_l1hdr.b_buf) {
3237 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3238 if (!mutex_tryenter(&buf->b_evict_lock)) {
3239 ARCSTAT_BUMP(arcstat_mutex_miss);
3240 break;
3241 }
3242 if (buf->b_data != NULL)
3243 bytes_evicted += HDR_GET_LSIZE(hdr);
3244 mutex_exit(&buf->b_evict_lock);
3245 arc_buf_destroy_impl(buf);
3246 }
3247
3248 if (HDR_HAS_L2HDR(hdr)) {
3249 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3250 } else {
3251 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3252 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3253 HDR_GET_LSIZE(hdr));
3254 } else {
3255 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3256 HDR_GET_LSIZE(hdr));
3257 }
3258 }
3259
3260 if (hdr->b_l1hdr.b_bufcnt == 0) {
3261 arc_cksum_free(hdr);
3262
3263 bytes_evicted += arc_hdr_size(hdr);
3264
3265 /*
3266 * If this hdr is being evicted and has a compressed
3267 * buffer then we discard it here before we change states.
3268 * This ensures that the accounting is updated correctly
3269 * in arc_free_data_impl().
3270 */
3271 arc_hdr_free_pabd(hdr);
3272
3273 arc_change_state(evicted_state, hdr, hash_lock);
3274 ASSERT(HDR_IN_HASH_TABLE(hdr));
3275 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3276 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3277 }
3278
3279 return (bytes_evicted);
3280 }
3281
3282 static uint64_t
3283 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3284 uint64_t spa, int64_t bytes)
3285 {
3286 multilist_sublist_t *mls;
3287 uint64_t bytes_evicted = 0;
3288 arc_buf_hdr_t *hdr;
3289 kmutex_t *hash_lock;
3290 int evict_count = 0;
3291
3292 ASSERT3P(marker, !=, NULL);
3293 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3294
3295 mls = multilist_sublist_lock(ml, idx);
3296
3297 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3298 hdr = multilist_sublist_prev(mls, marker)) {
3299 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3300 (evict_count >= zfs_arc_evict_batch_limit))
3301 break;
3302
3303 /*
3304 * To keep our iteration location, move the marker
3305 * forward. Since we're not holding hdr's hash lock, we
3306 * must be very careful and not remove 'hdr' from the
3307 * sublist. Otherwise, other consumers might mistake the
3308 * 'hdr' as not being on a sublist when they call the
3309 * multilist_link_active() function (they all rely on
3310 * the hash lock protecting concurrent insertions and
3311 * removals). multilist_sublist_move_forward() was
3312 * specifically implemented to ensure this is the case
3313 * (only 'marker' will be removed and re-inserted).
3314 */
3315 multilist_sublist_move_forward(mls, marker);
3316
3317 /*
3318 * The only case where the b_spa field should ever be
3319 * zero, is the marker headers inserted by
3320 * arc_evict_state(). It's possible for multiple threads
3321 * to be calling arc_evict_state() concurrently (e.g.
3322 * dsl_pool_close() and zio_inject_fault()), so we must
3323 * skip any markers we see from these other threads.
3324 */
3325 if (hdr->b_spa == 0)
3326 continue;
3327
3328 /* we're only interested in evicting buffers of a certain spa */
3329 if (spa != 0 && hdr->b_spa != spa) {
3330 ARCSTAT_BUMP(arcstat_evict_skip);
3331 continue;
3332 }
3333
3334 hash_lock = HDR_LOCK(hdr);
3335
3336 /*
3337 * We aren't calling this function from any code path
3338 * that would already be holding a hash lock, so we're
3339 * asserting on this assumption to be defensive in case
3340 * this ever changes. Without this check, it would be
3341 * possible to incorrectly increment arcstat_mutex_miss
3342 * below (e.g. if the code changed such that we called
3343 * this function with a hash lock held).
3344 */
3345 ASSERT(!MUTEX_HELD(hash_lock));
3346
3347 if (mutex_tryenter(hash_lock)) {
3348 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3349 mutex_exit(hash_lock);
3350
3351 bytes_evicted += evicted;
3352
3353 /*
3354 * If evicted is zero, arc_evict_hdr() must have
3355 * decided to skip this header, don't increment
3356 * evict_count in this case.
3357 */
3358 if (evicted != 0)
3359 evict_count++;
3360
3361 /*
3362 * If arc_size isn't overflowing, signal any
3363 * threads that might happen to be waiting.
3364 *
3365 * For each header evicted, we wake up a single
3366 * thread. If we used cv_broadcast, we could
3367 * wake up "too many" threads causing arc_size
3368 * to significantly overflow arc_c; since
3369 * arc_get_data_impl() doesn't check for overflow
3370 * when it's woken up (it doesn't because it's
3371 * possible for the ARC to be overflowing while
3372 * full of un-evictable buffers, and the
3373 * function should proceed in this case).
3374 *
3375 * If threads are left sleeping, due to not
3376 * using cv_broadcast, they will be woken up
3377 * just before arc_reclaim_thread() sleeps.
3378 */
3379 mutex_enter(&arc_reclaim_lock);
3380 if (!arc_is_overflowing())
3381 cv_signal(&arc_reclaim_waiters_cv);
3382 mutex_exit(&arc_reclaim_lock);
3383 } else {
3384 ARCSTAT_BUMP(arcstat_mutex_miss);
3385 }
3386 }
3387
3388 multilist_sublist_unlock(mls);
3389
3390 return (bytes_evicted);
3391 }
3392
3393 /*
3394 * Evict buffers from the given arc state, until we've removed the
3395 * specified number of bytes. Move the removed buffers to the
3396 * appropriate evict state.
3397 *
3398 * This function makes a "best effort". It skips over any buffers
3399 * it can't get a hash_lock on, and so, may not catch all candidates.
3400 * It may also return without evicting as much space as requested.
3401 *
3402 * If bytes is specified using the special value ARC_EVICT_ALL, this
3403 * will evict all available (i.e. unlocked and evictable) buffers from
3404 * the given arc state; which is used by arc_flush().
3405 */
3406 static uint64_t
3407 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3408 arc_buf_contents_t type)
3409 {
3410 uint64_t total_evicted = 0;
3411 multilist_t *ml = state->arcs_list[type];
3412 int num_sublists;
3413 arc_buf_hdr_t **markers;
3414
3415 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3416
3417 num_sublists = multilist_get_num_sublists(ml);
3418
3419 /*
3420 * If we've tried to evict from each sublist, made some
3421 * progress, but still have not hit the target number of bytes
3422 * to evict, we want to keep trying. The markers allow us to
3423 * pick up where we left off for each individual sublist, rather
3424 * than starting from the tail each time.
3425 */
3426 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3427 for (int i = 0; i < num_sublists; i++) {
3428 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3429
3430 /*
3431 * A b_spa of 0 is used to indicate that this header is
3432 * a marker. This fact is used in arc_adjust_type() and
3433 * arc_evict_state_impl().
3434 */
3435 markers[i]->b_spa = 0;
3436
3437 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3438 multilist_sublist_insert_tail(mls, markers[i]);
3439 multilist_sublist_unlock(mls);
3440 }
3441
3442 /*
3443 * While we haven't hit our target number of bytes to evict, or
3444 * we're evicting all available buffers.
3445 */
3446 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3447 /*
3448 * Start eviction using a randomly selected sublist,
3449 * this is to try and evenly balance eviction across all
3450 * sublists. Always starting at the same sublist
3451 * (e.g. index 0) would cause evictions to favor certain
3452 * sublists over others.
3453 */
3454 int sublist_idx = multilist_get_random_index(ml);
3455 uint64_t scan_evicted = 0;
3456
3457 for (int i = 0; i < num_sublists; i++) {
3458 uint64_t bytes_remaining;
3459 uint64_t bytes_evicted;
3460
3461 if (bytes == ARC_EVICT_ALL)
3462 bytes_remaining = ARC_EVICT_ALL;
3463 else if (total_evicted < bytes)
3464 bytes_remaining = bytes - total_evicted;
3465 else
3466 break;
3467
3468 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3469 markers[sublist_idx], spa, bytes_remaining);
3470
3471 scan_evicted += bytes_evicted;
3472 total_evicted += bytes_evicted;
3473
3474 /* we've reached the end, wrap to the beginning */
3475 if (++sublist_idx >= num_sublists)
3476 sublist_idx = 0;
3477 }
3478
3479 /*
3480 * If we didn't evict anything during this scan, we have
3481 * no reason to believe we'll evict more during another
3482 * scan, so break the loop.
3483 */
3484 if (scan_evicted == 0) {
3485 /* This isn't possible, let's make that obvious */
3486 ASSERT3S(bytes, !=, 0);
3487
3488 /*
3489 * When bytes is ARC_EVICT_ALL, the only way to
3490 * break the loop is when scan_evicted is zero.
3491 * In that case, we actually have evicted enough,
3492 * so we don't want to increment the kstat.
3493 */
3494 if (bytes != ARC_EVICT_ALL) {
3495 ASSERT3S(total_evicted, <, bytes);
3496 ARCSTAT_BUMP(arcstat_evict_not_enough);
3497 }
3498
3499 break;
3500 }
3501 }
3502
3503 for (int i = 0; i < num_sublists; i++) {
3504 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3505 multilist_sublist_remove(mls, markers[i]);
3506 multilist_sublist_unlock(mls);
3507
3508 kmem_cache_free(hdr_full_cache, markers[i]);
3509 }
3510 kmem_free(markers, sizeof (*markers) * num_sublists);
3511
3512 return (total_evicted);
3513 }
3514
3515 /*
3516 * Flush all "evictable" data of the given type from the arc state
3517 * specified. This will not evict any "active" buffers (i.e. referenced).
3518 *
3519 * When 'retry' is set to B_FALSE, the function will make a single pass
3520 * over the state and evict any buffers that it can. Since it doesn't
3521 * continually retry the eviction, it might end up leaving some buffers
3522 * in the ARC due to lock misses.
3523 *
3524 * When 'retry' is set to B_TRUE, the function will continually retry the
3525 * eviction until *all* evictable buffers have been removed from the
3526 * state. As a result, if concurrent insertions into the state are
3527 * allowed (e.g. if the ARC isn't shutting down), this function might
3528 * wind up in an infinite loop, continually trying to evict buffers.
3529 */
3530 static uint64_t
3531 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
3532 boolean_t retry)
3533 {
3534 uint64_t evicted = 0;
3535
3536 while (refcount_count(&state->arcs_esize[type]) != 0) {
3537 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
3538
3539 if (!retry)
3540 break;
3541 }
3542
3543 return (evicted);
3544 }
3545
3546 /*
3547 * Evict the specified number of bytes from the state specified,
3548 * restricting eviction to the spa and type given. This function
3549 * prevents us from trying to evict more from a state's list than
3550 * is "evictable", and to skip evicting altogether when passed a
3551 * negative value for "bytes". In contrast, arc_evict_state() will
3552 * evict everything it can, when passed a negative value for "bytes".
3553 */
3554 static uint64_t
3555 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3556 arc_buf_contents_t type)
3557 {
3558 int64_t delta;
3559
3560 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3561 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3562 return (arc_evict_state(state, spa, delta, type));
3563 }
3564
3565 return (0);
3566 }
3567
3568 /*
3569 * Evict metadata buffers from the cache, such that arc_meta_used is
3570 * capped by the arc_meta_limit tunable.
3571 */
3572 static uint64_t
3573 arc_adjust_meta(void)
3574 {
3575 uint64_t total_evicted = 0;
3576 int64_t target;
3577
3578 /*
3579 * If we're over the meta limit, we want to evict enough
3580 * metadata to get back under the meta limit. We don't want to
3581 * evict so much that we drop the MRU below arc_p, though. If
3582 * we're over the meta limit more than we're over arc_p, we
3583 * evict some from the MRU here, and some from the MFU below.
3584 */
3585 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3586 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3587 refcount_count(&arc_mru->arcs_size) - arc_p));
3588
3589 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3590
3591 /*
3592 * Similar to the above, we want to evict enough bytes to get us
3593 * below the meta limit, but not so much as to drop us below the
3594 * space allotted to the MFU (which is defined as arc_c - arc_p).
3595 */
3596 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
3597 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
3598
3599 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3600
3601 return (total_evicted);
3602 }
3603
3604 /*
3605 * Return the type of the oldest buffer in the given arc state
3606 *
3607 * This function will select a random sublist of type ARC_BUFC_DATA and
3608 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3609 * is compared, and the type which contains the "older" buffer will be
3610 * returned.
3611 */
3612 static arc_buf_contents_t
3613 arc_adjust_type(arc_state_t *state)
3614 {
3615 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
3616 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
3617 int data_idx = multilist_get_random_index(data_ml);
3618 int meta_idx = multilist_get_random_index(meta_ml);
3619 multilist_sublist_t *data_mls;
3620 multilist_sublist_t *meta_mls;
3621 arc_buf_contents_t type;
3622 arc_buf_hdr_t *data_hdr;
3623 arc_buf_hdr_t *meta_hdr;
3624
3625 /*
3626 * We keep the sublist lock until we're finished, to prevent
3627 * the headers from being destroyed via arc_evict_state().
3628 */
3629 data_mls = multilist_sublist_lock(data_ml, data_idx);
3630 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3631
3632 /*
3633 * These two loops are to ensure we skip any markers that
3634 * might be at the tail of the lists due to arc_evict_state().
3635 */
3636
3637 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3638 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3639 if (data_hdr->b_spa != 0)
3640 break;
3641 }
3642
3643 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3644 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3645 if (meta_hdr->b_spa != 0)
3646 break;
3647 }
3648
3649 if (data_hdr == NULL && meta_hdr == NULL) {
3650 type = ARC_BUFC_DATA;
3651 } else if (data_hdr == NULL) {
3652 ASSERT3P(meta_hdr, !=, NULL);
3653 type = ARC_BUFC_METADATA;
3654 } else if (meta_hdr == NULL) {
3655 ASSERT3P(data_hdr, !=, NULL);
3656 type = ARC_BUFC_DATA;
3657 } else {
3658 ASSERT3P(data_hdr, !=, NULL);
3659 ASSERT3P(meta_hdr, !=, NULL);
3660
3661 /* The headers can't be on the sublist without an L1 header */
3662 ASSERT(HDR_HAS_L1HDR(data_hdr));
3663 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3664
3665 if (data_hdr->b_l1hdr.b_arc_access <
3666 meta_hdr->b_l1hdr.b_arc_access) {
3667 type = ARC_BUFC_DATA;
3668 } else {
3669 type = ARC_BUFC_METADATA;
3670 }
3671 }
3672
3673 multilist_sublist_unlock(meta_mls);
3674 multilist_sublist_unlock(data_mls);
3675
3676 return (type);
3677 }
3678
3679 /*
3680 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3681 */
3682 static uint64_t
3683 arc_adjust(void)
3684 {
3685 uint64_t total_evicted = 0;
3686 uint64_t bytes;
3687 int64_t target;
3688
3689 /*
3690 * If we're over arc_meta_limit, we want to correct that before
3691 * potentially evicting data buffers below.
3692 */
3693 total_evicted += arc_adjust_meta();
3694
3695 /*
3696 * Adjust MRU size
3697 *
3698 * If we're over the target cache size, we want to evict enough
3699 * from the list to get back to our target size. We don't want
3700 * to evict too much from the MRU, such that it drops below
3701 * arc_p. So, if we're over our target cache size more than
3702 * the MRU is over arc_p, we'll evict enough to get back to
3703 * arc_p here, and then evict more from the MFU below.
3704 */
3705 target = MIN((int64_t)(arc_size - arc_c),
3706 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3707 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
3708
3709 /*
3710 * If we're below arc_meta_min, always prefer to evict data.
3711 * Otherwise, try to satisfy the requested number of bytes to
3712 * evict from the type which contains older buffers; in an
3713 * effort to keep newer buffers in the cache regardless of their
3714 * type. If we cannot satisfy the number of bytes from this
3715 * type, spill over into the next type.
3716 */
3717 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
3718 arc_meta_used > arc_meta_min) {
3719 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3720 total_evicted += bytes;
3721
3722 /*
3723 * If we couldn't evict our target number of bytes from
3724 * metadata, we try to get the rest from data.
3725 */
3726 target -= bytes;
3727
3728 total_evicted +=
3729 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3730 } else {
3731 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3732 total_evicted += bytes;
3733
3734 /*
3735 * If we couldn't evict our target number of bytes from
3736 * data, we try to get the rest from metadata.
