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 https://opensource.org/licenses/CDDL-1.0.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2018, Joyent, Inc.
24 * Copyright (c) 2011, 2020, Delphix. All rights reserved.
25 * Copyright (c) 2014, Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2017, Nexenta Systems, Inc. All rights reserved.
27 * Copyright (c) 2019, loli10K <ezomori.nozomu@gmail.com>. All rights reserved.
28 * Copyright (c) 2020, George Amanakis. All rights reserved.
29 * Copyright (c) 2019, 2023, Klara Inc.
30 * Copyright (c) 2019, Allan Jude
31 * Copyright (c) 2020, The FreeBSD Foundation [1]
32 * Copyright (c) 2021, 2024 by George Melikov. All rights reserved.
33 *
34 * [1] Portions of this software were developed by Allan Jude
35 * under sponsorship from the FreeBSD Foundation.
36 */
37
38 /*
39 * DVA-based Adjustable Replacement Cache
40 *
41 * While much of the theory of operation used here is
42 * based on the self-tuning, low overhead replacement cache
43 * presented by Megiddo and Modha at FAST 2003, there are some
44 * significant differences:
45 *
46 * 1. The Megiddo and Modha model assumes any page is evictable.
47 * Pages in its cache cannot be "locked" into memory. This makes
48 * the eviction algorithm simple: evict the last page in the list.
49 * This also make the performance characteristics easy to reason
50 * about. Our cache is not so simple. At any given moment, some
51 * subset of the blocks in the cache are un-evictable because we
52 * have handed out a reference to them. Blocks are only evictable
53 * when there are no external references active. This makes
54 * eviction far more problematic: we choose to evict the evictable
55 * blocks that are the "lowest" in the list.
56 *
57 * There are times when it is not possible to evict the requested
58 * space. In these circumstances we are unable to adjust the cache
59 * size. To prevent the cache growing unbounded at these times we
60 * implement a "cache throttle" that slows the flow of new data
61 * into the cache until we can make space available.
62 *
63 * 2. The Megiddo and Modha model assumes a fixed cache size.
64 * Pages are evicted when the cache is full and there is a cache
65 * miss. Our model has a variable sized cache. It grows with
66 * high use, but also tries to react to memory pressure from the
67 * operating system: decreasing its size when system memory is
68 * tight.
69 *
70 * 3. The Megiddo and Modha model assumes a fixed page size. All
71 * elements of the cache are therefore exactly the same size. So
72 * when adjusting the cache size following a cache miss, its simply
73 * a matter of choosing a single page to evict. In our model, we
74 * have variable sized cache blocks (ranging from 512 bytes to
75 * 128K bytes). We therefore choose a set of blocks to evict to make
76 * space for a cache miss that approximates as closely as possible
77 * the space used by the new block.
78 *
79 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
80 * by N. Megiddo & D. Modha, FAST 2003
81 */
82
83 /*
84 * The locking model:
85 *
86 * A new reference to a cache buffer can be obtained in two
87 * ways: 1) via a hash table lookup using the DVA as a key,
88 * or 2) via one of the ARC lists. The arc_read() interface
89 * uses method 1, while the internal ARC algorithms for
90 * adjusting the cache use method 2. We therefore provide two
91 * types of locks: 1) the hash table lock array, and 2) the
92 * ARC list locks.
93 *
94 * Buffers do not have their own mutexes, rather they rely on the
95 * hash table mutexes for the bulk of their protection (i.e. most
96 * fields in the arc_buf_hdr_t are protected by these mutexes).
97 *
98 * buf_hash_find() returns the appropriate mutex (held) when it
99 * locates the requested buffer in the hash table. It returns
100 * NULL for the mutex if the buffer was not in the table.
101 *
102 * buf_hash_remove() expects the appropriate hash mutex to be
103 * already held before it is invoked.
104 *
105 * Each ARC state also has a mutex which is used to protect the
106 * buffer list associated with the state. When attempting to
107 * obtain a hash table lock while holding an ARC list lock you
108 * must use: mutex_tryenter() to avoid deadlock. Also note that
109 * the active state mutex must be held before the ghost state mutex.
110 *
111 * It as also possible to register a callback which is run when the
112 * metadata limit is reached and no buffers can be safely evicted. In
113 * this case the arc user should drop a reference on some arc buffers so
114 * they can be reclaimed. For example, when using the ZPL each dentry
115 * holds a references on a znode. These dentries must be pruned before
116 * the arc buffer holding the znode can be safely evicted.
117 *
118 * Note that the majority of the performance stats are manipulated
119 * with atomic operations.
120 *
121 * The L2ARC uses the l2ad_mtx on each vdev for the following:
122 *
123 * - L2ARC buflist creation
124 * - L2ARC buflist eviction
125 * - L2ARC write completion, which walks L2ARC buflists
126 * - ARC header destruction, as it removes from L2ARC buflists
127 * - ARC header release, as it removes from L2ARC buflists
128 */
129
130 /*
131 * ARC operation:
132 *
133 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
134 * This structure can point either to a block that is still in the cache or to
135 * one that is only accessible in an L2 ARC device, or it can provide
136 * information about a block that was recently evicted. If a block is
137 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
138 * information to retrieve it from the L2ARC device. This information is
139 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
140 * that is in this state cannot access the data directly.
141 *
142 * Blocks that are actively being referenced or have not been evicted
143 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
144 * the arc_buf_hdr_t that will point to the data block in memory. A block can
145 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
146 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
147 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
148 *
149 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
150 * ability to store the physical data (b_pabd) associated with the DVA of the
151 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
152 * it will match its on-disk compression characteristics. This behavior can be
153 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
154 * compressed ARC functionality is disabled, the b_pabd will point to an
155 * uncompressed version of the on-disk data.
156 *
157 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
158 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
159 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
160 * consumer. The ARC will provide references to this data and will keep it
161 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
162 * data block and will evict any arc_buf_t that is no longer referenced. The
163 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
164 * "overhead_size" kstat.
165 *
166 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
167 * compressed form. The typical case is that consumers will want uncompressed
168 * data, and when that happens a new data buffer is allocated where the data is
169 * decompressed for them to use. Currently the only consumer who wants
170 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
171 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
172 * with the arc_buf_hdr_t.
173 *
174 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
175 * first one is owned by a compressed send consumer (and therefore references
176 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
177 * used by any other consumer (and has its own uncompressed copy of the data
178 * buffer).
179 *
180 * arc_buf_hdr_t
181 * +-----------+
182 * | fields |
183 * | common to |
184 * | L1- and |
185 * | L2ARC |
186 * +-----------+
187 * | l2arc_buf_hdr_t
188 * | |
189 * +-----------+
190 * | l1arc_buf_hdr_t
191 * | | arc_buf_t
192 * | b_buf +------------>+-----------+ arc_buf_t
193 * | b_pabd +-+ |b_next +---->+-----------+
194 * +-----------+ | |-----------| |b_next +-->NULL
195 * | |b_comp = T | +-----------+
196 * | |b_data +-+ |b_comp = F |
197 * | +-----------+ | |b_data +-+
198 * +->+------+ | +-----------+ |
199 * compressed | | | |
200 * data | |<--------------+ | uncompressed
201 * +------+ compressed, | data
202 * shared +-->+------+
203 * data | |
204 * | |
205 * +------+
206 *
207 * When a consumer reads a block, the ARC must first look to see if the
208 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
209 * arc_buf_t and either copies uncompressed data into a new data buffer from an
210 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
211 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
212 * hdr is compressed and the desired compression characteristics of the
213 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
214 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
215 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
216 * be anywhere in the hdr's list.
217 *
218 * The diagram below shows an example of an uncompressed ARC hdr that is
219 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
220 * the last element in the buf list):
221 *
222 * arc_buf_hdr_t
223 * +-----------+
224 * | |
225 * | |
226 * | |
227 * +-----------+
228 * l2arc_buf_hdr_t| |
229 * | |
230 * +-----------+
231 * l1arc_buf_hdr_t| |
232 * | | arc_buf_t (shared)
233 * | b_buf +------------>+---------+ arc_buf_t
234 * | | |b_next +---->+---------+
235 * | b_pabd +-+ |---------| |b_next +-->NULL
236 * +-----------+ | | | +---------+
237 * | |b_data +-+ | |
238 * | +---------+ | |b_data +-+
239 * +->+------+ | +---------+ |
240 * | | | |
241 * uncompressed | | | |
242 * data +------+ | |
243 * ^ +->+------+ |
244 * | uncompressed | | |
245 * | data | | |
246 * | +------+ |
247 * +---------------------------------+
248 *
249 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
250 * since the physical block is about to be rewritten. The new data contents
251 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
252 * it may compress the data before writing it to disk. The ARC will be called
253 * with the transformed data and will memcpy the transformed on-disk block into
254 * a newly allocated b_pabd. Writes are always done into buffers which have
255 * either been loaned (and hence are new and don't have other readers) or
256 * buffers which have been released (and hence have their own hdr, if there
257 * were originally other readers of the buf's original hdr). This ensures that
258 * the ARC only needs to update a single buf and its hdr after a write occurs.
259 *
260 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
261 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
262 * that when compressed ARC is enabled that the L2ARC blocks are identical
263 * to the on-disk block in the main data pool. This provides a significant
264 * advantage since the ARC can leverage the bp's checksum when reading from the
265 * L2ARC to determine if the contents are valid. However, if the compressed
266 * ARC is disabled, then the L2ARC's block must be transformed to look
267 * like the physical block in the main data pool before comparing the
268 * checksum and determining its validity.
269 *
270 * The L1ARC has a slightly different system for storing encrypted data.
271 * Raw (encrypted + possibly compressed) data has a few subtle differences from
272 * data that is just compressed. The biggest difference is that it is not
273 * possible to decrypt encrypted data (or vice-versa) if the keys aren't loaded.
274 * The other difference is that encryption cannot be treated as a suggestion.
275 * If a caller would prefer compressed data, but they actually wind up with
276 * uncompressed data the worst thing that could happen is there might be a
277 * performance hit. If the caller requests encrypted data, however, we must be
278 * sure they actually get it or else secret information could be leaked. Raw
279 * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore,
280 * may have both an encrypted version and a decrypted version of its data at
281 * once. When a caller needs a raw arc_buf_t, it is allocated and the data is
282 * copied out of this header. To avoid complications with b_pabd, raw buffers
283 * cannot be shared.
284 */
285
286 #include <sys/spa.h>
287 #include <sys/zio.h>
288 #include <sys/spa_impl.h>
289 #include <sys/zio_compress.h>
290 #include <sys/zio_checksum.h>
291 #include <sys/zfs_context.h>
292 #include <sys/arc.h>
293 #include <sys/zfs_refcount.h>
294 #include <sys/vdev.h>
295 #include <sys/vdev_impl.h>
296 #include <sys/dsl_pool.h>
297 #include <sys/multilist.h>
298 #include <sys/abd.h>
299 #include <sys/zil.h>
300 #include <sys/fm/fs/zfs.h>
301 #include <sys/callb.h>
302 #include <sys/kstat.h>
303 #include <sys/zthr.h>
304 #include <zfs_fletcher.h>
305 #include <sys/arc_impl.h>
306 #include <sys/trace_zfs.h>
307 #include <sys/aggsum.h>
308 #include <sys/wmsum.h>
309 #include <cityhash.h>
310 #include <sys/vdev_trim.h>
311 #include <sys/zfs_racct.h>
312 #include <sys/zstd/zstd.h>
313
314 #ifndef _KERNEL
315 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
316 boolean_t arc_watch = B_FALSE;
317 #endif
318
319 /*
320 * This thread's job is to keep enough free memory in the system, by
321 * calling arc_kmem_reap_soon() plus arc_reduce_target_size(), which improves
322 * arc_available_memory().
323 */
324 static zthr_t *arc_reap_zthr;
325
326 /*
327 * This thread's job is to keep arc_size under arc_c, by calling
328 * arc_evict(), which improves arc_is_overflowing().
329 */
330 static zthr_t *arc_evict_zthr;
331 static arc_buf_hdr_t **arc_state_evict_markers;
332 static int arc_state_evict_marker_count;
333
334 static kmutex_t arc_evict_lock;
335 static boolean_t arc_evict_needed = B_FALSE;
336 static clock_t arc_last_uncached_flush;
337
338 /*
339 * Count of bytes evicted since boot.
340 */
341 static uint64_t arc_evict_count;
342
343 /*
344 * List of arc_evict_waiter_t's, representing threads waiting for the
345 * arc_evict_count to reach specific values.
346 */
347 static list_t arc_evict_waiters;
348
349 /*
350 * When arc_is_overflowing(), arc_get_data_impl() waits for this percent of
351 * the requested amount of data to be evicted. For example, by default for
352 * every 2KB that's evicted, 1KB of it may be "reused" by a new allocation.
353 * Since this is above 100%, it ensures that progress is made towards getting
354 * arc_size under arc_c. Since this is finite, it ensures that allocations
355 * can still happen, even during the potentially long time that arc_size is
356 * more than arc_c.
357 */
358 static uint_t zfs_arc_eviction_pct = 200;
359
360 /*
361 * The number of headers to evict in arc_evict_state_impl() before
362 * dropping the sublist lock and evicting from another sublist. A lower
363 * value means we're more likely to evict the "correct" header (i.e. the
364 * oldest header in the arc state), but comes with higher overhead
365 * (i.e. more invocations of arc_evict_state_impl()).
366 */
367 static uint_t zfs_arc_evict_batch_limit = 10;
368
369 /* number of seconds before growing cache again */
370 uint_t arc_grow_retry = 5;
371
372 /*
373 * Minimum time between calls to arc_kmem_reap_soon().
374 */
375 static const int arc_kmem_cache_reap_retry_ms = 1000;
376
377 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
378 static int zfs_arc_overflow_shift = 8;
379
380 /* log2(fraction of arc to reclaim) */
381 uint_t arc_shrink_shift = 7;
382
383 /* percent of pagecache to reclaim arc to */
384 #ifdef _KERNEL
385 uint_t zfs_arc_pc_percent = 0;
386 #endif
387
388 /*
389 * log2(fraction of ARC which must be free to allow growing).
390 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
391 * when reading a new block into the ARC, we will evict an equal-sized block
392 * from the ARC.
393 *
394 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
395 * we will still not allow it to grow.
396 */
397 uint_t arc_no_grow_shift = 5;
398
399
400 /*
401 * minimum lifespan of a prefetch block in clock ticks
402 * (initialized in arc_init())
403 */
404 static uint_t arc_min_prefetch_ms;
405 static uint_t arc_min_prescient_prefetch_ms;
406
407 /*
408 * If this percent of memory is free, don't throttle.
409 */
410 uint_t arc_lotsfree_percent = 10;
411
412 /*
413 * The arc has filled available memory and has now warmed up.
414 */
415 boolean_t arc_warm;
416
417 /*
418 * These tunables are for performance analysis.
419 */
420 uint64_t zfs_arc_max = 0;
421 uint64_t zfs_arc_min = 0;
422 static uint64_t zfs_arc_dnode_limit = 0;
423 static uint_t zfs_arc_dnode_reduce_percent = 10;
424 static uint_t zfs_arc_grow_retry = 0;
425 static uint_t zfs_arc_shrink_shift = 0;
426 uint_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
427
428 /*
429 * ARC dirty data constraints for arc_tempreserve_space() throttle:
430 * * total dirty data limit
431 * * anon block dirty limit
432 * * each pool's anon allowance
433 */
434 static const unsigned long zfs_arc_dirty_limit_percent = 50;
435 static const unsigned long zfs_arc_anon_limit_percent = 25;
436 static const unsigned long zfs_arc_pool_dirty_percent = 20;
437
438 /*
439 * Enable or disable compressed arc buffers.
440 */
441 int zfs_compressed_arc_enabled = B_TRUE;
442
443 /*
444 * Balance between metadata and data on ghost hits. Values above 100
445 * increase metadata caching by proportionally reducing effect of ghost
446 * data hits on target data/metadata rate.
447 */
448 static uint_t zfs_arc_meta_balance = 500;
449
450 /*
451 * Percentage that can be consumed by dnodes of ARC meta buffers.
452 */
453 static uint_t zfs_arc_dnode_limit_percent = 10;
454
455 /*
456 * These tunables are Linux-specific
457 */
458 static uint64_t zfs_arc_sys_free = 0;
459 static uint_t zfs_arc_min_prefetch_ms = 0;
460 static uint_t zfs_arc_min_prescient_prefetch_ms = 0;
461 static uint_t zfs_arc_lotsfree_percent = 10;
462
463 /*
464 * Number of arc_prune threads
465 */
466 static int zfs_arc_prune_task_threads = 1;
467
468 /* The 7 states: */
469 arc_state_t ARC_anon;
470 arc_state_t ARC_mru;
471 arc_state_t ARC_mru_ghost;
472 arc_state_t ARC_mfu;
473 arc_state_t ARC_mfu_ghost;
474 arc_state_t ARC_l2c_only;
475 arc_state_t ARC_uncached;
476
477 arc_stats_t arc_stats = {
478 { "hits", KSTAT_DATA_UINT64 },
479 { "iohits", KSTAT_DATA_UINT64 },
480 { "misses", KSTAT_DATA_UINT64 },
481 { "demand_data_hits", KSTAT_DATA_UINT64 },
482 { "demand_data_iohits", KSTAT_DATA_UINT64 },
483 { "demand_data_misses", KSTAT_DATA_UINT64 },
484 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
485 { "demand_metadata_iohits", KSTAT_DATA_UINT64 },
486 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
487 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
488 { "prefetch_data_iohits", KSTAT_DATA_UINT64 },
489 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
490 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
491 { "prefetch_metadata_iohits", KSTAT_DATA_UINT64 },
492 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
493 { "mru_hits", KSTAT_DATA_UINT64 },
494 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
495 { "mfu_hits", KSTAT_DATA_UINT64 },
496 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
497 { "uncached_hits", KSTAT_DATA_UINT64 },
498 { "deleted", KSTAT_DATA_UINT64 },
499 { "mutex_miss", KSTAT_DATA_UINT64 },
500 { "access_skip", KSTAT_DATA_UINT64 },
501 { "evict_skip", KSTAT_DATA_UINT64 },
502 { "evict_not_enough", KSTAT_DATA_UINT64 },
503 { "evict_l2_cached", KSTAT_DATA_UINT64 },
504 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
505 { "evict_l2_eligible_mfu", KSTAT_DATA_UINT64 },
506 { "evict_l2_eligible_mru", KSTAT_DATA_UINT64 },
507 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
508 { "evict_l2_skip", KSTAT_DATA_UINT64 },
509 { "hash_elements", KSTAT_DATA_UINT64 },
510 { "hash_elements_max", KSTAT_DATA_UINT64 },
511 { "hash_collisions", KSTAT_DATA_UINT64 },
512 { "hash_chains", KSTAT_DATA_UINT64 },
513 { "hash_chain_max", KSTAT_DATA_UINT64 },
514 { "meta", KSTAT_DATA_UINT64 },
515 { "pd", KSTAT_DATA_UINT64 },
516 { "pm", KSTAT_DATA_UINT64 },
517 { "c", KSTAT_DATA_UINT64 },
518 { "c_min", KSTAT_DATA_UINT64 },
519 { "c_max", KSTAT_DATA_UINT64 },
520 { "size", KSTAT_DATA_UINT64 },
521 { "compressed_size", KSTAT_DATA_UINT64 },
522 { "uncompressed_size", KSTAT_DATA_UINT64 },
523 { "overhead_size", KSTAT_DATA_UINT64 },
524 { "hdr_size", KSTAT_DATA_UINT64 },
525 { "data_size", KSTAT_DATA_UINT64 },
526 { "metadata_size", KSTAT_DATA_UINT64 },
527 { "dbuf_size", KSTAT_DATA_UINT64 },
528 { "dnode_size", KSTAT_DATA_UINT64 },
529 { "bonus_size", KSTAT_DATA_UINT64 },
530 #if defined(COMPAT_FREEBSD11)
531 { "other_size", KSTAT_DATA_UINT64 },
532 #endif
533 { "anon_size", KSTAT_DATA_UINT64 },
534 { "anon_data", KSTAT_DATA_UINT64 },
535 { "anon_metadata", KSTAT_DATA_UINT64 },
536 { "anon_evictable_data", KSTAT_DATA_UINT64 },
537 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
538 { "mru_size", KSTAT_DATA_UINT64 },
539 { "mru_data", KSTAT_DATA_UINT64 },
540 { "mru_metadata", KSTAT_DATA_UINT64 },
541 { "mru_evictable_data", KSTAT_DATA_UINT64 },
542 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
543 { "mru_ghost_size", KSTAT_DATA_UINT64 },
544 { "mru_ghost_data", KSTAT_DATA_UINT64 },
545 { "mru_ghost_metadata", KSTAT_DATA_UINT64 },
546 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
547 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
548 { "mfu_size", KSTAT_DATA_UINT64 },
549 { "mfu_data", KSTAT_DATA_UINT64 },
550 { "mfu_metadata", KSTAT_DATA_UINT64 },
551 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
552 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
553 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
554 { "mfu_ghost_data", KSTAT_DATA_UINT64 },
555 { "mfu_ghost_metadata", KSTAT_DATA_UINT64 },
556 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
557 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
558 { "uncached_size", KSTAT_DATA_UINT64 },
559 { "uncached_data", KSTAT_DATA_UINT64 },
560 { "uncached_metadata", KSTAT_DATA_UINT64 },
561 { "uncached_evictable_data", KSTAT_DATA_UINT64 },
562 { "uncached_evictable_metadata", KSTAT_DATA_UINT64 },
563 { "l2_hits", KSTAT_DATA_UINT64 },
564 { "l2_misses", KSTAT_DATA_UINT64 },
565 { "l2_prefetch_asize", KSTAT_DATA_UINT64 },
566 { "l2_mru_asize", KSTAT_DATA_UINT64 },
567 { "l2_mfu_asize", KSTAT_DATA_UINT64 },
568 { "l2_bufc_data_asize", KSTAT_DATA_UINT64 },
569 { "l2_bufc_metadata_asize", KSTAT_DATA_UINT64 },
570 { "l2_feeds", KSTAT_DATA_UINT64 },
571 { "l2_rw_clash", KSTAT_DATA_UINT64 },
572 { "l2_read_bytes", KSTAT_DATA_UINT64 },
573 { "l2_write_bytes", KSTAT_DATA_UINT64 },
574 { "l2_writes_sent", KSTAT_DATA_UINT64 },
575 { "l2_writes_done", KSTAT_DATA_UINT64 },
576 { "l2_writes_error", KSTAT_DATA_UINT64 },
577 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
578 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
579 { "l2_evict_reading", KSTAT_DATA_UINT64 },
580 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
581 { "l2_free_on_write", KSTAT_DATA_UINT64 },
582 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
583 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
584 { "l2_io_error", KSTAT_DATA_UINT64 },
585 { "l2_size", KSTAT_DATA_UINT64 },
586 { "l2_asize", KSTAT_DATA_UINT64 },
587 { "l2_hdr_size", KSTAT_DATA_UINT64 },
588 { "l2_log_blk_writes", KSTAT_DATA_UINT64 },
589 { "l2_log_blk_avg_asize", KSTAT_DATA_UINT64 },
590 { "l2_log_blk_asize", KSTAT_DATA_UINT64 },
591 { "l2_log_blk_count", KSTAT_DATA_UINT64 },
592 { "l2_data_to_meta_ratio", KSTAT_DATA_UINT64 },
593 { "l2_rebuild_success", KSTAT_DATA_UINT64 },
594 { "l2_rebuild_unsupported", KSTAT_DATA_UINT64 },
595 { "l2_rebuild_io_errors", KSTAT_DATA_UINT64 },
596 { "l2_rebuild_dh_errors", KSTAT_DATA_UINT64 },
597 { "l2_rebuild_cksum_lb_errors", KSTAT_DATA_UINT64 },
598 { "l2_rebuild_lowmem", KSTAT_DATA_UINT64 },
599 { "l2_rebuild_size", KSTAT_DATA_UINT64 },
600 { "l2_rebuild_asize", KSTAT_DATA_UINT64 },
601 { "l2_rebuild_bufs", KSTAT_DATA_UINT64 },
602 { "l2_rebuild_bufs_precached", KSTAT_DATA_UINT64 },
603 { "l2_rebuild_log_blks", KSTAT_DATA_UINT64 },
604 { "memory_throttle_count", KSTAT_DATA_UINT64 },
605 { "memory_direct_count", KSTAT_DATA_UINT64 },
606 { "memory_indirect_count", KSTAT_DATA_UINT64 },
607 { "memory_all_bytes", KSTAT_DATA_UINT64 },
608 { "memory_free_bytes", KSTAT_DATA_UINT64 },
609 { "memory_available_bytes", KSTAT_DATA_INT64 },
610 { "arc_no_grow", KSTAT_DATA_UINT64 },
611 { "arc_tempreserve", KSTAT_DATA_UINT64 },
612 { "arc_loaned_bytes", KSTAT_DATA_UINT64 },
613 { "arc_prune", KSTAT_DATA_UINT64 },
614 { "arc_meta_used", KSTAT_DATA_UINT64 },
615 { "arc_dnode_limit", KSTAT_DATA_UINT64 },
616 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
617 { "predictive_prefetch", KSTAT_DATA_UINT64 },
618 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
619 { "demand_iohit_predictive_prefetch", KSTAT_DATA_UINT64 },
620 { "prescient_prefetch", KSTAT_DATA_UINT64 },
621 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
622 { "demand_iohit_prescient_prefetch", KSTAT_DATA_UINT64 },
623 { "arc_need_free", KSTAT_DATA_UINT64 },
624 { "arc_sys_free", KSTAT_DATA_UINT64 },
625 { "arc_raw_size", KSTAT_DATA_UINT64 },
626 { "cached_only_in_progress", KSTAT_DATA_UINT64 },
627 { "abd_chunk_waste_size", KSTAT_DATA_UINT64 },
628 };
629
630 arc_sums_t arc_sums;
631
632 #define ARCSTAT_MAX(stat, val) { \
633 uint64_t m; \
634 while ((val) > (m = arc_stats.stat.value.ui64) && \
635 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
636 continue; \
637 }
638
639 /*
640 * We define a macro to allow ARC hits/misses to be easily broken down by
641 * two separate conditions, giving a total of four different subtypes for
642 * each of hits and misses (so eight statistics total).
643 */
644 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
645 if (cond1) { \
646 if (cond2) { \
647 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
648 } else { \
649 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
650 } \
651 } else { \
652 if (cond2) { \
653 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
654 } else { \
655 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
656 } \
657 }
658
659 /*
660 * This macro allows us to use kstats as floating averages. Each time we
661 * update this kstat, we first factor it and the update value by
662 * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall
663 * average. This macro assumes that integer loads and stores are atomic, but
664 * is not safe for multiple writers updating the kstat in parallel (only the
665 * last writer's update will remain).
666 */
667 #define ARCSTAT_F_AVG_FACTOR 3
668 #define ARCSTAT_F_AVG(stat, value) \
669 do { \
670 uint64_t x = ARCSTAT(stat); \
671 x = x - x / ARCSTAT_F_AVG_FACTOR + \
672 (value) / ARCSTAT_F_AVG_FACTOR; \
673 ARCSTAT(stat) = x; \
674 } while (0)
675
676 static kstat_t *arc_ksp;
677
678 /*
679 * There are several ARC variables that are critical to export as kstats --
680 * but we don't want to have to grovel around in the kstat whenever we wish to
681 * manipulate them. For these variables, we therefore define them to be in
682 * terms of the statistic variable. This assures that we are not introducing
683 * the possibility of inconsistency by having shadow copies of the variables,
684 * while still allowing the code to be readable.
685 */
686 #define arc_tempreserve ARCSTAT(arcstat_tempreserve)
687 #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes)
688 #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */
689 #define arc_need_free ARCSTAT(arcstat_need_free) /* waiting to be evicted */
690
691 hrtime_t arc_growtime;
692 list_t arc_prune_list;
693 kmutex_t arc_prune_mtx;
694 taskq_t *arc_prune_taskq;
695
696 #define GHOST_STATE(state) \
697 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
698 (state) == arc_l2c_only)
699
700 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
701 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
702 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
703 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
704 #define HDR_PRESCIENT_PREFETCH(hdr) \
705 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
706 #define HDR_COMPRESSION_ENABLED(hdr) \
707 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
708
709 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
710 #define HDR_UNCACHED(hdr) ((hdr)->b_flags & ARC_FLAG_UNCACHED)
711 #define HDR_L2_READING(hdr) \
712 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
713 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
714 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
715 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
716 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
717 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
718 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
719 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
720
721 #define HDR_ISTYPE_METADATA(hdr) \
722 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
723 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
724
725 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
726 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
727 #define HDR_HAS_RABD(hdr) \
728 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
729 (hdr)->b_crypt_hdr.b_rabd != NULL)
730 #define HDR_ENCRYPTED(hdr) \
731 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
732 #define HDR_AUTHENTICATED(hdr) \
733 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
734
735 /* For storing compression mode in b_flags */
736 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
737
738 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
739 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
740 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
741 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
742
743 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
744 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
745 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
746 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
747
748 /*
749 * Other sizes
750 */
751
752 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
753 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
754
755 /*
756 * Hash table routines
757 */
758
759 #define BUF_LOCKS 2048
760 typedef struct buf_hash_table {
761 uint64_t ht_mask;
762 arc_buf_hdr_t **ht_table;
763 kmutex_t ht_locks[BUF_LOCKS] ____cacheline_aligned;
764 } buf_hash_table_t;
765
766 static buf_hash_table_t buf_hash_table;
767
768 #define BUF_HASH_INDEX(spa, dva, birth) \
769 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
770 #define BUF_HASH_LOCK(idx) (&buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
771 #define HDR_LOCK(hdr) \
772 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
773
774 uint64_t zfs_crc64_table[256];
775
776 /*
777 * Level 2 ARC
778 */
779
780 #define L2ARC_WRITE_SIZE (32 * 1024 * 1024) /* initial write max */
781 #define L2ARC_HEADROOM 8 /* num of writes */
782
783 /*
784 * If we discover during ARC scan any buffers to be compressed, we boost
785 * our headroom for the next scanning cycle by this percentage multiple.
786 */
787 #define L2ARC_HEADROOM_BOOST 200
788 #define L2ARC_FEED_SECS 1 /* caching interval secs */
789 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
790
791 /*
792 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
793 * and each of the state has two types: data and metadata.