3737 */
3738 target -= bytes;
3739
3740 total_evicted +=
3741 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3742 }
3743
3744 /*
3745 * Adjust MFU size
3746 *
3747 * Now that we've tried to evict enough from the MRU to get its
3748 * size back to arc_p, if we're still above the target cache
3749 * size, we evict the rest from the MFU.
3750 */
3751 target = arc_size - arc_c;
3752
3753 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
3754 arc_meta_used > arc_meta_min) {
3755 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3756 total_evicted += bytes;
3757
3758 /*
3759 * If we couldn't evict our target number of bytes from
3760 * metadata, we try to get the rest from data.
3761 */
3762 target -= bytes;
3763
3764 total_evicted +=
3765 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3766 } else {
3767 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3768 total_evicted += bytes;
3769
3770 /*
3771 * If we couldn't evict our target number of bytes from
3772 * data, we try to get the rest from data.
3773 */
3774 target -= bytes;
3775
3776 total_evicted +=
3777 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3778 }
3779
3780 /*
3781 * Adjust ghost lists
3782 *
3783 * In addition to the above, the ARC also defines target values
3784 * for the ghost lists. The sum of the mru list and mru ghost
3785 * list should never exceed the target size of the cache, and
3786 * the sum of the mru list, mfu list, mru ghost list, and mfu
3787 * ghost list should never exceed twice the target size of the
3788 * cache. The following logic enforces these limits on the ghost
3789 * caches, and evicts from them as needed.
3790 */
3791 target = refcount_count(&arc_mru->arcs_size) +
3792 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3793
3794 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3795 total_evicted += bytes;
3796
3797 target -= bytes;
3798
3799 total_evicted +=
3800 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3801
3802 /*
3803 * We assume the sum of the mru list and mfu list is less than
3804 * or equal to arc_c (we enforced this above), which means we
3805 * can use the simpler of the two equations below:
3806 *
3807 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3808 * mru ghost + mfu ghost <= arc_c
3809 */
3810 target = refcount_count(&arc_mru_ghost->arcs_size) +
3811 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3812
3813 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3814 total_evicted += bytes;
3815
3816 target -= bytes;
3817
3818 total_evicted +=
3819 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3820
3821 return (total_evicted);
3822 }
3823
3824 void
3825 arc_flush(spa_t *spa, boolean_t retry)
3826 {
3827 uint64_t guid = 0;
3828
3829 /*
3830 * If retry is B_TRUE, a spa must not be specified since we have
3831 * no good way to determine if all of a spa's buffers have been
3832 * evicted from an arc state.
3833 */
3834 ASSERT(!retry || spa == 0);
3835
3836 if (spa != NULL)
3837 guid = spa_load_guid(spa);
3838
3839 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3840 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3841
3842 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3843 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3844
3845 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3846 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3847
3848 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3849 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3850 }
3851
3852 void
3853 arc_shrink(int64_t to_free)
3854 {
3855 if (arc_c > arc_c_min) {
3856
3857 if (arc_c > arc_c_min + to_free)
3858 atomic_add_64(&arc_c, -to_free);
3859 else
3860 arc_c = arc_c_min;
3861
3862 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3863 if (arc_c > arc_size)
3864 arc_c = MAX(arc_size, arc_c_min);
3865 if (arc_p > arc_c)
3866 arc_p = (arc_c >> 1);
3867 ASSERT(arc_c >= arc_c_min);
3868 ASSERT((int64_t)arc_p >= 0);
3869 }
3870
3871 if (arc_size > arc_c)
3872 (void) arc_adjust();
3873 }
3874
3875 typedef enum free_memory_reason_t {
3876 FMR_UNKNOWN,
3877 FMR_NEEDFREE,
3878 FMR_LOTSFREE,
3879 FMR_SWAPFS_MINFREE,
3880 FMR_PAGES_PP_MAXIMUM,
3881 FMR_HEAP_ARENA,
3882 FMR_ZIO_ARENA,
3883 } free_memory_reason_t;
3884
3885 int64_t last_free_memory;
3886 free_memory_reason_t last_free_reason;
3887
3888 /*
3889 * Additional reserve of pages for pp_reserve.
3890 */
3891 int64_t arc_pages_pp_reserve = 64;
3892
3893 /*
3894 * Additional reserve of pages for swapfs.
3895 */
3896 int64_t arc_swapfs_reserve = 64;
3897
3898 /*
3899 * Return the amount of memory that can be consumed before reclaim will be
3900 * needed. Positive if there is sufficient free memory, negative indicates
3901 * the amount of memory that needs to be freed up.
3902 */
3903 static int64_t
3904 arc_available_memory(void)
3905 {
3906 int64_t lowest = INT64_MAX;
3907 int64_t n;
3908 free_memory_reason_t r = FMR_UNKNOWN;
3909
3910 #ifdef _KERNEL
3911 if (needfree > 0) {
3912 n = PAGESIZE * (-needfree);
3913 if (n < lowest) {
3914 lowest = n;
3915 r = FMR_NEEDFREE;
3916 }
3917 }
3918
3919 /*
3920 * check that we're out of range of the pageout scanner. It starts to
3921 * schedule paging if freemem is less than lotsfree and needfree.
3922 * lotsfree is the high-water mark for pageout, and needfree is the
3923 * number of needed free pages. We add extra pages here to make sure
3924 * the scanner doesn't start up while we're freeing memory.
3925 */
3926 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3927 if (n < lowest) {
3928 lowest = n;
3929 r = FMR_LOTSFREE;
3930 }
3931
3932 /*
3933 * check to make sure that swapfs has enough space so that anon
3934 * reservations can still succeed. anon_resvmem() checks that the
3935 * availrmem is greater than swapfs_minfree, and the number of reserved
3936 * swap pages. We also add a bit of extra here just to prevent
3937 * circumstances from getting really dire.
3938 */
3939 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3940 desfree - arc_swapfs_reserve);
3941 if (n < lowest) {
3942 lowest = n;
3943 r = FMR_SWAPFS_MINFREE;
3944 }
3945
3946
3947 /*
3948 * Check that we have enough availrmem that memory locking (e.g., via
3949 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3950 * stores the number of pages that cannot be locked; when availrmem
3951 * drops below pages_pp_maximum, page locking mechanisms such as
3952 * page_pp_lock() will fail.)
3953 */
3954 n = PAGESIZE * (availrmem - pages_pp_maximum -
3955 arc_pages_pp_reserve);
3956 if (n < lowest) {
3957 lowest = n;
3958 r = FMR_PAGES_PP_MAXIMUM;
3959 }
3960
3961 #if defined(__i386)
3962 /*
3963 * If we're on an i386 platform, it's possible that we'll exhaust the
3964 * kernel heap space before we ever run out of available physical
3965 * memory. Most checks of the size of the heap_area compare against
3966 * tune.t_minarmem, which is the minimum available real memory that we
3967 * can have in the system. However, this is generally fixed at 25 pages
3968 * which is so low that it's useless. In this comparison, we seek to
3969 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3970 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3971 * free)
3972 */
3973 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
3974 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3975 if (n < lowest) {
3976 lowest = n;
3977 r = FMR_HEAP_ARENA;
3978 }
3979 #endif
3980
3981 /*
3982 * If zio data pages are being allocated out of a separate heap segment,
3983 * then enforce that the size of available vmem for this arena remains
3984 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
3985 *
3986 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
3987 * memory (in the zio_arena) free, which can avoid memory
3988 * fragmentation issues.
3989 */
3990 if (zio_arena != NULL) {
3991 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
3992 (vmem_size(zio_arena, VMEM_ALLOC) >>
3993 arc_zio_arena_free_shift);
3994 if (n < lowest) {
3995 lowest = n;
3996 r = FMR_ZIO_ARENA;
3997 }
3998 }
3999 #else
4000 /* Every 100 calls, free a small amount */
4001 if (spa_get_random(100) == 0)
4002 lowest = -1024;
4003 #endif
4004
4005 last_free_memory = lowest;
4006 last_free_reason = r;
4007
4008 return (lowest);
4009 }
4010
4011
4012 /*
4013 * Determine if the system is under memory pressure and is asking
4014 * to reclaim memory. A return value of B_TRUE indicates that the system
4015 * is under memory pressure and that the arc should adjust accordingly.
4016 */
4017 static boolean_t
4018 arc_reclaim_needed(void)
4019 {
4020 return (arc_available_memory() < 0);
4021 }
4022
4023 static void
4024 arc_kmem_reap_now(void)
4025 {
4026 size_t i;
4027 kmem_cache_t *prev_cache = NULL;
4028 kmem_cache_t *prev_data_cache = NULL;
4029 extern kmem_cache_t *zio_buf_cache[];
4030 extern kmem_cache_t *zio_data_buf_cache[];
4031 extern kmem_cache_t *range_seg_cache;
4032 extern kmem_cache_t *abd_chunk_cache;
4033
4034 #ifdef _KERNEL
4035 if (arc_meta_used >= arc_meta_limit) {
4036 /*
4037 * We are exceeding our meta-data cache limit.
4038 * Purge some DNLC entries to release holds on meta-data.
4039 */
4040 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4041 }
4042 #if defined(__i386)
4043 /*
4044 * Reclaim unused memory from all kmem caches.
4045 */
4046 kmem_reap();
4047 #endif
4048 #endif
4049
4050 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4051 if (zio_buf_cache[i] != prev_cache) {
4052 prev_cache = zio_buf_cache[i];
4053 kmem_cache_reap_now(zio_buf_cache[i]);
4054 }
4055 if (zio_data_buf_cache[i] != prev_data_cache) {
4056 prev_data_cache = zio_data_buf_cache[i];
4057 kmem_cache_reap_now(zio_data_buf_cache[i]);
4058 }
4059 }
4060 kmem_cache_reap_now(abd_chunk_cache);
4061 kmem_cache_reap_now(buf_cache);
4062 kmem_cache_reap_now(hdr_full_cache);
4063 kmem_cache_reap_now(hdr_l2only_cache);
4064 kmem_cache_reap_now(range_seg_cache);
4065
4066 if (zio_arena != NULL) {
4067 /*
4068 * Ask the vmem arena to reclaim unused memory from its
4069 * quantum caches.
4070 */
4071 vmem_qcache_reap(zio_arena);
4072 }
4073 }
4074
4075 /*
4076 * Threads can block in arc_get_data_impl() waiting for this thread to evict
4077 * enough data and signal them to proceed. When this happens, the threads in
4078 * arc_get_data_impl() are sleeping while holding the hash lock for their
4079 * particular arc header. Thus, we must be careful to never sleep on a
4080 * hash lock in this thread. This is to prevent the following deadlock:
4081 *
4082 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L",
4083 * waiting for the reclaim thread to signal it.
4084 *
4085 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
4086 * fails, and goes to sleep forever.
4087 *
4088 * This possible deadlock is avoided by always acquiring a hash lock
4089 * using mutex_tryenter() from arc_reclaim_thread().
4090 */
4091 /* ARGSUSED */
4092 static void
4093 arc_reclaim_thread(void *unused)
4094 {
4095 hrtime_t growtime = 0;
4096 callb_cpr_t cpr;
4097
4098 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
4099
4100 mutex_enter(&arc_reclaim_lock);
4101 while (!arc_reclaim_thread_exit) {
4102 uint64_t evicted = 0;
4103
4104 /*
4105 * This is necessary in order for the mdb ::arc dcmd to
4106 * show up to date information. Since the ::arc command
4107 * does not call the kstat's update function, without
4108 * this call, the command may show stale stats for the
4109 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4110 * with this change, the data might be up to 1 second
4111 * out of date; but that should suffice. The arc_state_t
4112 * structures can be queried directly if more accurate
4113 * information is needed.
4114 */
4115 if (arc_ksp != NULL)
4116 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4117
4118 mutex_exit(&arc_reclaim_lock);
4119
4120 /*
4121 * We call arc_adjust() before (possibly) calling
4122 * arc_kmem_reap_now(), so that we can wake up
4123 * arc_get_data_impl() sooner.
4124 */
4125 evicted = arc_adjust();
4126
4127 int64_t free_memory = arc_available_memory();
4128 if (free_memory < 0) {
4129
4130 arc_no_grow = B_TRUE;
4131 arc_warm = B_TRUE;
4132
4133 /*
4134 * Wait at least zfs_grow_retry (default 60) seconds
4135 * before considering growing.
4136 */
4137 growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4138
4139 arc_kmem_reap_now();
4140
4141 /*
4142 * If we are still low on memory, shrink the ARC
4143 * so that we have arc_shrink_min free space.
4144 */
4145 free_memory = arc_available_memory();
4146
4147 int64_t to_free =
4148 (arc_c >> arc_shrink_shift) - free_memory;
4149 if (to_free > 0) {
4150 #ifdef _KERNEL
4151 to_free = MAX(to_free, ptob(needfree));
4152 #endif
4153 arc_shrink(to_free);
4154 }
4155 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4156 arc_no_grow = B_TRUE;
4157 } else if (gethrtime() >= growtime) {
4158 arc_no_grow = B_FALSE;
4159 }
4160
4161 mutex_enter(&arc_reclaim_lock);
4162
4163 /*
4164 * If evicted is zero, we couldn't evict anything via
4165 * arc_adjust(). This could be due to hash lock
4166 * collisions, but more likely due to the majority of
4167 * arc buffers being unevictable. Therefore, even if
4168 * arc_size is above arc_c, another pass is unlikely to
4169 * be helpful and could potentially cause us to enter an
4170 * infinite loop.
4171 */
4172 if (arc_size <= arc_c || evicted == 0) {
4173 /*
4174 * We're either no longer overflowing, or we
4175 * can't evict anything more, so we should wake
4176 * up any threads before we go to sleep.
4177 */
4178 cv_broadcast(&arc_reclaim_waiters_cv);
4179
4180 /*
4181 * Block until signaled, or after one second (we
4182 * might need to perform arc_kmem_reap_now()
4183 * even if we aren't being signalled)
4184 */
4185 CALLB_CPR_SAFE_BEGIN(&cpr);
4186 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4187 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4188 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4189 }
4190 }
4191
4192 arc_reclaim_thread_exit = B_FALSE;
4193 cv_broadcast(&arc_reclaim_thread_cv);
4194 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
4195 thread_exit();
4196 }
4197
4198 /*
4199 * Adapt arc info given the number of bytes we are trying to add and
4200 * the state that we are comming from. This function is only called
4201 * when we are adding new content to the cache.
4202 */
4203 static void
4204 arc_adapt(int bytes, arc_state_t *state)
4205 {
4206 int mult;
4207 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4208 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
4209 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
4210
4211 if (state == arc_l2c_only)
4212 return;
4213
4214 ASSERT(bytes > 0);
4215 /*
4216 * Adapt the target size of the MRU list:
4217 * - if we just hit in the MRU ghost list, then increase
4218 * the target size of the MRU list.
4219 * - if we just hit in the MFU ghost list, then increase
4220 * the target size of the MFU list by decreasing the
4221 * target size of the MRU list.
4222 */
4223 if (state == arc_mru_ghost) {
4224 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4225 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4226
4227 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4228 } else if (state == arc_mfu_ghost) {
4229 uint64_t delta;
4230
4231 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4232 mult = MIN(mult, 10);
4233
4234 delta = MIN(bytes * mult, arc_p);
4235 arc_p = MAX(arc_p_min, arc_p - delta);
4236 }
4237 ASSERT((int64_t)arc_p >= 0);
4238
4239 if (arc_reclaim_needed()) {
4240 cv_signal(&arc_reclaim_thread_cv);
4241 return;
4242 }
4243
4244 if (arc_no_grow)
4245 return;
4246
4247 if (arc_c >= arc_c_max)
4248 return;
4249
4250 /*
4251 * If we're within (2 * maxblocksize) bytes of the target
4252 * cache size, increment the target cache size
4253 */
4254 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
4255 atomic_add_64(&arc_c, (int64_t)bytes);
4256 if (arc_c > arc_c_max)
4257 arc_c = arc_c_max;
4258 else if (state == arc_anon)
4259 atomic_add_64(&arc_p, (int64_t)bytes);
4260 if (arc_p > arc_c)
4261 arc_p = arc_c;
4262 }
4263 ASSERT((int64_t)arc_p >= 0);
4264 }
4265
4266 /*
4267 * Check if arc_size has grown past our upper threshold, determined by
4268 * zfs_arc_overflow_shift.
4269 */
4270 static boolean_t
4271 arc_is_overflowing(void)
4272 {
4273 /* Always allow at least one block of overflow */
4274 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4275 arc_c >> zfs_arc_overflow_shift);
4276
4277 return (arc_size >= arc_c + overflow);
4278 }
4279
4280 static abd_t *
4281 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4282 {
4283 arc_buf_contents_t type = arc_buf_type(hdr);
4284
4285 arc_get_data_impl(hdr, size, tag);
4286 if (type == ARC_BUFC_METADATA) {
4287 return (abd_alloc(size, B_TRUE));
4288 } else {
4289 ASSERT(type == ARC_BUFC_DATA);
4290 return (abd_alloc(size, B_FALSE));
4291 }
4292 }
4293
4294 static void *
4295 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4296 {
4297 arc_buf_contents_t type = arc_buf_type(hdr);
4298
4299 arc_get_data_impl(hdr, size, tag);
4300 if (type == ARC_BUFC_METADATA) {
4301 return (zio_buf_alloc(size));
4302 } else {
4303 ASSERT(type == ARC_BUFC_DATA);
4304 return (zio_data_buf_alloc(size));
4305 }
4306 }
4307
4308 /*
4309 * Allocate a block and return it to the caller. If we are hitting the
4310 * hard limit for the cache size, we must sleep, waiting for the eviction
4311 * thread to catch up. If we're past the target size but below the hard
4312 * limit, we'll only signal the reclaim thread and continue on.
4313 */
4314 static void
4315 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4316 {
4317 arc_state_t *state = hdr->b_l1hdr.b_state;
4318 arc_buf_contents_t type = arc_buf_type(hdr);
4319
4320 arc_adapt(size, state);
4321
4322 /*
4323 * If arc_size is currently overflowing, and has grown past our
4324 * upper limit, we must be adding data faster than the evict
4325 * thread can evict. Thus, to ensure we don't compound the
4326 * problem by adding more data and forcing arc_size to grow even
4327 * further past it's target size, we halt and wait for the
4328 * eviction thread to catch up.
4329 *
4330 * It's also possible that the reclaim thread is unable to evict
4331 * enough buffers to get arc_size below the overflow limit (e.g.
4332 * due to buffers being un-evictable, or hash lock collisions).
4333 * In this case, we want to proceed regardless if we're
4334 * overflowing; thus we don't use a while loop here.
4335 */
4336 if (arc_is_overflowing()) {
4337 mutex_enter(&arc_reclaim_lock);
4338
4339 /*
4340 * Now that we've acquired the lock, we may no longer be
4341 * over the overflow limit, lets check.
4342 *
4343 * We're ignoring the case of spurious wake ups. If that
4344 * were to happen, it'd let this thread consume an ARC
4345 * buffer before it should have (i.e. before we're under
4346 * the overflow limit and were signalled by the reclaim
4347 * thread). As long as that is a rare occurrence, it
4348 * shouldn't cause any harm.
4349 */
4350 if (arc_is_overflowing()) {
4351 cv_signal(&arc_reclaim_thread_cv);
4352 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4353 }
4354
4355 mutex_exit(&arc_reclaim_lock);
4356 }
4357
4358 VERIFY3U(hdr->b_type, ==, type);
4359 if (type == ARC_BUFC_METADATA) {
4360 arc_space_consume(size, ARC_SPACE_META);
4361 } else {
4362 arc_space_consume(size, ARC_SPACE_DATA);
4363 }
4364
4365 /*
4366 * Update the state size. Note that ghost states have a
4367 * "ghost size" and so don't need to be updated.
4368 */
4369 if (!GHOST_STATE(state)) {
4370
4371 (void) refcount_add_many(&state->arcs_size, size, tag);
4372
4373 /*
4374 * If this is reached via arc_read, the link is
4375 * protected by the hash lock. If reached via
4376 * arc_buf_alloc, the header should not be accessed by
4377 * any other thread. And, if reached via arc_read_done,
4378 * the hash lock will protect it if it's found in the
4379 * hash table; otherwise no other thread should be
4380 * trying to [add|remove]_reference it.
4381 */
4382 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4383 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4384 (void) refcount_add_many(&state->arcs_esize[type],
4385 size, tag);
4386 }
4387
4388 /*
4389 * If we are growing the cache, and we are adding anonymous
4390 * data, and we have outgrown arc_p, update arc_p
4391 */
4392 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
4393 (refcount_count(&arc_anon->arcs_size) +
4394 refcount_count(&arc_mru->arcs_size) > arc_p))
4395 arc_p = MIN(arc_c, arc_p + size);
4396 }
4397 }
4398
4399 static void
4400 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
4401 {
4402 arc_free_data_impl(hdr, size, tag);
4403 abd_free(abd);
4404 }
4405
4406 static void
4407 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
4408 {
4409 arc_buf_contents_t type = arc_buf_type(hdr);
4410
4411 arc_free_data_impl(hdr, size, tag);
4412 if (type == ARC_BUFC_METADATA) {
4413 zio_buf_free(buf, size);
4414 } else {
4415 ASSERT(type == ARC_BUFC_DATA);
4416 zio_data_buf_free(buf, size);
4417 }
4418 }
4419
4420 /*
4421 * Free the arc data buffer.
4422 */
4423 static void
4424 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4425 {
4426 arc_state_t *state = hdr->b_l1hdr.b_state;
4427 arc_buf_contents_t type = arc_buf_type(hdr);
4428
4429 /* protected by hash lock, if in the hash table */
4430 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4431 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4432 ASSERT(state != arc_anon && state != arc_l2c_only);
4433
4434 (void) refcount_remove_many(&state->arcs_esize[type],
4435 size, tag);
4436 }
4437 (void) refcount_remove_many(&state->arcs_size, size, tag);
4438
4439 VERIFY3U(hdr->b_type, ==, type);
4440 if (type == ARC_BUFC_METADATA) {
4441 arc_space_return(size, ARC_SPACE_META);
4442 } else {
4443 ASSERT(type == ARC_BUFC_DATA);
4444 arc_space_return(size, ARC_SPACE_DATA);
4445 }
4446 }
4447
4448 /*
4449 * This routine is called whenever a buffer is accessed.
4450 * NOTE: the hash lock is dropped in this function.
4451 */
4452 static void
4453 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4454 {
4455 clock_t now;
4456
4457 ASSERT(MUTEX_HELD(hash_lock));
4458 ASSERT(HDR_HAS_L1HDR(hdr));
4459
4460 if (hdr->b_l1hdr.b_state == arc_anon) {
4461 /*
4462 * This buffer is not in the cache, and does not
4463 * appear in our "ghost" list. Add the new buffer
4464 * to the MRU state.
4465 */
4466
4467 ASSERT0(hdr->b_l1hdr.b_arc_access);
4468 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4469 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4470 arc_change_state(arc_mru, hdr, hash_lock);
4471
4472 } else if (hdr->b_l1hdr.b_state == arc_mru) {
4473 now = ddi_get_lbolt();
4474
4475 /*
4476 * If this buffer is here because of a prefetch, then either:
4477 * - clear the flag if this is a "referencing" read
4478 * (any subsequent access will bump this into the MFU state).
4479 * or
4480 * - move the buffer to the head of the list if this is
4481 * another prefetch (to make it less likely to be evicted).
4482 */
4483 if (HDR_PREFETCH(hdr)) {
4484 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4485 /* link protected by hash lock */
4486 ASSERT(multilist_link_active(
4487 &hdr->b_l1hdr.b_arc_node));
4488 } else {
4489 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4490 ARCSTAT_BUMP(arcstat_mru_hits);
4491 }
4492 hdr->b_l1hdr.b_arc_access = now;
4493 return;
4494 }
4495
4496 /*
4497 * This buffer has been "accessed" only once so far,
4498 * but it is still in the cache. Move it to the MFU
4499 * state.
4500 */
4501 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
4502 /*
4503 * More than 125ms have passed since we
4504 * instantiated this buffer. Move it to the
4505 * most frequently used state.
4506 */
4507 hdr->b_l1hdr.b_arc_access = now;
4508 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4509 arc_change_state(arc_mfu, hdr, hash_lock);
4510 }
4511 ARCSTAT_BUMP(arcstat_mru_hits);
4512 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
4513 arc_state_t *new_state;
4514 /*
4515 * This buffer has been "accessed" recently, but
4516 * was evicted from the cache. Move it to the
4517 * MFU state.
4518 */
4519
4520 if (HDR_PREFETCH(hdr)) {
4521 new_state = arc_mru;
4522 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
4523 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
4524 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
4525 } else {
4526 new_state = arc_mfu;
4527 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4528 }
4529
4530 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4531 arc_change_state(new_state, hdr, hash_lock);
4532
4533 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
4534 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
4535 /*
4536 * This buffer has been accessed more than once and is
4537 * still in the cache. Keep it in the MFU state.
4538 *
4539 * NOTE: an add_reference() that occurred when we did
4540 * the arc_read() will have kicked this off the list.
4541 * If it was a prefetch, we will explicitly move it to
4542 * the head of the list now.
4543 */
4544 if ((HDR_PREFETCH(hdr)) != 0) {
4545 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4546 /* link protected by hash_lock */
4547 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4548 }
4549 ARCSTAT_BUMP(arcstat_mfu_hits);
4550 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4551 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
4552 arc_state_t *new_state = arc_mfu;
4553 /*
4554 * This buffer has been accessed more than once but has
4555 * been evicted from the cache. Move it back to the
4556 * MFU state.
4557 */
4558
4559 if (HDR_PREFETCH(hdr)) {
4560 /*
4561 * This is a prefetch access...
4562 * move this block back to the MRU state.
4563 */
4564 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
4565 new_state = arc_mru;
4566 }
4567
4568 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4569 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4570 arc_change_state(new_state, hdr, hash_lock);
4571
4572 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4573 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4574 /*
4575 * This buffer is on the 2nd Level ARC.
4576 */
4577
4578 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4579 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4580 arc_change_state(arc_mfu, hdr, hash_lock);
4581 } else {
4582 ASSERT(!"invalid arc state");
4583 }
4584 }
4585
4586 /* a generic arc_done_func_t which you can use */
4587 /* ARGSUSED */
4588 void
4589 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4590 {
4591 if (zio == NULL || zio->io_error == 0)
4592 bcopy(buf->b_data, arg, arc_buf_size(buf));
4593 arc_buf_destroy(buf, arg);
4594 }
4595
4596 /* a generic arc_done_func_t */
4597 void
4598 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
4599 {
4600 arc_buf_t **bufp = arg;
4601 if (zio && zio->io_error) {
4602 arc_buf_destroy(buf, arg);
4603 *bufp = NULL;
4604 } else {
4605 *bufp = buf;
4606 ASSERT(buf->b_data);
4607 }
4608 }
4609
4610 static void
4611 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
4612 {
4613 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
4614 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
4615 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
4616 } else {
4617 if (HDR_COMPRESSION_ENABLED(hdr)) {
4618 ASSERT3U(HDR_GET_COMPRESS(hdr), ==,
4619 BP_GET_COMPRESS(bp));
4620 }
4621 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
4622 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
4623 }
4624 }
4625
4626 static void
4627 arc_read_done(zio_t *zio)
4628 {
4629 arc_buf_hdr_t *hdr = zio->io_private;
4630 kmutex_t *hash_lock = NULL;
4631 arc_callback_t *callback_list;
4632 arc_callback_t *acb;
4633 boolean_t freeable = B_FALSE;
4634 boolean_t no_zio_error = (zio->io_error == 0);
4635
4636 /*
4637 * The hdr was inserted into hash-table and removed from lists
4638 * prior to starting I/O. We should find this header, since
4639 * it's in the hash table, and it should be legit since it's
4640 * not possible to evict it during the I/O. The only possible
4641 * reason for it not to be found is if we were freed during the
4642 * read.
4643 */
4644 if (HDR_IN_HASH_TABLE(hdr)) {
4645 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
4646 ASSERT3U(hdr->b_dva.dva_word[0], ==,
4647 BP_IDENTITY(zio->io_bp)->dva_word[0]);
4648 ASSERT3U(hdr->b_dva.dva_word[1], ==,
4649 BP_IDENTITY(zio->io_bp)->dva_word[1]);
4650
4651 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
4652 &hash_lock);
4653
4654 ASSERT((found == hdr &&
4655 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
4656 (found == hdr && HDR_L2_READING(hdr)));
4657 ASSERT3P(hash_lock, !=, NULL);
4658 }
4659
4660 if (no_zio_error) {
4661 /* byteswap if necessary */
4662 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
4663 if (BP_GET_LEVEL(zio->io_bp) > 0) {
4664 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
4665 } else {
4666 hdr->b_l1hdr.b_byteswap =
4667 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
4668 }
4669 } else {
4670 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
4671 }
4672 }
4673
4674 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
4675 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
4676 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
4677
4678 callback_list = hdr->b_l1hdr.b_acb;
4679 ASSERT3P(callback_list, !=, NULL);
4680
4681 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) {
4682 /*
4683 * Only call arc_access on anonymous buffers. This is because
4684 * if we've issued an I/O for an evicted buffer, we've already
4685 * called arc_access (to prevent any simultaneous readers from
4686 * getting confused).
4687 */
4688 arc_access(hdr, hash_lock);
4689 }
4690
4691 /*
4692 * If a read request has a callback (i.e. acb_done is not NULL), then we
4693 * make a buf containing the data according to the parameters which were
4694 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
4695 * aren't needlessly decompressing the data multiple times.
4696 */
4697 int callback_cnt = 0;
4698 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
4699 if (!acb->acb_done)
4700 continue;
4701
4702 /* This is a demand read since prefetches don't use callbacks */
4703 callback_cnt++;
4704
4705 int error = arc_buf_alloc_impl(hdr, acb->acb_private,
4706 acb->acb_compressed, no_zio_error, &acb->acb_buf);
4707 if (no_zio_error) {
4708 zio->io_error = error;
4709 }
4710 }
4711 hdr->b_l1hdr.b_acb = NULL;
4712 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
4713 if (callback_cnt == 0) {
4714 ASSERT(HDR_PREFETCH(hdr));
4715 ASSERT0(hdr->b_l1hdr.b_bufcnt);
4716 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
4717 }
4718
4719 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
4720 callback_list != NULL);
4721
4722 if (no_zio_error) {
4723 arc_hdr_verify(hdr, zio->io_bp);
4724 } else {
4725 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
4726 if (hdr->b_l1hdr.b_state != arc_anon)
4727 arc_change_state(arc_anon, hdr, hash_lock);
4728 if (HDR_IN_HASH_TABLE(hdr))
4729 buf_hash_remove(hdr);
4730 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4731 }
4732
4733 /*
4734 * Broadcast before we drop the hash_lock to avoid the possibility
4735 * that the hdr (and hence the cv) might be freed before we get to
4736 * the cv_broadcast().
4737 */
4738 cv_broadcast(&hdr->b_l1hdr.b_cv);
4739
4740 if (hash_lock != NULL) {
4741 mutex_exit(hash_lock);
4742 } else {
4743 /*
4744 * This block was freed while we waited for the read to
4745 * complete. It has been removed from the hash table and
4746 * moved to the anonymous state (so that it won't show up
4747 * in the cache).
4748 */
4749 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
4750 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4751 }
4752
4753 /* execute each callback and free its structure */
4754 while ((acb = callback_list) != NULL) {
4755 if (acb->acb_done)
4756 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
4757
4758 if (acb->acb_zio_dummy != NULL) {
4759 acb->acb_zio_dummy->io_error = zio->io_error;
4760 zio_nowait(acb->acb_zio_dummy);
4761 }
4762
4763 callback_list = acb->acb_next;
4764 kmem_free(acb, sizeof (arc_callback_t));
4765 }
4766
4767 if (freeable)
4768 arc_hdr_destroy(hdr);
4769 }
4770
4771 /*
4772 * "Read" the block at the specified DVA (in bp) via the
4773 * cache. If the block is found in the cache, invoke the provided
4774 * callback immediately and return. Note that the `zio' parameter
4775 * in the callback will be NULL in this case, since no IO was
4776 * required. If the block is not in the cache pass the read request
4777 * on to the spa with a substitute callback function, so that the
4778 * requested block will be added to the cache.
4779 *
4780 * If a read request arrives for a block that has a read in-progress,
4781 * either wait for the in-progress read to complete (and return the
4782 * results); or, if this is a read with a "done" func, add a record
4783 * to the read to invoke the "done" func when the read completes,
4784 * and return; or just return.