794 */
795 #define L2ARC_FEED_TYPES 4
796
797 /* L2ARC Performance Tunables */
798 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */
799 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */
800 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */
801 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
802 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
803 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */
804 int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
805 int l2arc_feed_again = B_TRUE; /* turbo warmup */
806 int l2arc_norw = B_FALSE; /* no reads during writes */
807 static uint_t l2arc_meta_percent = 33; /* limit on headers size */
808
809 /*
810 * L2ARC Internals
811 */
812 static list_t L2ARC_dev_list; /* device list */
813 static list_t *l2arc_dev_list; /* device list pointer */
814 static kmutex_t l2arc_dev_mtx; /* device list mutex */
815 static l2arc_dev_t *l2arc_dev_last; /* last device used */
816 static list_t L2ARC_free_on_write; /* free after write buf list */
817 static list_t *l2arc_free_on_write; /* free after write list ptr */
818 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
819 static uint64_t l2arc_ndev; /* number of devices */
820
821 typedef struct l2arc_read_callback {
822 arc_buf_hdr_t *l2rcb_hdr; /* read header */
823 blkptr_t l2rcb_bp; /* original blkptr */
824 zbookmark_phys_t l2rcb_zb; /* original bookmark */
825 int l2rcb_flags; /* original flags */
826 abd_t *l2rcb_abd; /* temporary buffer */
827 } l2arc_read_callback_t;
828
829 typedef struct l2arc_data_free {
830 /* protected by l2arc_free_on_write_mtx */
831 abd_t *l2df_abd;
832 size_t l2df_size;
833 arc_buf_contents_t l2df_type;
834 list_node_t l2df_list_node;
835 } l2arc_data_free_t;
836
837 typedef enum arc_fill_flags {
838 ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */
839 ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */
840 ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */
841 ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */
842 ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */
843 } arc_fill_flags_t;
844
845 typedef enum arc_ovf_level {
846 ARC_OVF_NONE, /* ARC within target size. */
847 ARC_OVF_SOME, /* ARC is slightly overflowed. */
848 ARC_OVF_SEVERE /* ARC is severely overflowed. */
849 } arc_ovf_level_t;
850
851 static kmutex_t l2arc_feed_thr_lock;
852 static kcondvar_t l2arc_feed_thr_cv;
853 static uint8_t l2arc_thread_exit;
854
855 static kmutex_t l2arc_rebuild_thr_lock;
856 static kcondvar_t l2arc_rebuild_thr_cv;
857
858 enum arc_hdr_alloc_flags {
859 ARC_HDR_ALLOC_RDATA = 0x1,
860 ARC_HDR_USE_RESERVE = 0x4,
861 ARC_HDR_ALLOC_LINEAR = 0x8,
862 };
863
864
865 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, const void *, int);
866 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, const void *);
867 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, const void *, int);
868 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, const void *);
869 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, const void *);
870 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size,
871 const void *tag);
872 static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t);
873 static void arc_hdr_alloc_abd(arc_buf_hdr_t *, int);
874 static void arc_hdr_destroy(arc_buf_hdr_t *);
875 static void arc_access(arc_buf_hdr_t *, arc_flags_t, boolean_t);
876 static void arc_buf_watch(arc_buf_t *);
877 static void arc_change_state(arc_state_t *, arc_buf_hdr_t *);
878
879 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
880 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
881 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
882 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
883
884 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
885 static void l2arc_read_done(zio_t *);
886 static void l2arc_do_free_on_write(void);
887 static void l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
888 boolean_t state_only);
889
890 static void arc_prune_async(uint64_t adjust);
891
892 #define l2arc_hdr_arcstats_increment(hdr) \
893 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_FALSE)
894 #define l2arc_hdr_arcstats_decrement(hdr) \
895 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_FALSE)
896 #define l2arc_hdr_arcstats_increment_state(hdr) \
897 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_TRUE)
898 #define l2arc_hdr_arcstats_decrement_state(hdr) \
899 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_TRUE)
900
901 /*
902 * l2arc_exclude_special : A zfs module parameter that controls whether buffers
903 * present on special vdevs are eligibile for caching in L2ARC. If
904 * set to 1, exclude dbufs on special vdevs from being cached to
905 * L2ARC.
906 */
907 int l2arc_exclude_special = 0;
908
909 /*
910 * l2arc_mfuonly : A ZFS module parameter that controls whether only MFU
911 * metadata and data are cached from ARC into L2ARC.
912 */
913 static int l2arc_mfuonly = 0;
914
915 /*
916 * L2ARC TRIM
917 * l2arc_trim_ahead : A ZFS module parameter that controls how much ahead of
918 * the current write size (l2arc_write_max) we should TRIM if we
919 * have filled the device. It is defined as a percentage of the
920 * write size. If set to 100 we trim twice the space required to
921 * accommodate upcoming writes. A minimum of 64MB will be trimmed.
922 * It also enables TRIM of the whole L2ARC device upon creation or
923 * addition to an existing pool or if the header of the device is
924 * invalid upon importing a pool or onlining a cache device. The
925 * default is 0, which disables TRIM on L2ARC altogether as it can
926 * put significant stress on the underlying storage devices. This
927 * will vary depending of how well the specific device handles
928 * these commands.
929 */
930 static uint64_t l2arc_trim_ahead = 0;
931
932 /*
933 * Performance tuning of L2ARC persistence:
934 *
935 * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding
936 * an L2ARC device (either at pool import or later) will attempt
937 * to rebuild L2ARC buffer contents.
938 * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls
939 * whether log blocks are written to the L2ARC device. If the L2ARC
940 * device is less than 1GB, the amount of data l2arc_evict()
941 * evicts is significant compared to the amount of restored L2ARC
942 * data. In this case do not write log blocks in L2ARC in order
943 * not to waste space.
944 */
945 static int l2arc_rebuild_enabled = B_TRUE;
946 static uint64_t l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024;
947
948 /* L2ARC persistence rebuild control routines. */
949 void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen);
950 static __attribute__((noreturn)) void l2arc_dev_rebuild_thread(void *arg);
951 static int l2arc_rebuild(l2arc_dev_t *dev);
952
953 /* L2ARC persistence read I/O routines. */
954 static int l2arc_dev_hdr_read(l2arc_dev_t *dev);
955 static int l2arc_log_blk_read(l2arc_dev_t *dev,
956 const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp,
957 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
958 zio_t *this_io, zio_t **next_io);
959 static zio_t *l2arc_log_blk_fetch(vdev_t *vd,
960 const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb);
961 static void l2arc_log_blk_fetch_abort(zio_t *zio);
962
963 /* L2ARC persistence block restoration routines. */
964 static void l2arc_log_blk_restore(l2arc_dev_t *dev,
965 const l2arc_log_blk_phys_t *lb, uint64_t lb_asize);
966 static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le,
967 l2arc_dev_t *dev);
968
969 /* L2ARC persistence write I/O routines. */
970 static uint64_t l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
971 l2arc_write_callback_t *cb);
972
973 /* L2ARC persistence auxiliary routines. */
974 boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev,
975 const l2arc_log_blkptr_t *lbp);
976 static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev,
977 const arc_buf_hdr_t *ab);
978 boolean_t l2arc_range_check_overlap(uint64_t bottom,
979 uint64_t top, uint64_t check);
980 static void l2arc_blk_fetch_done(zio_t *zio);
981 static inline uint64_t
982 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev);
983
984 /*
985 * We use Cityhash for this. It's fast, and has good hash properties without
986 * requiring any large static buffers.
987 */
988 static uint64_t
buf_hash(uint64_t spa,const dva_t * dva,uint64_t birth)989 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
990 {
991 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
992 }
993
994 #define HDR_EMPTY(hdr) \
995 ((hdr)->b_dva.dva_word[0] == 0 && \
996 (hdr)->b_dva.dva_word[1] == 0)
997
998 #define HDR_EMPTY_OR_LOCKED(hdr) \
999 (HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr)))
1000
1001 #define HDR_EQUAL(spa, dva, birth, hdr) \
1002 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1003 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1004 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1005
1006 static void
buf_discard_identity(arc_buf_hdr_t * hdr)1007 buf_discard_identity(arc_buf_hdr_t *hdr)
1008 {
1009 hdr->b_dva.dva_word[0] = 0;
1010 hdr->b_dva.dva_word[1] = 0;
1011 hdr->b_birth = 0;
1012 }
1013
1014 static arc_buf_hdr_t *
buf_hash_find(uint64_t spa,const blkptr_t * bp,kmutex_t ** lockp)1015 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1016 {
1017 const dva_t *dva = BP_IDENTITY(bp);
1018 uint64_t birth = BP_GET_BIRTH(bp);
1019 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1020 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1021 arc_buf_hdr_t *hdr;
1022
1023 mutex_enter(hash_lock);
1024 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1025 hdr = hdr->b_hash_next) {
1026 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1027 *lockp = hash_lock;
1028 return (hdr);
1029 }
1030 }
1031 mutex_exit(hash_lock);
1032 *lockp = NULL;
1033 return (NULL);
1034 }
1035
1036 /*
1037 * Insert an entry into the hash table. If there is already an element
1038 * equal to elem in the hash table, then the already existing element
1039 * will be returned and the new element will not be inserted.
1040 * Otherwise returns NULL.
1041 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1042 */
1043 static arc_buf_hdr_t *
buf_hash_insert(arc_buf_hdr_t * hdr,kmutex_t ** lockp)1044 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1045 {
1046 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1047 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1048 arc_buf_hdr_t *fhdr;
1049 uint32_t i;
1050
1051 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1052 ASSERT(hdr->b_birth != 0);
1053 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1054
1055 if (lockp != NULL) {
1056 *lockp = hash_lock;
1057 mutex_enter(hash_lock);
1058 } else {
1059 ASSERT(MUTEX_HELD(hash_lock));
1060 }
1061
1062 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1063 fhdr = fhdr->b_hash_next, i++) {
1064 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1065 return (fhdr);
1066 }
1067
1068 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1069 buf_hash_table.ht_table[idx] = hdr;
1070 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1071
1072 /* collect some hash table performance data */
1073 if (i > 0) {
1074 ARCSTAT_BUMP(arcstat_hash_collisions);
1075 if (i == 1)
1076 ARCSTAT_BUMP(arcstat_hash_chains);
1077
1078 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1079 }
1080 uint64_t he = atomic_inc_64_nv(
1081 &arc_stats.arcstat_hash_elements.value.ui64);
1082 ARCSTAT_MAX(arcstat_hash_elements_max, he);
1083
1084 return (NULL);
1085 }
1086
1087 static void
buf_hash_remove(arc_buf_hdr_t * hdr)1088 buf_hash_remove(arc_buf_hdr_t *hdr)
1089 {
1090 arc_buf_hdr_t *fhdr, **hdrp;
1091 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1092
1093 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1094 ASSERT(HDR_IN_HASH_TABLE(hdr));
1095
1096 hdrp = &buf_hash_table.ht_table[idx];
1097 while ((fhdr = *hdrp) != hdr) {
1098 ASSERT3P(fhdr, !=, NULL);
1099 hdrp = &fhdr->b_hash_next;
1100 }
1101 *hdrp = hdr->b_hash_next;
1102 hdr->b_hash_next = NULL;
1103 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1104
1105 /* collect some hash table performance data */
1106 atomic_dec_64(&arc_stats.arcstat_hash_elements.value.ui64);
1107
1108 if (buf_hash_table.ht_table[idx] &&
1109 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1110 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1111 }
1112
1113 /*
1114 * Global data structures and functions for the buf kmem cache.
1115 */
1116
1117 static kmem_cache_t *hdr_full_cache;
1118 static kmem_cache_t *hdr_l2only_cache;
1119 static kmem_cache_t *buf_cache;
1120
1121 static void
buf_fini(void)1122 buf_fini(void)
1123 {
1124 #if defined(_KERNEL)
1125 /*
1126 * Large allocations which do not require contiguous pages
1127 * should be using vmem_free() in the linux kernel\
1128 */
1129 vmem_free(buf_hash_table.ht_table,
1130 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1131 #else
1132 kmem_free(buf_hash_table.ht_table,
1133 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1134 #endif
1135 for (int i = 0; i < BUF_LOCKS; i++)
1136 mutex_destroy(BUF_HASH_LOCK(i));
1137 kmem_cache_destroy(hdr_full_cache);
1138 kmem_cache_destroy(hdr_l2only_cache);
1139 kmem_cache_destroy(buf_cache);
1140 }
1141
1142 /*
1143 * Constructor callback - called when the cache is empty
1144 * and a new buf is requested.
1145 */
1146 static int
hdr_full_cons(void * vbuf,void * unused,int kmflag)1147 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1148 {
1149 (void) unused, (void) kmflag;
1150 arc_buf_hdr_t *hdr = vbuf;
1151
1152 memset(hdr, 0, HDR_FULL_SIZE);
1153 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
1154 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
1155 #ifdef ZFS_DEBUG
1156 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1157 #endif
1158 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1159 list_link_init(&hdr->b_l2hdr.b_l2node);
1160 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1161
1162 return (0);
1163 }
1164
1165 static int
hdr_l2only_cons(void * vbuf,void * unused,int kmflag)1166 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1167 {
1168 (void) unused, (void) kmflag;
1169 arc_buf_hdr_t *hdr = vbuf;
1170
1171 memset(hdr, 0, HDR_L2ONLY_SIZE);
1172 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1173
1174 return (0);
1175 }
1176
1177 static int
buf_cons(void * vbuf,void * unused,int kmflag)1178 buf_cons(void *vbuf, void *unused, int kmflag)
1179 {
1180 (void) unused, (void) kmflag;
1181 arc_buf_t *buf = vbuf;
1182
1183 memset(buf, 0, sizeof (arc_buf_t));
1184 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1185
1186 return (0);
1187 }
1188
1189 /*
1190 * Destructor callback - called when a cached buf is
1191 * no longer required.
1192 */
1193 static void
hdr_full_dest(void * vbuf,void * unused)1194 hdr_full_dest(void *vbuf, void *unused)
1195 {
1196 (void) unused;
1197 arc_buf_hdr_t *hdr = vbuf;
1198
1199 ASSERT(HDR_EMPTY(hdr));
1200 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1201 #ifdef ZFS_DEBUG
1202 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1203 #endif
1204 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1205 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1206 }
1207
1208 static void
hdr_l2only_dest(void * vbuf,void * unused)1209 hdr_l2only_dest(void *vbuf, void *unused)
1210 {
1211 (void) unused;
1212 arc_buf_hdr_t *hdr = vbuf;
1213
1214 ASSERT(HDR_EMPTY(hdr));
1215 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1216 }
1217
1218 static void
buf_dest(void * vbuf,void * unused)1219 buf_dest(void *vbuf, void *unused)
1220 {
1221 (void) unused;
1222 (void) vbuf;
1223
1224 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1225 }
1226
1227 static void
buf_init(void)1228 buf_init(void)
1229 {
1230 uint64_t *ct = NULL;
1231 uint64_t hsize = 1ULL << 12;
1232 int i, j;
1233
1234 /*
1235 * The hash table is big enough to fill all of physical memory
1236 * with an average block size of zfs_arc_average_blocksize (default 8K).
1237 * By default, the table will take up
1238 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1239 */
1240 while (hsize * zfs_arc_average_blocksize < arc_all_memory())
1241 hsize <<= 1;
1242 retry:
1243 buf_hash_table.ht_mask = hsize - 1;
1244 #if defined(_KERNEL)
1245 /*
1246 * Large allocations which do not require contiguous pages
1247 * should be using vmem_alloc() in the linux kernel
1248 */
1249 buf_hash_table.ht_table =
1250 vmem_zalloc(hsize * sizeof (void*), KM_SLEEP);
1251 #else
1252 buf_hash_table.ht_table =
1253 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1254 #endif
1255 if (buf_hash_table.ht_table == NULL) {
1256 ASSERT(hsize > (1ULL << 8));
1257 hsize >>= 1;
1258 goto retry;
1259 }
1260
1261 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1262 0, hdr_full_cons, hdr_full_dest, NULL, NULL, NULL, KMC_RECLAIMABLE);
1263 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1264 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, NULL,
1265 NULL, NULL, 0);
1266 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1267 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1268
1269 for (i = 0; i < 256; i++)
1270 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1271 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1272
1273 for (i = 0; i < BUF_LOCKS; i++)
1274 mutex_init(BUF_HASH_LOCK(i), NULL, MUTEX_DEFAULT, NULL);
1275 }
1276
1277 #define ARC_MINTIME (hz>>4) /* 62 ms */
1278
1279 /*
1280 * This is the size that the buf occupies in memory. If the buf is compressed,
1281 * it will correspond to the compressed size. You should use this method of
1282 * getting the buf size unless you explicitly need the logical size.
1283 */
1284 uint64_t
arc_buf_size(arc_buf_t * buf)1285 arc_buf_size(arc_buf_t *buf)
1286 {
1287 return (ARC_BUF_COMPRESSED(buf) ?
1288 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1289 }
1290
1291 uint64_t
arc_buf_lsize(arc_buf_t * buf)1292 arc_buf_lsize(arc_buf_t *buf)
1293 {
1294 return (HDR_GET_LSIZE(buf->b_hdr));
1295 }
1296
1297 /*
1298 * This function will return B_TRUE if the buffer is encrypted in memory.
1299 * This buffer can be decrypted by calling arc_untransform().
1300 */
1301 boolean_t
arc_is_encrypted(arc_buf_t * buf)1302 arc_is_encrypted(arc_buf_t *buf)
1303 {
1304 return (ARC_BUF_ENCRYPTED(buf) != 0);
1305 }
1306
1307 /*
1308 * Returns B_TRUE if the buffer represents data that has not had its MAC
1309 * verified yet.
1310 */
1311 boolean_t
arc_is_unauthenticated(arc_buf_t * buf)1312 arc_is_unauthenticated(arc_buf_t *buf)
1313 {
1314 return (HDR_NOAUTH(buf->b_hdr) != 0);
1315 }
1316
1317 void
arc_get_raw_params(arc_buf_t * buf,boolean_t * byteorder,uint8_t * salt,uint8_t * iv,uint8_t * mac)1318 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1319 uint8_t *iv, uint8_t *mac)
1320 {
1321 arc_buf_hdr_t *hdr = buf->b_hdr;
1322
1323 ASSERT(HDR_PROTECTED(hdr));
1324
1325 memcpy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
1326 memcpy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
1327 memcpy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
1328 *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1329 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1330 }
1331
1332 /*
1333 * Indicates how this buffer is compressed in memory. If it is not compressed
1334 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1335 * arc_untransform() as long as it is also unencrypted.
1336 */
1337 enum zio_compress
arc_get_compression(arc_buf_t * buf)1338 arc_get_compression(arc_buf_t *buf)
1339 {
1340 return (ARC_BUF_COMPRESSED(buf) ?
1341 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1342 }
1343
1344 /*
1345 * Return the compression algorithm used to store this data in the ARC. If ARC
1346 * compression is enabled or this is an encrypted block, this will be the same
1347 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1348 */
1349 static inline enum zio_compress
arc_hdr_get_compress(arc_buf_hdr_t * hdr)1350 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1351 {
1352 return (HDR_COMPRESSION_ENABLED(hdr) ?
1353 HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1354 }
1355
1356 uint8_t
arc_get_complevel(arc_buf_t * buf)1357 arc_get_complevel(arc_buf_t *buf)
1358 {
1359 return (buf->b_hdr->b_complevel);
1360 }
1361
1362 static inline boolean_t
arc_buf_is_shared(arc_buf_t * buf)1363 arc_buf_is_shared(arc_buf_t *buf)
1364 {
1365 boolean_t shared = (buf->b_data != NULL &&
1366 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1367 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1368 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1369 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1370 EQUIV(shared, ARC_BUF_SHARED(buf));
1371 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1372
1373 /*
1374 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1375 * already being shared" requirement prevents us from doing that.
1376 */
1377
1378 return (shared);
1379 }
1380
1381 /*
1382 * Free the checksum associated with this header. If there is no checksum, this
1383 * is a no-op.
1384 */
1385 static inline void
arc_cksum_free(arc_buf_hdr_t * hdr)1386 arc_cksum_free(arc_buf_hdr_t *hdr)
1387 {
1388 #ifdef ZFS_DEBUG
1389 ASSERT(HDR_HAS_L1HDR(hdr));
1390
1391 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1392 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1393 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1394 hdr->b_l1hdr.b_freeze_cksum = NULL;
1395 }
1396 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1397 #endif
1398 }
1399
1400 /*
1401 * Return true iff at least one of the bufs on hdr is not compressed.
1402 * Encrypted buffers count as compressed.
1403 */
1404 static boolean_t
arc_hdr_has_uncompressed_buf(arc_buf_hdr_t * hdr)1405 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1406 {
1407 ASSERT(hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY_OR_LOCKED(hdr));
1408
1409 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1410 if (!ARC_BUF_COMPRESSED(b)) {
1411 return (B_TRUE);
1412 }
1413 }
1414 return (B_FALSE);
1415 }
1416
1417
1418 /*
1419 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1420 * matches the checksum that is stored in the hdr. If there is no checksum,
1421 * or if the buf is compressed, this is a no-op.
1422 */
1423 static void
arc_cksum_verify(arc_buf_t * buf)1424 arc_cksum_verify(arc_buf_t *buf)
1425 {
1426 #ifdef ZFS_DEBUG
1427 arc_buf_hdr_t *hdr = buf->b_hdr;
1428 zio_cksum_t zc;
1429
1430 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1431 return;
1432
1433 if (ARC_BUF_COMPRESSED(buf))
1434 return;
1435
1436 ASSERT(HDR_HAS_L1HDR(hdr));
1437
1438 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1439
1440 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1441 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1442 return;
1443 }
1444
1445 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1446 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1447 panic("buffer modified while frozen!");
1448 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1449 #endif
1450 }
1451
1452 /*
1453 * This function makes the assumption that data stored in the L2ARC
1454 * will be transformed exactly as it is in the main pool. Because of
1455 * this we can verify the checksum against the reading process's bp.
1456 */
1457 static boolean_t
arc_cksum_is_equal(arc_buf_hdr_t * hdr,zio_t * zio)1458 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1459 {
1460 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1461 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1462
1463 /*
1464 * Block pointers always store the checksum for the logical data.
1465 * If the block pointer has the gang bit set, then the checksum
1466 * it represents is for the reconstituted data and not for an
1467 * individual gang member. The zio pipeline, however, must be able to
1468 * determine the checksum of each of the gang constituents so it
1469 * treats the checksum comparison differently than what we need
1470 * for l2arc blocks. This prevents us from using the
1471 * zio_checksum_error() interface directly. Instead we must call the
1472 * zio_checksum_error_impl() so that we can ensure the checksum is
1473 * generated using the correct checksum algorithm and accounts for the
1474 * logical I/O size and not just a gang fragment.
1475 */
1476 return (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1477 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1478 zio->io_offset, NULL) == 0);
1479 }
1480
1481 /*
1482 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1483 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1484 * isn't modified later on. If buf is compressed or there is already a checksum
1485 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1486 */
1487 static void
arc_cksum_compute(arc_buf_t * buf)1488 arc_cksum_compute(arc_buf_t *buf)
1489 {
1490 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1491 return;
1492
1493 #ifdef ZFS_DEBUG
1494 arc_buf_hdr_t *hdr = buf->b_hdr;
1495 ASSERT(HDR_HAS_L1HDR(hdr));
1496 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1497 if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
1498 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1499 return;
1500 }
1501
1502 ASSERT(!ARC_BUF_ENCRYPTED(buf));
1503 ASSERT(!ARC_BUF_COMPRESSED(buf));
1504 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1505 KM_SLEEP);
1506 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1507 hdr->b_l1hdr.b_freeze_cksum);
1508 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1509 #endif
1510 arc_buf_watch(buf);
1511 }
1512
1513 #ifndef _KERNEL
1514 void
arc_buf_sigsegv(int sig,siginfo_t * si,void * unused)1515 arc_buf_sigsegv(int sig, siginfo_t *si, void *unused)
1516 {
1517 (void) sig, (void) unused;
1518 panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr);
1519 }
1520 #endif
1521
1522 static void
arc_buf_unwatch(arc_buf_t * buf)1523 arc_buf_unwatch(arc_buf_t *buf)
1524 {
1525 #ifndef _KERNEL
1526 if (arc_watch) {
1527 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1528 PROT_READ | PROT_WRITE));
1529 }
1530 #else
1531 (void) buf;
1532 #endif
1533 }
1534
1535 static void
arc_buf_watch(arc_buf_t * buf)1536 arc_buf_watch(arc_buf_t *buf)
1537 {
1538 #ifndef _KERNEL
1539 if (arc_watch)
1540 ASSERT0(mprotect(buf->b_data, arc_buf_size(buf),
1541 PROT_READ));
1542 #else
1543 (void) buf;
1544 #endif
1545 }
1546
1547 static arc_buf_contents_t
arc_buf_type(arc_buf_hdr_t * hdr)1548 arc_buf_type(arc_buf_hdr_t *hdr)
1549 {
1550 arc_buf_contents_t type;
1551 if (HDR_ISTYPE_METADATA(hdr)) {
1552 type = ARC_BUFC_METADATA;
1553 } else {
1554 type = ARC_BUFC_DATA;
1555 }
1556 VERIFY3U(hdr->b_type, ==, type);
1557 return (type);
1558 }
1559
1560 boolean_t
arc_is_metadata(arc_buf_t * buf)1561 arc_is_metadata(arc_buf_t *buf)
1562 {
1563 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1564 }
1565
1566 static uint32_t
arc_bufc_to_flags(arc_buf_contents_t type)1567 arc_bufc_to_flags(arc_buf_contents_t type)
1568 {
1569 switch (type) {
1570 case ARC_BUFC_DATA:
1571 /* metadata field is 0 if buffer contains normal data */
1572 return (0);
1573 case ARC_BUFC_METADATA:
1574 return (ARC_FLAG_BUFC_METADATA);
1575 default:
1576 break;
1577 }
1578 panic("undefined ARC buffer type!");
1579 return ((uint32_t)-1);
1580 }
1581
1582 void
arc_buf_thaw(arc_buf_t * buf)1583 arc_buf_thaw(arc_buf_t *buf)
1584 {
1585 arc_buf_hdr_t *hdr = buf->b_hdr;
1586
1587 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1588 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1589
1590 arc_cksum_verify(buf);
1591
1592 /*
1593 * Compressed buffers do not manipulate the b_freeze_cksum.
1594 */
1595 if (ARC_BUF_COMPRESSED(buf))
1596 return;
1597
1598 ASSERT(HDR_HAS_L1HDR(hdr));
1599 arc_cksum_free(hdr);
1600 arc_buf_unwatch(buf);
1601 }
1602
1603 void
arc_buf_freeze(arc_buf_t * buf)1604 arc_buf_freeze(arc_buf_t *buf)
1605 {
1606 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1607 return;
1608
1609 if (ARC_BUF_COMPRESSED(buf))
1610 return;
1611
1612 ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
1613 arc_cksum_compute(buf);
1614 }
1615
1616 /*
1617 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1618 * the following functions should be used to ensure that the flags are
1619 * updated in a thread-safe way. When manipulating the flags either
1620 * the hash_lock must be held or the hdr must be undiscoverable. This
1621 * ensures that we're not racing with any other threads when updating
1622 * the flags.
1623 */
1624 static inline void
arc_hdr_set_flags(arc_buf_hdr_t * hdr,arc_flags_t flags)1625 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1626 {
1627 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1628 hdr->b_flags |= flags;
1629 }
1630
1631 static inline void
arc_hdr_clear_flags(arc_buf_hdr_t * hdr,arc_flags_t flags)1632 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1633 {
1634 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1635 hdr->b_flags &= ~flags;
1636 }
1637
1638 /*
1639 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1640 * done in a special way since we have to clear and set bits
1641 * at the same time. Consumers that wish to set the compression bits
1642 * must use this function to ensure that the flags are updated in
1643 * thread-safe manner.
1644 */
1645 static void
arc_hdr_set_compress(arc_buf_hdr_t * hdr,enum zio_compress cmp)1646 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1647 {
1648 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1649
1650 /*
1651 * Holes and embedded blocks will always have a psize = 0 so
1652 * we ignore the compression of the blkptr and set the
1653 * want to uncompress them. Mark them as uncompressed.
1654 */
1655 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1656 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1657 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1658 } else {
1659 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1660 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1661 }
1662
1663 HDR_SET_COMPRESS(hdr, cmp);
1664 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1665 }
1666
1667 /*
1668 * Looks for another buf on the same hdr which has the data decompressed, copies
1669 * from it, and returns true. If no such buf exists, returns false.
1670 */
1671 static boolean_t
arc_buf_try_copy_decompressed_data(arc_buf_t * buf)1672 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1673 {
1674 arc_buf_hdr_t *hdr = buf->b_hdr;
1675 boolean_t copied = B_FALSE;
1676
1677 ASSERT(HDR_HAS_L1HDR(hdr));
1678 ASSERT3P(buf->b_data, !=, NULL);
1679 ASSERT(!ARC_BUF_COMPRESSED(buf));
1680
1681 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1682 from = from->b_next) {
1683 /* can't use our own data buffer */
1684 if (from == buf) {
1685 continue;
1686 }
1687
1688 if (!ARC_BUF_COMPRESSED(from)) {
1689 memcpy(buf->b_data, from->b_data, arc_buf_size(buf));
1690 copied = B_TRUE;
1691 break;
1692 }
1693 }
1694
1695 #ifdef ZFS_DEBUG
1696 /*
1697 * There were no decompressed bufs, so there should not be a
1698 * checksum on the hdr either.
1699 */
1700 if (zfs_flags & ZFS_DEBUG_MODIFY)
1701 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1702 #endif
1703
1704 return (copied);
1705 }
1706
1707 /*
1708 * Allocates an ARC buf header that's in an evicted & L2-cached state.
1709 * This is used during l2arc reconstruction to make empty ARC buffers
1710 * which circumvent the regular disk->arc->l2arc path and instead come
1711 * into being in the reverse order, i.e. l2arc->arc.
1712 */
1713 static arc_buf_hdr_t *
arc_buf_alloc_l2only(size_t size,arc_buf_contents_t type,l2arc_dev_t * dev,dva_t dva,uint64_t daddr,int32_t psize,uint64_t birth,enum zio_compress compress,uint8_t complevel,boolean_t protected,boolean_t prefetch,arc_state_type_t arcs_state)1714 arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev,
1715 dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth,
1716 enum zio_compress compress, uint8_t complevel, boolean_t protected,
1717 boolean_t prefetch, arc_state_type_t arcs_state)
1718 {
1719 arc_buf_hdr_t *hdr;
1720
1721 ASSERT(size != 0);
1722 hdr = kmem_cache_alloc(hdr_l2only_cache, KM_SLEEP);
1723 hdr->b_birth = birth;
1724 hdr->b_type = type;
1725 hdr->b_flags = 0;
1726 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR);
1727 HDR_SET_LSIZE(hdr, size);
1728 HDR_SET_PSIZE(hdr, psize);
1729 arc_hdr_set_compress(hdr, compress);
1730 hdr->b_complevel = complevel;
1731 if (protected)
1732 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
1733 if (prefetch)
1734 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
1735 hdr->b_spa = spa_load_guid(dev->l2ad_vdev->vdev_spa);
1736
1737 hdr->b_dva = dva;
1738
1739 hdr->b_l2hdr.b_dev = dev;
1740 hdr->b_l2hdr.b_daddr = daddr;
1741 hdr->b_l2hdr.b_arcs_state = arcs_state;
1742
1743 return (hdr);
1744 }
1745
1746 /*
1747 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1748 */
1749 static uint64_t
arc_hdr_size(arc_buf_hdr_t * hdr)1750 arc_hdr_size(arc_buf_hdr_t *hdr)
1751 {
1752 uint64_t size;
1753
1754 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1755 HDR_GET_PSIZE(hdr) > 0) {
1756 size = HDR_GET_PSIZE(hdr);
1757 } else {
1758 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1759 size = HDR_GET_LSIZE(hdr);
1760 }
1761 return (size);
1762 }
1763
1764 static int
arc_hdr_authenticate(arc_buf_hdr_t * hdr,spa_t * spa,uint64_t dsobj)1765 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1766 {
1767 int ret;
1768 uint64_t csize;
1769 uint64_t lsize = HDR_GET_LSIZE(hdr);
1770 uint64_t psize = HDR_GET_PSIZE(hdr);
1771 abd_t *abd = hdr->b_l1hdr.b_pabd;
1772 boolean_t free_abd = B_FALSE;
1773
1774 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1775 ASSERT(HDR_AUTHENTICATED(hdr));
1776 ASSERT3P(abd, !=, NULL);
1777
1778 /*
1779 * The MAC is calculated on the compressed data that is stored on disk.
1780 * However, if compressed arc is disabled we will only have the
1781 * decompressed data available to us now. Compress it into a temporary
1782 * abd so we can verify the MAC. The performance overhead of this will
1783 * be relatively low, since most objects in an encrypted objset will
1784 * be encrypted (instead of authenticated) anyway.