4785 *
4786 * arc_read_done() will invoke all the requested "done" functions
4787 * for readers of this block.
4788 */
4789 int
4790 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
4791 void *private, zio_priority_t priority, int zio_flags,
4792 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
4793 {
4794 arc_buf_hdr_t *hdr = NULL;
4795 kmutex_t *hash_lock = NULL;
4796 zio_t *rzio;
4797 uint64_t guid = spa_load_guid(spa);
4798 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0;
4799
4800 ASSERT(!BP_IS_EMBEDDED(bp) ||
4801 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
4802
4803 top:
4804 if (!BP_IS_EMBEDDED(bp)) {
4805 /*
4806 * Embedded BP's have no DVA and require no I/O to "read".
4807 * Create an anonymous arc buf to back it.
4808 */
4809 hdr = buf_hash_find(guid, bp, &hash_lock);
4810 }
4811
4812 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) {
4813 arc_buf_t *buf = NULL;
4814 *arc_flags |= ARC_FLAG_CACHED;
4815
4816 if (HDR_IO_IN_PROGRESS(hdr)) {
4817
4818 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
4819 priority == ZIO_PRIORITY_SYNC_READ) {
4820 /*
4821 * This sync read must wait for an
4822 * in-progress async read (e.g. a predictive
4823 * prefetch). Async reads are queued
4824 * separately at the vdev_queue layer, so
4825 * this is a form of priority inversion.
4826 * Ideally, we would "inherit" the demand
4827 * i/o's priority by moving the i/o from
4828 * the async queue to the synchronous queue,
4829 * but there is currently no mechanism to do
4830 * so. Track this so that we can evaluate
4831 * the magnitude of this potential performance
4832 * problem.
4833 *
4834 * Note that if the prefetch i/o is already
4835 * active (has been issued to the device),
4836 * the prefetch improved performance, because
4837 * we issued it sooner than we would have
4838 * without the prefetch.
4839 */
4840 DTRACE_PROBE1(arc__sync__wait__for__async,
4841 arc_buf_hdr_t *, hdr);
4842 ARCSTAT_BUMP(arcstat_sync_wait_for_async);
4843 }
4844 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4845 arc_hdr_clear_flags(hdr,
4846 ARC_FLAG_PREDICTIVE_PREFETCH);
4847 }
4848
4849 if (*arc_flags & ARC_FLAG_WAIT) {
4850 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4851 mutex_exit(hash_lock);
4852 goto top;
4853 }
4854 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4855
4856 if (done) {
4857 arc_callback_t *acb = NULL;
4858
4859 acb = kmem_zalloc(sizeof (arc_callback_t),
4860 KM_SLEEP);
4861 acb->acb_done = done;
4862 acb->acb_private = private;
4863 acb->acb_compressed = compressed_read;
4864 if (pio != NULL)
4865 acb->acb_zio_dummy = zio_null(pio,
4866 spa, NULL, NULL, NULL, zio_flags);
4867
4868 ASSERT3P(acb->acb_done, !=, NULL);
4869 acb->acb_next = hdr->b_l1hdr.b_acb;
4870 hdr->b_l1hdr.b_acb = acb;
4871 mutex_exit(hash_lock);
4872 return (0);
4873 }
4874 mutex_exit(hash_lock);
4875 return (0);
4876 }
4877
4878 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4879 hdr->b_l1hdr.b_state == arc_mfu);
4880
4881 if (done) {
4882 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
4883 /*
4884 * This is a demand read which does not have to
4885 * wait for i/o because we did a predictive
4886 * prefetch i/o for it, which has completed.
4887 */
4888 DTRACE_PROBE1(
4889 arc__demand__hit__predictive__prefetch,
4890 arc_buf_hdr_t *, hdr);
4891 ARCSTAT_BUMP(
4892 arcstat_demand_hit_predictive_prefetch);
4893 arc_hdr_clear_flags(hdr,
4894 ARC_FLAG_PREDICTIVE_PREFETCH);
4895 }
4896 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
4897
4898 /* Get a buf with the desired data in it. */
4899 VERIFY0(arc_buf_alloc_impl(hdr, private,
4900 compressed_read, B_TRUE, &buf));
4901 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4902 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4903 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
4904 }
4905 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4906 arc_access(hdr, hash_lock);
4907 if (*arc_flags & ARC_FLAG_L2CACHE)
4908 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
4909 mutex_exit(hash_lock);
4910 ARCSTAT_BUMP(arcstat_hits);
4911 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4912 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4913 data, metadata, hits);
4914
4915 if (done)
4916 done(NULL, buf, private);
4917 } else {
4918 uint64_t lsize = BP_GET_LSIZE(bp);
4919 uint64_t psize = BP_GET_PSIZE(bp);
4920 arc_callback_t *acb;
4921 vdev_t *vd = NULL;
4922 uint64_t addr = 0;
4923 boolean_t devw = B_FALSE;
4924 uint64_t size;
4925
4926 if (hdr == NULL) {
4927 /* this block is not in the cache */
4928 arc_buf_hdr_t *exists = NULL;
4929 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4930 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
4931 BP_GET_COMPRESS(bp), type);
4932
4933 if (!BP_IS_EMBEDDED(bp)) {
4934 hdr->b_dva = *BP_IDENTITY(bp);
4935 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4936 exists = buf_hash_insert(hdr, &hash_lock);
4937 }
4938 if (exists != NULL) {
4939 /* somebody beat us to the hash insert */
4940 mutex_exit(hash_lock);
4941 buf_discard_identity(hdr);
4942 arc_hdr_destroy(hdr);
4943 goto top; /* restart the IO request */
4944 }
4945 } else {
4946 /*
4947 * This block is in the ghost cache. If it was L2-only
4948 * (and thus didn't have an L1 hdr), we realloc the
4949 * header to add an L1 hdr.
4950 */
4951 if (!HDR_HAS_L1HDR(hdr)) {
4952 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4953 hdr_full_cache);
4954 }
4955 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
4956 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4957 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4958 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4959 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
4960 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
4961
4962 /*
4963 * This is a delicate dance that we play here.
4964 * This hdr is in the ghost list so we access it
4965 * to move it out of the ghost list before we
4966 * initiate the read. If it's a prefetch then
4967 * it won't have a callback so we'll remove the
4968 * reference that arc_buf_alloc_impl() created. We
4969 * do this after we've called arc_access() to
4970 * avoid hitting an assert in remove_reference().
4971 */
4972 arc_access(hdr, hash_lock);
4973 arc_hdr_alloc_pabd(hdr);
4974 }
4975 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
4976 size = arc_hdr_size(hdr);
4977
4978 /*
4979 * If compression is enabled on the hdr, then will do
4980 * RAW I/O and will store the compressed data in the hdr's
4981 * data block. Otherwise, the hdr's data block will contain
4982 * the uncompressed data.
4983 */
4984 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
4985 zio_flags |= ZIO_FLAG_RAW;
4986 }
4987
4988 if (*arc_flags & ARC_FLAG_PREFETCH)
4989 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
4990 if (*arc_flags & ARC_FLAG_L2CACHE)
4991 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
4992 if (BP_GET_LEVEL(bp) > 0)
4993 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
4994 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
4995 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
4996 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
4997
4998 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
4999 acb->acb_done = done;
5000 acb->acb_private = private;
5001 acb->acb_compressed = compressed_read;
5002
5003 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5004 hdr->b_l1hdr.b_acb = acb;
5005 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5006
5007 if (HDR_HAS_L2HDR(hdr) &&
5008 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5009 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5010 addr = hdr->b_l2hdr.b_daddr;
5011 /*
5012 * Lock out L2ARC device removal.
5013 */
5014 if (vdev_is_dead(vd) ||
5015 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5016 vd = NULL;
5017 }
5018
5019 if (priority == ZIO_PRIORITY_ASYNC_READ)
5020 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5021 else
5022 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5023
5024 if (hash_lock != NULL)
5025 mutex_exit(hash_lock);
5026
5027 /*
5028 * At this point, we have a level 1 cache miss. Try again in
5029 * L2ARC if possible.
5030 */
5031 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5032
5033 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5034 uint64_t, lsize, zbookmark_phys_t *, zb);
5035 ARCSTAT_BUMP(arcstat_misses);
5036 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5037 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5038 data, metadata, misses);
5039
5040 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5041 /*
5042 * Read from the L2ARC if the following are true:
5043 * 1. The L2ARC vdev was previously cached.
5044 * 2. This buffer still has L2ARC metadata.
5045 * 3. This buffer isn't currently writing to the L2ARC.
5046 * 4. The L2ARC entry wasn't evicted, which may
5047 * also have invalidated the vdev.
5048 * 5. This isn't prefetch and l2arc_noprefetch is set.
5049 */
5050 if (HDR_HAS_L2HDR(hdr) &&
5051 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5052 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5053 l2arc_read_callback_t *cb;
5054 abd_t *abd;
5055 uint64_t asize;
5056
5057 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5058 ARCSTAT_BUMP(arcstat_l2_hits);
5059
5060 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5061 KM_SLEEP);
5062 cb->l2rcb_hdr = hdr;
5063 cb->l2rcb_bp = *bp;
5064 cb->l2rcb_zb = *zb;
5065 cb->l2rcb_flags = zio_flags;
5066
5067 asize = vdev_psize_to_asize(vd, size);
5068 if (asize != size) {
5069 abd = abd_alloc_for_io(asize,
5070 HDR_ISTYPE_METADATA(hdr));
5071 cb->l2rcb_abd = abd;
5072 } else {
5073 abd = hdr->b_l1hdr.b_pabd;
5074 }
5075
5076 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5077 addr + asize <= vd->vdev_psize -
5078 VDEV_LABEL_END_SIZE);
5079
5080 /*
5081 * l2arc read. The SCL_L2ARC lock will be
5082 * released by l2arc_read_done().
5083 * Issue a null zio if the underlying buffer
5084 * was squashed to zero size by compression.
5085 */
5086 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5087 ZIO_COMPRESS_EMPTY);
5088 rzio = zio_read_phys(pio, vd, addr,
5089 asize, abd,
5090 ZIO_CHECKSUM_OFF,
5091 l2arc_read_done, cb, priority,
5092 zio_flags | ZIO_FLAG_DONT_CACHE |
5093 ZIO_FLAG_CANFAIL |
5094 ZIO_FLAG_DONT_PROPAGATE |
5095 ZIO_FLAG_DONT_RETRY, B_FALSE);
5096 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5097 zio_t *, rzio);
5098 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5099
5100 if (*arc_flags & ARC_FLAG_NOWAIT) {
5101 zio_nowait(rzio);
5102 return (0);
5103 }
5104
5105 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5106 if (zio_wait(rzio) == 0)
5107 return (0);
5108
5109 /* l2arc read error; goto zio_read() */
5110 } else {
5111 DTRACE_PROBE1(l2arc__miss,
5112 arc_buf_hdr_t *, hdr);
5113 ARCSTAT_BUMP(arcstat_l2_misses);
5114 if (HDR_L2_WRITING(hdr))
5115 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5116 spa_config_exit(spa, SCL_L2ARC, vd);
5117 }
5118 } else {
5119 if (vd != NULL)
5120 spa_config_exit(spa, SCL_L2ARC, vd);
5121 if (l2arc_ndev != 0) {
5122 DTRACE_PROBE1(l2arc__miss,
5123 arc_buf_hdr_t *, hdr);
5124 ARCSTAT_BUMP(arcstat_l2_misses);
5125 }
5126 }
5127
5128 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size,
5129 arc_read_done, hdr, priority, zio_flags, zb);
5130
5131 if (*arc_flags & ARC_FLAG_WAIT)
5132 return (zio_wait(rzio));
5133
5134 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5135 zio_nowait(rzio);
5136 }
5137 return (0);
5138 }
5139
5140 /*
5141 * Notify the arc that a block was freed, and thus will never be used again.
5142 */
5143 void
5144 arc_freed(spa_t *spa, const blkptr_t *bp)
5145 {
5146 arc_buf_hdr_t *hdr;
5147 kmutex_t *hash_lock;
5148 uint64_t guid = spa_load_guid(spa);
5149
5150 ASSERT(!BP_IS_EMBEDDED(bp));
5151
5152 hdr = buf_hash_find(guid, bp, &hash_lock);
5153 if (hdr == NULL)
5154 return;
5155
5156 /*
5157 * We might be trying to free a block that is still doing I/O
5158 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5159 * dmu_sync-ed block). If this block is being prefetched, then it
5160 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5161 * until the I/O completes. A block may also have a reference if it is
5162 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
5163 * have written the new block to its final resting place on disk but
5164 * without the dedup flag set. This would have left the hdr in the MRU
5165 * state and discoverable. When the txg finally syncs it detects that
5166 * the block was overridden in open context and issues an override I/O.
5167 * Since this is a dedup block, the override I/O will determine if the
5168 * block is already in the DDT. If so, then it will replace the io_bp
5169 * with the bp from the DDT and allow the I/O to finish. When the I/O
5170 * reaches the done callback, dbuf_write_override_done, it will
5171 * check to see if the io_bp and io_bp_override are identical.
5172 * If they are not, then it indicates that the bp was replaced with
5173 * the bp in the DDT and the override bp is freed. This allows
5174 * us to arrive here with a reference on a block that is being
5175 * freed. So if we have an I/O in progress, or a reference to
5176 * this hdr, then we don't destroy the hdr.
5177 */
5178 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
5179 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
5180 arc_change_state(arc_anon, hdr, hash_lock);
5181 arc_hdr_destroy(hdr);
5182 mutex_exit(hash_lock);
5183 } else {
5184 mutex_exit(hash_lock);
5185 }
5186
5187 }
5188
5189 /*
5190 * Release this buffer from the cache, making it an anonymous buffer. This
5191 * must be done after a read and prior to modifying the buffer contents.
5192 * If the buffer has more than one reference, we must make
5193 * a new hdr for the buffer.
5194 */
5195 void
5196 arc_release(arc_buf_t *buf, void *tag)
5197 {
5198 arc_buf_hdr_t *hdr = buf->b_hdr;
5199
5200 /*
5201 * It would be nice to assert that if it's DMU metadata (level >
5202 * 0 || it's the dnode file), then it must be syncing context.
5203 * But we don't know that information at this level.
5204 */
5205
5206 mutex_enter(&buf->b_evict_lock);
5207
5208 ASSERT(HDR_HAS_L1HDR(hdr));
5209
5210 /*
5211 * We don't grab the hash lock prior to this check, because if
5212 * the buffer's header is in the arc_anon state, it won't be
5213 * linked into the hash table.
5214 */
5215 if (hdr->b_l1hdr.b_state == arc_anon) {
5216 mutex_exit(&buf->b_evict_lock);
5217 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5218 ASSERT(!HDR_IN_HASH_TABLE(hdr));
5219 ASSERT(!HDR_HAS_L2HDR(hdr));
5220 ASSERT(HDR_EMPTY(hdr));
5221
5222 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5223 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
5224 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
5225
5226 hdr->b_l1hdr.b_arc_access = 0;
5227
5228 /*
5229 * If the buf is being overridden then it may already
5230 * have a hdr that is not empty.
5231 */
5232 buf_discard_identity(hdr);
5233 arc_buf_thaw(buf);
5234
5235 return;
5236 }
5237
5238 kmutex_t *hash_lock = HDR_LOCK(hdr);
5239 mutex_enter(hash_lock);
5240
5241 /*
5242 * This assignment is only valid as long as the hash_lock is
5243 * held, we must be careful not to reference state or the
5244 * b_state field after dropping the lock.
5245 */
5246 arc_state_t *state = hdr->b_l1hdr.b_state;
5247 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5248 ASSERT3P(state, !=, arc_anon);
5249
5250 /* this buffer is not on any list */
5251 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
5252
5253 if (HDR_HAS_L2HDR(hdr)) {
5254 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5255
5256 /*
5257 * We have to recheck this conditional again now that
5258 * we're holding the l2ad_mtx to prevent a race with
5259 * another thread which might be concurrently calling
5260 * l2arc_evict(). In that case, l2arc_evict() might have
5261 * destroyed the header's L2 portion as we were waiting
5262 * to acquire the l2ad_mtx.
5263 */
5264 if (HDR_HAS_L2HDR(hdr))
5265 arc_hdr_l2hdr_destroy(hdr);
5266
5267 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
5268 }
5269
5270 /*
5271 * Do we have more than one buf?