1785 */
1786 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1787 !HDR_COMPRESSION_ENABLED(hdr)) {
1788 abd = NULL;
1789 csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1790 hdr->b_l1hdr.b_pabd, &abd, lsize, MIN(lsize, psize),
1791 hdr->b_complevel);
1792 if (csize >= lsize || csize > psize) {
1793 ret = SET_ERROR(EIO);
1794 return (ret);
1795 }
1796 ASSERT3P(abd, !=, NULL);
1797 abd_zero_off(abd, csize, psize - csize);
1798 free_abd = B_TRUE;
1799 }
1800
1801 /*
1802 * Authentication is best effort. We authenticate whenever the key is
1803 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1804 */
1805 if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1806 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1807 ASSERT3U(lsize, ==, psize);
1808 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1809 psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1810 } else {
1811 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1812 hdr->b_crypt_hdr.b_mac);
1813 }
1814
1815 if (ret == 0)
1816 arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1817 else if (ret == ENOENT)
1818 ret = 0;
1819
1820 if (free_abd)
1821 abd_free(abd);
1822
1823 return (ret);
1824 }
1825
1826 /*
1827 * This function will take a header that only has raw encrypted data in
1828 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1829 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1830 * also decompress the data.
1831 */
1832 static int
arc_hdr_decrypt(arc_buf_hdr_t * hdr,spa_t * spa,const zbookmark_phys_t * zb)1833 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
1834 {
1835 int ret;
1836 abd_t *cabd = NULL;
1837 boolean_t no_crypt = B_FALSE;
1838 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1839
1840 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1841 ASSERT(HDR_ENCRYPTED(hdr));
1842
1843 arc_hdr_alloc_abd(hdr, 0);
1844
1845 ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
1846 B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
1847 hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
1848 hdr->b_crypt_hdr.b_rabd, &no_crypt);
1849 if (ret != 0)
1850 goto error;
1851
1852 if (no_crypt) {
1853 abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1854 HDR_GET_PSIZE(hdr));
1855 }
1856
1857 /*
1858 * If this header has disabled arc compression but the b_pabd is
1859 * compressed after decrypting it, we need to decompress the newly
1860 * decrypted data.
1861 */
1862 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1863 !HDR_COMPRESSION_ENABLED(hdr)) {
1864 /*
1865 * We want to make sure that we are correctly honoring the
1866 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1867 * and then loan a buffer from it, rather than allocating a
1868 * linear buffer and wrapping it in an abd later.
1869 */
1870 cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr, 0);
1871
1872 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1873 hdr->b_l1hdr.b_pabd, cabd, HDR_GET_PSIZE(hdr),
1874 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
1875 if (ret != 0) {
1876 goto error;
1877 }
1878
1879 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
1880 arc_hdr_size(hdr), hdr);
1881 hdr->b_l1hdr.b_pabd = cabd;
1882 }
1883
1884 return (0);
1885
1886 error:
1887 arc_hdr_free_abd(hdr, B_FALSE);
1888 if (cabd != NULL)
1889 arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
1890
1891 return (ret);
1892 }
1893
1894 /*
1895 * This function is called during arc_buf_fill() to prepare the header's
1896 * abd plaintext pointer for use. This involves authenticated protected
1897 * data and decrypting encrypted data into the plaintext abd.
1898 */
1899 static int
arc_fill_hdr_crypt(arc_buf_hdr_t * hdr,kmutex_t * hash_lock,spa_t * spa,const zbookmark_phys_t * zb,boolean_t noauth)1900 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
1901 const zbookmark_phys_t *zb, boolean_t noauth)
1902 {
1903 int ret;
1904
1905 ASSERT(HDR_PROTECTED(hdr));
1906
1907 if (hash_lock != NULL)
1908 mutex_enter(hash_lock);
1909
1910 if (HDR_NOAUTH(hdr) && !noauth) {
1911 /*
1912 * The caller requested authenticated data but our data has
1913 * not been authenticated yet. Verify the MAC now if we can.
1914 */
1915 ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
1916 if (ret != 0)
1917 goto error;
1918 } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
1919 /*
1920 * If we only have the encrypted version of the data, but the
1921 * unencrypted version was requested we take this opportunity
1922 * to store the decrypted version in the header for future use.
1923 */
1924 ret = arc_hdr_decrypt(hdr, spa, zb);
1925 if (ret != 0)
1926 goto error;
1927 }
1928
1929 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1930
1931 if (hash_lock != NULL)
1932 mutex_exit(hash_lock);
1933
1934 return (0);
1935
1936 error:
1937 if (hash_lock != NULL)
1938 mutex_exit(hash_lock);
1939
1940 return (ret);
1941 }
1942
1943 /*
1944 * This function is used by the dbuf code to decrypt bonus buffers in place.
1945 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
1946 * block, so we use the hash lock here to protect against concurrent calls to
1947 * arc_buf_fill().
1948 */
1949 static void
arc_buf_untransform_in_place(arc_buf_t * buf)1950 arc_buf_untransform_in_place(arc_buf_t *buf)
1951 {
1952 arc_buf_hdr_t *hdr = buf->b_hdr;
1953
1954 ASSERT(HDR_ENCRYPTED(hdr));
1955 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1956 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1957 ASSERT3PF(hdr->b_l1hdr.b_pabd, !=, NULL, "hdr %px buf %px", hdr, buf);
1958
1959 zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
1960 arc_buf_size(buf));
1961 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
1962 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1963 }
1964
1965 /*
1966 * Given a buf that has a data buffer attached to it, this function will
1967 * efficiently fill the buf with data of the specified compression setting from
1968 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1969 * are already sharing a data buf, no copy is performed.
1970 *
1971 * If the buf is marked as compressed but uncompressed data was requested, this
1972 * will allocate a new data buffer for the buf, remove that flag, and fill the
1973 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1974 * uncompressed data, and (since we haven't added support for it yet) if you
1975 * want compressed data your buf must already be marked as compressed and have
1976 * the correct-sized data buffer.
1977 */
1978 static int
arc_buf_fill(arc_buf_t * buf,spa_t * spa,const zbookmark_phys_t * zb,arc_fill_flags_t flags)1979 arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
1980 arc_fill_flags_t flags)
1981 {
1982 int error = 0;
1983 arc_buf_hdr_t *hdr = buf->b_hdr;
1984 boolean_t hdr_compressed =
1985 (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
1986 boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
1987 boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
1988 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
1989 kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
1990
1991 ASSERT3P(buf->b_data, !=, NULL);
1992 IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
1993 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
1994 IMPLY(encrypted, HDR_ENCRYPTED(hdr));
1995 IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
1996 IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
1997 IMPLY(encrypted, !arc_buf_is_shared(buf));
1998
1999 /*
2000 * If the caller wanted encrypted data we just need to copy it from
2001 * b_rabd and potentially byteswap it. We won't be able to do any
2002 * further transforms on it.
2003 */
2004 if (encrypted) {
2005 ASSERT(HDR_HAS_RABD(hdr));
2006 abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
2007 HDR_GET_PSIZE(hdr));
2008 goto byteswap;
2009 }
2010
2011 /*
2012 * Adjust encrypted and authenticated headers to accommodate
2013 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
2014 * allowed to fail decryption due to keys not being loaded
2015 * without being marked as an IO error.
2016 */
2017 if (HDR_PROTECTED(hdr)) {
2018 error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
2019 zb, !!(flags & ARC_FILL_NOAUTH));
2020 if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
2021 return (error);
2022 } else if (error != 0) {
2023 if (hash_lock != NULL)
2024 mutex_enter(hash_lock);
2025 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2026 if (hash_lock != NULL)
2027 mutex_exit(hash_lock);
2028 return (error);
2029 }
2030 }
2031
2032 /*
2033 * There is a special case here for dnode blocks which are
2034 * decrypting their bonus buffers. These blocks may request to
2035 * be decrypted in-place. This is necessary because there may
2036 * be many dnodes pointing into this buffer and there is
2037 * currently no method to synchronize replacing the backing
2038 * b_data buffer and updating all of the pointers. Here we use
2039 * the hash lock to ensure there are no races. If the need
2040 * arises for other types to be decrypted in-place, they must
2041 * add handling here as well.
2042 */
2043 if ((flags & ARC_FILL_IN_PLACE) != 0) {
2044 ASSERT(!hdr_compressed);
2045 ASSERT(!compressed);
2046 ASSERT(!encrypted);
2047
2048 if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
2049 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
2050
2051 if (hash_lock != NULL)
2052 mutex_enter(hash_lock);
2053 arc_buf_untransform_in_place(buf);
2054 if (hash_lock != NULL)
2055 mutex_exit(hash_lock);
2056
2057 /* Compute the hdr's checksum if necessary */
2058 arc_cksum_compute(buf);
2059 }
2060
2061 return (0);
2062 }
2063
2064 if (hdr_compressed == compressed) {
2065 if (ARC_BUF_SHARED(buf)) {
2066 ASSERT(arc_buf_is_shared(buf));
2067 } else {
2068 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
2069 arc_buf_size(buf));
2070 }
2071 } else {
2072 ASSERT(hdr_compressed);
2073 ASSERT(!compressed);
2074
2075 /*
2076 * If the buf is sharing its data with the hdr, unlink it and
2077 * allocate a new data buffer for the buf.
2078 */
2079 if (ARC_BUF_SHARED(buf)) {
2080 ASSERTF(ARC_BUF_COMPRESSED(buf),
2081 "buf %p was uncompressed", buf);
2082
2083 /* We need to give the buf its own b_data */
2084 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2085 buf->b_data =
2086 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2087 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2088
2089 /* Previously overhead was 0; just add new overhead */
2090 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2091 } else if (ARC_BUF_COMPRESSED(buf)) {
2092 ASSERT(!arc_buf_is_shared(buf));
2093
2094 /* We need to reallocate the buf's b_data */
2095 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2096 buf);
2097 buf->b_data =
2098 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2099
2100 /* We increased the size of b_data; update overhead */
2101 ARCSTAT_INCR(arcstat_overhead_size,
2102 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2103 }
2104
2105 /*
2106 * Regardless of the buf's previous compression settings, it
2107 * should not be compressed at the end of this function.
2108 */
2109 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2110
2111 /*
2112 * Try copying the data from another buf which already has a
2113 * decompressed version. If that's not possible, it's time to
2114 * bite the bullet and decompress the data from the hdr.
2115 */
2116 if (arc_buf_try_copy_decompressed_data(buf)) {
2117 /* Skip byteswapping and checksumming (already done) */
2118 return (0);
2119 } else {
2120 abd_t dabd;
2121 abd_get_from_buf_struct(&dabd, buf->b_data,
2122 HDR_GET_LSIZE(hdr));
2123 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2124 hdr->b_l1hdr.b_pabd, &dabd,
2125 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr),
2126 &hdr->b_complevel);
2127 abd_free(&dabd);
2128
2129 /*
2130 * Absent hardware errors or software bugs, this should
2131 * be impossible, but log it anyway so we can debug it.
2132 */
2133 if (error != 0) {
2134 zfs_dbgmsg(
2135 "hdr %px, compress %d, psize %d, lsize %d",
2136 hdr, arc_hdr_get_compress(hdr),
2137 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2138 if (hash_lock != NULL)
2139 mutex_enter(hash_lock);
2140 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2141 if (hash_lock != NULL)
2142 mutex_exit(hash_lock);
2143 return (SET_ERROR(EIO));
2144 }
2145 }
2146 }
2147
2148 byteswap:
2149 /* Byteswap the buf's data if necessary */
2150 if (bswap != DMU_BSWAP_NUMFUNCS) {
2151 ASSERT(!HDR_SHARED_DATA(hdr));
2152 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2153 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2154 }
2155
2156 /* Compute the hdr's checksum if necessary */
2157 arc_cksum_compute(buf);
2158
2159 return (0);
2160 }
2161
2162 /*
2163 * If this function is being called to decrypt an encrypted buffer or verify an
2164 * authenticated one, the key must be loaded and a mapping must be made
2165 * available in the keystore via spa_keystore_create_mapping() or one of its
2166 * callers.
2167 */
2168 int
arc_untransform(arc_buf_t * buf,spa_t * spa,const zbookmark_phys_t * zb,boolean_t in_place)2169 arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2170 boolean_t in_place)
2171 {
2172 int ret;
2173 arc_fill_flags_t flags = 0;
2174
2175 if (in_place)
2176 flags |= ARC_FILL_IN_PLACE;
2177
2178 ret = arc_buf_fill(buf, spa, zb, flags);
2179 if (ret == ECKSUM) {
2180 /*
2181 * Convert authentication and decryption errors to EIO
2182 * (and generate an ereport) before leaving the ARC.
2183 */
2184 ret = SET_ERROR(EIO);
2185 spa_log_error(spa, zb, buf->b_hdr->b_birth);
2186 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
2187 spa, NULL, zb, NULL, 0);
2188 }
2189
2190 return (ret);
2191 }
2192
2193 /*
2194 * Increment the amount of evictable space in the arc_state_t's refcount.
2195 * We account for the space used by the hdr and the arc buf individually
2196 * so that we can add and remove them from the refcount individually.
2197 */
2198 static void
arc_evictable_space_increment(arc_buf_hdr_t * hdr,arc_state_t * state)2199 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2200 {
2201 arc_buf_contents_t type = arc_buf_type(hdr);
2202
2203 ASSERT(HDR_HAS_L1HDR(hdr));
2204
2205 if (GHOST_STATE(state)) {
2206 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2207 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2208 ASSERT(!HDR_HAS_RABD(hdr));
2209 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2210 HDR_GET_LSIZE(hdr), hdr);
2211 return;
2212 }
2213
2214 if (hdr->b_l1hdr.b_pabd != NULL) {
2215 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2216 arc_hdr_size(hdr), hdr);
2217 }
2218 if (HDR_HAS_RABD(hdr)) {
2219 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2220 HDR_GET_PSIZE(hdr), hdr);
2221 }
2222
2223 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2224 buf = buf->b_next) {
2225 if (ARC_BUF_SHARED(buf))
2226 continue;
2227 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2228 arc_buf_size(buf), buf);
2229 }
2230 }
2231
2232 /*
2233 * Decrement the amount of evictable space in the arc_state_t's refcount.
2234 * We account for the space used by the hdr and the arc buf individually
2235 * so that we can add and remove them from the refcount individually.
2236 */
2237 static void
arc_evictable_space_decrement(arc_buf_hdr_t * hdr,arc_state_t * state)2238 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2239 {
2240 arc_buf_contents_t type = arc_buf_type(hdr);
2241
2242 ASSERT(HDR_HAS_L1HDR(hdr));
2243
2244 if (GHOST_STATE(state)) {
2245 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2246 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2247 ASSERT(!HDR_HAS_RABD(hdr));
2248 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2249 HDR_GET_LSIZE(hdr), hdr);
2250 return;
2251 }
2252
2253 if (hdr->b_l1hdr.b_pabd != NULL) {
2254 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2255 arc_hdr_size(hdr), hdr);
2256 }
2257 if (HDR_HAS_RABD(hdr)) {
2258 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2259 HDR_GET_PSIZE(hdr), hdr);
2260 }
2261
2262 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2263 buf = buf->b_next) {
2264 if (ARC_BUF_SHARED(buf))
2265 continue;
2266 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2267 arc_buf_size(buf), buf);
2268 }
2269 }
2270
2271 /*
2272 * Add a reference to this hdr indicating that someone is actively
2273 * referencing that memory. When the refcount transitions from 0 to 1,
2274 * we remove it from the respective arc_state_t list to indicate that
2275 * it is not evictable.
2276 */
2277 static void
add_reference(arc_buf_hdr_t * hdr,const void * tag)2278 add_reference(arc_buf_hdr_t *hdr, const void *tag)
2279 {
2280 arc_state_t *state = hdr->b_l1hdr.b_state;
2281
2282 ASSERT(HDR_HAS_L1HDR(hdr));
2283 if (!HDR_EMPTY(hdr) && !MUTEX_HELD(HDR_LOCK(hdr))) {
2284 ASSERT(state == arc_anon);
2285 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2286 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2287 }
2288
2289 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2290 state != arc_anon && state != arc_l2c_only) {
2291 /* We don't use the L2-only state list. */
2292 multilist_remove(&state->arcs_list[arc_buf_type(hdr)], hdr);
2293 arc_evictable_space_decrement(hdr, state);
2294 }
2295 }
2296
2297 /*
2298 * Remove a reference from this hdr. When the reference transitions from
2299 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2300 * list making it eligible for eviction.
2301 */
2302 static int
remove_reference(arc_buf_hdr_t * hdr,const void * tag)2303 remove_reference(arc_buf_hdr_t *hdr, const void *tag)
2304 {
2305 int cnt;
2306 arc_state_t *state = hdr->b_l1hdr.b_state;
2307
2308 ASSERT(HDR_HAS_L1HDR(hdr));
2309 ASSERT(state == arc_anon || MUTEX_HELD(HDR_LOCK(hdr)));
2310 ASSERT(!GHOST_STATE(state)); /* arc_l2c_only counts as a ghost. */
2311
2312 if ((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) != 0)
2313 return (cnt);
2314
2315 if (state == arc_anon) {
2316 arc_hdr_destroy(hdr);
2317 return (0);
2318 }
2319 if (state == arc_uncached && !HDR_PREFETCH(hdr)) {
2320 arc_change_state(arc_anon, hdr);
2321 arc_hdr_destroy(hdr);
2322 return (0);
2323 }
2324 multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr);
2325 arc_evictable_space_increment(hdr, state);
2326 return (0);
2327 }
2328
2329 /*
2330 * Returns detailed information about a specific arc buffer. When the
2331 * state_index argument is set the function will calculate the arc header
2332 * list position for its arc state. Since this requires a linear traversal
2333 * callers are strongly encourage not to do this. However, it can be helpful
2334 * for targeted analysis so the functionality is provided.
2335 */
2336 void
arc_buf_info(arc_buf_t * ab,arc_buf_info_t * abi,int state_index)2337 arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index)
2338 {
2339 (void) state_index;
2340 arc_buf_hdr_t *hdr = ab->b_hdr;
2341 l1arc_buf_hdr_t *l1hdr = NULL;
2342 l2arc_buf_hdr_t *l2hdr = NULL;
2343 arc_state_t *state = NULL;
2344
2345 memset(abi, 0, sizeof (arc_buf_info_t));
2346
2347 if (hdr == NULL)
2348 return;
2349
2350 abi->abi_flags = hdr->b_flags;
2351
2352 if (HDR_HAS_L1HDR(hdr)) {
2353 l1hdr = &hdr->b_l1hdr;
2354 state = l1hdr->b_state;
2355 }
2356 if (HDR_HAS_L2HDR(hdr))
2357 l2hdr = &hdr->b_l2hdr;
2358
2359 if (l1hdr) {
2360 abi->abi_bufcnt = 0;
2361 for (arc_buf_t *buf = l1hdr->b_buf; buf; buf = buf->b_next)
2362 abi->abi_bufcnt++;
2363 abi->abi_access = l1hdr->b_arc_access;
2364 abi->abi_mru_hits = l1hdr->b_mru_hits;
2365 abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits;
2366 abi->abi_mfu_hits = l1hdr->b_mfu_hits;
2367 abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits;
2368 abi->abi_holds = zfs_refcount_count(&l1hdr->b_refcnt);
2369 }
2370
2371 if (l2hdr) {
2372 abi->abi_l2arc_dattr = l2hdr->b_daddr;
2373 abi->abi_l2arc_hits = l2hdr->b_hits;
2374 }
2375
2376 abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON;
2377 abi->abi_state_contents = arc_buf_type(hdr);
2378 abi->abi_size = arc_hdr_size(hdr);
2379 }
2380
2381 /*
2382 * Move the supplied buffer to the indicated state. The hash lock
2383 * for the buffer must be held by the caller.
2384 */
2385 static void
arc_change_state(arc_state_t * new_state,arc_buf_hdr_t * hdr)2386 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr)
2387 {
2388 arc_state_t *old_state;
2389 int64_t refcnt;
2390 boolean_t update_old, update_new;
2391 arc_buf_contents_t type = arc_buf_type(hdr);
2392
2393 /*
2394 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2395 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2396 * L1 hdr doesn't always exist when we change state to arc_anon before
2397 * destroying a header, in which case reallocating to add the L1 hdr is
2398 * pointless.
2399 */
2400 if (HDR_HAS_L1HDR(hdr)) {
2401 old_state = hdr->b_l1hdr.b_state;
2402 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2403 update_old = (hdr->b_l1hdr.b_buf != NULL ||
2404 hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
2405
2406 IMPLY(GHOST_STATE(old_state), hdr->b_l1hdr.b_buf == NULL);
2407 IMPLY(GHOST_STATE(new_state), hdr->b_l1hdr.b_buf == NULL);
2408 IMPLY(old_state == arc_anon, hdr->b_l1hdr.b_buf == NULL ||
2409 ARC_BUF_LAST(hdr->b_l1hdr.b_buf));
2410 } else {
2411 old_state = arc_l2c_only;
2412 refcnt = 0;
2413 update_old = B_FALSE;
2414 }
2415 update_new = update_old;
2416 if (GHOST_STATE(old_state))
2417 update_old = B_TRUE;
2418 if (GHOST_STATE(new_state))
2419 update_new = B_TRUE;
2420
2421 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
2422 ASSERT3P(new_state, !=, old_state);
2423
2424 /*
2425 * If this buffer is evictable, transfer it from the
2426 * old state list to the new state list.
2427 */
2428 if (refcnt == 0) {
2429 if (old_state != arc_anon && old_state != arc_l2c_only) {
2430 ASSERT(HDR_HAS_L1HDR(hdr));
2431 /* remove_reference() saves on insert. */
2432 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2433 multilist_remove(&old_state->arcs_list[type],
2434 hdr);
2435 arc_evictable_space_decrement(hdr, old_state);
2436 }
2437 }
2438 if (new_state != arc_anon && new_state != arc_l2c_only) {
2439 /*
2440 * An L1 header always exists here, since if we're
2441 * moving to some L1-cached state (i.e. not l2c_only or
2442 * anonymous), we realloc the header to add an L1hdr
2443 * beforehand.
2444 */
2445 ASSERT(HDR_HAS_L1HDR(hdr));
2446 multilist_insert(&new_state->arcs_list[type], hdr);
2447 arc_evictable_space_increment(hdr, new_state);
2448 }
2449 }
2450
2451 ASSERT(!HDR_EMPTY(hdr));
2452 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2453 buf_hash_remove(hdr);
2454
2455 /* adjust state sizes (ignore arc_l2c_only) */
2456
2457 if (update_new && new_state != arc_l2c_only) {
2458 ASSERT(HDR_HAS_L1HDR(hdr));
2459 if (GHOST_STATE(new_state)) {
2460
2461 /*
2462 * When moving a header to a ghost state, we first
2463 * remove all arc buffers. Thus, we'll have no arc
2464 * buffer to use for the reference. As a result, we
2465 * use the arc header pointer for the reference.
2466 */
2467 (void) zfs_refcount_add_many(
2468 &new_state->arcs_size[type],
2469 HDR_GET_LSIZE(hdr), hdr);
2470 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2471 ASSERT(!HDR_HAS_RABD(hdr));
2472 } else {
2473
2474 /*
2475 * Each individual buffer holds a unique reference,
2476 * thus we must remove each of these references one
2477 * at a time.
2478 */
2479 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2480 buf = buf->b_next) {
2481
2482 /*
2483 * When the arc_buf_t is sharing the data
2484 * block with the hdr, the owner of the
2485 * reference belongs to the hdr. Only
2486 * add to the refcount if the arc_buf_t is
2487 * not shared.
2488 */
2489 if (ARC_BUF_SHARED(buf))
2490 continue;
2491
2492 (void) zfs_refcount_add_many(
2493 &new_state->arcs_size[type],
2494 arc_buf_size(buf), buf);
2495 }
2496
2497 if (hdr->b_l1hdr.b_pabd != NULL) {
2498 (void) zfs_refcount_add_many(
2499 &new_state->arcs_size[type],
2500 arc_hdr_size(hdr), hdr);
2501 }
2502
2503 if (HDR_HAS_RABD(hdr)) {
2504 (void) zfs_refcount_add_many(
2505 &new_state->arcs_size[type],
2506 HDR_GET_PSIZE(hdr), hdr);
2507 }
2508 }
2509 }
2510
2511 if (update_old && old_state != arc_l2c_only) {
2512 ASSERT(HDR_HAS_L1HDR(hdr));
2513 if (GHOST_STATE(old_state)) {
2514 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2515 ASSERT(!HDR_HAS_RABD(hdr));
2516
2517 /*
2518 * When moving a header off of a ghost state,
2519 * the header will not contain any arc buffers.
2520 * We use the arc header pointer for the reference
2521 * which is exactly what we did when we put the
2522 * header on the ghost state.
2523 */
2524
2525 (void) zfs_refcount_remove_many(
2526 &old_state->arcs_size[type],
2527 HDR_GET_LSIZE(hdr), hdr);
2528 } else {
2529
2530 /*
2531 * Each individual buffer holds a unique reference,
2532 * thus we must remove each of these references one
2533 * at a time.
2534 */
2535 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2536 buf = buf->b_next) {
2537
2538 /*
2539 * When the arc_buf_t is sharing the data
2540 * block with the hdr, the owner of the
2541 * reference belongs to the hdr. Only
2542 * add to the refcount if the arc_buf_t is
2543 * not shared.
2544 */
2545 if (ARC_BUF_SHARED(buf))
2546 continue;
2547
2548 (void) zfs_refcount_remove_many(
2549 &old_state->arcs_size[type],
2550 arc_buf_size(buf), buf);
2551 }
2552 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2553 HDR_HAS_RABD(hdr));
2554
2555 if (hdr->b_l1hdr.b_pabd != NULL) {
2556 (void) zfs_refcount_remove_many(
2557 &old_state->arcs_size[type],
2558 arc_hdr_size(hdr), hdr);
2559 }
2560
2561 if (HDR_HAS_RABD(hdr)) {
2562 (void) zfs_refcount_remove_many(
2563 &old_state->arcs_size[type],
2564 HDR_GET_PSIZE(hdr), hdr);
2565 }
2566 }
2567 }
2568
2569 if (HDR_HAS_L1HDR(hdr)) {
2570 hdr->b_l1hdr.b_state = new_state;
2571
2572 if (HDR_HAS_L2HDR(hdr) && new_state != arc_l2c_only) {
2573 l2arc_hdr_arcstats_decrement_state(hdr);
2574 hdr->b_l2hdr.b_arcs_state = new_state->arcs_state;
2575 l2arc_hdr_arcstats_increment_state(hdr);
2576 }
2577 }
2578 }
2579
2580 void
arc_space_consume(uint64_t space,arc_space_type_t type)2581 arc_space_consume(uint64_t space, arc_space_type_t type)
2582 {
2583 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2584
2585 switch (type) {
2586 default:
2587 break;
2588 case ARC_SPACE_DATA:
2589 ARCSTAT_INCR(arcstat_data_size, space);
2590 break;
2591 case ARC_SPACE_META:
2592 ARCSTAT_INCR(arcstat_metadata_size, space);
2593 break;
2594 case ARC_SPACE_BONUS:
2595 ARCSTAT_INCR(arcstat_bonus_size, space);
2596 break;
2597 case ARC_SPACE_DNODE:
2598 ARCSTAT_INCR(arcstat_dnode_size, space);
2599 break;
2600 case ARC_SPACE_DBUF:
2601 ARCSTAT_INCR(arcstat_dbuf_size, space);
2602 break;
2603 case ARC_SPACE_HDRS:
2604 ARCSTAT_INCR(arcstat_hdr_size, space);
2605 break;
2606 case ARC_SPACE_L2HDRS:
2607 aggsum_add(&arc_sums.arcstat_l2_hdr_size, space);
2608 break;
2609 case ARC_SPACE_ABD_CHUNK_WASTE:
2610 /*
2611 * Note: this includes space wasted by all scatter ABD's, not
2612 * just those allocated by the ARC. But the vast majority of
2613 * scatter ABD's come from the ARC, because other users are
2614 * very short-lived.
2615 */
2616 ARCSTAT_INCR(arcstat_abd_chunk_waste_size, space);
2617 break;
2618 }
2619
2620 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE)
2621 ARCSTAT_INCR(arcstat_meta_used, space);
2622
2623 aggsum_add(&arc_sums.arcstat_size, space);
2624 }
2625
2626 void
arc_space_return(uint64_t space,arc_space_type_t type)2627 arc_space_return(uint64_t space, arc_space_type_t type)
2628 {
2629 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2630
2631 switch (type) {
2632 default:
2633 break;
2634 case ARC_SPACE_DATA:
2635 ARCSTAT_INCR(arcstat_data_size, -space);
2636 break;
2637 case ARC_SPACE_META:
2638 ARCSTAT_INCR(arcstat_metadata_size, -space);
2639 break;
2640 case ARC_SPACE_BONUS:
2641 ARCSTAT_INCR(arcstat_bonus_size, -space);
2642 break;
2643 case ARC_SPACE_DNODE:
2644 ARCSTAT_INCR(arcstat_dnode_size, -space);
2645 break;
2646 case ARC_SPACE_DBUF:
2647 ARCSTAT_INCR(arcstat_dbuf_size, -space);
2648 break;
2649 case ARC_SPACE_HDRS:
2650 ARCSTAT_INCR(arcstat_hdr_size, -space);
2651 break;
2652 case ARC_SPACE_L2HDRS:
2653 aggsum_add(&arc_sums.arcstat_l2_hdr_size, -space);
2654 break;
2655 case ARC_SPACE_ABD_CHUNK_WASTE:
2656 ARCSTAT_INCR(arcstat_abd_chunk_waste_size, -space);
2657 break;
2658 }
2659
2660 if (type != ARC_SPACE_DATA && type != ARC_SPACE_ABD_CHUNK_WASTE)
2661 ARCSTAT_INCR(arcstat_meta_used, -space);
2662
2663 ASSERT(aggsum_compare(&arc_sums.arcstat_size, space) >= 0);
2664 aggsum_add(&arc_sums.arcstat_size, -space);
2665 }
2666
2667 /*
2668 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2669 * with the hdr's b_pabd.
2670 */
2671 static boolean_t
arc_can_share(arc_buf_hdr_t * hdr,arc_buf_t * buf)2672 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2673 {
2674 /*
2675 * The criteria for sharing a hdr's data are:
2676 * 1. the buffer is not encrypted
2677 * 2. the hdr's compression matches the buf's compression
2678 * 3. the hdr doesn't need to be byteswapped
2679 * 4. the hdr isn't already being shared
2680 * 5. the buf is either compressed or it is the last buf in the hdr list
2681 *
2682 * Criterion #5 maintains the invariant that shared uncompressed
2683 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2684 * might ask, "if a compressed buf is allocated first, won't that be the
2685 * last thing in the list?", but in that case it's impossible to create
2686 * a shared uncompressed buf anyway (because the hdr must be compressed
2687 * to have the compressed buf). You might also think that #3 is
2688 * sufficient to make this guarantee, however it's possible
2689 * (specifically in the rare L2ARC write race mentioned in
2690 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2691 * is shareable, but wasn't at the time of its allocation. Rather than
2692 * allow a new shared uncompressed buf to be created and then shuffle
2693 * the list around to make it the last element, this simply disallows
2694 * sharing if the new buf isn't the first to be added.
2695 */
2696 ASSERT3P(buf->b_hdr, ==, hdr);
2697 boolean_t hdr_compressed =
2698 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF;
2699 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2700 return (!ARC_BUF_ENCRYPTED(buf) &&
2701 buf_compressed == hdr_compressed &&
2702 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2703 !HDR_SHARED_DATA(hdr) &&
2704 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2705 }
2706
2707 /*
2708 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2709 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2710 * copy was made successfully, or an error code otherwise.