5272 */
5273 if (hdr->b_l1hdr.b_bufcnt > 1) {
5274 arc_buf_hdr_t *nhdr;
5275 uint64_t spa = hdr->b_spa;
5276 uint64_t psize = HDR_GET_PSIZE(hdr);
5277 uint64_t lsize = HDR_GET_LSIZE(hdr);
5278 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
5279 arc_buf_contents_t type = arc_buf_type(hdr);
5280 VERIFY3U(hdr->b_type, ==, type);
5281
5282 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
5283 (void) remove_reference(hdr, hash_lock, tag);
5284
5285 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
5286 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
5287 ASSERT(ARC_BUF_LAST(buf));
5288 }
5289
5290 /*
5291 * Pull the data off of this hdr and attach it to
5292 * a new anonymous hdr. Also find the last buffer
5293 * in the hdr's buffer list.
5294 */
5295 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
5296 ASSERT3P(lastbuf, !=, NULL);
5297
5298 /*
5299 * If the current arc_buf_t and the hdr are sharing their data
5300 * buffer, then we must stop sharing that block.
5301 */
5302 if (arc_buf_is_shared(buf)) {
5303 VERIFY(!arc_buf_is_shared(lastbuf));
5304
5305 /*
5306 * First, sever the block sharing relationship between
5307 * buf and the arc_buf_hdr_t.
5308 */
5309 arc_unshare_buf(hdr, buf);
5310
5311 /*
5312 * Now we need to recreate the hdr's b_pabd. Since we
5313 * have lastbuf handy, we try to share with it, but if
5314 * we can't then we allocate a new b_pabd and copy the
5315 * data from buf into it.
5316 */
5317 if (arc_can_share(hdr, lastbuf)) {
5318 arc_share_buf(hdr, lastbuf);
5319 } else {
5320 arc_hdr_alloc_pabd(hdr);
5321 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
5322 buf->b_data, psize);
5323 }
5324 VERIFY3P(lastbuf->b_data, !=, NULL);
5325 } else if (HDR_SHARED_DATA(hdr)) {
5326 /*
5327 * Uncompressed shared buffers are always at the end
5328 * of the list. Compressed buffers don't have the
5329 * same requirements. This makes it hard to
5330 * simply assert that the lastbuf is shared so
5331 * we rely on the hdr's compression flags to determine
5332 * if we have a compressed, shared buffer.
5333 */
5334 ASSERT(arc_buf_is_shared(lastbuf) ||
5335 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF);
5336 ASSERT(!ARC_BUF_SHARED(buf));
5337 }
5338 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5339 ASSERT3P(state, !=, arc_l2c_only);
5340
5341 (void) refcount_remove_many(&state->arcs_size,
5342 arc_buf_size(buf), buf);
5343
5344 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5345 ASSERT3P(state, !=, arc_l2c_only);
5346 (void) refcount_remove_many(&state->arcs_esize[type],
5347 arc_buf_size(buf), buf);
5348 }
5349
5350 hdr->b_l1hdr.b_bufcnt -= 1;
5351 arc_cksum_verify(buf);
5352 arc_buf_unwatch(buf);
5353
5354 mutex_exit(hash_lock);
5355
5356 /*
5357 * Allocate a new hdr. The new hdr will contain a b_pabd
5358 * buffer which will be freed in arc_write().
5359 */
5360 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5361 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5362 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5363 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5364 VERIFY3U(nhdr->b_type, ==, type);
5365 ASSERT(!HDR_SHARED_DATA(nhdr));
5366
5367 nhdr->b_l1hdr.b_buf = buf;
5368 nhdr->b_l1hdr.b_bufcnt = 1;
5369 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5370 buf->b_hdr = nhdr;
5371
5372 mutex_exit(&buf->b_evict_lock);
5373 (void) refcount_add_many(&arc_anon->arcs_size,
5374 arc_buf_size(buf), buf);
5375 } else {
5376 mutex_exit(&buf->b_evict_lock);
5377 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5378 /* protected by hash lock, or hdr is on arc_anon */
5379 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5380 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5381 arc_change_state(arc_anon, hdr, hash_lock);
5382 hdr->b_l1hdr.b_arc_access = 0;
5383 mutex_exit(hash_lock);
5384
5385 buf_discard_identity(hdr);
5386 arc_buf_thaw(buf);
5387 }
5388 }
5389
5390 int
5391 arc_released(arc_buf_t *buf)
5392 {
5393 int released;
5394
5395 mutex_enter(&buf->b_evict_lock);
5396 released = (buf->b_data != NULL &&
5397 buf->b_hdr->b_l1hdr.b_state == arc_anon);
5398 mutex_exit(&buf->b_evict_lock);
5399 return (released);
5400 }
5401
5402 #ifdef ZFS_DEBUG
5403 int
5404 arc_referenced(arc_buf_t *buf)
5405 {
5406 int referenced;
5407
5408 mutex_enter(&buf->b_evict_lock);
5409 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
5410 mutex_exit(&buf->b_evict_lock);
5411 return (referenced);
5412 }
5413 #endif
5414
5415 static void
5416 arc_write_ready(zio_t *zio)
5417 {
5418 arc_write_callback_t *callback = zio->io_private;
5419 arc_buf_t *buf = callback->awcb_buf;
5420 arc_buf_hdr_t *hdr = buf->b_hdr;
5421 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp);
5422
5423 ASSERT(HDR_HAS_L1HDR(hdr));
5424 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
5425 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
5426
5427 /*
5428 * If we're reexecuting this zio because the pool suspended, then
5429 * cleanup any state that was previously set the first time the
5430 * callback was invoked.
5431 */
5432 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
5433 arc_cksum_free(hdr);
5434 arc_buf_unwatch(buf);
5435 if (hdr->b_l1hdr.b_pabd != NULL) {
5436 if (arc_buf_is_shared(buf)) {
5437 arc_unshare_buf(hdr, buf);
5438 } else {
5439 arc_hdr_free_pabd(hdr);
5440 }
5441 }
5442 }
5443 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5444 ASSERT(!HDR_SHARED_DATA(hdr));
5445 ASSERT(!arc_buf_is_shared(buf));
5446
5447 callback->awcb_ready(zio, buf, callback->awcb_private);
5448
5449 if (HDR_IO_IN_PROGRESS(hdr))
5450 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
5451
5452 arc_cksum_compute(buf);
5453 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5454
5455 enum zio_compress compress;
5456 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5457 compress = ZIO_COMPRESS_OFF;
5458 } else {
5459 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp));
5460 compress = BP_GET_COMPRESS(zio->io_bp);
5461 }
5462 HDR_SET_PSIZE(hdr, psize);
5463 arc_hdr_set_compress(hdr, compress);
5464
5465
5466 /*
5467 * Fill the hdr with data. If the hdr is compressed, the data we want
5468 * is available from the zio, otherwise we can take it from the buf.
5469 *
5470 * We might be able to share the buf's data with the hdr here. However,
5471 * doing so would cause the ARC to be full of linear ABDs if we write a
5472 * lot of shareable data. As a compromise, we check whether scattered
5473 * ABDs are allowed, and assume that if they are then the user wants
5474 * the ARC to be primarily filled with them regardless of the data being
5475 * written. Therefore, if they're allowed then we allocate one and copy
5476 * the data into it; otherwise, we share the data directly if we can.
5477 */
5478 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
5479 arc_hdr_alloc_pabd(hdr);
5480
5481 /*
5482 * Ideally, we would always copy the io_abd into b_pabd, but the
5483 * user may have disabled compressed ARC, thus we must check the
5484 * hdr's compression setting rather than the io_bp's.
5485 */
5486 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) {
5487 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=,
5488 ZIO_COMPRESS_OFF);
5489 ASSERT3U(psize, >, 0);
5490
5491 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
5492 } else {
5493 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
5494
5495 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
5496 arc_buf_size(buf));
5497 }
5498 } else {
5499 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
5500 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
5501 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
5502
5503 arc_share_buf(hdr, buf);
5504 }
5505
5506 arc_hdr_verify(hdr, zio->io_bp);
5507 }
5508
5509 static void
5510 arc_write_children_ready(zio_t *zio)
5511 {
5512 arc_write_callback_t *callback = zio->io_private;
5513 arc_buf_t *buf = callback->awcb_buf;
5514
5515 callback->awcb_children_ready(zio, buf, callback->awcb_private);
5516 }
5517
5518 /*
5519 * The SPA calls this callback for each physical write that happens on behalf
5520 * of a logical write. See the comment in dbuf_write_physdone() for details.
5521 */
5522 static void
5523 arc_write_physdone(zio_t *zio)
5524 {
5525 arc_write_callback_t *cb = zio->io_private;
5526 if (cb->awcb_physdone != NULL)
5527 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
5528 }
5529
5530 static void
5531 arc_write_done(zio_t *zio)
5532 {
5533 arc_write_callback_t *callback = zio->io_private;
5534 arc_buf_t *buf = callback->awcb_buf;
5535 arc_buf_hdr_t *hdr = buf->b_hdr;
5536
5537 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5538
5539 if (zio->io_error == 0) {
5540 arc_hdr_verify(hdr, zio->io_bp);
5541
5542 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
5543 buf_discard_identity(hdr);
5544 } else {
5545 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
5546 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
5547 }
5548 } else {
5549 ASSERT(HDR_EMPTY(hdr));
5550 }
5551
5552 /*
5553 * If the block to be written was all-zero or compressed enough to be
5554 * embedded in the BP, no write was performed so there will be no
5555 * dva/birth/checksum. The buffer must therefore remain anonymous
5556 * (and uncached).
5557 */
5558 if (!HDR_EMPTY(hdr)) {
5559 arc_buf_hdr_t *exists;
5560 kmutex_t *hash_lock;
5561
5562 ASSERT3U(zio->io_error, ==, 0);
5563
5564 arc_cksum_verify(buf);
5565
5566 exists = buf_hash_insert(hdr, &hash_lock);
5567 if (exists != NULL) {
5568 /*
5569 * This can only happen if we overwrite for
5570 * sync-to-convergence, because we remove
5571 * buffers from the hash table when we arc_free().
5572 */
5573 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
5574 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5575 panic("bad overwrite, hdr=%p exists=%p",
5576 (void *)hdr, (void *)exists);
5577 ASSERT(refcount_is_zero(
5578 &exists->b_l1hdr.b_refcnt));
5579 arc_change_state(arc_anon, exists, hash_lock);
5580 mutex_exit(hash_lock);
5581 arc_hdr_destroy(exists);
5582 exists = buf_hash_insert(hdr, &hash_lock);
5583 ASSERT3P(exists, ==, NULL);
5584 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
5585 /* nopwrite */
5586 ASSERT(zio->io_prop.zp_nopwrite);
5587 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5588 panic("bad nopwrite, hdr=%p exists=%p",
5589 (void *)hdr, (void *)exists);
5590 } else {
5591 /* Dedup */
5592 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
5593 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
5594 ASSERT(BP_GET_DEDUP(zio->io_bp));
5595 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
5596 }
5597 }
5598 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5599 /* if it's not anon, we are doing a scrub */
5600 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
5601 arc_access(hdr, hash_lock);
5602 mutex_exit(hash_lock);
5603 } else {
5604 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5605 }
5606
5607 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5608 callback->awcb_done(zio, buf, callback->awcb_private);
5609
5610 abd_put(zio->io_abd);
5611 kmem_free(callback, sizeof (arc_write_callback_t));
5612 }
5613
5614 zio_t *
5615 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
5616 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
5617 arc_done_func_t *children_ready, arc_done_func_t *physdone,
5618 arc_done_func_t *done, void *private, zio_priority_t priority,
5619 int zio_flags, const zbookmark_phys_t *zb)
5620 {
5621 arc_buf_hdr_t *hdr = buf->b_hdr;
5622 arc_write_callback_t *callback;
5623 zio_t *zio;
5624 zio_prop_t localprop = *zp;
5625
5626 ASSERT3P(ready, !=, NULL);
5627 ASSERT3P(done, !=, NULL);
5628 ASSERT(!HDR_IO_ERROR(hdr));
5629 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5630 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5631 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
5632 if (l2arc)
5633 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5634 if (ARC_BUF_COMPRESSED(buf)) {
5635 /*
5636 * We're writing a pre-compressed buffer. Make the
5637 * compression algorithm requested by the zio_prop_t match
5638 * the pre-compressed buffer's compression algorithm.
5639 */
5640 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
5641
5642 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
5643 zio_flags |= ZIO_FLAG_RAW;
5644 }
5645 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
5646 callback->awcb_ready = ready;
5647 callback->awcb_children_ready = children_ready;
5648 callback->awcb_physdone = physdone;
5649 callback->awcb_done = done;
5650 callback->awcb_private = private;
5651 callback->awcb_buf = buf;
5652
5653 /*
5654 * The hdr's b_pabd is now stale, free it now. A new data block
5655 * will be allocated when the zio pipeline calls arc_write_ready().
5656 */
5657 if (hdr->b_l1hdr.b_pabd != NULL) {
5658 /*
5659 * If the buf is currently sharing the data block with
5660 * the hdr then we need to break that relationship here.
5661 * The hdr will remain with a NULL data pointer and the
5662 * buf will take sole ownership of the block.
5663 */
5664 if (arc_buf_is_shared(buf)) {
5665 arc_unshare_buf(hdr, buf);
5666 } else {
5667 arc_hdr_free_pabd(hdr);
5668 }
5669 VERIFY3P(buf->b_data, !=, NULL);
5670 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
5671 }
5672 ASSERT(!arc_buf_is_shared(buf));
5673 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5674
5675 zio = zio_write(pio, spa, txg, bp,
5676 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
5677 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
5678 (children_ready != NULL) ? arc_write_children_ready : NULL,
5679 arc_write_physdone, arc_write_done, callback,
5680 priority, zio_flags, zb);
5681
5682 return (zio);
5683 }
5684
5685 static int
5686 arc_memory_throttle(uint64_t reserve, uint64_t txg)
5687 {
5688 #ifdef _KERNEL
5689 uint64_t available_memory = ptob(freemem);
5690 static uint64_t page_load = 0;
5691 static uint64_t last_txg = 0;
5692
5693 #if defined(__i386)
5694 available_memory =
5695 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
5696 #endif
5697
5698 if (freemem > physmem * arc_lotsfree_percent / 100)
5699 return (0);
5700
5701 if (txg > last_txg) {
5702 last_txg = txg;
5703 page_load = 0;
5704 }
5705 /*
5706 * If we are in pageout, we know that memory is already tight,
5707 * the arc is already going to be evicting, so we just want to
5708 * continue to let page writes occur as quickly as possible.
5709 */
5710 if (curproc == proc_pageout) {
5711 if (page_load > MAX(ptob(minfree), available_memory) / 4)
5712 return (SET_ERROR(ERESTART));
5713 /* Note: reserve is inflated, so we deflate */
5714 page_load += reserve / 8;
5715 return (0);
5716 } else if (page_load > 0 && arc_reclaim_needed()) {
5717 /* memory is low, delay before restarting */
5718 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
5719 return (SET_ERROR(EAGAIN));
5720 }
5721 page_load = 0;
5722 #endif
5723 return (0);
5724 }
5725
5726 void
5727 arc_tempreserve_clear(uint64_t reserve)
5728 {
5729 atomic_add_64(&arc_tempreserve, -reserve);
5730 ASSERT((int64_t)arc_tempreserve >= 0);
5731 }
5732
5733 int
5734 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
5735 {
5736 int error;
5737 uint64_t anon_size;
5738
5739 if (reserve > arc_c/4 && !arc_no_grow)
5740 arc_c = MIN(arc_c_max, reserve * 4);
5741 if (reserve > arc_c)
5742 return (SET_ERROR(ENOMEM));
5743
5744 /*
5745 * Don't count loaned bufs as in flight dirty data to prevent long
5746 * network delays from blocking transactions that are ready to be
5747 * assigned to a txg.
5748 */
5749
5750 /* assert that it has not wrapped around */
5751 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
5752
5753 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
5754 arc_loaned_bytes), 0);
5755
5756 /*
5757 * Writes will, almost always, require additional memory allocations
5758 * in order to compress/encrypt/etc the data. We therefore need to
5759 * make sure that there is sufficient available memory for this.
5760 */
5761 error = arc_memory_throttle(reserve, txg);
5762 if (error != 0)
5763 return (error);
5764
5765 /*
5766 * Throttle writes when the amount of dirty data in the cache
5767 * gets too large. We try to keep the cache less than half full
5768 * of dirty blocks so that our sync times don't grow too large.
5769 * Note: if two requests come in concurrently, we might let them
5770 * both succeed, when one of them should fail. Not a huge deal.
5771 */
5772
5773 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
5774 anon_size > arc_c / 4) {
5775 uint64_t meta_esize =
5776 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
5777 uint64_t data_esize =
5778 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
5779 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5780 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5781 arc_tempreserve >> 10, meta_esize >> 10,
5782 data_esize >> 10, reserve >> 10, arc_c >> 10);
5783 return (SET_ERROR(ERESTART));
5784 }
5785 atomic_add_64(&arc_tempreserve, reserve);
5786 return (0);
5787 }
5788
5789 static void
5790 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
5791 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
5792 {
5793 size->value.ui64 = refcount_count(&state->arcs_size);
5794 evict_data->value.ui64 =
5795 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
5796 evict_metadata->value.ui64 =
5797 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
5798 }
5799
5800 static int
5801 arc_kstat_update(kstat_t *ksp, int rw)
5802 {
5803 arc_stats_t *as = ksp->ks_data;
5804
5805 if (rw == KSTAT_WRITE) {
5806 return (EACCES);
5807 } else {
5808 arc_kstat_update_state(arc_anon,
5809 &as->arcstat_anon_size,
5810 &as->arcstat_anon_evictable_data,
5811 &as->arcstat_anon_evictable_metadata);
5812 arc_kstat_update_state(arc_mru,
5813 &as->arcstat_mru_size,
5814 &as->arcstat_mru_evictable_data,
5815 &as->arcstat_mru_evictable_metadata);
5816 arc_kstat_update_state(arc_mru_ghost,
5817 &as->arcstat_mru_ghost_size,
5818 &as->arcstat_mru_ghost_evictable_data,
5819 &as->arcstat_mru_ghost_evictable_metadata);
5820 arc_kstat_update_state(arc_mfu,
5821 &as->arcstat_mfu_size,
5822 &as->arcstat_mfu_evictable_data,
5823 &as->arcstat_mfu_evictable_metadata);
5824 arc_kstat_update_state(arc_mfu_ghost,
5825 &as->arcstat_mfu_ghost_size,
5826 &as->arcstat_mfu_ghost_evictable_data,
5827 &as->arcstat_mfu_ghost_evictable_metadata);
5828 }
5829
5830 return (0);
5831 }
5832
5833 /*
5834 * This function *must* return indices evenly distributed between all
5835 * sublists of the multilist. This is needed due to how the ARC eviction
5836 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5837 * distributed between all sublists and uses this assumption when
5838 * deciding which sublist to evict from and how much to evict from it.
5839 */
5840 unsigned int
5841 arc_state_multilist_index_func(multilist_t *ml, void *obj)
5842 {
5843 arc_buf_hdr_t *hdr = obj;
5844
5845 /*
5846 * We rely on b_dva to generate evenly distributed index
5847 * numbers using buf_hash below. So, as an added precaution,
5848 * let's make sure we never add empty buffers to the arc lists.
5849 */
5850 ASSERT(!HDR_EMPTY(hdr));
5851
5852 /*
5853 * The assumption here, is the hash value for a given
5854 * arc_buf_hdr_t will remain constant throughout it's lifetime
5855 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
5856 * Thus, we don't need to store the header's sublist index
5857 * on insertion, as this index can be recalculated on removal.
5858 *
5859 * Also, the low order bits of the hash value are thought to be
5860 * distributed evenly. Otherwise, in the case that the multilist
5861 * has a power of two number of sublists, each sublists' usage
5862 * would not be evenly distributed.
5863 */
5864 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
5865 multilist_get_num_sublists(ml));
5866 }
5867
5868 static void
5869 arc_state_init(void)
5870 {
5871 arc_anon = &ARC_anon;
5872 arc_mru = &ARC_mru;
5873 arc_mru_ghost = &ARC_mru_ghost;
5874 arc_mfu = &ARC_mfu;
5875 arc_mfu_ghost = &ARC_mfu_ghost;
5876 arc_l2c_only = &ARC_l2c_only;
5877
5878 arc_mru->arcs_list[ARC_BUFC_METADATA] =
5879 multilist_create(sizeof (arc_buf_hdr_t),
5880 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5881 arc_state_multilist_index_func);
5882 arc_mru->arcs_list[ARC_BUFC_DATA] =
5883 multilist_create(sizeof (arc_buf_hdr_t),
5884 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5885 arc_state_multilist_index_func);
5886 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
5887 multilist_create(sizeof (arc_buf_hdr_t),
5888 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5889 arc_state_multilist_index_func);
5890 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
5891 multilist_create(sizeof (arc_buf_hdr_t),
5892 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5893 arc_state_multilist_index_func);
5894 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
5895 multilist_create(sizeof (arc_buf_hdr_t),
5896 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5897 arc_state_multilist_index_func);
5898 arc_mfu->arcs_list[ARC_BUFC_DATA] =
5899 multilist_create(sizeof (arc_buf_hdr_t),
5900 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5901 arc_state_multilist_index_func);
5902 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
5903 multilist_create(sizeof (arc_buf_hdr_t),
5904 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5905 arc_state_multilist_index_func);
5906 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
5907 multilist_create(sizeof (arc_buf_hdr_t),
5908 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5909 arc_state_multilist_index_func);
5910 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
5911 multilist_create(sizeof (arc_buf_hdr_t),
5912 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5913 arc_state_multilist_index_func);
5914 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
5915 multilist_create(sizeof (arc_buf_hdr_t),
5916 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5917 arc_state_multilist_index_func);
5918
5919 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
5920 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
5921 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
5922 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
5923 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
5924 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
5925 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
5926 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
5927 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
5928 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
5929 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
5930 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
5931
5932 refcount_create(&arc_anon->arcs_size);
5933 refcount_create(&arc_mru->arcs_size);
5934 refcount_create(&arc_mru_ghost->arcs_size);
5935 refcount_create(&arc_mfu->arcs_size);
5936 refcount_create(&arc_mfu_ghost->arcs_size);
5937 refcount_create(&arc_l2c_only->arcs_size);
5938 }
5939
5940 static void
5941 arc_state_fini(void)
5942 {
5943 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
5944 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
5945 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
5946 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
5947 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
5948 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
5949 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
5950 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
5951 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
5952 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
5953 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
5954 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
5955
5956 refcount_destroy(&arc_anon->arcs_size);
5957 refcount_destroy(&arc_mru->arcs_size);
5958 refcount_destroy(&arc_mru_ghost->arcs_size);
5959 refcount_destroy(&arc_mfu->arcs_size);
5960 refcount_destroy(&arc_mfu_ghost->arcs_size);
5961 refcount_destroy(&arc_l2c_only->arcs_size);
5962
5963 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
5964 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
5965 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
5966 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
5967 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
5968 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
5969 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
5970 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
5971 }
5972
5973 uint64_t
5974 arc_max_bytes(void)
5975 {
5976 return (arc_c_max);
5977 }
5978
5979 void
5980 arc_init(void)
5981 {
5982 /*
5983 * allmem is "all memory that we could possibly use".
5984 */
5985 #ifdef _KERNEL
5986 uint64_t allmem = ptob(physmem - swapfs_minfree);
5987 #else
5988 uint64_t allmem = (physmem * PAGESIZE) / 2;
5989 #endif
5990
5991 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
5992 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
5993 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
5994
5995 /* Convert seconds to clock ticks */
5996 arc_min_prefetch_lifespan = 1 * hz;
5997
5998 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
5999 arc_c_min = MAX(allmem / 32, 64 << 20);
6000 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
6001 if (allmem >= 1 << 30)
6002 arc_c_max = allmem - (1 << 30);
6003 else
6004 arc_c_max = arc_c_min;
6005 arc_c_max = MAX(allmem * 3 / 4, arc_c_max);
6006
6007 /*
6008 * In userland, there's only the memory pressure that we artificially
6009 * create (see arc_available_memory()). Don't let arc_c get too
6010 * small, because it can cause transactions to be larger than
6011 * arc_c, causing arc_tempreserve_space() to fail.
6012 */
6013 #ifndef _KERNEL
6014 arc_c_min = arc_c_max / 2;
6015 #endif
6016
6017 /*
6018 * Allow the tunables to override our calculations if they are
6019 * reasonable (ie. over 64MB)
6020 */
6021 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) {
6022 arc_c_max = zfs_arc_max;
6023 arc_c_min = MIN(arc_c_min, arc_c_max);
6024 }
6025 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
6026 arc_c_min = zfs_arc_min;
6027
6028 arc_c = arc_c_max;
6029 arc_p = (arc_c >> 1);
6030 arc_size = 0;
6031
6032 /* limit meta-data to 1/4 of the arc capacity */
6033 arc_meta_limit = arc_c_max / 4;
6034
6035 #ifdef _KERNEL
6036 /*
6037 * Metadata is stored in the kernel's heap. Don't let us
6038 * use more than half the heap for the ARC.
6039 */
6040 arc_meta_limit = MIN(arc_meta_limit,
6041 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
6042 #endif
6043
6044 /* Allow the tunable to override if it is reasonable */
6045 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6046 arc_meta_limit = zfs_arc_meta_limit;
6047
6048 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6049 arc_c_min = arc_meta_limit / 2;
6050
6051 if (zfs_arc_meta_min > 0) {
6052 arc_meta_min = zfs_arc_meta_min;
6053 } else {
6054 arc_meta_min = arc_c_min / 2;
6055 }
6056
6057 if (zfs_arc_grow_retry > 0)
6058 arc_grow_retry = zfs_arc_grow_retry;
6059
6060 if (zfs_arc_shrink_shift > 0)
6061 arc_shrink_shift = zfs_arc_shrink_shift;
6062
6063 /*
6064 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6065 */
6066 if (arc_no_grow_shift >= arc_shrink_shift)
6067 arc_no_grow_shift = arc_shrink_shift - 1;
6068
6069 if (zfs_arc_p_min_shift > 0)
6070 arc_p_min_shift = zfs_arc_p_min_shift;
6071
6072 /* if kmem_flags are set, lets try to use less memory */
6073 if (kmem_debugging())
6074 arc_c = arc_c / 2;
6075 if (arc_c < arc_c_min)
6076 arc_c = arc_c_min;
6077
6078 arc_state_init();
6079 buf_init();
6080
6081 arc_reclaim_thread_exit = B_FALSE;
6082
6083 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
6084 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
6085
6086 if (arc_ksp != NULL) {
6087 arc_ksp->ks_data = &arc_stats;
6088 arc_ksp->ks_update = arc_kstat_update;
6089 kstat_install(arc_ksp);
6090 }
6091
6092 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
6093 TS_RUN, minclsyspri);
6094
6095 arc_dead = B_FALSE;
6096 arc_warm = B_FALSE;
6097
6098 /*
6099 * Calculate maximum amount of dirty data per pool.
6100 *
6101 * If it has been set by /etc/system, take that.
6102 * Otherwise, use a percentage of physical memory defined by
6103 * zfs_dirty_data_max_percent (default 10%) with a cap at
6104 * zfs_dirty_data_max_max (default 4GB).
6105 */
6106 if (zfs_dirty_data_max == 0) {
6107 zfs_dirty_data_max = physmem * PAGESIZE *
6108 zfs_dirty_data_max_percent / 100;
6109 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
6110 zfs_dirty_data_max_max);
6111 }
6112 }
6113
6114 void
6115 arc_fini(void)
6116 {
6117 mutex_enter(&arc_reclaim_lock);
6118 arc_reclaim_thread_exit = B_TRUE;
6119 /*
6120 * The reclaim thread will set arc_reclaim_thread_exit back to
6121 * B_FALSE when it is finished exiting; we're waiting for that.
6122 */
6123 while (arc_reclaim_thread_exit) {
6124 cv_signal(&arc_reclaim_thread_cv);
6125 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6126 }
6127 mutex_exit(&arc_reclaim_lock);
6128
6129 /* Use B_TRUE to ensure *all* buffers are evicted */
6130 arc_flush(NULL, B_TRUE);
6131
6132 arc_dead = B_TRUE;
6133
6134 if (arc_ksp != NULL) {
6135 kstat_delete(arc_ksp);
6136 arc_ksp = NULL;
6137 }
6138
6139 mutex_destroy(&arc_reclaim_lock);
6140 cv_destroy(&arc_reclaim_thread_cv);
6141 cv_destroy(&arc_reclaim_waiters_cv);
6142
6143 arc_state_fini();
6144 buf_fini();
6145
6146 ASSERT0(arc_loaned_bytes);
6147 }
6148
6149 /*
6150 * Level 2 ARC
6151 *
6152 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6153 * It uses dedicated storage devices to hold cached data, which are populated
6154 * using large infrequent writes. The main role of this cache is to boost
6155 * the performance of random read workloads. The intended L2ARC devices
6156 * include short-stroked disks, solid state disks, and other media with
6157 * substantially faster read latency than disk.
6158 *
6159 * +-----------------------+
6160 * | ARC |
6161 * +-----------------------+
6162 * | ^ ^
6163 * | | |
6164 * l2arc_feed_thread() arc_read()
6165 * | | |
6166 * | l2arc read |
6167 * V | |
6168 * +---------------+ |
6169 * | L2ARC | |
6170 * +---------------+ |
6171 * | ^ |
6172 * l2arc_write() | |
6173 * | | |
6174 * V | |
6175 * +-------+ +-------+
6176 * | vdev | | vdev |
6177 * | cache | | cache |
6178 * +-------+ +-------+
6179 * +=========+ .-----.
6180 * : L2ARC : |-_____-|
6181 * : devices : | Disks |
6182 * +=========+ `-_____-'
6183 *
6184 * Read requests are satisfied from the following sources, in order:
6185 *
6186 * 1) ARC
6187 * 2) vdev cache of L2ARC devices
6188 * 3) L2ARC devices
6189 * 4) vdev cache of disks
6190 * 5) disks
6191 *
6192 * Some L2ARC device types exhibit extremely slow write performance.
6193 * To accommodate for this there are some significant differences between
6194 * the L2ARC and traditional cache design:
6195 *
6196 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
6197 * the ARC behave as usual, freeing buffers and placing headers on ghost
6198 * lists. The ARC does not send buffers to the L2ARC during eviction as
6199 * this would add inflated write latencies for all ARC memory pressure.
6200 *
6201 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
6202 * It does this by periodically scanning buffers from the eviction-end of
6203 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
6204 * not already there. It scans until a headroom of buffers is satisfied,
6205 * which itself is a buffer for ARC eviction. If a compressible buffer is
6206 * found during scanning and selected for writing to an L2ARC device, we
6207 * temporarily boost scanning headroom during the next scan cycle to make
6208 * sure we adapt to compression effects (which might significantly reduce
6209 * the data volume we write to L2ARC). The thread that does this is
6210 * l2arc_feed_thread(), illustrated below; example sizes are included to
6211 * provide a better sense of ratio than this diagram:
6212 *
6213 * head --> tail
6214 * +---------------------+----------+
6215 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
6216 * +---------------------+----------+ | o L2ARC eligible
6217 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
6218 * +---------------------+----------+ |
6219 * 15.9 Gbytes ^ 32 Mbytes |
6220 * headroom |
6221 * l2arc_feed_thread()
6222 * |
6223 * l2arc write hand <--[oooo]--'
6224 * | 8 Mbyte
6225 * | write max
6226 * V
6227 * +==============================+
6228 * L2ARC dev |####|#|###|###| |####| ... |
6229 * +==============================+
6230 * 32 Gbytes
6231 *
6232 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
6233 * evicted, then the L2ARC has cached a buffer much sooner than it probably
6234 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
6235 * safe to say that this is an uncommon case, since buffers at the end of
6236 * the ARC lists have moved there due to inactivity.
6237 *
6238 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
6239 * then the L2ARC simply misses copying some buffers. This serves as a
6240 * pressure valve to prevent heavy read workloads from both stalling the ARC
6241 * with waits and clogging the L2ARC with writes. This also helps prevent
6242 * the potential for the L2ARC to churn if it attempts to cache content too
6243 * quickly, such as during backups of the entire pool.
6244 *
6245 * 5. After system boot and before the ARC has filled main memory, there are
6246 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
6247 * lists can remain mostly static. Instead of searching from tail of these
6248 * lists as pictured, the l2arc_feed_thread() will search from the list heads
6249 * for eligible buffers, greatly increasing its chance of finding them.