2711 */
2712 static int
arc_buf_alloc_impl(arc_buf_hdr_t * hdr,spa_t * spa,const zbookmark_phys_t * zb,const void * tag,boolean_t encrypted,boolean_t compressed,boolean_t noauth,boolean_t fill,arc_buf_t ** ret)2713 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
2714 const void *tag, boolean_t encrypted, boolean_t compressed,
2715 boolean_t noauth, boolean_t fill, arc_buf_t **ret)
2716 {
2717 arc_buf_t *buf;
2718 arc_fill_flags_t flags = ARC_FILL_LOCKED;
2719
2720 ASSERT(HDR_HAS_L1HDR(hdr));
2721 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2722 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2723 hdr->b_type == ARC_BUFC_METADATA);
2724 ASSERT3P(ret, !=, NULL);
2725 ASSERT3P(*ret, ==, NULL);
2726 IMPLY(encrypted, compressed);
2727
2728 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2729 buf->b_hdr = hdr;
2730 buf->b_data = NULL;
2731 buf->b_next = hdr->b_l1hdr.b_buf;
2732 buf->b_flags = 0;
2733
2734 add_reference(hdr, tag);
2735
2736 /*
2737 * We're about to change the hdr's b_flags. We must either
2738 * hold the hash_lock or be undiscoverable.
2739 */
2740 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2741
2742 /*
2743 * Only honor requests for compressed bufs if the hdr is actually
2744 * compressed. This must be overridden if the buffer is encrypted since
2745 * encrypted buffers cannot be decompressed.
2746 */
2747 if (encrypted) {
2748 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2749 buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2750 flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2751 } else if (compressed &&
2752 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2753 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2754 flags |= ARC_FILL_COMPRESSED;
2755 }
2756
2757 if (noauth) {
2758 ASSERT0(encrypted);
2759 flags |= ARC_FILL_NOAUTH;
2760 }
2761
2762 /*
2763 * If the hdr's data can be shared then we share the data buffer and
2764 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2765 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new
2766 * buffer to store the buf's data.
2767 *
2768 * There are two additional restrictions here because we're sharing
2769 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2770 * actively involved in an L2ARC write, because if this buf is used by
2771 * an arc_write() then the hdr's data buffer will be released when the
2772 * write completes, even though the L2ARC write might still be using it.
2773 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2774 * need to be ABD-aware. It must be allocated via
2775 * zio_[data_]buf_alloc(), not as a page, because we need to be able
2776 * to abd_release_ownership_of_buf(), which isn't allowed on "linear
2777 * page" buffers because the ABD code needs to handle freeing them
2778 * specially.
2779 */
2780 boolean_t can_share = arc_can_share(hdr, buf) &&
2781 !HDR_L2_WRITING(hdr) &&
2782 hdr->b_l1hdr.b_pabd != NULL &&
2783 abd_is_linear(hdr->b_l1hdr.b_pabd) &&
2784 !abd_is_linear_page(hdr->b_l1hdr.b_pabd);
2785
2786 /* Set up b_data and sharing */
2787 if (can_share) {
2788 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2789 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2790 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2791 } else {
2792 buf->b_data =
2793 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2794 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2795 }
2796 VERIFY3P(buf->b_data, !=, NULL);
2797
2798 hdr->b_l1hdr.b_buf = buf;
2799
2800 /*
2801 * If the user wants the data from the hdr, we need to either copy or
2802 * decompress the data.
2803 */
2804 if (fill) {
2805 ASSERT3P(zb, !=, NULL);
2806 return (arc_buf_fill(buf, spa, zb, flags));
2807 }
2808
2809 return (0);
2810 }
2811
2812 static const char *arc_onloan_tag = "onloan";
2813
2814 static inline void
arc_loaned_bytes_update(int64_t delta)2815 arc_loaned_bytes_update(int64_t delta)
2816 {
2817 atomic_add_64(&arc_loaned_bytes, delta);
2818
2819 /* assert that it did not wrap around */
2820 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2821 }
2822
2823 /*
2824 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2825 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2826 * buffers must be returned to the arc before they can be used by the DMU or
2827 * freed.
2828 */
2829 arc_buf_t *
arc_loan_buf(spa_t * spa,boolean_t is_metadata,int size)2830 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2831 {
2832 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2833 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2834
2835 arc_loaned_bytes_update(arc_buf_size(buf));
2836
2837 return (buf);
2838 }
2839
2840 arc_buf_t *
arc_loan_compressed_buf(spa_t * spa,uint64_t psize,uint64_t lsize,enum zio_compress compression_type,uint8_t complevel)2841 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2842 enum zio_compress compression_type, uint8_t complevel)
2843 {
2844 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2845 psize, lsize, compression_type, complevel);
2846
2847 arc_loaned_bytes_update(arc_buf_size(buf));
2848
2849 return (buf);
2850 }
2851
2852 arc_buf_t *
arc_loan_raw_buf(spa_t * spa,uint64_t dsobj,boolean_t byteorder,const uint8_t * salt,const uint8_t * iv,const uint8_t * mac,dmu_object_type_t ot,uint64_t psize,uint64_t lsize,enum zio_compress compression_type,uint8_t complevel)2853 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
2854 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
2855 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
2856 enum zio_compress compression_type, uint8_t complevel)
2857 {
2858 arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
2859 byteorder, salt, iv, mac, ot, psize, lsize, compression_type,
2860 complevel);
2861
2862 atomic_add_64(&arc_loaned_bytes, psize);
2863 return (buf);
2864 }
2865
2866
2867 /*
2868 * Return a loaned arc buffer to the arc.
2869 */
2870 void
arc_return_buf(arc_buf_t * buf,const void * tag)2871 arc_return_buf(arc_buf_t *buf, const void *tag)
2872 {
2873 arc_buf_hdr_t *hdr = buf->b_hdr;
2874
2875 ASSERT3P(buf->b_data, !=, NULL);
2876 ASSERT(HDR_HAS_L1HDR(hdr));
2877 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2878 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2879
2880 arc_loaned_bytes_update(-arc_buf_size(buf));
2881 }
2882
2883 /* Detach an arc_buf from a dbuf (tag) */
2884 void
arc_loan_inuse_buf(arc_buf_t * buf,const void * tag)2885 arc_loan_inuse_buf(arc_buf_t *buf, const void *tag)
2886 {
2887 arc_buf_hdr_t *hdr = buf->b_hdr;
2888
2889 ASSERT3P(buf->b_data, !=, NULL);
2890 ASSERT(HDR_HAS_L1HDR(hdr));
2891 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2892 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2893
2894 arc_loaned_bytes_update(arc_buf_size(buf));
2895 }
2896
2897 static void
l2arc_free_abd_on_write(abd_t * abd,size_t size,arc_buf_contents_t type)2898 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2899 {
2900 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2901
2902 df->l2df_abd = abd;
2903 df->l2df_size = size;
2904 df->l2df_type = type;
2905 mutex_enter(&l2arc_free_on_write_mtx);
2906 list_insert_head(l2arc_free_on_write, df);
2907 mutex_exit(&l2arc_free_on_write_mtx);
2908 }
2909
2910 static void
arc_hdr_free_on_write(arc_buf_hdr_t * hdr,boolean_t free_rdata)2911 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
2912 {
2913 arc_state_t *state = hdr->b_l1hdr.b_state;
2914 arc_buf_contents_t type = arc_buf_type(hdr);
2915 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
2916
2917 /* protected by hash lock, if in the hash table */
2918 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2919 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2920 ASSERT(state != arc_anon && state != arc_l2c_only);
2921
2922 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2923 size, hdr);
2924 }
2925 (void) zfs_refcount_remove_many(&state->arcs_size[type], size, hdr);
2926 if (type == ARC_BUFC_METADATA) {
2927 arc_space_return(size, ARC_SPACE_META);
2928 } else {
2929 ASSERT(type == ARC_BUFC_DATA);
2930 arc_space_return(size, ARC_SPACE_DATA);
2931 }
2932
2933 if (free_rdata) {
2934 l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
2935 } else {
2936 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2937 }
2938 }
2939
2940 /*
2941 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2942 * data buffer, we transfer the refcount ownership to the hdr and update
2943 * the appropriate kstats.
2944 */
2945 static void
arc_share_buf(arc_buf_hdr_t * hdr,arc_buf_t * buf)2946 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2947 {
2948 ASSERT(arc_can_share(hdr, buf));
2949 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2950 ASSERT(!ARC_BUF_ENCRYPTED(buf));
2951 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2952
2953 /*
2954 * Start sharing the data buffer. We transfer the
2955 * refcount ownership to the hdr since it always owns
2956 * the refcount whenever an arc_buf_t is shared.
2957 */
2958 zfs_refcount_transfer_ownership_many(
2959 &hdr->b_l1hdr.b_state->arcs_size[arc_buf_type(hdr)],
2960 arc_hdr_size(hdr), buf, hdr);
2961 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2962 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2963 HDR_ISTYPE_METADATA(hdr));
2964 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2965 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2966
2967 /*
2968 * Since we've transferred ownership to the hdr we need
2969 * to increment its compressed and uncompressed kstats and
2970 * decrement the overhead size.
2971 */
2972 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2973 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2974 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2975 }
2976
2977 static void
arc_unshare_buf(arc_buf_hdr_t * hdr,arc_buf_t * buf)2978 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2979 {
2980 ASSERT(arc_buf_is_shared(buf));
2981 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2982 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2983
2984 /*
2985 * We are no longer sharing this buffer so we need
2986 * to transfer its ownership to the rightful owner.
2987 */
2988 zfs_refcount_transfer_ownership_many(
2989 &hdr->b_l1hdr.b_state->arcs_size[arc_buf_type(hdr)],
2990 arc_hdr_size(hdr), hdr, buf);
2991 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2992 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
2993 abd_free(hdr->b_l1hdr.b_pabd);
2994 hdr->b_l1hdr.b_pabd = NULL;
2995 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2996
2997 /*
2998 * Since the buffer is no longer shared between
2999 * the arc buf and the hdr, count it as overhead.
3000 */
3001 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3002 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3003 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
3004 }
3005
3006 /*
3007 * Remove an arc_buf_t from the hdr's buf list and return the last
3008 * arc_buf_t on the list. If no buffers remain on the list then return
3009 * NULL.
3010 */
3011 static arc_buf_t *
arc_buf_remove(arc_buf_hdr_t * hdr,arc_buf_t * buf)3012 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3013 {
3014 ASSERT(HDR_HAS_L1HDR(hdr));
3015 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3016
3017 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
3018 arc_buf_t *lastbuf = NULL;
3019
3020 /*
3021 * Remove the buf from the hdr list and locate the last
3022 * remaining buffer on the list.
3023 */
3024 while (*bufp != NULL) {
3025 if (*bufp == buf)
3026 *bufp = buf->b_next;
3027
3028 /*
3029 * If we've removed a buffer in the middle of
3030 * the list then update the lastbuf and update
3031 * bufp.
3032 */
3033 if (*bufp != NULL) {
3034 lastbuf = *bufp;
3035 bufp = &(*bufp)->b_next;
3036 }
3037 }
3038 buf->b_next = NULL;
3039 ASSERT3P(lastbuf, !=, buf);
3040 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
3041
3042 return (lastbuf);
3043 }
3044
3045 /*
3046 * Free up buf->b_data and pull the arc_buf_t off of the arc_buf_hdr_t's
3047 * list and free it.
3048 */
3049 static void
arc_buf_destroy_impl(arc_buf_t * buf)3050 arc_buf_destroy_impl(arc_buf_t *buf)
3051 {
3052 arc_buf_hdr_t *hdr = buf->b_hdr;
3053
3054 /*
3055 * Free up the data associated with the buf but only if we're not
3056 * sharing this with the hdr. If we are sharing it with the hdr, the
3057 * hdr is responsible for doing the free.
3058 */
3059 if (buf->b_data != NULL) {
3060 /*
3061 * We're about to change the hdr's b_flags. We must either
3062 * hold the hash_lock or be undiscoverable.
3063 */
3064 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
3065
3066 arc_cksum_verify(buf);
3067 arc_buf_unwatch(buf);
3068
3069 if (ARC_BUF_SHARED(buf)) {
3070 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
3071 } else {
3072 ASSERT(!arc_buf_is_shared(buf));
3073 uint64_t size = arc_buf_size(buf);
3074 arc_free_data_buf(hdr, buf->b_data, size, buf);
3075 ARCSTAT_INCR(arcstat_overhead_size, -size);
3076 }
3077 buf->b_data = NULL;
3078
3079 /*
3080 * If we have no more encrypted buffers and we've already
3081 * gotten a copy of the decrypted data we can free b_rabd
3082 * to save some space.
3083 */
3084 if (ARC_BUF_ENCRYPTED(buf) && HDR_HAS_RABD(hdr) &&
3085 hdr->b_l1hdr.b_pabd != NULL && !HDR_IO_IN_PROGRESS(hdr)) {
3086 arc_buf_t *b;
3087 for (b = hdr->b_l1hdr.b_buf; b; b = b->b_next) {
3088 if (b != buf && ARC_BUF_ENCRYPTED(b))
3089 break;
3090 }
3091 if (b == NULL)
3092 arc_hdr_free_abd(hdr, B_TRUE);
3093 }
3094 }
3095
3096 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3097
3098 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3099 /*
3100 * If the current arc_buf_t is sharing its data buffer with the
3101 * hdr, then reassign the hdr's b_pabd to share it with the new
3102 * buffer at the end of the list. The shared buffer is always
3103 * the last one on the hdr's buffer list.
3104 *
3105 * There is an equivalent case for compressed bufs, but since
3106 * they aren't guaranteed to be the last buf in the list and
3107 * that is an exceedingly rare case, we just allow that space be
3108 * wasted temporarily. We must also be careful not to share
3109 * encrypted buffers, since they cannot be shared.
3110 */
3111 if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
3112 /* Only one buf can be shared at once */
3113 ASSERT(!arc_buf_is_shared(lastbuf));
3114 /* hdr is uncompressed so can't have compressed buf */
3115 ASSERT(!ARC_BUF_COMPRESSED(lastbuf));
3116
3117 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3118 arc_hdr_free_abd(hdr, B_FALSE);
3119
3120 /*
3121 * We must setup a new shared block between the
3122 * last buffer and the hdr. The data would have
3123 * been allocated by the arc buf so we need to transfer
3124 * ownership to the hdr since it's now being shared.
3125 */
3126 arc_share_buf(hdr, lastbuf);
3127 }
3128 } else if (HDR_SHARED_DATA(hdr)) {
3129 /*
3130 * Uncompressed shared buffers are always at the end
3131 * of the list. Compressed buffers don't have the
3132 * same requirements. This makes it hard to
3133 * simply assert that the lastbuf is shared so
3134 * we rely on the hdr's compression flags to determine
3135 * if we have a compressed, shared buffer.
3136 */
3137 ASSERT3P(lastbuf, !=, NULL);
3138 ASSERT(arc_buf_is_shared(lastbuf) ||
3139 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3140 }
3141
3142 /*
3143 * Free the checksum if we're removing the last uncompressed buf from
3144 * this hdr.
3145 */
3146 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3147 arc_cksum_free(hdr);
3148 }
3149
3150 /* clean up the buf */
3151 buf->b_hdr = NULL;
3152 kmem_cache_free(buf_cache, buf);
3153 }
3154
3155 static void
arc_hdr_alloc_abd(arc_buf_hdr_t * hdr,int alloc_flags)3156 arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, int alloc_flags)
3157 {
3158 uint64_t size;
3159 boolean_t alloc_rdata = ((alloc_flags & ARC_HDR_ALLOC_RDATA) != 0);
3160
3161 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3162 ASSERT(HDR_HAS_L1HDR(hdr));
3163 ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3164 IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3165
3166 if (alloc_rdata) {
3167 size = HDR_GET_PSIZE(hdr);
3168 ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3169 hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr,
3170 alloc_flags);
3171 ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3172 ARCSTAT_INCR(arcstat_raw_size, size);
3173 } else {
3174 size = arc_hdr_size(hdr);
3175 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3176 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr,
3177 alloc_flags);
3178 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3179 }
3180
3181 ARCSTAT_INCR(arcstat_compressed_size, size);
3182 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3183 }
3184
3185 static void
arc_hdr_free_abd(arc_buf_hdr_t * hdr,boolean_t free_rdata)3186 arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3187 {
3188 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3189
3190 ASSERT(HDR_HAS_L1HDR(hdr));
3191 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3192 IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3193
3194 /*
3195 * If the hdr is currently being written to the l2arc then
3196 * we defer freeing the data by adding it to the l2arc_free_on_write
3197 * list. The l2arc will free the data once it's finished
3198 * writing it to the l2arc device.
3199 */
3200 if (HDR_L2_WRITING(hdr)) {
3201 arc_hdr_free_on_write(hdr, free_rdata);
3202 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3203 } else if (free_rdata) {
3204 arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3205 } else {
3206 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr);
3207 }
3208
3209 if (free_rdata) {
3210 hdr->b_crypt_hdr.b_rabd = NULL;
3211 ARCSTAT_INCR(arcstat_raw_size, -size);
3212 } else {
3213 hdr->b_l1hdr.b_pabd = NULL;
3214 }
3215
3216 if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3217 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3218
3219 ARCSTAT_INCR(arcstat_compressed_size, -size);
3220 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3221 }
3222
3223 /*
3224 * Allocate empty anonymous ARC header. The header will get its identity
3225 * assigned and buffers attached later as part of read or write operations.
3226 *
3227 * In case of read arc_read() assigns header its identify (b_dva + b_birth),
3228 * inserts it into ARC hash to become globally visible and allocates physical
3229 * (b_pabd) or raw (b_rabd) ABD buffer to read into from disk. On disk read
3230 * completion arc_read_done() allocates ARC buffer(s) as needed, potentially
3231 * sharing one of them with the physical ABD buffer.
3232 *
3233 * In case of write arc_alloc_buf() allocates ARC buffer to be filled with
3234 * data. Then after compression and/or encryption arc_write_ready() allocates
3235 * and fills (or potentially shares) physical (b_pabd) or raw (b_rabd) ABD
3236 * buffer. On disk write completion arc_write_done() assigns the header its
3237 * new identity (b_dva + b_birth) and inserts into ARC hash.
3238 *
3239 * In case of partial overwrite the old data is read first as described. Then
3240 * arc_release() either allocates new anonymous ARC header and moves the ARC
3241 * buffer to it, or reuses the old ARC header by discarding its identity and
3242 * removing it from ARC hash. After buffer modification normal write process
3243 * follows as described.
3244 */
3245 static arc_buf_hdr_t *
arc_hdr_alloc(uint64_t spa,int32_t psize,int32_t lsize,boolean_t protected,enum zio_compress compression_type,uint8_t complevel,arc_buf_contents_t type)3246 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3247 boolean_t protected, enum zio_compress compression_type, uint8_t complevel,
3248 arc_buf_contents_t type)
3249 {
3250 arc_buf_hdr_t *hdr;
3251
3252 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3253 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3254
3255 ASSERT(HDR_EMPTY(hdr));
3256 #ifdef ZFS_DEBUG
3257 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3258 #endif
3259 HDR_SET_PSIZE(hdr, psize);
3260 HDR_SET_LSIZE(hdr, lsize);
3261 hdr->b_spa = spa;
3262 hdr->b_type = type;
3263 hdr->b_flags = 0;
3264 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3265 arc_hdr_set_compress(hdr, compression_type);
3266 hdr->b_complevel = complevel;
3267 if (protected)
3268 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3269
3270 hdr->b_l1hdr.b_state = arc_anon;
3271 hdr->b_l1hdr.b_arc_access = 0;
3272 hdr->b_l1hdr.b_mru_hits = 0;
3273 hdr->b_l1hdr.b_mru_ghost_hits = 0;
3274 hdr->b_l1hdr.b_mfu_hits = 0;
3275 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
3276 hdr->b_l1hdr.b_buf = NULL;
3277
3278 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3279
3280 return (hdr);
3281 }
3282
3283 /*
3284 * Transition between the two allocation states for the arc_buf_hdr struct.
3285 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3286 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3287 * version is used when a cache buffer is only in the L2ARC in order to reduce
3288 * memory usage.
3289 */
3290 static arc_buf_hdr_t *
arc_hdr_realloc(arc_buf_hdr_t * hdr,kmem_cache_t * old,kmem_cache_t * new)3291 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3292 {
3293 ASSERT(HDR_HAS_L2HDR(hdr));
3294
3295 arc_buf_hdr_t *nhdr;
3296 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3297
3298 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3299 (old == hdr_l2only_cache && new == hdr_full_cache));
3300
3301 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3302
3303 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3304 buf_hash_remove(hdr);
3305
3306 memcpy(nhdr, hdr, HDR_L2ONLY_SIZE);
3307
3308 if (new == hdr_full_cache) {
3309 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3310 /*
3311 * arc_access and arc_change_state need to be aware that a
3312 * header has just come out of L2ARC, so we set its state to
3313 * l2c_only even though it's about to change.
3314 */
3315 nhdr->b_l1hdr.b_state = arc_l2c_only;
3316
3317 /* Verify previous threads set to NULL before freeing */
3318 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3319 ASSERT(!HDR_HAS_RABD(hdr));
3320 } else {
3321 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3322 #ifdef ZFS_DEBUG
3323 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3324 #endif
3325
3326 /*
3327 * If we've reached here, We must have been called from
3328 * arc_evict_hdr(), as such we should have already been
3329 * removed from any ghost list we were previously on
3330 * (which protects us from racing with arc_evict_state),
3331 * thus no locking is needed during this check.
3332 */
3333 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3334
3335 /*
3336 * A buffer must not be moved into the arc_l2c_only
3337 * state if it's not finished being written out to the
3338 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3339 * might try to be accessed, even though it was removed.
3340 */
3341 VERIFY(!HDR_L2_WRITING(hdr));
3342 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3343 ASSERT(!HDR_HAS_RABD(hdr));
3344
3345 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3346 }
3347 /*
3348 * The header has been reallocated so we need to re-insert it into any
3349 * lists it was on.
3350 */
3351 (void) buf_hash_insert(nhdr, NULL);
3352
3353 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3354
3355 mutex_enter(&dev->l2ad_mtx);
3356
3357 /*
3358 * We must place the realloc'ed header back into the list at
3359 * the same spot. Otherwise, if it's placed earlier in the list,
3360 * l2arc_write_buffers() could find it during the function's
3361 * write phase, and try to write it out to the l2arc.
3362 */
3363 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3364 list_remove(&dev->l2ad_buflist, hdr);
3365
3366 mutex_exit(&dev->l2ad_mtx);
3367
3368 /*
3369 * Since we're using the pointer address as the tag when
3370 * incrementing and decrementing the l2ad_alloc refcount, we
3371 * must remove the old pointer (that we're about to destroy) and
3372 * add the new pointer to the refcount. Otherwise we'd remove
3373 * the wrong pointer address when calling arc_hdr_destroy() later.
3374 */
3375
3376 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
3377 arc_hdr_size(hdr), hdr);
3378 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
3379 arc_hdr_size(nhdr), nhdr);
3380
3381 buf_discard_identity(hdr);
3382 kmem_cache_free(old, hdr);
3383
3384 return (nhdr);
3385 }
3386
3387 /*
3388 * This function is used by the send / receive code to convert a newly
3389 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3390 * is also used to allow the root objset block to be updated without altering
3391 * its embedded MACs. Both block types will always be uncompressed so we do not
3392 * have to worry about compression type or psize.
3393 */
3394 void
arc_convert_to_raw(arc_buf_t * buf,uint64_t dsobj,boolean_t byteorder,dmu_object_type_t ot,const uint8_t * salt,const uint8_t * iv,const uint8_t * mac)3395 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3396 dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3397 const uint8_t *mac)
3398 {
3399 arc_buf_hdr_t *hdr = buf->b_hdr;
3400
3401 ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3402 ASSERT(HDR_HAS_L1HDR(hdr));
3403 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3404
3405 buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3406 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3407 hdr->b_crypt_hdr.b_dsobj = dsobj;
3408 hdr->b_crypt_hdr.b_ot = ot;
3409 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3410 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3411 if (!arc_hdr_has_uncompressed_buf(hdr))
3412 arc_cksum_free(hdr);
3413
3414 if (salt != NULL)
3415 memcpy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
3416 if (iv != NULL)
3417 memcpy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
3418 if (mac != NULL)
3419 memcpy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
3420 }
3421
3422 /*
3423 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3424 * The buf is returned thawed since we expect the consumer to modify it.
3425 */
3426 arc_buf_t *
arc_alloc_buf(spa_t * spa,const void * tag,arc_buf_contents_t type,int32_t size)3427 arc_alloc_buf(spa_t *spa, const void *tag, arc_buf_contents_t type,
3428 int32_t size)
3429 {
3430 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3431 B_FALSE, ZIO_COMPRESS_OFF, 0, type);
3432
3433 arc_buf_t *buf = NULL;
3434 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
3435 B_FALSE, B_FALSE, &buf));
3436 arc_buf_thaw(buf);
3437
3438 return (buf);
3439 }
3440
3441 /*
3442 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3443 * for bufs containing metadata.
3444 */
3445 arc_buf_t *
arc_alloc_compressed_buf(spa_t * spa,const void * tag,uint64_t psize,uint64_t lsize,enum zio_compress compression_type,uint8_t complevel)3446 arc_alloc_compressed_buf(spa_t *spa, const void *tag, uint64_t psize,
3447 uint64_t lsize, enum zio_compress compression_type, uint8_t complevel)
3448 {
3449 ASSERT3U(lsize, >, 0);
3450 ASSERT3U(lsize, >=, psize);
3451 ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3452 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3453
3454 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3455 B_FALSE, compression_type, complevel, ARC_BUFC_DATA);
3456
3457 arc_buf_t *buf = NULL;
3458 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
3459 B_TRUE, B_FALSE, B_FALSE, &buf));
3460 arc_buf_thaw(buf);
3461
3462 /*
3463 * To ensure that the hdr has the correct data in it if we call
3464 * arc_untransform() on this buf before it's been written to disk,
3465 * it's easiest if we just set up sharing between the buf and the hdr.
3466 */
3467 arc_share_buf(hdr, buf);
3468
3469 return (buf);
3470 }
3471
3472 arc_buf_t *
arc_alloc_raw_buf(spa_t * spa,const void * tag,uint64_t dsobj,boolean_t byteorder,const uint8_t * salt,const uint8_t * iv,const uint8_t * mac,dmu_object_type_t ot,uint64_t psize,uint64_t lsize,enum zio_compress compression_type,uint8_t complevel)3473 arc_alloc_raw_buf(spa_t *spa, const void *tag, uint64_t dsobj,
3474 boolean_t byteorder, const uint8_t *salt, const uint8_t *iv,
3475 const uint8_t *mac, dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3476 enum zio_compress compression_type, uint8_t complevel)
3477 {
3478 arc_buf_hdr_t *hdr;
3479 arc_buf_t *buf;
3480 arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3481 ARC_BUFC_METADATA : ARC_BUFC_DATA;
3482
3483 ASSERT3U(lsize, >, 0);
3484 ASSERT3U(lsize, >=, psize);
3485 ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3486 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3487
3488 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3489 compression_type, complevel, type);
3490
3491 hdr->b_crypt_hdr.b_dsobj = dsobj;
3492 hdr->b_crypt_hdr.b_ot = ot;
3493 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3494 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3495 memcpy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
3496 memcpy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
3497 memcpy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
3498
3499 /*
3500 * This buffer will be considered encrypted even if the ot is not an
3501 * encrypted type. It will become authenticated instead in
3502 * arc_write_ready().
3503 */
3504 buf = NULL;
3505 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
3506 B_FALSE, B_FALSE, &buf));
3507 arc_buf_thaw(buf);
3508
3509 return (buf);
3510 }
3511
3512 static void
l2arc_hdr_arcstats_update(arc_buf_hdr_t * hdr,boolean_t incr,boolean_t state_only)3513 l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
3514 boolean_t state_only)
3515 {
3516 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3517 l2arc_dev_t *dev = l2hdr->b_dev;
3518 uint64_t lsize = HDR_GET_LSIZE(hdr);
3519 uint64_t psize = HDR_GET_PSIZE(hdr);
3520 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3521 arc_buf_contents_t type = hdr->b_type;
3522 int64_t lsize_s;
3523 int64_t psize_s;
3524 int64_t asize_s;
3525
3526 if (incr) {
3527 lsize_s = lsize;
3528 psize_s = psize;
3529 asize_s = asize;
3530 } else {
3531 lsize_s = -lsize;
3532 psize_s = -psize;
3533 asize_s = -asize;
3534 }
3535
3536 /* If the buffer is a prefetch, count it as such. */
3537 if (HDR_PREFETCH(hdr)) {
3538 ARCSTAT_INCR(arcstat_l2_prefetch_asize, asize_s);
3539 } else {
3540 /*
3541 * We use the value stored in the L2 header upon initial
3542 * caching in L2ARC. This value will be updated in case
3543 * an MRU/MRU_ghost buffer transitions to MFU but the L2ARC
3544 * metadata (log entry) cannot currently be updated. Having
3545 * the ARC state in the L2 header solves the problem of a
3546 * possibly absent L1 header (apparent in buffers restored
3547 * from persistent L2ARC).
3548 */
3549 switch (hdr->b_l2hdr.b_arcs_state) {
3550 case ARC_STATE_MRU_GHOST:
3551 case ARC_STATE_MRU:
3552 ARCSTAT_INCR(arcstat_l2_mru_asize, asize_s);
3553 break;
3554 case ARC_STATE_MFU_GHOST:
3555 case ARC_STATE_MFU:
3556 ARCSTAT_INCR(arcstat_l2_mfu_asize, asize_s);
3557 break;
3558 default:
3559 break;
3560 }
3561 }
3562
3563 if (state_only)
3564 return;
3565
3566 ARCSTAT_INCR(arcstat_l2_psize, psize_s);
3567 ARCSTAT_INCR(arcstat_l2_lsize, lsize_s);
3568
3569 switch (type) {
3570 case ARC_BUFC_DATA:
3571 ARCSTAT_INCR(arcstat_l2_bufc_data_asize, asize_s);
3572 break;
3573 case ARC_BUFC_METADATA:
3574 ARCSTAT_INCR(arcstat_l2_bufc_metadata_asize, asize_s);
3575 break;
3576 default:
3577 break;
3578 }
3579 }
3580
3581
3582 static void
arc_hdr_l2hdr_destroy(arc_buf_hdr_t * hdr)3583 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3584 {
3585 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3586 l2arc_dev_t *dev = l2hdr->b_dev;
3587 uint64_t psize = HDR_GET_PSIZE(hdr);
3588 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3589
3590 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3591 ASSERT(HDR_HAS_L2HDR(hdr));
3592
3593 list_remove(&dev->l2ad_buflist, hdr);
3594
3595 l2arc_hdr_arcstats_decrement(hdr);
3596 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3597
3598 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3599 hdr);
3600 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3601 }
3602
3603 static void
arc_hdr_destroy(arc_buf_hdr_t * hdr)3604 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3605 {
3606 if (HDR_HAS_L1HDR(hdr)) {
3607 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3608 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3609 }
3610 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3611 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3612
3613 if (HDR_HAS_L2HDR(hdr)) {
3614 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3615 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3616
3617 if (!buflist_held)
3618 mutex_enter(&dev->l2ad_mtx);
3619
3620 /*
3621 * Even though we checked this conditional above, we
3622 * need to check this again now that we have the
3623 * l2ad_mtx. This is because we could be racing with
3624 * another thread calling l2arc_evict() which might have
3625 * destroyed this header's L2 portion as we were waiting
3626 * to acquire the l2ad_mtx. If that happens, we don't
3627 * want to re-destroy the header's L2 portion.
3628 */
3629 if (HDR_HAS_L2HDR(hdr)) {
3630
3631 if (!HDR_EMPTY(hdr))
3632 buf_discard_identity(hdr);
3633
3634 arc_hdr_l2hdr_destroy(hdr);
3635 }
3636
3637 if (!buflist_held)
3638 mutex_exit(&dev->l2ad_mtx);
3639 }
3640
3641 /*
3642 * The header's identify can only be safely discarded once it is no
3643 * longer discoverable. This requires removing it from the hash table
3644 * and the l2arc header list. After this point the hash lock can not
3645 * be used to protect the header.
3646 */
3647 if (!HDR_EMPTY(hdr))
3648 buf_discard_identity(hdr);
3649
3650 if (HDR_HAS_L1HDR(hdr)) {
3651 arc_cksum_free(hdr);
3652
3653 while (hdr->b_l1hdr.b_buf != NULL)
3654 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3655
3656 if (hdr->b_l1hdr.b_pabd != NULL)
3657 arc_hdr_free_abd(hdr, B_FALSE);
3658
3659 if (HDR_HAS_RABD(hdr))
3660 arc_hdr_free_abd(hdr, B_TRUE);
3661 }
3662
3663 ASSERT3P(hdr->b_hash_next, ==, NULL);
3664 if (HDR_HAS_L1HDR(hdr)) {
3665 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3666 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3667 #ifdef ZFS_DEBUG
3668 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3669 #endif
3670 kmem_cache_free(hdr_full_cache, hdr);
3671 } else {
3672 kmem_cache_free(hdr_l2only_cache, hdr);
3673 }
3674 }
3675
3676 void
arc_buf_destroy(arc_buf_t * buf,const void * tag)3677 arc_buf_destroy(arc_buf_t *buf, const void *tag)
3678 {
3679 arc_buf_hdr_t *hdr = buf->b_hdr;
3680
3681 if (hdr->b_l1hdr.b_state == arc_anon) {
3682 ASSERT3P(hdr->b_l1hdr.b_buf, ==, buf);
3683 ASSERT(ARC_BUF_LAST(buf));
3684 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3685 VERIFY0(remove_reference(hdr, tag));
3686 return;
3687 }
3688
3689 kmutex_t *hash_lock = HDR_LOCK(hdr);
3690 mutex_enter(hash_lock);
3691
3692 ASSERT3P(hdr, ==, buf->b_hdr);
3693 ASSERT3P(hdr->b_l1hdr.b_buf, !=, NULL);
3694 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3695 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3696 ASSERT3P(buf->b_data, !=, NULL);
3697
3698 arc_buf_destroy_impl(buf);
3699 (void) remove_reference(hdr, tag);
3700 mutex_exit(hash_lock);
3701 }
3702
3703 /*
3704 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3705 * state of the header is dependent on its state prior to entering this
3706 * function. The following transitions are possible:
3707 *
3708 * - arc_mru -> arc_mru_ghost
3709 * - arc_mfu -> arc_mfu_ghost
3710 * - arc_mru_ghost -> arc_l2c_only
3711 * - arc_mru_ghost -> deleted
3712 * - arc_mfu_ghost -> arc_l2c_only
3713 * - arc_mfu_ghost -> deleted
3714 * - arc_uncached -> deleted
3715 *
3716 * Return total size of evicted data buffers for eviction progress tracking.
3717 * When evicting from ghost states return logical buffer size to make eviction
3718 * progress at the same (or at least comparable) rate as from non-ghost states.
3719 *
3720 * Return *real_evicted for actual ARC size reduction to wake up threads
3721 * waiting for it. For non-ghost states it includes size of evicted data
3722 * buffers (the headers are not freed there). For ghost states it includes
3723 * only the evicted headers size.
3724 */
3725 static int64_t
arc_evict_hdr(arc_buf_hdr_t * hdr,uint64_t * real_evicted)3726 arc_evict_hdr(arc_buf_hdr_t *hdr, uint64_t *real_evicted)
3727 {
3728 arc_state_t *evicted_state, *state;
3729 int64_t bytes_evicted = 0;
3730 uint_t min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3731 arc_min_prescient_prefetch_ms : arc_min_prefetch_ms;
3732
3733 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3734 ASSERT(HDR_HAS_L1HDR(hdr));
3735 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3736 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3737 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3738
3739 *real_evicted = 0;
3740 state = hdr->b_l1hdr.b_state;
3741 if (GHOST_STATE(state)) {
3742
3743 /*
3744 * l2arc_write_buffers() relies on a header's L1 portion
3745 * (i.e. its b_pabd field) during it's write phase.
3746 * Thus, we cannot push a header onto the arc_l2c_only
3747 * state (removing its L1 piece) until the header is
3748 * done being written to the l2arc.
3749 */
3750 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3751 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3752 return (bytes_evicted);
3753 }
3754
3755 ARCSTAT_BUMP(arcstat_deleted);
3756 bytes_evicted += HDR_GET_LSIZE(hdr);
3757
3758 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3759
3760 if (HDR_HAS_L2HDR(hdr)) {
3761 ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3762 ASSERT(!HDR_HAS_RABD(hdr));
3763 /*
3764 * This buffer is cached on the 2nd Level ARC;
3765 * don't destroy the header.
3766 */
3767 arc_change_state(arc_l2c_only, hdr);
3768 /*
3769 * dropping from L1+L2 cached to L2-only,
3770 * realloc to remove the L1 header.
3771 */
3772 (void) arc_hdr_realloc(hdr, hdr_full_cache,
3773 hdr_l2only_cache);
3774 *real_evicted += HDR_FULL_SIZE - HDR_L2ONLY_SIZE;
3775 } else {
3776 arc_change_state(arc_anon, hdr);
3777 arc_hdr_destroy(hdr);
3778 *real_evicted += HDR_FULL_SIZE;
3779 }
3780 return (bytes_evicted);
3781 }
3782
3783 ASSERT(state == arc_mru || state == arc_mfu || state == arc_uncached);
3784 evicted_state = (state == arc_uncached) ? arc_anon :
3785 ((state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost);
3786
3787 /* prefetch buffers have a minimum lifespan */
3788 if ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3789 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
3790 MSEC_TO_TICK(min_lifetime)) {
3791 ARCSTAT_BUMP(arcstat_evict_skip);
3792 return (bytes_evicted);
3793 }
3794
3795 if (HDR_HAS_L2HDR(hdr)) {
3796 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3797 } else {
3798 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3799 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3800 HDR_GET_LSIZE(hdr));
3801
3802 switch (state->arcs_state) {
3803 case ARC_STATE_MRU:
3804 ARCSTAT_INCR(
3805 arcstat_evict_l2_eligible_mru,
3806 HDR_GET_LSIZE(hdr));
3807 break;
3808 case ARC_STATE_MFU:
3809 ARCSTAT_INCR(
3810 arcstat_evict_l2_eligible_mfu,
3811 HDR_GET_LSIZE(hdr));
3812 break;
3813 default:
3814 break;
3815 }
3816 } else {
3817 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3818 HDR_GET_LSIZE(hdr));
3819 }
3820 }
3821
3822 bytes_evicted += arc_hdr_size(hdr);
3823 *real_evicted += arc_hdr_size(hdr);
3824
3825 /*
3826 * If this hdr is being evicted and has a compressed buffer then we
3827 * discard it here before we change states. This ensures that the
3828 * accounting is updated correctly in arc_free_data_impl().
3829 */
3830 if (hdr->b_l1hdr.b_pabd != NULL)
3831 arc_hdr_free_abd(hdr, B_FALSE);
3832
3833 if (HDR_HAS_RABD(hdr))
3834 arc_hdr_free_abd(hdr, B_TRUE);
3835
3836 arc_change_state(evicted_state, hdr);
3837 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3838 if (evicted_state == arc_anon) {
3839 arc_hdr_destroy(hdr);
3840 *real_evicted += HDR_FULL_SIZE;
3841 } else {
3842 ASSERT(HDR_IN_HASH_TABLE(hdr));
3843 }
3844
3845 return (bytes_evicted);
3846 }
3847
3848 static void
arc_set_need_free(void)3849 arc_set_need_free(void)
3850 {
3851 ASSERT(MUTEX_HELD(&arc_evict_lock));
3852 int64_t remaining = arc_free_memory() - arc_sys_free / 2;
3853 arc_evict_waiter_t *aw = list_tail(&arc_evict_waiters);
3854 if (aw == NULL) {
3855 arc_need_free = MAX(-remaining, 0);
3856 } else {
3857 arc_need_free =
3858 MAX(-remaining, (int64_t)(aw->aew_count - arc_evict_count));
3859 }
3860 }
3861
3862 static uint64_t
arc_evict_state_impl(multilist_t * ml,int idx,arc_buf_hdr_t * marker,uint64_t spa,uint64_t bytes)3863 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3864 uint64_t spa, uint64_t bytes)
3865 {
3866 multilist_sublist_t *mls;
3867 uint64_t bytes_evicted = 0, real_evicted = 0;
3868 arc_buf_hdr_t *hdr;
3869 kmutex_t *hash_lock;
3870 uint_t evict_count = zfs_arc_evict_batch_limit;
3871
3872 ASSERT3P(marker, !=, NULL);
3873
3874 mls = multilist_sublist_lock_idx(ml, idx);
3875
3876 for (hdr = multilist_sublist_prev(mls, marker); likely(hdr != NULL);
3877 hdr = multilist_sublist_prev(mls, marker)) {
3878 if ((evict_count == 0) || (bytes_evicted >= bytes))
3879 break;
3880
3881 /*
3882 * To keep our iteration location, move the marker
3883 * forward. Since we're not holding hdr's hash lock, we
3884 * must be very careful and not remove 'hdr' from the
3885 * sublist. Otherwise, other consumers might mistake the
3886 * 'hdr' as not being on a sublist when they call the
3887 * multilist_link_active() function (they all rely on
3888 * the hash lock protecting concurrent insertions and
3889 * removals). multilist_sublist_move_forward() was
3890 * specifically implemented to ensure this is the case
3891 * (only 'marker' will be removed and re-inserted).
3892 */
3893 multilist_sublist_move_forward(mls, marker);
3894
3895 /*
3896 * The only case where the b_spa field should ever be
3897 * zero, is the marker headers inserted by
3898 * arc_evict_state(). It's possible for multiple threads
3899 * to be calling arc_evict_state() concurrently (e.g.
3900 * dsl_pool_close() and zio_inject_fault()), so we must
3901 * skip any markers we see from these other threads.
3902 */
3903 if (hdr->b_spa == 0)
3904 continue;
3905
3906 /* we're only interested in evicting buffers of a certain spa */
3907 if (spa != 0 && hdr->b_spa != spa) {
3908 ARCSTAT_BUMP(arcstat_evict_skip);
3909 continue;
3910 }
3911
3912 hash_lock = HDR_LOCK(hdr);
3913
3914 /*
3915 * We aren't calling this function from any code path
3916 * that would already be holding a hash lock, so we're
3917 * asserting on this assumption to be defensive in case
3918 * this ever changes. Without this check, it would be
3919 * possible to incorrectly increment arcstat_mutex_miss
3920 * below (e.g. if the code changed such that we called
3921 * this function with a hash lock held).
3922 */
3923 ASSERT(!MUTEX_HELD(hash_lock));
3924
3925 if (mutex_tryenter(hash_lock)) {
3926 uint64_t revicted;
3927 uint64_t evicted = arc_evict_hdr(hdr, &revicted);
3928 mutex_exit(hash_lock);
3929
3930 bytes_evicted += evicted;
3931 real_evicted += revicted;
3932
3933 /*
3934 * If evicted is zero, arc_evict_hdr() must have
3935 * decided to skip this header, don't increment
3936 * evict_count in this case.
3937 */
3938 if (evicted != 0)
3939 evict_count--;
3940
3941 } else {
3942 ARCSTAT_BUMP(arcstat_mutex_miss);
3943 }
3944 }
3945
3946 multilist_sublist_unlock(mls);
3947
3948 /*
3949 * Increment the count of evicted bytes, and wake up any threads that
3950 * are waiting for the count to reach this value. Since the list is
3951 * ordered by ascending aew_count, we pop off the beginning of the
3952 * list until we reach the end, or a waiter that's past the current
3953 * "count". Doing this outside the loop reduces the number of times
3954 * we need to acquire the global arc_evict_lock.
3955 *
3956 * Only wake when there's sufficient free memory in the system
3957 * (specifically, arc_sys_free/2, which by default is a bit more than
3958 * 1/64th of RAM). See the comments in arc_wait_for_eviction().
3959 */
3960 mutex_enter(&arc_evict_lock);
3961 arc_evict_count += real_evicted;
3962
3963 if (arc_free_memory() > arc_sys_free / 2) {
3964 arc_evict_waiter_t *aw;
3965 while ((aw = list_head(&arc_evict_waiters)) != NULL &&
3966 aw->aew_count <= arc_evict_count) {
3967 list_remove(&arc_evict_waiters, aw);
3968 cv_broadcast(&aw->aew_cv);
3969 }
3970 }
3971 arc_set_need_free();
3972 mutex_exit(&arc_evict_lock);
3973
3974 /*
3975 * If the ARC size is reduced from arc_c_max to arc_c_min (especially
3976 * if the average cached block is small), eviction can be on-CPU for
3977 * many seconds. To ensure that other threads that may be bound to
3978 * this CPU are able to make progress, make a voluntary preemption
3979 * call here.
3980 */
3981 kpreempt(KPREEMPT_SYNC);
3982
3983 return (bytes_evicted);
3984 }
3985
3986 static arc_buf_hdr_t *
arc_state_alloc_marker(void)3987 arc_state_alloc_marker(void)
3988 {
3989 arc_buf_hdr_t *marker = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3990
3991 /*
3992 * A b_spa of 0 is used to indicate that this header is
3993 * a marker. This fact is used in arc_evict_state_impl().
3994 */
3995 marker->b_spa = 0;
3996
3997 return (marker);
3998 }
3999
4000 static void
arc_state_free_marker(arc_buf_hdr_t * marker)4001 arc_state_free_marker(arc_buf_hdr_t *marker)
4002 {
4003 kmem_cache_free(hdr_full_cache, marker);
4004 }
4005
4006 /*
4007 * Allocate an array of buffer headers used as placeholders during arc state
4008 * eviction.
4009 */
4010 static arc_buf_hdr_t **
arc_state_alloc_markers(int count)4011 arc_state_alloc_markers(int count)
4012 {
4013 arc_buf_hdr_t **markers;
4014
4015 markers = kmem_zalloc(sizeof (*markers) * count, KM_SLEEP);
4016 for (int i = 0; i < count; i++)
4017 markers[i] = arc_state_alloc_marker();
4018 return (markers);
4019 }
4020
4021 static void
arc_state_free_markers(arc_buf_hdr_t ** markers,int count)4022 arc_state_free_markers(arc_buf_hdr_t **markers, int count)
4023 {
4024 for (int i = 0; i < count; i++)
4025 arc_state_free_marker(markers[i]);
4026 kmem_free(markers, sizeof (*markers) * count);
4027 }
4028
4029 /*
4030 * Evict buffers from the given arc state, until we've removed the
4031 * specified number of bytes. Move the removed buffers to the
4032 * appropriate evict state.
4033 *
4034 * This function makes a "best effort". It skips over any buffers
4035 * it can't get a hash_lock on, and so, may not catch all candidates.
4036 * It may also return without evicting as much space as requested.
4037 *
4038 * If bytes is specified using the special value ARC_EVICT_ALL, this
4039 * will evict all available (i.e. unlocked and evictable) buffers from
4040 * the given arc state; which is used by arc_flush().
4041 */
4042 static uint64_t
arc_evict_state(arc_state_t * state,arc_buf_contents_t type,uint64_t spa,uint64_t bytes)4043 arc_evict_state(arc_state_t *state, arc_buf_contents_t type, uint64_t spa,
4044 uint64_t bytes)
4045 {
4046 uint64_t total_evicted = 0;
4047 multilist_t *ml = &state->arcs_list[type];
4048 int num_sublists;
4049 arc_buf_hdr_t **markers;
4050
4051 num_sublists = multilist_get_num_sublists(ml);
4052
4053 /*
4054 * If we've tried to evict from each sublist, made some
4055 * progress, but still have not hit the target number of bytes
4056 * to evict, we want to keep trying. The markers allow us to
4057 * pick up where we left off for each individual sublist, rather
4058 * than starting from the tail each time.
4059 */
4060 if (zthr_iscurthread(arc_evict_zthr)) {
4061 markers = arc_state_evict_markers;
4062 ASSERT3S(num_sublists, <=, arc_state_evict_marker_count);
4063 } else {
4064 markers = arc_state_alloc_markers(num_sublists);
4065 }
4066 for (int i = 0; i < num_sublists; i++) {
4067 multilist_sublist_t *mls;
4068
4069 mls = multilist_sublist_lock_idx(ml, i);
4070 multilist_sublist_insert_tail(mls, markers[i]);
4071 multilist_sublist_unlock(mls);
4072 }
4073
4074 /*
4075 * While we haven't hit our target number of bytes to evict, or
4076 * we're evicting all available buffers.
4077 */
4078 while (total_evicted < bytes) {
4079 int sublist_idx = multilist_get_random_index(ml);
4080 uint64_t scan_evicted = 0;
4081
4082 /*
4083 * Start eviction using a randomly selected sublist,
4084 * this is to try and evenly balance eviction across all
4085 * sublists. Always starting at the same sublist
4086 * (e.g. index 0) would cause evictions to favor certain
4087 * sublists over others.
4088 */
4089 for (int i = 0; i < num_sublists; i++) {
4090 uint64_t bytes_remaining;
4091 uint64_t bytes_evicted;
4092
4093 if (total_evicted < bytes)
4094 bytes_remaining = bytes - total_evicted;
4095 else
4096 break;
4097
4098 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4099 markers[sublist_idx], spa, bytes_remaining);
4100
4101 scan_evicted += bytes_evicted;
4102 total_evicted += bytes_evicted;
4103
4104 /* we've reached the end, wrap to the beginning */
4105 if (++sublist_idx >= num_sublists)
4106 sublist_idx = 0;
4107 }
4108
4109 /*
4110 * If we didn't evict anything during this scan, we have
4111 * no reason to believe we'll evict more during another
4112 * scan, so break the loop.
4113 */
4114 if (scan_evicted == 0) {
4115 /* This isn't possible, let's make that obvious */
4116 ASSERT3S(bytes, !=, 0);
4117
4118 /*
4119 * When bytes is ARC_EVICT_ALL, the only way to
4120 * break the loop is when scan_evicted is zero.
4121 * In that case, we actually have evicted enough,
4122 * so we don't want to increment the kstat.
4123 */
4124 if (bytes != ARC_EVICT_ALL) {
4125 ASSERT3S(total_evicted, <, bytes);
4126 ARCSTAT_BUMP(arcstat_evict_not_enough);
4127 }
4128
4129 break;
4130 }
4131 }
4132
4133 for (int i = 0; i < num_sublists; i++) {
4134 multilist_sublist_t *mls = multilist_sublist_lock_idx(ml, i);
4135 multilist_sublist_remove(mls, markers[i]);
4136 multilist_sublist_unlock(mls);
4137 }
4138 if (markers != arc_state_evict_markers)
4139 arc_state_free_markers(markers, num_sublists);
4140
4141 return (total_evicted);
4142 }
4143
4144 /*
4145 * Flush all "evictable" data of the given type from the arc state
4146 * specified. This will not evict any "active" buffers (i.e. referenced).
4147 *
4148 * When 'retry' is set to B_FALSE, the function will make a single pass
4149 * over the state and evict any buffers that it can. Since it doesn't
4150 * continually retry the eviction, it might end up leaving some buffers
4151 * in the ARC due to lock misses.
4152 *
4153 * When 'retry' is set to B_TRUE, the function will continually retry the
4154 * eviction until *all* evictable buffers have been removed from the
4155 * state. As a result, if concurrent insertions into the state are
4156 * allowed (e.g. if the ARC isn't shutting down), this function might
4157 * wind up in an infinite loop, continually trying to evict buffers.
4158 */
4159 static uint64_t
arc_flush_state(arc_state_t * state,uint64_t spa,arc_buf_contents_t type,boolean_t retry)4160 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4161 boolean_t retry)
4162 {
4163 uint64_t evicted = 0;
4164
4165 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
4166 evicted += arc_evict_state(state, type, spa, ARC_EVICT_ALL);
4167
4168 if (!retry)
4169 break;
4170 }
4171
4172 return (evicted);
4173 }
4174
4175 /*
4176 * Evict the specified number of bytes from the state specified. This
4177 * function prevents us from trying to evict more from a state's list
4178 * than is "evictable", and to skip evicting altogether when passed a
4179 * negative value for "bytes". In contrast, arc_evict_state() will
4180 * evict everything it can, when passed a negative value for "bytes".
4181 */
4182 static uint64_t
arc_evict_impl(arc_state_t * state,arc_buf_contents_t type,int64_t bytes)4183 arc_evict_impl(arc_state_t *state, arc_buf_contents_t type, int64_t bytes)
4184 {
4185 uint64_t delta;
4186
4187 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
4188 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
4189 bytes);
4190 return (arc_evict_state(state, type, 0, delta));
4191 }
4192
4193 return (0);
4194 }
4195
4196 /*
4197 * Adjust specified fraction, taking into account initial ghost state(s) size,
4198 * ghost hit bytes towards increasing the fraction, ghost hit bytes towards
4199 * decreasing it, plus a balance factor, controlling the decrease rate, used
4200 * to balance metadata vs data.
4201 */
4202 static uint64_t
arc_evict_adj(uint64_t frac,uint64_t total,uint64_t up,uint64_t down,uint_t balance)4203 arc_evict_adj(uint64_t frac, uint64_t total, uint64_t up, uint64_t down,
4204 uint_t balance)
4205 {
4206 if (total < 8 || up + down == 0)
4207 return (frac);
4208
4209 /*
4210 * We should not have more ghost hits than ghost size, but they
4211 * may get close. Restrict maximum adjustment in that case.
4212 */
4213 if (up + down >= total / 4) {
4214 uint64_t scale = (up + down) / (total / 8);
4215 up /= scale;
4216 down /= scale;
4217 }
4218
4219 /* Get maximal dynamic range by choosing optimal shifts. */
4220 int s = highbit64(total);
4221 s = MIN(64 - s, 32);
4222
4223 uint64_t ofrac = (1ULL << 32) - frac;
4224
4225 if (frac >= 4 * ofrac)
4226 up /= frac / (2 * ofrac + 1);
4227 up = (up << s) / (total >> (32 - s));
4228 if (ofrac >= 4 * frac)
4229 down /= ofrac / (2 * frac + 1);
4230 down = (down << s) / (total >> (32 - s));
4231 down = down * 100 / balance;
4232
4233 return (frac + up - down);
4234 }
4235
4236 /*
4237 * Calculate (x * multiplier / divisor) without unnecesary overflows.
4238 */
4239 static uint64_t
arc_mf(uint64_t x,uint64_t multiplier,uint64_t divisor)4240 arc_mf(uint64_t x, uint64_t multiplier, uint64_t divisor)
4241 {
4242 uint64_t q = (x / divisor);
4243 uint64_t r = (x % divisor);
4244
4245 return ((q * multiplier) + ((r * multiplier) / divisor));
4246 }
4247
4248 /*
4249 * Evict buffers from the cache, such that arcstat_size is capped by arc_c.
4250 */
4251 static uint64_t
arc_evict(void)4252 arc_evict(void)
4253 {
4254 uint64_t bytes, total_evicted = 0;
4255 int64_t e, mrud, mrum, mfud, mfum, w;
4256 static uint64_t ogrd, ogrm, ogfd, ogfm;
4257 static uint64_t gsrd, gsrm, gsfd, gsfm;
4258 uint64_t ngrd, ngrm, ngfd, ngfm;
4259
4260 /* Get current size of ARC states we can evict from. */
4261 mrud = zfs_refcount_count(&arc_mru->arcs_size[ARC_BUFC_DATA]) +
4262 zfs_refcount_count(&arc_anon->arcs_size[ARC_BUFC_DATA]);
4263 mrum = zfs_refcount_count(&arc_mru->arcs_size[ARC_BUFC_METADATA]) +
4264 zfs_refcount_count(&arc_anon->arcs_size[ARC_BUFC_METADATA]);
4265 mfud = zfs_refcount_count(&arc_mfu->arcs_size[ARC_BUFC_DATA]);
4266 mfum = zfs_refcount_count(&arc_mfu->arcs_size[ARC_BUFC_METADATA]);
4267 uint64_t d = mrud + mfud;
4268 uint64_t m = mrum + mfum;
4269 uint64_t t = d + m;
4270
4271 /* Get ARC ghost hits since last eviction. */
4272 ngrd = wmsum_value(&arc_mru_ghost->arcs_hits[ARC_BUFC_DATA]);
4273 uint64_t grd = ngrd - ogrd;
4274 ogrd = ngrd;
4275 ngrm = wmsum_value(&arc_mru_ghost->arcs_hits[ARC_BUFC_METADATA]);
4276 uint64_t grm = ngrm - ogrm;
4277 ogrm = ngrm;
4278 ngfd = wmsum_value(&arc_mfu_ghost->arcs_hits[ARC_BUFC_DATA]);
4279 uint64_t gfd = ngfd - ogfd;
4280 ogfd = ngfd;
4281 ngfm = wmsum_value(&arc_mfu_ghost->arcs_hits[ARC_BUFC_METADATA]);
4282 uint64_t gfm = ngfm - ogfm;
4283 ogfm = ngfm;
4284
4285 /* Adjust ARC states balance based on ghost hits. */
4286 arc_meta = arc_evict_adj(arc_meta, gsrd + gsrm + gsfd + gsfm,
4287 grm + gfm, grd + gfd, zfs_arc_meta_balance);
4288 arc_pd = arc_evict_adj(arc_pd, gsrd + gsfd, grd, gfd, 100);
4289 arc_pm = arc_evict_adj(arc_pm, gsrm + gsfm, grm, gfm, 100);
4290
4291 uint64_t asize = aggsum_value(&arc_sums.arcstat_size);
4292 uint64_t ac = arc_c;
4293 int64_t wt = t - (asize - ac);
4294
4295 /*
4296 * Try to reduce pinned dnodes if more than 3/4 of wanted metadata
4297 * target is not evictable or if they go over arc_dnode_limit.
4298 */
4299 int64_t prune = 0;
4300 int64_t dn = wmsum_value(&arc_sums.arcstat_dnode_size);
4301 int64_t nem = zfs_refcount_count(&arc_mru->arcs_size[ARC_BUFC_METADATA])
4302 + zfs_refcount_count(&arc_mfu->arcs_size[ARC_BUFC_METADATA])
4303 - zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA])
4304 - zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
4305 w = wt * (int64_t)(arc_meta >> 16) >> 16;
4306 if (nem > w * 3 / 4) {
4307 prune = dn / sizeof (dnode_t) *
4308 zfs_arc_dnode_reduce_percent / 100;
4309 if (nem < w && w > 4)
4310 prune = arc_mf(prune, nem - w * 3 / 4, w / 4);
4311 }
4312 if (dn > arc_dnode_limit) {
4313 prune = MAX(prune, (dn - arc_dnode_limit) / sizeof (dnode_t) *
4314 zfs_arc_dnode_reduce_percent / 100);
4315 }
4316 if (prune > 0)
4317 arc_prune_async(prune);
4318
4319 /* Evict MRU metadata. */
4320 w = wt * (int64_t)(arc_meta * arc_pm >> 48) >> 16;
4321 e = MIN((int64_t)(asize - ac), (int64_t)(mrum - w));
4322 bytes = arc_evict_impl(arc_mru, ARC_BUFC_METADATA, e);
4323 total_evicted += bytes;
4324 mrum -= bytes;
4325 asize -= bytes;
4326
4327 /* Evict MFU metadata. */
4328 w = wt * (int64_t)(arc_meta >> 16) >> 16;
4329 e = MIN((int64_t)(asize - ac), (int64_t)(m - bytes - w));
4330 bytes = arc_evict_impl(arc_mfu, ARC_BUFC_METADATA, e);
4331 total_evicted += bytes;
4332 mfum -= bytes;
4333 asize -= bytes;
4334
4335 /* Evict MRU data. */
4336 wt -= m - total_evicted;
4337 w = wt * (int64_t)(arc_pd >> 16) >> 16;
4338 e = MIN((int64_t)(asize - ac), (int64_t)(mrud - w));
4339 bytes = arc_evict_impl(arc_mru, ARC_BUFC_DATA, e);
4340 total_evicted += bytes;
4341 mrud -= bytes;
4342 asize -= bytes;
4343
4344 /* Evict MFU data. */
4345 e = asize - ac;
4346 bytes = arc_evict_impl(arc_mfu, ARC_BUFC_DATA, e);
4347 mfud -= bytes;
4348 total_evicted += bytes;
4349
4350 /*
4351 * Evict ghost lists
4352 *
4353 * Size of each state's ghost list represents how much that state
4354 * may grow by shrinking the other states. Would it need to shrink
4355 * other states to zero (that is unlikely), its ghost size would be
4356 * equal to sum of other three state sizes. But excessive ghost
4357 * size may result in false ghost hits (too far back), that may
4358 * never result in real cache hits if several states are competing.
4359 * So choose some arbitraty point of 1/2 of other state sizes.
4360 */
4361 gsrd = (mrum + mfud + mfum) / 2;
4362 e = zfs_refcount_count(&arc_mru_ghost->arcs_size[ARC_BUFC_DATA]) -
4363 gsrd;
4364 (void) arc_evict_impl(arc_mru_ghost, ARC_BUFC_DATA, e);
4365
4366 gsrm = (mrud + mfud + mfum) / 2;
4367 e = zfs_refcount_count(&arc_mru_ghost->arcs_size[ARC_BUFC_METADATA]) -
4368 gsrm;
4369 (void) arc_evict_impl(arc_mru_ghost, ARC_BUFC_METADATA, e);
4370
4371 gsfd = (mrud + mrum + mfum) / 2;
4372 e = zfs_refcount_count(&arc_mfu_ghost->arcs_size[ARC_BUFC_DATA]) -
4373 gsfd;
4374 (void) arc_evict_impl(arc_mfu_ghost, ARC_BUFC_DATA, e);
4375
4376 gsfm = (mrud + mrum + mfud) / 2;
4377 e = zfs_refcount_count(&arc_mfu_ghost->arcs_size[ARC_BUFC_METADATA]) -
4378 gsfm;
4379 (void) arc_evict_impl(arc_mfu_ghost, ARC_BUFC_METADATA, e);
4380
4381 return (total_evicted);
4382 }
4383
4384 void
arc_flush(spa_t * spa,boolean_t retry)4385 arc_flush(spa_t *spa, boolean_t retry)
4386 {
4387 uint64_t guid = 0;
4388
4389 /*
4390 * If retry is B_TRUE, a spa must not be specified since we have
4391 * no good way to determine if all of a spa's buffers have been
4392 * evicted from an arc state.
4393 */
4394 ASSERT(!retry || spa == NULL);
4395
4396 if (spa != NULL)
4397 guid = spa_load_guid(spa);
4398
4399 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4400 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4401
4402 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4403 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4404
4405 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4406 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4407
4408 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4409 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4410
4411 (void) arc_flush_state(arc_uncached, guid, ARC_BUFC_DATA, retry);
4412 (void) arc_flush_state(arc_uncached, guid, ARC_BUFC_METADATA, retry);
4413 }
4414
4415 uint64_t
arc_reduce_target_size(uint64_t to_free)4416 arc_reduce_target_size(uint64_t to_free)
4417 {
4418 /*
4419 * Get the actual arc size. Even if we don't need it, this updates
4420 * the aggsum lower bound estimate for arc_is_overflowing().