6250 *
6251 * The L2ARC device write speed is also boosted during this time so that
6252 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
6253 * there are no L2ARC reads, and no fear of degrading read performance
6254 * through increased writes.
6255 *
6256 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
6257 * the vdev queue can aggregate them into larger and fewer writes. Each
6258 * device is written to in a rotor fashion, sweeping writes through
6259 * available space then repeating.
6260 *
6261 * 7. The L2ARC does not store dirty content. It never needs to flush
6262 * write buffers back to disk based storage.
6263 *
6264 * 8. If an ARC buffer is written (and dirtied) which also exists in the
6265 * L2ARC, the now stale L2ARC buffer is immediately dropped.
6266 *
6267 * The performance of the L2ARC can be tweaked by a number of tunables, which
6268 * may be necessary for different workloads:
6269 *
6270 * l2arc_write_max max write bytes per interval
6271 * l2arc_write_boost extra write bytes during device warmup
6272 * l2arc_noprefetch skip caching prefetched buffers
6273 * l2arc_headroom number of max device writes to precache
6274 * l2arc_headroom_boost when we find compressed buffers during ARC
6275 * scanning, we multiply headroom by this
6276 * percentage factor for the next scan cycle,
6277 * since more compressed buffers are likely to
6278 * be present
6279 * l2arc_feed_secs seconds between L2ARC writing
6280 *
6281 * Tunables may be removed or added as future performance improvements are
6282 * integrated, and also may become zpool properties.
6283 *
6284 * There are three key functions that control how the L2ARC warms up:
6285 *
6286 * l2arc_write_eligible() check if a buffer is eligible to cache
6287 * l2arc_write_size() calculate how much to write
6288 * l2arc_write_interval() calculate sleep delay between writes
6289 *
6290 * These three functions determine what to write, how much, and how quickly
6291 * to send writes.
6292 */
6293
6294 static boolean_t
6295 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
6296 {
6297 /*
6298 * A buffer is *not* eligible for the L2ARC if it:
6299 * 1. belongs to a different spa.
6300 * 2. is already cached on the L2ARC.
6301 * 3. has an I/O in progress (it may be an incomplete read).
6302 * 4. is flagged not eligible (zfs property).
6303 */
6304 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
6305 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
6306 return (B_FALSE);
6307
6308 return (B_TRUE);
6309 }
6310
6311 static uint64_t
6312 l2arc_write_size(void)
6313 {
6314 uint64_t size;
6315
6316 /*
6317 * Make sure our globals have meaningful values in case the user
6318 * altered them.
6319 */
6320 size = l2arc_write_max;
6321 if (size == 0) {
6322 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
6323 "be greater than zero, resetting it to the default (%d)",
6324 L2ARC_WRITE_SIZE);
6325 size = l2arc_write_max = L2ARC_WRITE_SIZE;
6326 }
6327
6328 if (arc_warm == B_FALSE)
6329 size += l2arc_write_boost;
6330
6331 return (size);
6332
6333 }
6334
6335 static clock_t
6336 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
6337 {
6338 clock_t interval, next, now;
6339
6340 /*
6341 * If the ARC lists are busy, increase our write rate; if the
6342 * lists are stale, idle back. This is achieved by checking
6343 * how much we previously wrote - if it was more than half of
6344 * what we wanted, schedule the next write much sooner.
6345 */
6346 if (l2arc_feed_again && wrote > (wanted / 2))
6347 interval = (hz * l2arc_feed_min_ms) / 1000;
6348 else
6349 interval = hz * l2arc_feed_secs;
6350
6351 now = ddi_get_lbolt();
6352 next = MAX(now, MIN(now + interval, began + interval));
6353
6354 return (next);
6355 }
6356
6357 /*
6358 * Cycle through L2ARC devices. This is how L2ARC load balances.
6359 * If a device is returned, this also returns holding the spa config lock.
6360 */
6361 static l2arc_dev_t *
6362 l2arc_dev_get_next(void)
6363 {
6364 l2arc_dev_t *first, *next = NULL;
6365
6366 /*
6367 * Lock out the removal of spas (spa_namespace_lock), then removal
6368 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6369 * both locks will be dropped and a spa config lock held instead.
6370 */
6371 mutex_enter(&spa_namespace_lock);
6372 mutex_enter(&l2arc_dev_mtx);
6373
6374 /* if there are no vdevs, there is nothing to do */
6375 if (l2arc_ndev == 0)
6376 goto out;
6377
6378 first = NULL;
6379 next = l2arc_dev_last;
6380 do {
6381 /* loop around the list looking for a non-faulted vdev */
6382 if (next == NULL) {
6383 next = list_head(l2arc_dev_list);
6384 } else {
6385 next = list_next(l2arc_dev_list, next);
6386 if (next == NULL)
6387 next = list_head(l2arc_dev_list);
6388 }
6389
6390 /* if we have come back to the start, bail out */
6391 if (first == NULL)
6392 first = next;
6393 else if (next == first)
6394 break;
6395
6396 } while (vdev_is_dead(next->l2ad_vdev));
6397
6398 /* if we were unable to find any usable vdevs, return NULL */
6399 if (vdev_is_dead(next->l2ad_vdev))
6400 next = NULL;
6401
6402 l2arc_dev_last = next;
6403
6404 out:
6405 mutex_exit(&l2arc_dev_mtx);
6406
6407 /*
6408 * Grab the config lock to prevent the 'next' device from being
6409 * removed while we are writing to it.
6410 */
6411 if (next != NULL)
6412 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
6413 mutex_exit(&spa_namespace_lock);
6414
6415 return (next);
6416 }
6417
6418 /*
6419 * Free buffers that were tagged for destruction.
6420 */
6421 static void
6422 l2arc_do_free_on_write()
6423 {
6424 list_t *buflist;
6425 l2arc_data_free_t *df, *df_prev;
6426
6427 mutex_enter(&l2arc_free_on_write_mtx);
6428 buflist = l2arc_free_on_write;
6429
6430 for (df = list_tail(buflist); df; df = df_prev) {
6431 df_prev = list_prev(buflist, df);
6432 ASSERT3P(df->l2df_abd, !=, NULL);
6433 abd_free(df->l2df_abd);
6434 list_remove(buflist, df);
6435 kmem_free(df, sizeof (l2arc_data_free_t));
6436 }
6437
6438 mutex_exit(&l2arc_free_on_write_mtx);
6439 }
6440
6441 /*
6442 * A write to a cache device has completed. Update all headers to allow
6443 * reads from these buffers to begin.
6444 */
6445 static void
6446 l2arc_write_done(zio_t *zio)
6447 {
6448 l2arc_write_callback_t *cb;
6449 l2arc_dev_t *dev;
6450 list_t *buflist;
6451 arc_buf_hdr_t *head, *hdr, *hdr_prev;
6452 kmutex_t *hash_lock;
6453 int64_t bytes_dropped = 0;
6454
6455 cb = zio->io_private;
6456 ASSERT3P(cb, !=, NULL);
6457 dev = cb->l2wcb_dev;
6458 ASSERT3P(dev, !=, NULL);
6459 head = cb->l2wcb_head;
6460 ASSERT3P(head, !=, NULL);
6461 buflist = &dev->l2ad_buflist;
6462 ASSERT3P(buflist, !=, NULL);
6463 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
6464 l2arc_write_callback_t *, cb);
6465
6466 if (zio->io_error != 0)
6467 ARCSTAT_BUMP(arcstat_l2_writes_error);
6468
6469 /*
6470 * All writes completed, or an error was hit.
6471 */
6472 top:
6473 mutex_enter(&dev->l2ad_mtx);
6474 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
6475 hdr_prev = list_prev(buflist, hdr);
6476
6477 hash_lock = HDR_LOCK(hdr);
6478
6479 /*
6480 * We cannot use mutex_enter or else we can deadlock
6481 * with l2arc_write_buffers (due to swapping the order
6482 * the hash lock and l2ad_mtx are taken).
6483 */
6484 if (!mutex_tryenter(hash_lock)) {
6485 /*
6486 * Missed the hash lock. We must retry so we
6487 * don't leave the ARC_FLAG_L2_WRITING bit set.
6488 */
6489 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
6490
6491 /*
6492 * We don't want to rescan the headers we've
6493 * already marked as having been written out, so
6494 * we reinsert the head node so we can pick up
6495 * where we left off.
6496 */
6497 list_remove(buflist, head);
6498 list_insert_after(buflist, hdr, head);
6499
6500 mutex_exit(&dev->l2ad_mtx);
6501
6502 /*
6503 * We wait for the hash lock to become available
6504 * to try and prevent busy waiting, and increase
6505 * the chance we'll be able to acquire the lock
6506 * the next time around.
6507 */
6508 mutex_enter(hash_lock);
6509 mutex_exit(hash_lock);
6510 goto top;
6511 }
6512
6513 /*
6514 * We could not have been moved into the arc_l2c_only
6515 * state while in-flight due to our ARC_FLAG_L2_WRITING
6516 * bit being set. Let's just ensure that's being enforced.
6517 */
6518 ASSERT(HDR_HAS_L1HDR(hdr));
6519
6520 if (zio->io_error != 0) {
6521 /*
6522 * Error - drop L2ARC entry.
6523 */
6524 list_remove(buflist, hdr);
6525 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
6526
6527 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr));
6528 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
6529
6530 bytes_dropped += arc_hdr_size(hdr);
6531 (void) refcount_remove_many(&dev->l2ad_alloc,
6532 arc_hdr_size(hdr), hdr);
6533 }
6534
6535 /*
6536 * Allow ARC to begin reads and ghost list evictions to
6537 * this L2ARC entry.
6538 */
6539 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
6540
6541 mutex_exit(hash_lock);
6542 }
6543
6544 atomic_inc_64(&l2arc_writes_done);
6545 list_remove(buflist, head);
6546 ASSERT(!HDR_HAS_L1HDR(head));
6547 kmem_cache_free(hdr_l2only_cache, head);
6548 mutex_exit(&dev->l2ad_mtx);
6549
6550 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
6551
6552 l2arc_do_free_on_write();
6553
6554 kmem_free(cb, sizeof (l2arc_write_callback_t));
6555 }
6556
6557 /*
6558 * A read to a cache device completed. Validate buffer contents before
6559 * handing over to the regular ARC routines.
6560 */
6561 static void
6562 l2arc_read_done(zio_t *zio)
6563 {
6564 l2arc_read_callback_t *cb;
6565 arc_buf_hdr_t *hdr;
6566 kmutex_t *hash_lock;
6567 boolean_t valid_cksum;
6568
6569 ASSERT3P(zio->io_vd, !=, NULL);
6570 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
6571
6572 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
6573
6574 cb = zio->io_private;
6575 ASSERT3P(cb, !=, NULL);
6576 hdr = cb->l2rcb_hdr;
6577 ASSERT3P(hdr, !=, NULL);
6578
6579 hash_lock = HDR_LOCK(hdr);
6580 mutex_enter(hash_lock);
6581 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6582
6583 /*
6584 * If the data was read into a temporary buffer,
6585 * move it and free the buffer.
6586 */
6587 if (cb->l2rcb_abd != NULL) {
6588 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
6589 if (zio->io_error == 0) {
6590 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd,
6591 arc_hdr_size(hdr));
6592 }
6593
6594 /*
6595 * The following must be done regardless of whether
6596 * there was an error:
6597 * - free the temporary buffer
6598 * - point zio to the real ARC buffer
6599 * - set zio size accordingly
6600 * These are required because zio is either re-used for
6601 * an I/O of the block in the case of the error
6602 * or the zio is passed to arc_read_done() and it
6603 * needs real data.
6604 */
6605 abd_free(cb->l2rcb_abd);
6606 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
6607 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
6608 }
6609
6610 ASSERT3P(zio->io_abd, !=, NULL);
6611
6612 /*
6613 * Check this survived the L2ARC journey.
6614 */
6615 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd);
6616 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
6617 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
6618
6619 valid_cksum = arc_cksum_is_equal(hdr, zio);
6620 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
6621 mutex_exit(hash_lock);
6622 zio->io_private = hdr;
6623 arc_read_done(zio);
6624 } else {
6625 mutex_exit(hash_lock);
6626 /*
6627 * Buffer didn't survive caching. Increment stats and
6628 * reissue to the original storage device.
6629 */
6630 if (zio->io_error != 0) {
6631 ARCSTAT_BUMP(arcstat_l2_io_error);
6632 } else {
6633 zio->io_error = SET_ERROR(EIO);
6634 }
6635 if (!valid_cksum)
6636 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
6637
6638 /*
6639 * If there's no waiter, issue an async i/o to the primary
6640 * storage now. If there *is* a waiter, the caller must
6641 * issue the i/o in a context where it's OK to block.
6642 */
6643 if (zio->io_waiter == NULL) {
6644 zio_t *pio = zio_unique_parent(zio);
6645
6646 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
6647
6648 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
6649 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
6650 hdr, zio->io_priority, cb->l2rcb_flags,
6651 &cb->l2rcb_zb));
6652 }
6653 }
6654
6655 kmem_free(cb, sizeof (l2arc_read_callback_t));
6656 }
6657
6658 /*
6659 * This is the list priority from which the L2ARC will search for pages to
6660 * cache. This is used within loops (0..3) to cycle through lists in the
6661 * desired order. This order can have a significant effect on cache
6662 * performance.
6663 *
6664 * Currently the metadata lists are hit first, MFU then MRU, followed by
6665 * the data lists. This function returns a locked list, and also returns
6666 * the lock pointer.
6667 */
6668 static multilist_sublist_t *
6669 l2arc_sublist_lock(int list_num)
6670 {
6671 multilist_t *ml = NULL;
6672 unsigned int idx;
6673
6674 ASSERT(list_num >= 0 && list_num <= 3);
6675
6676 switch (list_num) {
6677 case 0:
6678 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
6679 break;
6680 case 1:
6681 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
6682 break;
6683 case 2:
6684 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
6685 break;
6686 case 3:
6687 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
6688 break;
6689 }
6690
6691 /*
6692 * Return a randomly-selected sublist. This is acceptable
6693 * because the caller feeds only a little bit of data for each
6694 * call (8MB). Subsequent calls will result in different
6695 * sublists being selected.
6696 */
6697 idx = multilist_get_random_index(ml);
6698 return (multilist_sublist_lock(ml, idx));
6699 }
6700
6701 /*
6702 * Evict buffers from the device write hand to the distance specified in
6703 * bytes. This distance may span populated buffers, it may span nothing.
6704 * This is clearing a region on the L2ARC device ready for writing.
6705 * If the 'all' boolean is set, every buffer is evicted.
6706 */
6707 static void
6708 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
6709 {
6710 list_t *buflist;
6711 arc_buf_hdr_t *hdr, *hdr_prev;
6712 kmutex_t *hash_lock;
6713 uint64_t taddr;
6714
6715 buflist = &dev->l2ad_buflist;
6716
6717 if (!all && dev->l2ad_first) {
6718 /*
6719 * This is the first sweep through the device. There is
6720 * nothing to evict.
6721 */
6722 return;
6723 }
6724
6725 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
6726 /*
6727 * When nearing the end of the device, evict to the end
6728 * before the device write hand jumps to the start.
6729 */
6730 taddr = dev->l2ad_end;
6731 } else {
6732 taddr = dev->l2ad_hand + distance;
6733 }
6734 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
6735 uint64_t, taddr, boolean_t, all);
6736
6737 top:
6738 mutex_enter(&dev->l2ad_mtx);
6739 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
6740 hdr_prev = list_prev(buflist, hdr);
6741
6742 hash_lock = HDR_LOCK(hdr);
6743
6744 /*
6745 * We cannot use mutex_enter or else we can deadlock
6746 * with l2arc_write_buffers (due to swapping the order
6747 * the hash lock and l2ad_mtx are taken).
6748 */
6749 if (!mutex_tryenter(hash_lock)) {
6750 /*
6751 * Missed the hash lock. Retry.
6752 */
6753 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
6754 mutex_exit(&dev->l2ad_mtx);
6755 mutex_enter(hash_lock);
6756 mutex_exit(hash_lock);
6757 goto top;
6758 }
6759
6760 /*
6761 * A header can't be on this list if it doesn't have L2 header.
6762 */
6763 ASSERT(HDR_HAS_L2HDR(hdr));
6764
6765 /* Ensure this header has finished being written. */
6766 ASSERT(!HDR_L2_WRITING(hdr));
6767 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
6768
6769 if (!all && (hdr->b_l2hdr.b_daddr >= taddr ||
6770 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
6771 /*
6772 * We've evicted to the target address,
6773 * or the end of the device.