4421 */
4422 uint64_t asize = aggsum_value(&arc_sums.arcstat_size);
4423
4424 /*
4425 * All callers want the ARC to actually evict (at least) this much
4426 * memory. Therefore we reduce from the lower of the current size and
4427 * the target size. This way, even if arc_c is much higher than
4428 * arc_size (as can be the case after many calls to arc_freed(), we will
4429 * immediately have arc_c < arc_size and therefore the arc_evict_zthr
4430 * will evict.
4431 */
4432 uint64_t c = arc_c;
4433 if (c > arc_c_min) {
4434 c = MIN(c, MAX(asize, arc_c_min));
4435 to_free = MIN(to_free, c - arc_c_min);
4436 arc_c = c - to_free;
4437 } else {
4438 to_free = 0;
4439 }
4440
4441 /*
4442 * Whether or not we reduced the target size, request eviction if the
4443 * current size is over it now, since caller obviously wants some RAM.
4444 */
4445 if (asize > arc_c) {
4446 /* See comment in arc_evict_cb_check() on why lock+flag */
4447 mutex_enter(&arc_evict_lock);
4448 arc_evict_needed = B_TRUE;
4449 mutex_exit(&arc_evict_lock);
4450 zthr_wakeup(arc_evict_zthr);
4451 }
4452
4453 return (to_free);
4454 }
4455
4456 /*
4457 * Determine if the system is under memory pressure and is asking
4458 * to reclaim memory. A return value of B_TRUE indicates that the system
4459 * is under memory pressure and that the arc should adjust accordingly.
4460 */
4461 boolean_t
arc_reclaim_needed(void)4462 arc_reclaim_needed(void)
4463 {
4464 return (arc_available_memory() < 0);
4465 }
4466
4467 void
arc_kmem_reap_soon(void)4468 arc_kmem_reap_soon(void)
4469 {
4470 size_t i;
4471 kmem_cache_t *prev_cache = NULL;
4472 kmem_cache_t *prev_data_cache = NULL;
4473
4474 #ifdef _KERNEL
4475 #if defined(_ILP32)
4476 /*
4477 * Reclaim unused memory from all kmem caches.
4478 */
4479 kmem_reap();
4480 #endif
4481 #endif
4482
4483 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4484 #if defined(_ILP32)
4485 /* reach upper limit of cache size on 32-bit */
4486 if (zio_buf_cache[i] == NULL)
4487 break;
4488 #endif
4489 if (zio_buf_cache[i] != prev_cache) {
4490 prev_cache = zio_buf_cache[i];
4491 kmem_cache_reap_now(zio_buf_cache[i]);
4492 }
4493 if (zio_data_buf_cache[i] != prev_data_cache) {
4494 prev_data_cache = zio_data_buf_cache[i];
4495 kmem_cache_reap_now(zio_data_buf_cache[i]);
4496 }
4497 }
4498 kmem_cache_reap_now(buf_cache);
4499 kmem_cache_reap_now(hdr_full_cache);
4500 kmem_cache_reap_now(hdr_l2only_cache);
4501 kmem_cache_reap_now(zfs_btree_leaf_cache);
4502 abd_cache_reap_now();
4503 }
4504
4505 static boolean_t
arc_evict_cb_check(void * arg,zthr_t * zthr)4506 arc_evict_cb_check(void *arg, zthr_t *zthr)
4507 {
4508 (void) arg, (void) zthr;
4509
4510 #ifdef ZFS_DEBUG
4511 /*
4512 * This is necessary in order to keep the kstat information
4513 * up to date for tools that display kstat data such as the
4514 * mdb ::arc dcmd and the Linux crash utility. These tools
4515 * typically do not call kstat's update function, but simply
4516 * dump out stats from the most recent update. Without
4517 * this call, these commands may show stale stats for the
4518 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4519 * with this call, the data might be out of date if the
4520 * evict thread hasn't been woken recently; but that should
4521 * suffice. The arc_state_t structures can be queried
4522 * directly if more accurate information is needed.
4523 */
4524 if (arc_ksp != NULL)
4525 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4526 #endif
4527
4528 /*
4529 * We have to rely on arc_wait_for_eviction() to tell us when to
4530 * evict, rather than checking if we are overflowing here, so that we
4531 * are sure to not leave arc_wait_for_eviction() waiting on aew_cv.
4532 * If we have become "not overflowing" since arc_wait_for_eviction()
4533 * checked, we need to wake it up. We could broadcast the CV here,
4534 * but arc_wait_for_eviction() may have not yet gone to sleep. We
4535 * would need to use a mutex to ensure that this function doesn't
4536 * broadcast until arc_wait_for_eviction() has gone to sleep (e.g.
4537 * the arc_evict_lock). However, the lock ordering of such a lock
4538 * would necessarily be incorrect with respect to the zthr_lock,
4539 * which is held before this function is called, and is held by
4540 * arc_wait_for_eviction() when it calls zthr_wakeup().
4541 */
4542 if (arc_evict_needed)
4543 return (B_TRUE);
4544
4545 /*
4546 * If we have buffers in uncached state, evict them periodically.
4547 */
4548 return ((zfs_refcount_count(&arc_uncached->arcs_esize[ARC_BUFC_DATA]) +
4549 zfs_refcount_count(&arc_uncached->arcs_esize[ARC_BUFC_METADATA]) &&
4550 ddi_get_lbolt() - arc_last_uncached_flush >
4551 MSEC_TO_TICK(arc_min_prefetch_ms / 2)));
4552 }
4553
4554 /*
4555 * Keep arc_size under arc_c by running arc_evict which evicts data
4556 * from the ARC.
4557 */
4558 static void
arc_evict_cb(void * arg,zthr_t * zthr)4559 arc_evict_cb(void *arg, zthr_t *zthr)
4560 {
4561 (void) arg;
4562
4563 uint64_t evicted = 0;
4564 fstrans_cookie_t cookie = spl_fstrans_mark();
4565
4566 /* Always try to evict from uncached state. */
4567 arc_last_uncached_flush = ddi_get_lbolt();
4568 evicted += arc_flush_state(arc_uncached, 0, ARC_BUFC_DATA, B_FALSE);
4569 evicted += arc_flush_state(arc_uncached, 0, ARC_BUFC_METADATA, B_FALSE);
4570
4571 /* Evict from other states only if told to. */
4572 if (arc_evict_needed)
4573 evicted += arc_evict();
4574
4575 /*
4576 * If evicted is zero, we couldn't evict anything
4577 * via arc_evict(). This could be due to hash lock
4578 * collisions, but more likely due to the majority of
4579 * arc buffers being unevictable. Therefore, even if
4580 * arc_size is above arc_c, another pass is unlikely to
4581 * be helpful and could potentially cause us to enter an
4582 * infinite loop. Additionally, zthr_iscancelled() is
4583 * checked here so that if the arc is shutting down, the
4584 * broadcast will wake any remaining arc evict waiters.
4585 *
4586 * Note we cancel using zthr instead of arc_evict_zthr
4587 * because the latter may not yet be initializd when the
4588 * callback is first invoked.
4589 */
4590 mutex_enter(&arc_evict_lock);
4591 arc_evict_needed = !zthr_iscancelled(zthr) &&
4592 evicted > 0 && aggsum_compare(&arc_sums.arcstat_size, arc_c) > 0;
4593 if (!arc_evict_needed) {
4594 /*
4595 * We're either no longer overflowing, or we
4596 * can't evict anything more, so we should wake
4597 * arc_get_data_impl() sooner.
4598 */
4599 arc_evict_waiter_t *aw;
4600 while ((aw = list_remove_head(&arc_evict_waiters)) != NULL) {
4601 cv_broadcast(&aw->aew_cv);
4602 }
4603 arc_set_need_free();
4604 }
4605 mutex_exit(&arc_evict_lock);
4606 spl_fstrans_unmark(cookie);
4607 }
4608
4609 static boolean_t
arc_reap_cb_check(void * arg,zthr_t * zthr)4610 arc_reap_cb_check(void *arg, zthr_t *zthr)
4611 {
4612 (void) arg, (void) zthr;
4613
4614 int64_t free_memory = arc_available_memory();
4615 static int reap_cb_check_counter = 0;
4616
4617 /*
4618 * If a kmem reap is already active, don't schedule more. We must
4619 * check for this because kmem_cache_reap_soon() won't actually
4620 * block on the cache being reaped (this is to prevent callers from
4621 * becoming implicitly blocked by a system-wide kmem reap -- which,
4622 * on a system with many, many full magazines, can take minutes).
4623 */
4624 if (!kmem_cache_reap_active() && free_memory < 0) {
4625
4626 arc_no_grow = B_TRUE;
4627 arc_warm = B_TRUE;
4628 /*
4629 * Wait at least zfs_grow_retry (default 5) seconds
4630 * before considering growing.
4631 */
4632 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4633 return (B_TRUE);
4634 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4635 arc_no_grow = B_TRUE;
4636 } else if (gethrtime() >= arc_growtime) {
4637 arc_no_grow = B_FALSE;
4638 }
4639
4640 /*
4641 * Called unconditionally every 60 seconds to reclaim unused
4642 * zstd compression and decompression context. This is done
4643 * here to avoid the need for an independent thread.
4644 */
4645 if (!((reap_cb_check_counter++) % 60))
4646 zfs_zstd_cache_reap_now();
4647
4648 return (B_FALSE);
4649 }
4650
4651 /*
4652 * Keep enough free memory in the system by reaping the ARC's kmem
4653 * caches. To cause more slabs to be reapable, we may reduce the
4654 * target size of the cache (arc_c), causing the arc_evict_cb()
4655 * to free more buffers.
4656 */
4657 static void
arc_reap_cb(void * arg,zthr_t * zthr)4658 arc_reap_cb(void *arg, zthr_t *zthr)
4659 {
4660 int64_t can_free, free_memory, to_free;
4661
4662 (void) arg, (void) zthr;
4663 fstrans_cookie_t cookie = spl_fstrans_mark();
4664
4665 /*
4666 * Kick off asynchronous kmem_reap()'s of all our caches.
4667 */
4668 arc_kmem_reap_soon();
4669
4670 /*
4671 * Wait at least arc_kmem_cache_reap_retry_ms between
4672 * arc_kmem_reap_soon() calls. Without this check it is possible to
4673 * end up in a situation where we spend lots of time reaping
4674 * caches, while we're near arc_c_min. Waiting here also gives the
4675 * subsequent free memory check a chance of finding that the
4676 * asynchronous reap has already freed enough memory, and we don't
4677 * need to call arc_reduce_target_size().
4678 */
4679 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
4680
4681 /*
4682 * Reduce the target size as needed to maintain the amount of free
4683 * memory in the system at a fraction of the arc_size (1/128th by
4684 * default). If oversubscribed (free_memory < 0) then reduce the
4685 * target arc_size by the deficit amount plus the fractional
4686 * amount. If free memory is positive but less than the fractional
4687 * amount, reduce by what is needed to hit the fractional amount.
4688 */
4689 free_memory = arc_available_memory();
4690 can_free = arc_c - arc_c_min;
4691 to_free = (MAX(can_free, 0) >> arc_shrink_shift) - free_memory;
4692 if (to_free > 0)
4693 arc_reduce_target_size(to_free);
4694 spl_fstrans_unmark(cookie);
4695 }
4696
4697 #ifdef _KERNEL
4698 /*
4699 * Determine the amount of memory eligible for eviction contained in the
4700 * ARC. All clean data reported by the ghost lists can always be safely
4701 * evicted. Due to arc_c_min, the same does not hold for all clean data
4702 * contained by the regular mru and mfu lists.
4703 *
4704 * In the case of the regular mru and mfu lists, we need to report as
4705 * much clean data as possible, such that evicting that same reported
4706 * data will not bring arc_size below arc_c_min. Thus, in certain
4707 * circumstances, the total amount of clean data in the mru and mfu
4708 * lists might not actually be evictable.
4709 *
4710 * The following two distinct cases are accounted for:
4711 *
4712 * 1. The sum of the amount of dirty data contained by both the mru and
4713 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4714 * is greater than or equal to arc_c_min.
4715 * (i.e. amount of dirty data >= arc_c_min)
4716 *
4717 * This is the easy case; all clean data contained by the mru and mfu
4718 * lists is evictable. Evicting all clean data can only drop arc_size
4719 * to the amount of dirty data, which is greater than arc_c_min.
4720 *
4721 * 2. The sum of the amount of dirty data contained by both the mru and
4722 * mfu lists, plus the ARC's other accounting (e.g. the anon list),
4723 * is less than arc_c_min.
4724 * (i.e. arc_c_min > amount of dirty data)
4725 *
4726 * 2.1. arc_size is greater than or equal arc_c_min.
4727 * (i.e. arc_size >= arc_c_min > amount of dirty data)
4728 *
4729 * In this case, not all clean data from the regular mru and mfu
4730 * lists is actually evictable; we must leave enough clean data
4731 * to keep arc_size above arc_c_min. Thus, the maximum amount of
4732 * evictable data from the two lists combined, is exactly the
4733 * difference between arc_size and arc_c_min.
4734 *
4735 * 2.2. arc_size is less than arc_c_min
4736 * (i.e. arc_c_min > arc_size > amount of dirty data)
4737 *
4738 * In this case, none of the data contained in the mru and mfu
4739 * lists is evictable, even if it's clean. Since arc_size is
4740 * already below arc_c_min, evicting any more would only
4741 * increase this negative difference.
4742 */
4743
4744 #endif /* _KERNEL */
4745
4746 /*
4747 * Adapt arc info given the number of bytes we are trying to add and
4748 * the state that we are coming from. This function is only called
4749 * when we are adding new content to the cache.
4750 */
4751 static void
arc_adapt(uint64_t bytes)4752 arc_adapt(uint64_t bytes)
4753 {
4754 /*
4755 * Wake reap thread if we do not have any available memory
4756 */
4757 if (arc_reclaim_needed()) {
4758 zthr_wakeup(arc_reap_zthr);
4759 return;
4760 }
4761
4762 if (arc_no_grow)
4763 return;
4764
4765 if (arc_c >= arc_c_max)
4766 return;
4767
4768 /*
4769 * If we're within (2 * maxblocksize) bytes of the target
4770 * cache size, increment the target cache size
4771 */
4772 if (aggsum_upper_bound(&arc_sums.arcstat_size) +
4773 2 * SPA_MAXBLOCKSIZE >= arc_c) {
4774 uint64_t dc = MAX(bytes, SPA_OLD_MAXBLOCKSIZE);
4775 if (atomic_add_64_nv(&arc_c, dc) > arc_c_max)
4776 arc_c = arc_c_max;
4777 }
4778 }
4779
4780 /*
4781 * Check if ARC current size has grown past our upper thresholds.
4782 */
4783 static arc_ovf_level_t
arc_is_overflowing(boolean_t lax,boolean_t use_reserve)4784 arc_is_overflowing(boolean_t lax, boolean_t use_reserve)
4785 {
4786 /*
4787 * We just compare the lower bound here for performance reasons. Our
4788 * primary goals are to make sure that the arc never grows without
4789 * bound, and that it can reach its maximum size. This check
4790 * accomplishes both goals. The maximum amount we could run over by is
4791 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4792 * in the ARC. In practice, that's in the tens of MB, which is low
4793 * enough to be safe.
4794 */
4795 int64_t over = aggsum_lower_bound(&arc_sums.arcstat_size) - arc_c -
4796 zfs_max_recordsize;
4797
4798 /* Always allow at least one block of overflow. */
4799 if (over < 0)
4800 return (ARC_OVF_NONE);
4801
4802 /* If we are under memory pressure, report severe overflow. */
4803 if (!lax)
4804 return (ARC_OVF_SEVERE);
4805
4806 /* We are not under pressure, so be more or less relaxed. */
4807 int64_t overflow = (arc_c >> zfs_arc_overflow_shift) / 2;
4808 if (use_reserve)
4809 overflow *= 3;
4810 return (over < overflow ? ARC_OVF_SOME : ARC_OVF_SEVERE);
4811 }
4812
4813 static abd_t *
arc_get_data_abd(arc_buf_hdr_t * hdr,uint64_t size,const void * tag,int alloc_flags)4814 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, const void *tag,
4815 int alloc_flags)
4816 {
4817 arc_buf_contents_t type = arc_buf_type(hdr);
4818
4819 arc_get_data_impl(hdr, size, tag, alloc_flags);
4820 if (alloc_flags & ARC_HDR_ALLOC_LINEAR)
4821 return (abd_alloc_linear(size, type == ARC_BUFC_METADATA));
4822 else
4823 return (abd_alloc(size, type == ARC_BUFC_METADATA));
4824 }
4825
4826 static void *
arc_get_data_buf(arc_buf_hdr_t * hdr,uint64_t size,const void * tag)4827 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, const void *tag)
4828 {
4829 arc_buf_contents_t type = arc_buf_type(hdr);
4830
4831 arc_get_data_impl(hdr, size, tag, 0);
4832 if (type == ARC_BUFC_METADATA) {
4833 return (zio_buf_alloc(size));
4834 } else {
4835 ASSERT(type == ARC_BUFC_DATA);
4836 return (zio_data_buf_alloc(size));
4837 }
4838 }
4839
4840 /*
4841 * Wait for the specified amount of data (in bytes) to be evicted from the
4842 * ARC, and for there to be sufficient free memory in the system.
4843 * The lax argument specifies that caller does not have a specific reason
4844 * to wait, not aware of any memory pressure. Low memory handlers though
4845 * should set it to B_FALSE to wait for all required evictions to complete.
4846 * The use_reserve argument allows some callers to wait less than others
4847 * to not block critical code paths, possibly blocking other resources.
4848 */
4849 void
arc_wait_for_eviction(uint64_t amount,boolean_t lax,boolean_t use_reserve)4850 arc_wait_for_eviction(uint64_t amount, boolean_t lax, boolean_t use_reserve)
4851 {
4852 switch (arc_is_overflowing(lax, use_reserve)) {
4853 case ARC_OVF_NONE:
4854 return;
4855 case ARC_OVF_SOME:
4856 /*
4857 * This is a bit racy without taking arc_evict_lock, but the
4858 * worst that can happen is we either call zthr_wakeup() extra
4859 * time due to race with other thread here, or the set flag
4860 * get cleared by arc_evict_cb(), which is unlikely due to
4861 * big hysteresis, but also not important since at this level
4862 * of overflow the eviction is purely advisory. Same time
4863 * taking the global lock here every time without waiting for
4864 * the actual eviction creates a significant lock contention.
4865 */
4866 if (!arc_evict_needed) {
4867 arc_evict_needed = B_TRUE;
4868 zthr_wakeup(arc_evict_zthr);
4869 }
4870 return;
4871 case ARC_OVF_SEVERE:
4872 default:
4873 {
4874 arc_evict_waiter_t aw;
4875 list_link_init(&aw.aew_node);
4876 cv_init(&aw.aew_cv, NULL, CV_DEFAULT, NULL);
4877
4878 uint64_t last_count = 0;
4879 mutex_enter(&arc_evict_lock);
4880 if (!list_is_empty(&arc_evict_waiters)) {
4881 arc_evict_waiter_t *last =
4882 list_tail(&arc_evict_waiters);
4883 last_count = last->aew_count;
4884 } else if (!arc_evict_needed) {
4885 arc_evict_needed = B_TRUE;
4886 zthr_wakeup(arc_evict_zthr);
4887 }
4888 /*
4889 * Note, the last waiter's count may be less than
4890 * arc_evict_count if we are low on memory in which
4891 * case arc_evict_state_impl() may have deferred
4892 * wakeups (but still incremented arc_evict_count).
4893 */
4894 aw.aew_count = MAX(last_count, arc_evict_count) + amount;
4895
4896 list_insert_tail(&arc_evict_waiters, &aw);
4897
4898 arc_set_need_free();
4899
4900 DTRACE_PROBE3(arc__wait__for__eviction,
4901 uint64_t, amount,
4902 uint64_t, arc_evict_count,
4903 uint64_t, aw.aew_count);
4904
4905 /*
4906 * We will be woken up either when arc_evict_count reaches
4907 * aew_count, or when the ARC is no longer overflowing and
4908 * eviction completes.
4909 * In case of "false" wakeup, we will still be on the list.
4910 */
4911 do {
4912 cv_wait(&aw.aew_cv, &arc_evict_lock);
4913 } while (list_link_active(&aw.aew_node));
4914 mutex_exit(&arc_evict_lock);
4915
4916 cv_destroy(&aw.aew_cv);
4917 }
4918 }
4919 }
4920
4921 /*
4922 * Allocate a block and return it to the caller. If we are hitting the
4923 * hard limit for the cache size, we must sleep, waiting for the eviction
4924 * thread to catch up. If we're past the target size but below the hard
4925 * limit, we'll only signal the reclaim thread and continue on.
4926 */
4927 static void
arc_get_data_impl(arc_buf_hdr_t * hdr,uint64_t size,const void * tag,int alloc_flags)4928 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, const void *tag,
4929 int alloc_flags)
4930 {
4931 arc_adapt(size);
4932
4933 /*
4934 * If arc_size is currently overflowing, we must be adding data
4935 * faster than we are evicting. To ensure we don't compound the
4936 * problem by adding more data and forcing arc_size to grow even
4937 * further past it's target size, we wait for the eviction thread to
4938 * make some progress. We also wait for there to be sufficient free
4939 * memory in the system, as measured by arc_free_memory().
4940 *
4941 * Specifically, we wait for zfs_arc_eviction_pct percent of the
4942 * requested size to be evicted. This should be more than 100%, to
4943 * ensure that that progress is also made towards getting arc_size
4944 * under arc_c. See the comment above zfs_arc_eviction_pct.
4945 */
4946 arc_wait_for_eviction(size * zfs_arc_eviction_pct / 100,
4947 B_TRUE, alloc_flags & ARC_HDR_USE_RESERVE);
4948
4949 arc_buf_contents_t type = arc_buf_type(hdr);
4950 if (type == ARC_BUFC_METADATA) {
4951 arc_space_consume(size, ARC_SPACE_META);
4952 } else {
4953 arc_space_consume(size, ARC_SPACE_DATA);
4954 }
4955
4956 /*
4957 * Update the state size. Note that ghost states have a
4958 * "ghost size" and so don't need to be updated.
4959 */
4960 arc_state_t *state = hdr->b_l1hdr.b_state;
4961 if (!GHOST_STATE(state)) {
4962
4963 (void) zfs_refcount_add_many(&state->arcs_size[type], size,
4964 tag);
4965
4966 /*
4967 * If this is reached via arc_read, the link is
4968 * protected by the hash lock. If reached via
4969 * arc_buf_alloc, the header should not be accessed by
4970 * any other thread. And, if reached via arc_read_done,
4971 * the hash lock will protect it if it's found in the
4972 * hash table; otherwise no other thread should be
4973 * trying to [add|remove]_reference it.
4974 */
4975 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4976 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4977 (void) zfs_refcount_add_many(&state->arcs_esize[type],
4978 size, tag);
4979 }
4980 }
4981 }
4982
4983 static void
arc_free_data_abd(arc_buf_hdr_t * hdr,abd_t * abd,uint64_t size,const void * tag)4984 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size,
4985 const void *tag)
4986 {
4987 arc_free_data_impl(hdr, size, tag);
4988 abd_free(abd);
4989 }
4990
4991 static void
arc_free_data_buf(arc_buf_hdr_t * hdr,void * buf,uint64_t size,const void * tag)4992 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, const void *tag)
4993 {
4994 arc_buf_contents_t type = arc_buf_type(hdr);
4995
4996 arc_free_data_impl(hdr, size, tag);
4997 if (type == ARC_BUFC_METADATA) {
4998 zio_buf_free(buf, size);
4999 } else {
5000 ASSERT(type == ARC_BUFC_DATA);
5001 zio_data_buf_free(buf, size);
5002 }
5003 }
5004
5005 /*
5006 * Free the arc data buffer.
5007 */
5008 static void
arc_free_data_impl(arc_buf_hdr_t * hdr,uint64_t size,const void * tag)5009 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, const void *tag)
5010 {
5011 arc_state_t *state = hdr->b_l1hdr.b_state;
5012 arc_buf_contents_t type = arc_buf_type(hdr);
5013
5014 /* protected by hash lock, if in the hash table */
5015 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5016 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5017 ASSERT(state != arc_anon && state != arc_l2c_only);
5018
5019 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
5020 size, tag);
5021 }
5022 (void) zfs_refcount_remove_many(&state->arcs_size[type], size, tag);
5023
5024 VERIFY3U(hdr->b_type, ==, type);
5025 if (type == ARC_BUFC_METADATA) {
5026 arc_space_return(size, ARC_SPACE_META);
5027 } else {
5028 ASSERT(type == ARC_BUFC_DATA);
5029 arc_space_return(size, ARC_SPACE_DATA);
5030 }
5031 }
5032
5033 /*
5034 * This routine is called whenever a buffer is accessed.
5035 */
5036 static void
arc_access(arc_buf_hdr_t * hdr,arc_flags_t arc_flags,boolean_t hit)5037 arc_access(arc_buf_hdr_t *hdr, arc_flags_t arc_flags, boolean_t hit)
5038 {
5039 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
5040 ASSERT(HDR_HAS_L1HDR(hdr));
5041
5042 /*
5043 * Update buffer prefetch status.
5044 */
5045 boolean_t was_prefetch = HDR_PREFETCH(hdr);
5046 boolean_t now_prefetch = arc_flags & ARC_FLAG_PREFETCH;
5047 if (was_prefetch != now_prefetch) {
5048 if (was_prefetch) {
5049 ARCSTAT_CONDSTAT(hit, demand_hit, demand_iohit,
5050 HDR_PRESCIENT_PREFETCH(hdr), prescient, predictive,
5051 prefetch);
5052 }
5053 if (HDR_HAS_L2HDR(hdr))
5054 l2arc_hdr_arcstats_decrement_state(hdr);
5055 if (was_prefetch) {
5056 arc_hdr_clear_flags(hdr,
5057 ARC_FLAG_PREFETCH | ARC_FLAG_PRESCIENT_PREFETCH);
5058 } else {
5059 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5060 }
5061 if (HDR_HAS_L2HDR(hdr))
5062 l2arc_hdr_arcstats_increment_state(hdr);
5063 }
5064 if (now_prefetch) {
5065 if (arc_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
5066 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5067 ARCSTAT_BUMP(arcstat_prescient_prefetch);
5068 } else {
5069 ARCSTAT_BUMP(arcstat_predictive_prefetch);
5070 }
5071 }
5072 if (arc_flags & ARC_FLAG_L2CACHE)
5073 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5074
5075 clock_t now = ddi_get_lbolt();
5076 if (hdr->b_l1hdr.b_state == arc_anon) {
5077 arc_state_t *new_state;
5078 /*
5079 * This buffer is not in the cache, and does not appear in
5080 * our "ghost" lists. Add it to the MRU or uncached state.
5081 */
5082 ASSERT0(hdr->b_l1hdr.b_arc_access);
5083 hdr->b_l1hdr.b_arc_access = now;
5084 if (HDR_UNCACHED(hdr)) {
5085 new_state = arc_uncached;
5086 DTRACE_PROBE1(new_state__uncached, arc_buf_hdr_t *,
5087 hdr);
5088 } else {
5089 new_state = arc_mru;
5090 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5091 }
5092 arc_change_state(new_state, hdr);
5093 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5094 /*
5095 * This buffer has been accessed once recently and either
5096 * its read is still in progress or it is in the cache.
5097 */
5098 if (HDR_IO_IN_PROGRESS(hdr)) {
5099 hdr->b_l1hdr.b_arc_access = now;
5100 return;
5101 }
5102 hdr->b_l1hdr.b_mru_hits++;
5103 ARCSTAT_BUMP(arcstat_mru_hits);
5104
5105 /*
5106 * If the previous access was a prefetch, then it already
5107 * handled possible promotion, so nothing more to do for now.
5108 */
5109 if (was_prefetch) {
5110 hdr->b_l1hdr.b_arc_access = now;
5111 return;
5112 }
5113
5114 /*
5115 * If more than ARC_MINTIME have passed from the previous
5116 * hit, promote the buffer to the MFU state.
5117 */
5118 if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access +
5119 ARC_MINTIME)) {
5120 hdr->b_l1hdr.b_arc_access = now;
5121 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5122 arc_change_state(arc_mfu, hdr);
5123 }
5124 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5125 arc_state_t *new_state;
5126 /*
5127 * This buffer has been accessed once recently, but was
5128 * evicted from the cache. Would we have bigger MRU, it
5129 * would be an MRU hit, so handle it the same way, except
5130 * we don't need to check the previous access time.
5131 */
5132 hdr->b_l1hdr.b_mru_ghost_hits++;
5133 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5134 hdr->b_l1hdr.b_arc_access = now;
5135 wmsum_add(&arc_mru_ghost->arcs_hits[arc_buf_type(hdr)],
5136 arc_hdr_size(hdr));
5137 if (was_prefetch) {
5138 new_state = arc_mru;
5139 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5140 } else {
5141 new_state = arc_mfu;
5142 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5143 }
5144 arc_change_state(new_state, hdr);
5145 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5146 /*
5147 * This buffer has been accessed more than once and either
5148 * still in the cache or being restored from one of ghosts.
5149 */
5150 if (!HDR_IO_IN_PROGRESS(hdr)) {
5151 hdr->b_l1hdr.b_mfu_hits++;
5152 ARCSTAT_BUMP(arcstat_mfu_hits);
5153 }
5154 hdr->b_l1hdr.b_arc_access = now;
5155 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5156 /*
5157 * This buffer has been accessed more than once recently, but
5158 * has been evicted from the cache. Would we have bigger MFU
5159 * it would stay in cache, so move it back to MFU state.
5160 */
5161 hdr->b_l1hdr.b_mfu_ghost_hits++;
5162 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5163 hdr->b_l1hdr.b_arc_access = now;
5164 wmsum_add(&arc_mfu_ghost->arcs_hits[arc_buf_type(hdr)],
5165 arc_hdr_size(hdr));
5166 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5167 arc_change_state(arc_mfu, hdr);
5168 } else if (hdr->b_l1hdr.b_state == arc_uncached) {
5169 /*
5170 * This buffer is uncacheable, but we got a hit. Probably
5171 * a demand read after prefetch. Nothing more to do here.
5172 */
5173 if (!HDR_IO_IN_PROGRESS(hdr))
5174 ARCSTAT_BUMP(arcstat_uncached_hits);
5175 hdr->b_l1hdr.b_arc_access = now;
5176 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5177 /*
5178 * This buffer is on the 2nd Level ARC and was not accessed
5179 * for a long time, so treat it as new and put into MRU.
5180 */
5181 hdr->b_l1hdr.b_arc_access = now;
5182 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5183 arc_change_state(arc_mru, hdr);
5184 } else {
5185 cmn_err(CE_PANIC, "invalid arc state 0x%p",
5186 hdr->b_l1hdr.b_state);
5187 }
5188 }
5189
5190 /*
5191 * This routine is called by dbuf_hold() to update the arc_access() state
5192 * which otherwise would be skipped for entries in the dbuf cache.
5193 */
5194 void
arc_buf_access(arc_buf_t * buf)5195 arc_buf_access(arc_buf_t *buf)
5196 {
5197 arc_buf_hdr_t *hdr = buf->b_hdr;
5198
5199 /*
5200 * Avoid taking the hash_lock when possible as an optimization.
5201 * The header must be checked again under the hash_lock in order
5202 * to handle the case where it is concurrently being released.
5203 */
5204 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr))
5205 return;
5206
5207 kmutex_t *hash_lock = HDR_LOCK(hdr);
5208 mutex_enter(hash_lock);
5209
5210 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5211 mutex_exit(hash_lock);
5212 ARCSTAT_BUMP(arcstat_access_skip);
5213 return;
5214 }
5215
5216 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5217 hdr->b_l1hdr.b_state == arc_mfu ||
5218 hdr->b_l1hdr.b_state == arc_uncached);
5219
5220 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5221 arc_access(hdr, 0, B_TRUE);
5222 mutex_exit(hash_lock);
5223
5224 ARCSTAT_BUMP(arcstat_hits);
5225 ARCSTAT_CONDSTAT(B_TRUE /* demand */, demand, prefetch,
5226 !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5227 }
5228
5229 /* a generic arc_read_done_func_t which you can use */
5230 void
arc_bcopy_func(zio_t * zio,const zbookmark_phys_t * zb,const blkptr_t * bp,arc_buf_t * buf,void * arg)5231 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5232 arc_buf_t *buf, void *arg)
5233 {
5234 (void) zio, (void) zb, (void) bp;
5235
5236 if (buf == NULL)
5237 return;
5238
5239 memcpy(arg, buf->b_data, arc_buf_size(buf));
5240 arc_buf_destroy(buf, arg);
5241 }
5242
5243 /* a generic arc_read_done_func_t */
5244 void
arc_getbuf_func(zio_t * zio,const zbookmark_phys_t * zb,const blkptr_t * bp,arc_buf_t * buf,void * arg)5245 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5246 arc_buf_t *buf, void *arg)
5247 {
5248 (void) zb, (void) bp;
5249 arc_buf_t **bufp = arg;
5250
5251 if (buf == NULL) {
5252 ASSERT(zio == NULL || zio->io_error != 0);
5253 *bufp = NULL;
5254 } else {
5255 ASSERT(zio == NULL || zio->io_error == 0);
5256 *bufp = buf;
5257 ASSERT(buf->b_data != NULL);
5258 }
5259 }
5260
5261 static void
arc_hdr_verify(arc_buf_hdr_t * hdr,blkptr_t * bp)5262 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp)
5263 {
5264 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5265 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5266 ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5267 } else {
5268 if (HDR_COMPRESSION_ENABLED(hdr)) {
5269 ASSERT3U(arc_hdr_get_compress(hdr), ==,
5270 BP_GET_COMPRESS(bp));
5271 }
5272 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5273 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5274 ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5275 }
5276 }
5277
5278 static void
arc_read_done(zio_t * zio)5279 arc_read_done(zio_t *zio)
5280 {
5281 blkptr_t *bp = zio->io_bp;
5282 arc_buf_hdr_t *hdr = zio->io_private;
5283 kmutex_t *hash_lock = NULL;
5284 arc_callback_t *callback_list;
5285 arc_callback_t *acb;
5286
5287 /*
5288 * The hdr was inserted into hash-table and removed from lists
5289 * prior to starting I/O. We should find this header, since
5290 * it's in the hash table, and it should be legit since it's
5291 * not possible to evict it during the I/O. The only possible
5292 * reason for it not to be found is if we were freed during the
5293 * read.
5294 */
5295 if (HDR_IN_HASH_TABLE(hdr)) {
5296 arc_buf_hdr_t *found;
5297
5298 ASSERT3U(hdr->b_birth, ==, BP_GET_BIRTH(zio->io_bp));
5299 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5300 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5301 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5302 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5303
5304 found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock);
5305
5306 ASSERT((found == hdr &&
5307 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5308 (found == hdr && HDR_L2_READING(hdr)));
5309 ASSERT3P(hash_lock, !=, NULL);
5310 }
5311
5312 if (BP_IS_PROTECTED(bp)) {
5313 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5314 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5315 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5316 hdr->b_crypt_hdr.b_iv);
5317
5318 if (zio->io_error == 0) {
5319 if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5320 void *tmpbuf;
5321
5322 tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5323 sizeof (zil_chain_t));
5324 zio_crypt_decode_mac_zil(tmpbuf,
5325 hdr->b_crypt_hdr.b_mac);
5326 abd_return_buf(zio->io_abd, tmpbuf,
5327 sizeof (zil_chain_t));
5328 } else {
5329 zio_crypt_decode_mac_bp(bp,
5330 hdr->b_crypt_hdr.b_mac);
5331 }
5332 }
5333 }
5334
5335 if (zio->io_error == 0) {
5336 /* byteswap if necessary */
5337 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5338 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5339 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5340 } else {
5341 hdr->b_l1hdr.b_byteswap =
5342 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5343 }
5344 } else {
5345 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5346 }
5347 if (!HDR_L2_READING(hdr)) {
5348 hdr->b_complevel = zio->io_prop.zp_complevel;
5349 }
5350 }
5351
5352 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5353 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5354 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5355
5356 callback_list = hdr->b_l1hdr.b_acb;
5357 ASSERT3P(callback_list, !=, NULL);
5358 hdr->b_l1hdr.b_acb = NULL;
5359
5360 /*
5361 * If a read request has a callback (i.e. acb_done is not NULL), then we
5362 * make a buf containing the data according to the parameters which were
5363 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5364 * aren't needlessly decompressing the data multiple times.
5365 */
5366 int callback_cnt = 0;
5367 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5368
5369 /* We need the last one to call below in original order. */
5370 callback_list = acb;
5371
5372 if (!acb->acb_done || acb->acb_nobuf)
5373 continue;
5374
5375 callback_cnt++;
5376
5377 if (zio->io_error != 0)
5378 continue;
5379
5380 int error = arc_buf_alloc_impl(hdr, zio->io_spa,
5381 &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
5382 acb->acb_compressed, acb->acb_noauth, B_TRUE,
5383 &acb->acb_buf);
5384
5385 /*
5386 * Assert non-speculative zios didn't fail because an
5387 * encryption key wasn't loaded
5388 */
5389 ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
5390 error != EACCES);
5391
5392 /*
5393 * If we failed to decrypt, report an error now (as the zio
5394 * layer would have done if it had done the transforms).
5395 */
5396 if (error == ECKSUM) {
5397 ASSERT(BP_IS_PROTECTED(bp));
5398 error = SET_ERROR(EIO);
5399 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5400 spa_log_error(zio->io_spa, &acb->acb_zb,
5401 BP_GET_LOGICAL_BIRTH(zio->io_bp));
5402 (void) zfs_ereport_post(
5403 FM_EREPORT_ZFS_AUTHENTICATION,
5404 zio->io_spa, NULL, &acb->acb_zb, zio, 0);
5405 }
5406 }
5407
5408 if (error != 0) {
5409 /*
5410 * Decompression or decryption failed. Set
5411 * io_error so that when we call acb_done
5412 * (below), we will indicate that the read
5413 * failed. Note that in the unusual case
5414 * where one callback is compressed and another
5415 * uncompressed, we will mark all of them
5416 * as failed, even though the uncompressed
5417 * one can't actually fail. In this case,
5418 * the hdr will not be anonymous, because
5419 * if there are multiple callbacks, it's
5420 * because multiple threads found the same
5421 * arc buf in the hash table.
5422 */
5423 zio->io_error = error;
5424 }
5425 }
5426
5427 /*
5428 * If there are multiple callbacks, we must have the hash lock,
5429 * because the only way for multiple threads to find this hdr is
5430 * in the hash table. This ensures that if there are multiple
5431 * callbacks, the hdr is not anonymous. If it were anonymous,
5432 * we couldn't use arc_buf_destroy() in the error case below.
5433 */
5434 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5435
5436 if (zio->io_error == 0) {
5437 arc_hdr_verify(hdr, zio->io_bp);
5438 } else {
5439 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5440 if (hdr->b_l1hdr.b_state != arc_anon)
5441 arc_change_state(arc_anon, hdr);
5442 if (HDR_IN_HASH_TABLE(hdr))
5443 buf_hash_remove(hdr);
5444 }
5445
5446 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5447 (void) remove_reference(hdr, hdr);
5448
5449 if (hash_lock != NULL)
5450 mutex_exit(hash_lock);
5451
5452 /* execute each callback and free its structure */
5453 while ((acb = callback_list) != NULL) {
5454 if (acb->acb_done != NULL) {
5455 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5456 /*
5457 * If arc_buf_alloc_impl() fails during
5458 * decompression, the buf will still be
5459 * allocated, and needs to be freed here.
5460 */
5461 arc_buf_destroy(acb->acb_buf,
5462 acb->acb_private);
5463 acb->acb_buf = NULL;
5464 }
5465 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5466 acb->acb_buf, acb->acb_private);
5467 }
5468
5469 if (acb->acb_zio_dummy != NULL) {
5470 acb->acb_zio_dummy->io_error = zio->io_error;
5471 zio_nowait(acb->acb_zio_dummy);
5472 }
5473
5474 callback_list = acb->acb_prev;
5475 if (acb->acb_wait) {
5476 mutex_enter(&acb->acb_wait_lock);
5477 acb->acb_wait_error = zio->io_error;
5478 acb->acb_wait = B_FALSE;
5479 cv_signal(&acb->acb_wait_cv);
5480 mutex_exit(&acb->acb_wait_lock);
5481 /* acb will be freed by the waiting thread. */
5482 } else {
5483 kmem_free(acb, sizeof (arc_callback_t));
5484 }
5485 }
5486 }
5487
5488 /*
5489 * Lookup the block at the specified DVA (in bp), and return the manner in
5490 * which the block is cached. A zero return indicates not cached.
5491 */
5492 int
arc_cached(spa_t * spa,const blkptr_t * bp)5493 arc_cached(spa_t *spa, const blkptr_t *bp)
5494 {
5495 arc_buf_hdr_t *hdr = NULL;
5496 kmutex_t *hash_lock = NULL;
5497 uint64_t guid = spa_load_guid(spa);
5498 int flags = 0;
5499
5500 if (BP_IS_EMBEDDED(bp))
5501 return (ARC_CACHED_EMBEDDED);
5502
5503 hdr = buf_hash_find(guid, bp, &hash_lock);
5504 if (hdr == NULL)
5505 return (0);
5506
5507 if (HDR_HAS_L1HDR(hdr)) {
5508 arc_state_t *state = hdr->b_l1hdr.b_state;
5509 /*
5510 * We switch to ensure that any future arc_state_type_t
5511 * changes are handled. This is just a shift to promote
5512 * more compile-time checking.
5513 */
5514 switch (state->arcs_state) {
5515 case ARC_STATE_ANON:
5516 break;
5517 case ARC_STATE_MRU:
5518 flags |= ARC_CACHED_IN_MRU | ARC_CACHED_IN_L1;
5519 break;
5520 case ARC_STATE_MFU:
5521 flags |= ARC_CACHED_IN_MFU | ARC_CACHED_IN_L1;
5522 break;
5523 case ARC_STATE_UNCACHED:
5524 /* The header is still in L1, probably not for long */
5525 flags |= ARC_CACHED_IN_L1;
5526 break;
5527 default:
5528 break;
5529 }
5530 }
5531 if (HDR_HAS_L2HDR(hdr))
5532 flags |= ARC_CACHED_IN_L2;
5533
5534 mutex_exit(hash_lock);
5535
5536 return (flags);
5537 }
5538
5539 /*
5540 * "Read" the block at the specified DVA (in bp) via the
5541 * cache. If the block is found in the cache, invoke the provided
5542 * callback immediately and return. Note that the `zio' parameter
5543 * in the callback will be NULL in this case, since no IO was
5544 * required. If the block is not in the cache pass the read request
5545 * on to the spa with a substitute callback function, so that the
5546 * requested block will be added to the cache.
5547 *
5548 * If a read request arrives for a block that has a read in-progress,
5549 * either wait for the in-progress read to complete (and return the
5550 * results); or, if this is a read with a "done" func, add a record
5551 * to the read to invoke the "done" func when the read completes,
5552 * and return; or just return.
5553 *
5554 * arc_read_done() will invoke all the requested "done" functions
5555 * for readers of this block.
5556 */
5557 int
arc_read(zio_t * pio,spa_t * spa,const blkptr_t * bp,arc_read_done_func_t * done,void * private,zio_priority_t priority,int zio_flags,arc_flags_t * arc_flags,const zbookmark_phys_t * zb)5558 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
5559 arc_read_done_func_t *done, void *private, zio_priority_t priority,
5560 int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5561 {
5562 arc_buf_hdr_t *hdr = NULL;
5563 kmutex_t *hash_lock = NULL;
5564 zio_t *rzio;
5565 uint64_t guid = spa_load_guid(spa);
5566 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
5567 boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
5568 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5569 boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
5570 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5571 boolean_t embedded_bp = !!BP_IS_EMBEDDED(bp);
5572 boolean_t no_buf = *arc_flags & ARC_FLAG_NO_BUF;
5573 arc_buf_t *buf = NULL;
5574 int rc = 0;
5575
5576 ASSERT(!embedded_bp ||
5577 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5578 ASSERT(!BP_IS_HOLE(bp));
5579 ASSERT(!BP_IS_REDACTED(bp));
5580
5581 /*
5582 * Normally SPL_FSTRANS will already be set since kernel threads which
5583 * expect to call the DMU interfaces will set it when created. System
5584 * calls are similarly handled by setting/cleaning the bit in the
5585 * registered callback (module/os/.../zfs/zpl_*).
5586 *
5587 * External consumers such as Lustre which call the exported DMU
5588 * interfaces may not have set SPL_FSTRANS. To avoid a deadlock
5589 * on the hash_lock always set and clear the bit.
5590 */
5591 fstrans_cookie_t cookie = spl_fstrans_mark();
5592 top:
5593 if (!embedded_bp) {
5594 /*
5595 * Embedded BP's have no DVA and require no I/O to "read".
5596 * Create an anonymous arc buf to back it.
5597 */
5598 hdr = buf_hash_find(guid, bp, &hash_lock);
5599 }
5600
5601 /*
5602 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5603 * we maintain encrypted data separately from compressed / uncompressed
5604 * data. If the user is requesting raw encrypted data and we don't have
5605 * that in the header we will read from disk to guarantee that we can
5606 * get it even if the encryption keys aren't loaded.
5607 */
5608 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
5609 (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
5610 boolean_t is_data = !HDR_ISTYPE_METADATA(hdr);
5611
5612 /*
5613 * Verify the block pointer contents are reasonable. This
5614 * should always be the case since the blkptr is protected by
5615 * a checksum.
5616 */
5617 if (!zfs_blkptr_verify(spa, bp, BLK_CONFIG_SKIP,
5618 BLK_VERIFY_LOG)) {
5619 mutex_exit(hash_lock);
5620 rc = SET_ERROR(ECKSUM);
5621 goto done;
5622 }
5623
5624 if (HDR_IO_IN_PROGRESS(hdr)) {
5625 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
5626 mutex_exit(hash_lock);
5627 ARCSTAT_BUMP(arcstat_cached_only_in_progress);
5628 rc = SET_ERROR(ENOENT);
5629 goto done;
5630 }
5631
5632 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5633 ASSERT3P(head_zio, !=, NULL);
5634 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5635 priority == ZIO_PRIORITY_SYNC_READ) {
5636 /*
5637 * This is a sync read that needs to wait for
5638 * an in-flight async read. Request that the
5639 * zio have its priority upgraded.
5640 */
5641 zio_change_priority(head_zio, priority);
5642 DTRACE_PROBE1(arc__async__upgrade__sync,
5643 arc_buf_hdr_t *, hdr);
5644 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5645 }
5646
5647 DTRACE_PROBE1(arc__iohit, arc_buf_hdr_t *, hdr);
5648 arc_access(hdr, *arc_flags, B_FALSE);
5649
5650 /*
5651 * If there are multiple threads reading the same block
5652 * and that block is not yet in the ARC, then only one
5653 * thread will do the physical I/O and all other
5654 * threads will wait until that I/O completes.
5655 * Synchronous reads use the acb_wait_cv whereas nowait
5656 * reads register a callback. Both are signalled/called
5657 * in arc_read_done.
5658 *
5659 * Errors of the physical I/O may need to be propagated.
5660 * Synchronous read errors are returned here from
5661 * arc_read_done via acb_wait_error. Nowait reads
5662 * attach the acb_zio_dummy zio to pio and
5663 * arc_read_done propagates the physical I/O's io_error
5664 * to acb_zio_dummy, and thereby to pio.
5665 */
5666 arc_callback_t *acb = NULL;
5667 if (done || pio || *arc_flags & ARC_FLAG_WAIT) {
5668 acb = kmem_zalloc(sizeof (arc_callback_t),
5669 KM_SLEEP);
5670 acb->acb_done = done;
5671 acb->acb_private = private;
5672 acb->acb_compressed = compressed_read;
5673 acb->acb_encrypted = encrypted_read;
5674 acb->acb_noauth = noauth_read;
5675 acb->acb_nobuf = no_buf;
5676 if (*arc_flags & ARC_FLAG_WAIT) {
5677 acb->acb_wait = B_TRUE;
5678 mutex_init(&acb->acb_wait_lock, NULL,
5679 MUTEX_DEFAULT, NULL);
5680 cv_init(&acb->acb_wait_cv, NULL,
5681 CV_DEFAULT, NULL);
5682 }
5683 acb->acb_zb = *zb;
5684 if (pio != NULL) {
5685 acb->acb_zio_dummy = zio_null(pio,
5686 spa, NULL, NULL, NULL, zio_flags);
5687 }
5688 acb->acb_zio_head = head_zio;
5689 acb->acb_next = hdr->b_l1hdr.b_acb;
5690 hdr->b_l1hdr.b_acb->acb_prev = acb;
5691 hdr->b_l1hdr.b_acb = acb;
5692 }
5693 mutex_exit(hash_lock);
5694
5695 ARCSTAT_BUMP(arcstat_iohits);
5696 ARCSTAT_CONDSTAT(!(*arc_flags & ARC_FLAG_PREFETCH),
5697 demand, prefetch, is_data, data, metadata, iohits);
5698
5699 if (*arc_flags & ARC_FLAG_WAIT) {
5700 mutex_enter(&acb->acb_wait_lock);
5701 while (acb->acb_wait) {
5702 cv_wait(&acb->acb_wait_cv,
5703 &acb->acb_wait_lock);
5704 }
5705 rc = acb->acb_wait_error;
5706 mutex_exit(&acb->acb_wait_lock);
5707 mutex_destroy(&acb->acb_wait_lock);
5708 cv_destroy(&acb->acb_wait_cv);
5709 kmem_free(acb, sizeof (arc_callback_t));
5710 }
5711 goto out;
5712 }
5713
5714 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5715 hdr->b_l1hdr.b_state == arc_mfu ||
5716 hdr->b_l1hdr.b_state == arc_uncached);
5717
5718 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5719 arc_access(hdr, *arc_flags, B_TRUE);
5720
5721 if (done && !no_buf) {
5722 ASSERT(!embedded_bp || !BP_IS_HOLE(bp));
5723
5724 /* Get a buf with the desired data in it. */
5725 rc = arc_buf_alloc_impl(hdr, spa, zb, private,
5726 encrypted_read, compressed_read, noauth_read,
5727 B_TRUE, &buf);
5728 if (rc == ECKSUM) {
5729 /*
5730 * Convert authentication and decryption errors
5731 * to EIO (and generate an ereport if needed)
5732 * before leaving the ARC.
5733 */
5734 rc = SET_ERROR(EIO);
5735 if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5736 spa_log_error(spa, zb, hdr->b_birth);
5737 (void) zfs_ereport_post(
5738 FM_EREPORT_ZFS_AUTHENTICATION,
5739 spa, NULL, zb, NULL, 0);
5740 }
5741 }
5742 if (rc != 0) {
5743 arc_buf_destroy_impl(buf);
5744 buf = NULL;
5745 (void) remove_reference(hdr, private);
5746 }
5747
5748 /* assert any errors weren't due to unloaded keys */
5749 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
5750 rc != EACCES);
5751 }
5752 mutex_exit(hash_lock);
5753 ARCSTAT_BUMP(arcstat_hits);
5754 ARCSTAT_CONDSTAT(!(*arc_flags & ARC_FLAG_PREFETCH),
5755 demand, prefetch, is_data, data, metadata, hits);
5756 *arc_flags |= ARC_FLAG_CACHED;
5757 goto done;
5758 } else {
5759 uint64_t lsize = BP_GET_LSIZE(bp);
5760 uint64_t psize = BP_GET_PSIZE(bp);
5761 arc_callback_t *acb;
5762 vdev_t *vd = NULL;
5763 uint64_t addr = 0;
5764 boolean_t devw = B_FALSE;
5765 uint64_t size;
5766 abd_t *hdr_abd;
5767 int alloc_flags = encrypted_read ? ARC_HDR_ALLOC_RDATA : 0;
5768 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5769
5770 if (*arc_flags & ARC_FLAG_CACHED_ONLY) {
5771 if (hash_lock != NULL)
5772 mutex_exit(hash_lock);
5773 rc = SET_ERROR(ENOENT);
5774 goto done;
5775 }
5776
5777 /*
5778 * Verify the block pointer contents are reasonable. This
5779 * should always be the case since the blkptr is protected by
5780 * a checksum.
5781 */
5782 if (!zfs_blkptr_verify(spa, bp,
5783 (zio_flags & ZIO_FLAG_CONFIG_WRITER) ?
5784 BLK_CONFIG_HELD : BLK_CONFIG_NEEDED, BLK_VERIFY_LOG)) {
5785 if (hash_lock != NULL)
5786 mutex_exit(hash_lock);
5787 rc = SET_ERROR(ECKSUM);
5788 goto done;
5789 }
5790
5791 if (hdr == NULL) {
5792 /*
5793 * This block is not in the cache or it has
5794 * embedded data.
5795 */
5796 arc_buf_hdr_t *exists = NULL;
5797 hdr = arc_hdr_alloc(guid, psize, lsize,
5798 BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), 0, type);
5799
5800 if (!embedded_bp) {
5801 hdr->b_dva = *BP_IDENTITY(bp);
5802 hdr->b_birth = BP_GET_BIRTH(bp);
5803 exists = buf_hash_insert(hdr, &hash_lock);
5804 }
5805 if (exists != NULL) {
5806 /* somebody beat us to the hash insert */
5807 mutex_exit(hash_lock);
5808 buf_discard_identity(hdr);
5809 arc_hdr_destroy(hdr);
5810 goto top; /* restart the IO request */
5811 }
5812 } else {
5813 /*
5814 * This block is in the ghost cache or encrypted data
5815 * was requested and we didn't have it. If it was
5816 * L2-only (and thus didn't have an L1 hdr),
5817 * we realloc the header to add an L1 hdr.
5818 */
5819 if (!HDR_HAS_L1HDR(hdr)) {
5820 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5821 hdr_full_cache);
5822 }
5823
5824 if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
5825 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5826 ASSERT(!HDR_HAS_RABD(hdr));
5827 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5828 ASSERT0(zfs_refcount_count(
5829 &hdr->b_l1hdr.b_refcnt));
5830 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5831 #ifdef ZFS_DEBUG
5832 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5833 #endif
5834 } else if (HDR_IO_IN_PROGRESS(hdr)) {
5835 /*
5836 * If this header already had an IO in progress
5837 * and we are performing another IO to fetch
5838 * encrypted data we must wait until the first
5839 * IO completes so as not to confuse
5840 * arc_read_done(). This should be very rare
5841 * and so the performance impact shouldn't
5842 * matter.
5843 */
5844 arc_callback_t *acb = kmem_zalloc(
5845 sizeof (arc_callback_t), KM_SLEEP);
5846 acb->acb_wait = B_TRUE;
5847 mutex_init(&acb->acb_wait_lock, NULL,
5848 MUTEX_DEFAULT, NULL);
5849 cv_init(&acb->acb_wait_cv, NULL, CV_DEFAULT,
5850 NULL);
5851 acb->acb_zio_head =
5852 hdr->b_l1hdr.b_acb->acb_zio_head;
5853 acb->acb_next = hdr->b_l1hdr.b_acb;
5854 hdr->b_l1hdr.b_acb->acb_prev = acb;
5855 hdr->b_l1hdr.b_acb = acb;
5856 mutex_exit(hash_lock);
5857 mutex_enter(&acb->acb_wait_lock);
5858 while (acb->acb_wait) {
5859 cv_wait(&acb->acb_wait_cv,
5860 &acb->acb_wait_lock);
5861 }
5862 mutex_exit(&acb->acb_wait_lock);
5863 mutex_destroy(&acb->acb_wait_lock);
5864 cv_destroy(&acb->acb_wait_cv);
5865 kmem_free(acb, sizeof (arc_callback_t));
5866 goto top;
5867 }
5868 }
5869 if (*arc_flags & ARC_FLAG_UNCACHED) {
5870 arc_hdr_set_flags(hdr, ARC_FLAG_UNCACHED);
5871 if (!encrypted_read)
5872 alloc_flags |= ARC_HDR_ALLOC_LINEAR;
5873 }
5874
5875 /*
5876 * Take additional reference for IO_IN_PROGRESS. It stops
5877 * arc_access() from putting this header without any buffers
5878 * and so other references but obviously nonevictable onto
5879 * the evictable list of MRU or MFU state.
5880 */
5881 add_reference(hdr, hdr);
5882 if (!embedded_bp)
5883 arc_access(hdr, *arc_flags, B_FALSE);
5884 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5885 arc_hdr_alloc_abd(hdr, alloc_flags);
5886 if (encrypted_read) {
5887 ASSERT(HDR_HAS_RABD(hdr));
5888 size = HDR_GET_PSIZE(hdr);
5889 hdr_abd = hdr->b_crypt_hdr.b_rabd;
5890 zio_flags |= ZIO_FLAG_RAW;
5891 } else {
5892 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5893 size = arc_hdr_size(hdr);
5894 hdr_abd = hdr->b_l1hdr.b_pabd;
5895
5896 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
5897 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
5898 }
5899
5900 /*
5901 * For authenticated bp's, we do not ask the ZIO layer
5902 * to authenticate them since this will cause the entire
5903 * IO to fail if the key isn't loaded. Instead, we
5904 * defer authentication until arc_buf_fill(), which will
5905 * verify the data when the key is available.
5906 */
5907 if (BP_IS_AUTHENTICATED(bp))
5908 zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
5909 }
5910
5911 if (BP_IS_AUTHENTICATED(bp))
5912 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
5913 if (BP_GET_LEVEL(bp) > 0)
5914 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5915 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5916
5917 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5918 acb->acb_done = done;
5919 acb->acb_private = private;
5920 acb->acb_compressed = compressed_read;
5921 acb->acb_encrypted = encrypted_read;
5922 acb->acb_noauth = noauth_read;
5923 acb->acb_zb = *zb;
5924
5925 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5926 hdr->b_l1hdr.b_acb = acb;
5927
5928 if (HDR_HAS_L2HDR(hdr) &&
5929 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5930 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5931 addr = hdr->b_l2hdr.b_daddr;
5932 /*
5933 * Lock out L2ARC device removal.
5934 */
5935 if (vdev_is_dead(vd) ||
5936 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5937 vd = NULL;
5938 }
5939
5940 /*
5941 * We count both async reads and scrub IOs as asynchronous so
5942 * that both can be upgraded in the event of a cache hit while
5943 * the read IO is still in-flight.
5944 */
5945 if (priority == ZIO_PRIORITY_ASYNC_READ ||
5946 priority == ZIO_PRIORITY_SCRUB)
5947 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5948 else
5949 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5950
5951 /*
5952 * At this point, we have a level 1 cache miss or a blkptr
5953 * with embedded data. Try again in L2ARC if possible.
5954 */
5955 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5956
5957 /*
5958 * Skip ARC stat bump for block pointers with embedded
5959 * data. The data are read from the blkptr itself via
5960 * decode_embedded_bp_compressed().
5961 */
5962 if (!embedded_bp) {
5963 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr,
5964 blkptr_t *, bp, uint64_t, lsize,
5965 zbookmark_phys_t *, zb);
5966 ARCSTAT_BUMP(arcstat_misses);
5967 ARCSTAT_CONDSTAT(!(*arc_flags & ARC_FLAG_PREFETCH),
5968 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data,
5969 metadata, misses);
5970 zfs_racct_read(spa, size, 1, 0);
5971 }
5972
5973 /* Check if the spa even has l2 configured */
5974 const boolean_t spa_has_l2 = l2arc_ndev != 0 &&
5975 spa->spa_l2cache.sav_count > 0;
5976
5977 if (vd != NULL && spa_has_l2 && !(l2arc_norw && devw)) {
5978 /*
5979 * Read from the L2ARC if the following are true:
5980 * 1. The L2ARC vdev was previously cached.
5981 * 2. This buffer still has L2ARC metadata.
5982 * 3. This buffer isn't currently writing to the L2ARC.
5983 * 4. The L2ARC entry wasn't evicted, which may
5984 * also have invalidated the vdev.
5985 */
5986 if (HDR_HAS_L2HDR(hdr) &&
5987 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr)) {
5988 l2arc_read_callback_t *cb;
5989 abd_t *abd;
5990 uint64_t asize;
5991
5992 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5993 ARCSTAT_BUMP(arcstat_l2_hits);
5994 hdr->b_l2hdr.b_hits++;
5995
5996 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5997 KM_SLEEP);
5998 cb->l2rcb_hdr = hdr;
5999 cb->l2rcb_bp = *bp;
6000 cb->l2rcb_zb = *zb;
6001 cb->l2rcb_flags = zio_flags;
6002
6003 /*
6004 * When Compressed ARC is disabled, but the
6005 * L2ARC block is compressed, arc_hdr_size()
6006 * will have returned LSIZE rather than PSIZE.
6007 */
6008 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
6009 !HDR_COMPRESSION_ENABLED(hdr) &&
6010 HDR_GET_PSIZE(hdr) != 0) {
6011 size = HDR_GET_PSIZE(hdr);
6012 }
6013
6014 asize = vdev_psize_to_asize(vd, size);
6015 if (asize != size) {
6016 abd = abd_alloc_for_io(asize,
6017 HDR_ISTYPE_METADATA(hdr));
6018 cb->l2rcb_abd = abd;
6019 } else {
6020 abd = hdr_abd;
6021 }
6022
6023 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6024 addr + asize <= vd->vdev_psize -
6025 VDEV_LABEL_END_SIZE);
6026
6027 /*
6028 * l2arc read. The SCL_L2ARC lock will be
6029 * released by l2arc_read_done().
6030 * Issue a null zio if the underlying buffer
6031 * was squashed to zero size by compression.
6032 */
6033 ASSERT3U(arc_hdr_get_compress(hdr), !=,
6034 ZIO_COMPRESS_EMPTY);
6035 rzio = zio_read_phys(pio, vd, addr,
6036 asize, abd,
6037 ZIO_CHECKSUM_OFF,
6038 l2arc_read_done, cb, priority,
6039 zio_flags | ZIO_FLAG_CANFAIL |
6040 ZIO_FLAG_DONT_PROPAGATE |
6041 ZIO_FLAG_DONT_RETRY, B_FALSE);
6042 acb->acb_zio_head = rzio;
6043
6044 if (hash_lock != NULL)
6045 mutex_exit(hash_lock);
6046
6047 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6048 zio_t *, rzio);
6049 ARCSTAT_INCR(arcstat_l2_read_bytes,
6050 HDR_GET_PSIZE(hdr));
6051
6052 if (*arc_flags & ARC_FLAG_NOWAIT) {
6053 zio_nowait(rzio);
6054 goto out;
6055 }
6056
6057 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6058 if (zio_wait(rzio) == 0)
6059 goto out;
6060
6061 /* l2arc read error; goto zio_read() */
6062 if (hash_lock != NULL)
6063 mutex_enter(hash_lock);
6064 } else {
6065 DTRACE_PROBE1(l2arc__miss,
6066 arc_buf_hdr_t *, hdr);
6067 ARCSTAT_BUMP(arcstat_l2_misses);
6068 if (HDR_L2_WRITING(hdr))
6069 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6070 spa_config_exit(spa, SCL_L2ARC, vd);
6071 }
6072 } else {
6073 if (vd != NULL)
6074 spa_config_exit(spa, SCL_L2ARC, vd);
6075
6076 /*
6077 * Only a spa with l2 should contribute to l2
6078 * miss stats. (Including the case of having a
6079 * faulted cache device - that's also a miss.)