6774 */
6775 mutex_exit(hash_lock);
6776 break;
6777 }
6778
6779 if (!HDR_HAS_L1HDR(hdr)) {
6780 ASSERT(!HDR_L2_READING(hdr));
6781 /*
6782 * This doesn't exist in the ARC. Destroy.
6783 * arc_hdr_destroy() will call list_remove()
6784 * and decrement arcstat_l2_lsize.
6785 */
6786 arc_change_state(arc_anon, hdr, hash_lock);
6787 arc_hdr_destroy(hdr);
6788 } else {
6789 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
6790 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
6791 /*
6792 * Invalidate issued or about to be issued
6793 * reads, since we may be about to write
6794 * over this location.
6795 */
6796 if (HDR_L2_READING(hdr)) {
6797 ARCSTAT_BUMP(arcstat_l2_evict_reading);
6798 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
6799 }
6800
6801 arc_hdr_l2hdr_destroy(hdr);
6802 }
6803 mutex_exit(hash_lock);
6804 }
6805 mutex_exit(&dev->l2ad_mtx);
6806 }
6807
6808 /*
6809 * Find and write ARC buffers to the L2ARC device.
6810 *
6811 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
6812 * for reading until they have completed writing.
6813 * The headroom_boost is an in-out parameter used to maintain headroom boost
6814 * state between calls to this function.
6815 *
6816 * Returns the number of bytes actually written (which may be smaller than
6817 * the delta by which the device hand has changed due to alignment).
6818 */
6819 static uint64_t
6820 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
6821 {
6822 arc_buf_hdr_t *hdr, *hdr_prev, *head;
6823 uint64_t write_asize, write_psize, write_lsize, headroom;
6824 boolean_t full;
6825 l2arc_write_callback_t *cb;
6826 zio_t *pio, *wzio;
6827 uint64_t guid = spa_load_guid(spa);
6828
6829 ASSERT3P(dev->l2ad_vdev, !=, NULL);
6830
6831 pio = NULL;
6832 write_lsize = write_asize = write_psize = 0;
6833 full = B_FALSE;
6834 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
6835 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
6836
6837 /*
6838 * Copy buffers for L2ARC writing.
6839 */
6840 for (int try = 0; try <= 3; try++) {
6841 multilist_sublist_t *mls = l2arc_sublist_lock(try);
6842 uint64_t passed_sz = 0;
6843
6844 /*
6845 * L2ARC fast warmup.
6846 *
6847 * Until the ARC is warm and starts to evict, read from the
6848 * head of the ARC lists rather than the tail.
6849 */
6850 if (arc_warm == B_FALSE)
6851 hdr = multilist_sublist_head(mls);
6852 else
6853 hdr = multilist_sublist_tail(mls);
6854
6855 headroom = target_sz * l2arc_headroom;
6856 if (zfs_compressed_arc_enabled)
6857 headroom = (headroom * l2arc_headroom_boost) / 100;
6858
6859 for (; hdr; hdr = hdr_prev) {
6860 kmutex_t *hash_lock;
6861
6862 if (arc_warm == B_FALSE)
6863 hdr_prev = multilist_sublist_next(mls, hdr);
6864 else
6865 hdr_prev = multilist_sublist_prev(mls, hdr);
6866
6867 hash_lock = HDR_LOCK(hdr);
6868 if (!mutex_tryenter(hash_lock)) {
6869 /*
6870 * Skip this buffer rather than waiting.
6871 */
6872 continue;
6873 }
6874
6875 passed_sz += HDR_GET_LSIZE(hdr);
6876 if (passed_sz > headroom) {
6877 /*
6878 * Searched too far.
6879 */
6880 mutex_exit(hash_lock);
6881 break;
6882 }
6883
6884 if (!l2arc_write_eligible(guid, hdr)) {
6885 mutex_exit(hash_lock);
6886 continue;
6887 }
6888
6889 /*
6890 * We rely on the L1 portion of the header below, so
6891 * it's invalid for this header to have been evicted out
6892 * of the ghost cache, prior to being written out. The
6893 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6894 */
6895 ASSERT(HDR_HAS_L1HDR(hdr));
6896
6897 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
6898 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6899 ASSERT3U(arc_hdr_size(hdr), >, 0);
6900 uint64_t psize = arc_hdr_size(hdr);
6901 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
6902 psize);
6903
6904 if ((write_asize + asize) > target_sz) {
6905 full = B_TRUE;
6906 mutex_exit(hash_lock);
6907 break;
6908 }
6909
6910 if (pio == NULL) {
6911 /*
6912 * Insert a dummy header on the buflist so
6913 * l2arc_write_done() can find where the
6914 * write buffers begin without searching.
6915 */
6916 mutex_enter(&dev->l2ad_mtx);
6917 list_insert_head(&dev->l2ad_buflist, head);
6918 mutex_exit(&dev->l2ad_mtx);
6919
6920 cb = kmem_alloc(
6921 sizeof (l2arc_write_callback_t), KM_SLEEP);
6922 cb->l2wcb_dev = dev;
6923 cb->l2wcb_head = head;
6924 pio = zio_root(spa, l2arc_write_done, cb,
6925 ZIO_FLAG_CANFAIL);
6926 }
6927
6928 hdr->b_l2hdr.b_dev = dev;
6929 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
6930 arc_hdr_set_flags(hdr,
6931 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
6932
6933 mutex_enter(&dev->l2ad_mtx);
6934 list_insert_head(&dev->l2ad_buflist, hdr);
6935 mutex_exit(&dev->l2ad_mtx);
6936
6937 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
6938
6939 /*
6940 * Normally the L2ARC can use the hdr's data, but if
6941 * we're sharing data between the hdr and one of its
6942 * bufs, L2ARC needs its own copy of the data so that
6943 * the ZIO below can't race with the buf consumer.
6944 * Another case where we need to create a copy of the
6945 * data is when the buffer size is not device-aligned
6946 * and we need to pad the block to make it such.
6947 * That also keeps the clock hand suitably aligned.
6948 *
6949 * To ensure that the copy will be available for the
6950 * lifetime of the ZIO and be cleaned up afterwards, we
6951 * add it to the l2arc_free_on_write queue.
6952 */
6953 abd_t *to_write;
6954 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
6955 to_write = hdr->b_l1hdr.b_pabd;
6956 } else {
6957 to_write = abd_alloc_for_io(asize,
6958 HDR_ISTYPE_METADATA(hdr));
6959 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
6960 if (asize != psize) {
6961 abd_zero_off(to_write, psize,
6962 asize - psize);
6963 }
6964 l2arc_free_abd_on_write(to_write, asize,
6965 arc_buf_type(hdr));
6966 }
6967 wzio = zio_write_phys(pio, dev->l2ad_vdev,
6968 hdr->b_l2hdr.b_daddr, asize, to_write,
6969 ZIO_CHECKSUM_OFF, NULL, hdr,
6970 ZIO_PRIORITY_ASYNC_WRITE,
6971 ZIO_FLAG_CANFAIL, B_FALSE);
6972
6973 write_lsize += HDR_GET_LSIZE(hdr);
6974 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
6975 zio_t *, wzio);
6976
6977 write_psize += psize;
6978 write_asize += asize;
6979 dev->l2ad_hand += asize;
6980
6981 mutex_exit(hash_lock);
6982
6983 (void) zio_nowait(wzio);
6984 }
6985
6986 multilist_sublist_unlock(mls);
6987
6988 if (full == B_TRUE)
6989 break;
6990 }
6991
6992 /* No buffers selected for writing? */
6993 if (pio == NULL) {
6994 ASSERT0(write_lsize);
6995 ASSERT(!HDR_HAS_L1HDR(head));
6996 kmem_cache_free(hdr_l2only_cache, head);
6997 return (0);
6998 }
6999
7000 ASSERT3U(write_asize, <=, target_sz);
7001 ARCSTAT_BUMP(arcstat_l2_writes_sent);
7002 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
7003 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
7004 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
7005 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
7006
7007 /*
7008 * Bump device hand to the device start if it is approaching the end.
7009 * l2arc_evict() will already have evicted ahead for this case.
7010 */
7011 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7012 dev->l2ad_hand = dev->l2ad_start;
7013 dev->l2ad_first = B_FALSE;
7014 }
7015
7016 dev->l2ad_writing = B_TRUE;
7017 (void) zio_wait(pio);
7018 dev->l2ad_writing = B_FALSE;
7019
7020 return (write_asize);
7021 }
7022
7023 /*
7024 * This thread feeds the L2ARC at regular intervals. This is the beating
7025 * heart of the L2ARC.
7026 */
7027 /* ARGSUSED */
7028 static void
7029 l2arc_feed_thread(void *unused)
7030 {
7031 callb_cpr_t cpr;
7032 l2arc_dev_t *dev;
7033 spa_t *spa;
7034 uint64_t size, wrote;
7035 clock_t begin, next = ddi_get_lbolt();
7036
7037 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7038
7039 mutex_enter(&l2arc_feed_thr_lock);
7040
7041 while (l2arc_thread_exit == 0) {
7042 CALLB_CPR_SAFE_BEGIN(&cpr);
7043 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7044 next);
7045 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7046 next = ddi_get_lbolt() + hz;
7047
7048 /*
7049 * Quick check for L2ARC devices.
7050 */
7051 mutex_enter(&l2arc_dev_mtx);
7052 if (l2arc_ndev == 0) {
7053 mutex_exit(&l2arc_dev_mtx);
7054 continue;
7055 }
7056 mutex_exit(&l2arc_dev_mtx);
7057 begin = ddi_get_lbolt();
7058
7059 /*
7060 * This selects the next l2arc device to write to, and in
7061 * doing so the next spa to feed from: dev->l2ad_spa. This
7062 * will return NULL if there are now no l2arc devices or if
7063 * they are all faulted.
7064 *
7065 * If a device is returned, its spa's config lock is also
7066 * held to prevent device removal. l2arc_dev_get_next()
7067 * will grab and release l2arc_dev_mtx.
7068 */
7069 if ((dev = l2arc_dev_get_next()) == NULL)
7070 continue;
7071
7072 spa = dev->l2ad_spa;
7073 ASSERT3P(spa, !=, NULL);
7074
7075 /*
7076 * If the pool is read-only then force the feed thread to
7077 * sleep a little longer.
7078 */
7079 if (!spa_writeable(spa)) {
7080 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7081 spa_config_exit(spa, SCL_L2ARC, dev);
7082 continue;
7083 }
7084
7085 /*
7086 * Avoid contributing to memory pressure.
7087 */
7088 if (arc_reclaim_needed()) {
7089 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7090 spa_config_exit(spa, SCL_L2ARC, dev);
7091 continue;
7092 }
7093
7094 ARCSTAT_BUMP(arcstat_l2_feeds);
7095
7096 size = l2arc_write_size();
7097
7098 /*
7099 * Evict L2ARC buffers that will be overwritten.
7100 */
7101 l2arc_evict(dev, size, B_FALSE);
7102
7103 /*
7104 * Write ARC buffers.
7105 */
7106 wrote = l2arc_write_buffers(spa, dev, size);
7107
7108 /*
7109 * Calculate interval between writes.
7110 */
7111 next = l2arc_write_interval(begin, size, wrote);
7112 spa_config_exit(spa, SCL_L2ARC, dev);
7113 }
7114
7115 l2arc_thread_exit = 0;
7116 cv_broadcast(&l2arc_feed_thr_cv);
7117 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7118 thread_exit();
7119 }
7120
7121 boolean_t
7122 l2arc_vdev_present(vdev_t *vd)
7123 {
7124 l2arc_dev_t *dev;
7125
7126 mutex_enter(&l2arc_dev_mtx);
7127 for (dev = list_head(l2arc_dev_list); dev != NULL;
7128 dev = list_next(l2arc_dev_list, dev)) {
7129 if (dev->l2ad_vdev == vd)
7130 break;
7131 }
7132 mutex_exit(&l2arc_dev_mtx);
7133
7134 return (dev != NULL);
7135 }
7136
7137 /*
7138 * Add a vdev for use by the L2ARC. By this point the spa has already
7139 * validated the vdev and opened it.
7140 */
7141 void
7142 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7143 {
7144 l2arc_dev_t *adddev;
7145
7146 ASSERT(!l2arc_vdev_present(vd));
7147
7148 /*
7149 * Create a new l2arc device entry.
7150 */
7151 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7152 adddev->l2ad_spa = spa;
7153 adddev->l2ad_vdev = vd;
7154 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7155 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7156 adddev->l2ad_hand = adddev->l2ad_start;
7157 adddev->l2ad_first = B_TRUE;
7158 adddev->l2ad_writing = B_FALSE;
7159
7160 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7161 /*
7162 * This is a list of all ARC buffers that are still valid on the
7163 * device.
7164 */
7165 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7166 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7167
7168 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7169 refcount_create(&adddev->l2ad_alloc);
7170
7171 /*
7172 * Add device to global list
7173 */
7174 mutex_enter(&l2arc_dev_mtx);
7175 list_insert_head(l2arc_dev_list, adddev);
7176 atomic_inc_64(&l2arc_ndev);
7177 mutex_exit(&l2arc_dev_mtx);
7178 }
7179
7180 /*
7181 * Remove a vdev from the L2ARC.
7182 */
7183 void
7184 l2arc_remove_vdev(vdev_t *vd)
7185 {
7186 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7187
7188 /*
7189 * Find the device by vdev
7190 */
7191 mutex_enter(&l2arc_dev_mtx);
7192 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7193 nextdev = list_next(l2arc_dev_list, dev);
7194 if (vd == dev->l2ad_vdev) {
7195 remdev = dev;
7196 break;
7197 }
7198 }
7199 ASSERT3P(remdev, !=, NULL);
7200
7201 /*
7202 * Remove device from global list
7203 */
7204 list_remove(l2arc_dev_list, remdev);
7205 l2arc_dev_last = NULL; /* may have been invalidated */
7206 atomic_dec_64(&l2arc_ndev);
7207 mutex_exit(&l2arc_dev_mtx);
7208
7209 /*
7210 * Clear all buflists and ARC references. L2ARC device flush.
7211 */
7212 l2arc_evict(remdev, 0, B_TRUE);
7213 list_destroy(&remdev->l2ad_buflist);
7214 mutex_destroy(&remdev->l2ad_mtx);
7215 refcount_destroy(&remdev->l2ad_alloc);
7216 kmem_free(remdev, sizeof (l2arc_dev_t));
7217 }
7218
7219 void
7220 l2arc_init(void)
7221 {
7222 l2arc_thread_exit = 0;
7223 l2arc_ndev = 0;
7224 l2arc_writes_sent = 0;
7225 l2arc_writes_done = 0;
7226
7227 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7228 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7229 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7230 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7231
7232 l2arc_dev_list = &L2ARC_dev_list;
7233 l2arc_free_on_write = &L2ARC_free_on_write;
7234 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7235 offsetof(l2arc_dev_t, l2ad_node));
7236 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7237 offsetof(l2arc_data_free_t, l2df_list_node));
7238 }
7239
7240 void
7241 l2arc_fini(void)
7242 {
7243 /*
7244 * This is called from dmu_fini(), which is called from spa_fini();
7245 * Because of this, we can assume that all l2arc devices have
7246 * already been removed when the pools themselves were removed.
7247 */
7248
7249 l2arc_do_free_on_write();
7250
7251 mutex_destroy(&l2arc_feed_thr_lock);
7252 cv_destroy(&l2arc_feed_thr_cv);
7253 mutex_destroy(&l2arc_dev_mtx);
7254 mutex_destroy(&l2arc_free_on_write_mtx);
7255
7256 list_destroy(l2arc_dev_list);
7257 list_destroy(l2arc_free_on_write);
7258 }
7259
7260 void
7261 l2arc_start(void)
7262 {
7263 if (!(spa_mode_global & FWRITE))
7264 return;
7265
7266 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7267 TS_RUN, minclsyspri);
7268 }
7269
7270 void
7271 l2arc_stop(void)
7272 {
7273 if (!(spa_mode_global & FWRITE))
7274 return;
7275
7276 mutex_enter(&l2arc_feed_thr_lock);
7277 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7278 l2arc_thread_exit = 1;
7279 while (l2arc_thread_exit != 0)
7280 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7281 mutex_exit(&l2arc_feed_thr_lock);
7282 }