6080 */
6081 if (spa_has_l2) {
6082 /*
6083 * Skip ARC stat bump for block pointers with
6084 * embedded data. The data are read from the
6085 * blkptr itself via
6086 * decode_embedded_bp_compressed().
6087 */
6088 if (!embedded_bp) {
6089 DTRACE_PROBE1(l2arc__miss,
6090 arc_buf_hdr_t *, hdr);
6091 ARCSTAT_BUMP(arcstat_l2_misses);
6092 }
6093 }
6094 }
6095
6096 rzio = zio_read(pio, spa, bp, hdr_abd, size,
6097 arc_read_done, hdr, priority, zio_flags, zb);
6098 acb->acb_zio_head = rzio;
6099
6100 if (hash_lock != NULL)
6101 mutex_exit(hash_lock);
6102
6103 if (*arc_flags & ARC_FLAG_WAIT) {
6104 rc = zio_wait(rzio);
6105 goto out;
6106 }
6107
6108 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6109 zio_nowait(rzio);
6110 }
6111
6112 out:
6113 /* embedded bps don't actually go to disk */
6114 if (!embedded_bp)
6115 spa_read_history_add(spa, zb, *arc_flags);
6116 spl_fstrans_unmark(cookie);
6117 return (rc);
6118
6119 done:
6120 if (done)
6121 done(NULL, zb, bp, buf, private);
6122 if (pio && rc != 0) {
6123 zio_t *zio = zio_null(pio, spa, NULL, NULL, NULL, zio_flags);
6124 zio->io_error = rc;
6125 zio_nowait(zio);
6126 }
6127 goto out;
6128 }
6129
6130 arc_prune_t *
arc_add_prune_callback(arc_prune_func_t * func,void * private)6131 arc_add_prune_callback(arc_prune_func_t *func, void *private)
6132 {
6133 arc_prune_t *p;
6134
6135 p = kmem_alloc(sizeof (*p), KM_SLEEP);
6136 p->p_pfunc = func;
6137 p->p_private = private;
6138 list_link_init(&p->p_node);
6139 zfs_refcount_create(&p->p_refcnt);
6140
6141 mutex_enter(&arc_prune_mtx);
6142 zfs_refcount_add(&p->p_refcnt, &arc_prune_list);
6143 list_insert_head(&arc_prune_list, p);
6144 mutex_exit(&arc_prune_mtx);
6145
6146 return (p);
6147 }
6148
6149 void
arc_remove_prune_callback(arc_prune_t * p)6150 arc_remove_prune_callback(arc_prune_t *p)
6151 {
6152 boolean_t wait = B_FALSE;
6153 mutex_enter(&arc_prune_mtx);
6154 list_remove(&arc_prune_list, p);
6155 if (zfs_refcount_remove(&p->p_refcnt, &arc_prune_list) > 0)
6156 wait = B_TRUE;
6157 mutex_exit(&arc_prune_mtx);
6158
6159 /* wait for arc_prune_task to finish */
6160 if (wait)
6161 taskq_wait_outstanding(arc_prune_taskq, 0);
6162 ASSERT0(zfs_refcount_count(&p->p_refcnt));
6163 zfs_refcount_destroy(&p->p_refcnt);
6164 kmem_free(p, sizeof (*p));
6165 }
6166
6167 /*
6168 * Helper function for arc_prune_async() it is responsible for safely
6169 * handling the execution of a registered arc_prune_func_t.
6170 */
6171 static void
arc_prune_task(void * ptr)6172 arc_prune_task(void *ptr)
6173 {
6174 arc_prune_t *ap = (arc_prune_t *)ptr;
6175 arc_prune_func_t *func = ap->p_pfunc;
6176
6177 if (func != NULL)
6178 func(ap->p_adjust, ap->p_private);
6179
6180 (void) zfs_refcount_remove(&ap->p_refcnt, func);
6181 }
6182
6183 /*
6184 * Notify registered consumers they must drop holds on a portion of the ARC
6185 * buffers they reference. This provides a mechanism to ensure the ARC can
6186 * honor the metadata limit and reclaim otherwise pinned ARC buffers.
6187 *
6188 * This operation is performed asynchronously so it may be safely called
6189 * in the context of the arc_reclaim_thread(). A reference is taken here
6190 * for each registered arc_prune_t and the arc_prune_task() is responsible
6191 * for releasing it once the registered arc_prune_func_t has completed.
6192 */
6193 static void
arc_prune_async(uint64_t adjust)6194 arc_prune_async(uint64_t adjust)
6195 {
6196 arc_prune_t *ap;
6197
6198 mutex_enter(&arc_prune_mtx);
6199 for (ap = list_head(&arc_prune_list); ap != NULL;
6200 ap = list_next(&arc_prune_list, ap)) {
6201
6202 if (zfs_refcount_count(&ap->p_refcnt) >= 2)
6203 continue;
6204
6205 zfs_refcount_add(&ap->p_refcnt, ap->p_pfunc);
6206 ap->p_adjust = adjust;
6207 if (taskq_dispatch(arc_prune_taskq, arc_prune_task,
6208 ap, TQ_SLEEP) == TASKQID_INVALID) {
6209 (void) zfs_refcount_remove(&ap->p_refcnt, ap->p_pfunc);
6210 continue;
6211 }
6212 ARCSTAT_BUMP(arcstat_prune);
6213 }
6214 mutex_exit(&arc_prune_mtx);
6215 }
6216
6217 /*
6218 * Notify the arc that a block was freed, and thus will never be used again.
6219 */
6220 void
arc_freed(spa_t * spa,const blkptr_t * bp)6221 arc_freed(spa_t *spa, const blkptr_t *bp)
6222 {
6223 arc_buf_hdr_t *hdr;
6224 kmutex_t *hash_lock;
6225 uint64_t guid = spa_load_guid(spa);
6226
6227 ASSERT(!BP_IS_EMBEDDED(bp));
6228
6229 hdr = buf_hash_find(guid, bp, &hash_lock);
6230 if (hdr == NULL)
6231 return;
6232
6233 /*
6234 * We might be trying to free a block that is still doing I/O
6235 * (i.e. prefetch) or has some other reference (i.e. a dedup-ed,
6236 * dmu_sync-ed block). A block may also have a reference if it is
6237 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6238 * have written the new block to its final resting place on disk but
6239 * without the dedup flag set. This would have left the hdr in the MRU
6240 * state and discoverable. When the txg finally syncs it detects that
6241 * the block was overridden in open context and issues an override I/O.
6242 * Since this is a dedup block, the override I/O will determine if the
6243 * block is already in the DDT. If so, then it will replace the io_bp
6244 * with the bp from the DDT and allow the I/O to finish. When the I/O
6245 * reaches the done callback, dbuf_write_override_done, it will
6246 * check to see if the io_bp and io_bp_override are identical.
6247 * If they are not, then it indicates that the bp was replaced with
6248 * the bp in the DDT and the override bp is freed. This allows
6249 * us to arrive here with a reference on a block that is being
6250 * freed. So if we have an I/O in progress, or a reference to
6251 * this hdr, then we don't destroy the hdr.
6252 */
6253 if (!HDR_HAS_L1HDR(hdr) ||
6254 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6255 arc_change_state(arc_anon, hdr);
6256 arc_hdr_destroy(hdr);
6257 mutex_exit(hash_lock);
6258 } else {
6259 mutex_exit(hash_lock);
6260 }
6261
6262 }
6263
6264 /*
6265 * Release this buffer from the cache, making it an anonymous buffer. This
6266 * must be done after a read and prior to modifying the buffer contents.
6267 * If the buffer has more than one reference, we must make
6268 * a new hdr for the buffer.
6269 */
6270 void
arc_release(arc_buf_t * buf,const void * tag)6271 arc_release(arc_buf_t *buf, const void *tag)
6272 {
6273 arc_buf_hdr_t *hdr = buf->b_hdr;
6274
6275 /*
6276 * It would be nice to assert that if its DMU metadata (level >
6277 * 0 || it's the dnode file), then it must be syncing context.
6278 * But we don't know that information at this level.
6279 */
6280
6281 ASSERT(HDR_HAS_L1HDR(hdr));
6282
6283 /*
6284 * We don't grab the hash lock prior to this check, because if
6285 * the buffer's header is in the arc_anon state, it won't be
6286 * linked into the hash table.
6287 */
6288 if (hdr->b_l1hdr.b_state == arc_anon) {
6289 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6290 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6291 ASSERT(!HDR_HAS_L2HDR(hdr));
6292
6293 ASSERT3P(hdr->b_l1hdr.b_buf, ==, buf);
6294 ASSERT(ARC_BUF_LAST(buf));
6295 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6296 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6297
6298 hdr->b_l1hdr.b_arc_access = 0;
6299
6300 /*
6301 * If the buf is being overridden then it may already
6302 * have a hdr that is not empty.
6303 */
6304 buf_discard_identity(hdr);
6305 arc_buf_thaw(buf);
6306
6307 return;
6308 }
6309
6310 kmutex_t *hash_lock = HDR_LOCK(hdr);
6311 mutex_enter(hash_lock);
6312
6313 /*
6314 * This assignment is only valid as long as the hash_lock is
6315 * held, we must be careful not to reference state or the
6316 * b_state field after dropping the lock.
6317 */
6318 arc_state_t *state = hdr->b_l1hdr.b_state;
6319 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6320 ASSERT3P(state, !=, arc_anon);
6321
6322 /* this buffer is not on any list */
6323 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6324
6325 if (HDR_HAS_L2HDR(hdr)) {
6326 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6327
6328 /*
6329 * We have to recheck this conditional again now that
6330 * we're holding the l2ad_mtx to prevent a race with
6331 * another thread which might be concurrently calling
6332 * l2arc_evict(). In that case, l2arc_evict() might have
6333 * destroyed the header's L2 portion as we were waiting
6334 * to acquire the l2ad_mtx.
6335 */
6336 if (HDR_HAS_L2HDR(hdr))
6337 arc_hdr_l2hdr_destroy(hdr);
6338
6339 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6340 }
6341
6342 /*
6343 * Do we have more than one buf?
6344 */
6345 if (hdr->b_l1hdr.b_buf != buf || !ARC_BUF_LAST(buf)) {
6346 arc_buf_hdr_t *nhdr;
6347 uint64_t spa = hdr->b_spa;
6348 uint64_t psize = HDR_GET_PSIZE(hdr);
6349 uint64_t lsize = HDR_GET_LSIZE(hdr);
6350 boolean_t protected = HDR_PROTECTED(hdr);
6351 enum zio_compress compress = arc_hdr_get_compress(hdr);
6352 arc_buf_contents_t type = arc_buf_type(hdr);
6353 VERIFY3U(hdr->b_type, ==, type);
6354
6355 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6356 VERIFY3S(remove_reference(hdr, tag), >, 0);
6357
6358 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
6359 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6360 ASSERT(ARC_BUF_LAST(buf));
6361 }
6362
6363 /*
6364 * Pull the data off of this hdr and attach it to
6365 * a new anonymous hdr. Also find the last buffer
6366 * in the hdr's buffer list.
6367 */
6368 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6369 ASSERT3P(lastbuf, !=, NULL);
6370
6371 /*
6372 * If the current arc_buf_t and the hdr are sharing their data
6373 * buffer, then we must stop sharing that block.
6374 */
6375 if (ARC_BUF_SHARED(buf)) {
6376 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6377 ASSERT(!arc_buf_is_shared(lastbuf));
6378
6379 /*
6380 * First, sever the block sharing relationship between
6381 * buf and the arc_buf_hdr_t.
6382 */
6383 arc_unshare_buf(hdr, buf);
6384
6385 /*
6386 * Now we need to recreate the hdr's b_pabd. Since we
6387 * have lastbuf handy, we try to share with it, but if
6388 * we can't then we allocate a new b_pabd and copy the
6389 * data from buf into it.
6390 */
6391 if (arc_can_share(hdr, lastbuf)) {
6392 arc_share_buf(hdr, lastbuf);
6393 } else {
6394 arc_hdr_alloc_abd(hdr, 0);
6395 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6396 buf->b_data, psize);
6397 }
6398 VERIFY3P(lastbuf->b_data, !=, NULL);
6399 } else if (HDR_SHARED_DATA(hdr)) {
6400 /*
6401 * Uncompressed shared buffers are always at the end
6402 * of the list. Compressed buffers don't have the
6403 * same requirements. This makes it hard to
6404 * simply assert that the lastbuf is shared so
6405 * we rely on the hdr's compression flags to determine
6406 * if we have a compressed, shared buffer.
6407 */
6408 ASSERT(arc_buf_is_shared(lastbuf) ||
6409 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6410 ASSERT(!arc_buf_is_shared(buf));
6411 }
6412
6413 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6414 ASSERT3P(state, !=, arc_l2c_only);
6415
6416 (void) zfs_refcount_remove_many(&state->arcs_size[type],
6417 arc_buf_size(buf), buf);
6418
6419 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6420 ASSERT3P(state, !=, arc_l2c_only);
6421 (void) zfs_refcount_remove_many(
6422 &state->arcs_esize[type],
6423 arc_buf_size(buf), buf);
6424 }
6425
6426 arc_cksum_verify(buf);
6427 arc_buf_unwatch(buf);
6428
6429 /* if this is the last uncompressed buf free the checksum */
6430 if (!arc_hdr_has_uncompressed_buf(hdr))
6431 arc_cksum_free(hdr);
6432
6433 mutex_exit(hash_lock);
6434
6435 nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6436 compress, hdr->b_complevel, type);
6437 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6438 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
6439 VERIFY3U(nhdr->b_type, ==, type);
6440 ASSERT(!HDR_SHARED_DATA(nhdr));
6441
6442 nhdr->b_l1hdr.b_buf = buf;
6443 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6444 buf->b_hdr = nhdr;
6445
6446 (void) zfs_refcount_add_many(&arc_anon->arcs_size[type],
6447 arc_buf_size(buf), buf);
6448 } else {
6449 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6450 /* protected by hash lock, or hdr is on arc_anon */
6451 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6452 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6453 hdr->b_l1hdr.b_mru_hits = 0;
6454 hdr->b_l1hdr.b_mru_ghost_hits = 0;
6455 hdr->b_l1hdr.b_mfu_hits = 0;
6456 hdr->b_l1hdr.b_mfu_ghost_hits = 0;
6457 arc_change_state(arc_anon, hdr);
6458 hdr->b_l1hdr.b_arc_access = 0;
6459
6460 mutex_exit(hash_lock);
6461 buf_discard_identity(hdr);
6462 arc_buf_thaw(buf);
6463 }
6464 }
6465
6466 int
arc_released(arc_buf_t * buf)6467 arc_released(arc_buf_t *buf)
6468 {
6469 return (buf->b_data != NULL &&
6470 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6471 }
6472
6473 #ifdef ZFS_DEBUG
6474 int
arc_referenced(arc_buf_t * buf)6475 arc_referenced(arc_buf_t *buf)
6476 {
6477 return (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6478 }
6479 #endif
6480
6481 static void
arc_write_ready(zio_t * zio)6482 arc_write_ready(zio_t *zio)
6483 {
6484 arc_write_callback_t *callback = zio->io_private;
6485 arc_buf_t *buf = callback->awcb_buf;
6486 arc_buf_hdr_t *hdr = buf->b_hdr;
6487 blkptr_t *bp = zio->io_bp;
6488 uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6489 fstrans_cookie_t cookie = spl_fstrans_mark();
6490
6491 ASSERT(HDR_HAS_L1HDR(hdr));
6492 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6493 ASSERT3P(hdr->b_l1hdr.b_buf, !=, NULL);
6494
6495 /*
6496 * If we're reexecuting this zio because the pool suspended, then
6497 * cleanup any state that was previously set the first time the
6498 * callback was invoked.
6499 */
6500 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6501 arc_cksum_free(hdr);
6502 arc_buf_unwatch(buf);
6503 if (hdr->b_l1hdr.b_pabd != NULL) {
6504 if (ARC_BUF_SHARED(buf)) {
6505 arc_unshare_buf(hdr, buf);
6506 } else {
6507 ASSERT(!arc_buf_is_shared(buf));
6508 arc_hdr_free_abd(hdr, B_FALSE);
6509 }
6510 }
6511
6512 if (HDR_HAS_RABD(hdr))
6513 arc_hdr_free_abd(hdr, B_TRUE);
6514 }
6515 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6516 ASSERT(!HDR_HAS_RABD(hdr));
6517 ASSERT(!HDR_SHARED_DATA(hdr));
6518 ASSERT(!arc_buf_is_shared(buf));
6519
6520 callback->awcb_ready(zio, buf, callback->awcb_private);
6521
6522 if (HDR_IO_IN_PROGRESS(hdr)) {
6523 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6524 } else {
6525 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6526 add_reference(hdr, hdr); /* For IO_IN_PROGRESS. */
6527 }
6528
6529 if (BP_IS_PROTECTED(bp)) {
6530 /* ZIL blocks are written through zio_rewrite */
6531 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
6532
6533 if (BP_SHOULD_BYTESWAP(bp)) {
6534 if (BP_GET_LEVEL(bp) > 0) {
6535 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
6536 } else {
6537 hdr->b_l1hdr.b_byteswap =
6538 DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
6539 }
6540 } else {
6541 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
6542 }
6543
6544 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
6545 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
6546 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
6547 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
6548 hdr->b_crypt_hdr.b_iv);
6549 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
6550 } else {
6551 arc_hdr_clear_flags(hdr, ARC_FLAG_PROTECTED);
6552 }
6553
6554 /*
6555 * If this block was written for raw encryption but the zio layer
6556 * ended up only authenticating it, adjust the buffer flags now.
6557 */
6558 if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
6559 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6560 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6561 if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
6562 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6563 } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
6564 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6565 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6566 }
6567
6568 /* this must be done after the buffer flags are adjusted */
6569 arc_cksum_compute(buf);
6570
6571 enum zio_compress compress;
6572 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
6573 compress = ZIO_COMPRESS_OFF;
6574 } else {
6575 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
6576 compress = BP_GET_COMPRESS(bp);
6577 }
6578 HDR_SET_PSIZE(hdr, psize);
6579 arc_hdr_set_compress(hdr, compress);
6580 hdr->b_complevel = zio->io_prop.zp_complevel;
6581
6582 if (zio->io_error != 0 || psize == 0)
6583 goto out;
6584
6585 /*
6586 * Fill the hdr with data. If the buffer is encrypted we have no choice
6587 * but to copy the data into b_radb. If the hdr is compressed, the data
6588 * we want is available from the zio, otherwise we can take it from
6589 * the buf.
6590 *
6591 * We might be able to share the buf's data with the hdr here. However,
6592 * doing so would cause the ARC to be full of linear ABDs if we write a
6593 * lot of shareable data. As a compromise, we check whether scattered
6594 * ABDs are allowed, and assume that if they are then the user wants
6595 * the ARC to be primarily filled with them regardless of the data being
6596 * written. Therefore, if they're allowed then we allocate one and copy
6597 * the data into it; otherwise, we share the data directly if we can.
6598 */
6599 if (ARC_BUF_ENCRYPTED(buf)) {
6600 ASSERT3U(psize, >, 0);
6601 ASSERT(ARC_BUF_COMPRESSED(buf));
6602 arc_hdr_alloc_abd(hdr, ARC_HDR_ALLOC_RDATA |
6603 ARC_HDR_USE_RESERVE);
6604 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6605 } else if (!(HDR_UNCACHED(hdr) ||
6606 abd_size_alloc_linear(arc_buf_size(buf))) ||
6607 !arc_can_share(hdr, buf)) {
6608 /*
6609 * Ideally, we would always copy the io_abd into b_pabd, but the
6610 * user may have disabled compressed ARC, thus we must check the
6611 * hdr's compression setting rather than the io_bp's.
6612 */
6613 if (BP_IS_ENCRYPTED(bp)) {
6614 ASSERT3U(psize, >, 0);
6615 arc_hdr_alloc_abd(hdr, ARC_HDR_ALLOC_RDATA |
6616 ARC_HDR_USE_RESERVE);
6617 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6618 } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
6619 !ARC_BUF_COMPRESSED(buf)) {
6620 ASSERT3U(psize, >, 0);
6621 arc_hdr_alloc_abd(hdr, ARC_HDR_USE_RESERVE);
6622 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6623 } else {
6624 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6625 arc_hdr_alloc_abd(hdr, ARC_HDR_USE_RESERVE);
6626 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6627 arc_buf_size(buf));
6628 }
6629 } else {
6630 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6631 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6632 ASSERT3P(hdr->b_l1hdr.b_buf, ==, buf);
6633 ASSERT(ARC_BUF_LAST(buf));
6634
6635 arc_share_buf(hdr, buf);
6636 }
6637
6638 out:
6639 arc_hdr_verify(hdr, bp);
6640 spl_fstrans_unmark(cookie);
6641 }
6642
6643 static void
arc_write_children_ready(zio_t * zio)6644 arc_write_children_ready(zio_t *zio)
6645 {
6646 arc_write_callback_t *callback = zio->io_private;
6647 arc_buf_t *buf = callback->awcb_buf;
6648
6649 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6650 }
6651
6652 static void
arc_write_done(zio_t * zio)6653 arc_write_done(zio_t *zio)
6654 {
6655 arc_write_callback_t *callback = zio->io_private;
6656 arc_buf_t *buf = callback->awcb_buf;
6657 arc_buf_hdr_t *hdr = buf->b_hdr;
6658
6659 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6660
6661 if (zio->io_error == 0) {
6662 arc_hdr_verify(hdr, zio->io_bp);
6663
6664 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6665 buf_discard_identity(hdr);
6666 } else {
6667 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6668 hdr->b_birth = BP_GET_BIRTH(zio->io_bp);
6669 }
6670 } else {
6671 ASSERT(HDR_EMPTY(hdr));
6672 }
6673
6674 /*
6675 * If the block to be written was all-zero or compressed enough to be
6676 * embedded in the BP, no write was performed so there will be no
6677 * dva/birth/checksum. The buffer must therefore remain anonymous
6678 * (and uncached).
6679 */
6680 if (!HDR_EMPTY(hdr)) {
6681 arc_buf_hdr_t *exists;
6682 kmutex_t *hash_lock;
6683
6684 ASSERT3U(zio->io_error, ==, 0);
6685
6686 arc_cksum_verify(buf);
6687
6688 exists = buf_hash_insert(hdr, &hash_lock);
6689 if (exists != NULL) {
6690 /*
6691 * This can only happen if we overwrite for
6692 * sync-to-convergence, because we remove
6693 * buffers from the hash table when we arc_free().
6694 */
6695 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6696 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6697 panic("bad overwrite, hdr=%p exists=%p",
6698 (void *)hdr, (void *)exists);
6699 ASSERT(zfs_refcount_is_zero(
6700 &exists->b_l1hdr.b_refcnt));
6701 arc_change_state(arc_anon, exists);
6702 arc_hdr_destroy(exists);
6703 mutex_exit(hash_lock);
6704 exists = buf_hash_insert(hdr, &hash_lock);
6705 ASSERT3P(exists, ==, NULL);
6706 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6707 /* nopwrite */
6708 ASSERT(zio->io_prop.zp_nopwrite);
6709 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6710 panic("bad nopwrite, hdr=%p exists=%p",
6711 (void *)hdr, (void *)exists);
6712 } else {
6713 /* Dedup */
6714 ASSERT3P(hdr->b_l1hdr.b_buf, !=, NULL);
6715 ASSERT(ARC_BUF_LAST(hdr->b_l1hdr.b_buf));
6716 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6717 ASSERT(BP_GET_DEDUP(zio->io_bp));
6718 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6719 }
6720 }
6721 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6722 VERIFY3S(remove_reference(hdr, hdr), >, 0);
6723 /* if it's not anon, we are doing a scrub */
6724 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6725 arc_access(hdr, 0, B_FALSE);
6726 mutex_exit(hash_lock);
6727 } else {
6728 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6729 VERIFY3S(remove_reference(hdr, hdr), >, 0);
6730 }
6731
6732 callback->awcb_done(zio, buf, callback->awcb_private);
6733
6734 abd_free(zio->io_abd);
6735 kmem_free(callback, sizeof (arc_write_callback_t));
6736 }
6737
6738 zio_t *
arc_write(zio_t * pio,spa_t * spa,uint64_t txg,blkptr_t * bp,arc_buf_t * buf,boolean_t uncached,boolean_t l2arc,const zio_prop_t * zp,arc_write_done_func_t * ready,arc_write_done_func_t * children_ready,arc_write_done_func_t * done,void * private,zio_priority_t priority,int zio_flags,const zbookmark_phys_t * zb)6739 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
6740 blkptr_t *bp, arc_buf_t *buf, boolean_t uncached, boolean_t l2arc,
6741 const zio_prop_t *zp, arc_write_done_func_t *ready,
6742 arc_write_done_func_t *children_ready, arc_write_done_func_t *done,
6743 void *private, zio_priority_t priority, int zio_flags,
6744 const zbookmark_phys_t *zb)
6745 {
6746 arc_buf_hdr_t *hdr = buf->b_hdr;
6747 arc_write_callback_t *callback;
6748 zio_t *zio;
6749 zio_prop_t localprop = *zp;
6750
6751 ASSERT3P(ready, !=, NULL);
6752 ASSERT3P(done, !=, NULL);
6753 ASSERT(!HDR_IO_ERROR(hdr));
6754 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6755 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6756 ASSERT3P(hdr->b_l1hdr.b_buf, !=, NULL);
6757 if (uncached)
6758 arc_hdr_set_flags(hdr, ARC_FLAG_UNCACHED);
6759 else if (l2arc)
6760 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6761
6762 if (ARC_BUF_ENCRYPTED(buf)) {
6763 ASSERT(ARC_BUF_COMPRESSED(buf));
6764 localprop.zp_encrypt = B_TRUE;
6765 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6766 localprop.zp_complevel = hdr->b_complevel;
6767 localprop.zp_byteorder =
6768 (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
6769 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
6770 memcpy(localprop.zp_salt, hdr->b_crypt_hdr.b_salt,
6771 ZIO_DATA_SALT_LEN);
6772 memcpy(localprop.zp_iv, hdr->b_crypt_hdr.b_iv,
6773 ZIO_DATA_IV_LEN);
6774 memcpy(localprop.zp_mac, hdr->b_crypt_hdr.b_mac,
6775 ZIO_DATA_MAC_LEN);
6776 if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
6777 localprop.zp_nopwrite = B_FALSE;
6778 localprop.zp_copies =
6779 MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
6780 }
6781 zio_flags |= ZIO_FLAG_RAW;
6782 } else if (ARC_BUF_COMPRESSED(buf)) {
6783 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6784 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6785 localprop.zp_complevel = hdr->b_complevel;
6786 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6787 }
6788 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6789 callback->awcb_ready = ready;
6790 callback->awcb_children_ready = children_ready;
6791 callback->awcb_done = done;
6792 callback->awcb_private = private;
6793 callback->awcb_buf = buf;
6794
6795 /*
6796 * The hdr's b_pabd is now stale, free it now. A new data block
6797 * will be allocated when the zio pipeline calls arc_write_ready().
6798 */
6799 if (hdr->b_l1hdr.b_pabd != NULL) {
6800 /*
6801 * If the buf is currently sharing the data block with
6802 * the hdr then we need to break that relationship here.
6803 * The hdr will remain with a NULL data pointer and the
6804 * buf will take sole ownership of the block.
6805 */
6806 if (ARC_BUF_SHARED(buf)) {
6807 arc_unshare_buf(hdr, buf);
6808 } else {
6809 ASSERT(!arc_buf_is_shared(buf));
6810 arc_hdr_free_abd(hdr, B_FALSE);
6811 }
6812 VERIFY3P(buf->b_data, !=, NULL);
6813 }
6814
6815 if (HDR_HAS_RABD(hdr))
6816 arc_hdr_free_abd(hdr, B_TRUE);
6817
6818 if (!(zio_flags & ZIO_FLAG_RAW))
6819 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6820
6821 ASSERT(!arc_buf_is_shared(buf));
6822 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6823
6824 zio = zio_write(pio, spa, txg, bp,
6825 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6826 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6827 (children_ready != NULL) ? arc_write_children_ready : NULL,
6828 arc_write_done, callback, priority, zio_flags, zb);
6829
6830 return (zio);
6831 }
6832
6833 void
arc_tempreserve_clear(uint64_t reserve)6834 arc_tempreserve_clear(uint64_t reserve)
6835 {
6836 atomic_add_64(&arc_tempreserve, -reserve);
6837 ASSERT((int64_t)arc_tempreserve >= 0);
6838 }
6839
6840 int
arc_tempreserve_space(spa_t * spa,uint64_t reserve,uint64_t txg)6841 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
6842 {
6843 int error;
6844 uint64_t anon_size;
6845
6846 if (!arc_no_grow &&
6847 reserve > arc_c/4 &&
6848 reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT))
6849 arc_c = MIN(arc_c_max, reserve * 4);
6850
6851 /*
6852 * Throttle when the calculated memory footprint for the TXG
6853 * exceeds the target ARC size.
6854 */
6855 if (reserve > arc_c) {
6856 DMU_TX_STAT_BUMP(dmu_tx_memory_reserve);
6857 return (SET_ERROR(ERESTART));
6858 }
6859
6860 /*
6861 * Don't count loaned bufs as in flight dirty data to prevent long
6862 * network delays from blocking transactions that are ready to be
6863 * assigned to a txg.
6864 */
6865
6866 /* assert that it has not wrapped around */
6867 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6868
6869 anon_size = MAX((int64_t)
6870 (zfs_refcount_count(&arc_anon->arcs_size[ARC_BUFC_DATA]) +
6871 zfs_refcount_count(&arc_anon->arcs_size[ARC_BUFC_METADATA]) -
6872 arc_loaned_bytes), 0);
6873
6874 /*
6875 * Writes will, almost always, require additional memory allocations
6876 * in order to compress/encrypt/etc the data. We therefore need to
6877 * make sure that there is sufficient available memory for this.
6878 */
6879 error = arc_memory_throttle(spa, reserve, txg);
6880 if (error != 0)
6881 return (error);
6882
6883 /*
6884 * Throttle writes when the amount of dirty data in the cache
6885 * gets too large. We try to keep the cache less than half full
6886 * of dirty blocks so that our sync times don't grow too large.
6887 *
6888 * In the case of one pool being built on another pool, we want
6889 * to make sure we don't end up throttling the lower (backing)
6890 * pool when the upper pool is the majority contributor to dirty
6891 * data. To insure we make forward progress during throttling, we
6892 * also check the current pool's net dirty data and only throttle
6893 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
6894 * data in the cache.
6895 *
6896 * Note: if two requests come in concurrently, we might let them
6897 * both succeed, when one of them should fail. Not a huge deal.
6898 */
6899 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
6900 uint64_t spa_dirty_anon = spa_dirty_data(spa);
6901 uint64_t rarc_c = arc_warm ? arc_c : arc_c_max;
6902 if (total_dirty > rarc_c * zfs_arc_dirty_limit_percent / 100 &&
6903 anon_size > rarc_c * zfs_arc_anon_limit_percent / 100 &&
6904 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
6905 #ifdef ZFS_DEBUG
6906 uint64_t meta_esize = zfs_refcount_count(
6907 &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6908 uint64_t data_esize =
6909 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6910 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6911 "anon_data=%lluK tempreserve=%lluK rarc_c=%lluK\n",
6912 (u_longlong_t)arc_tempreserve >> 10,
6913 (u_longlong_t)meta_esize >> 10,
6914 (u_longlong_t)data_esize >> 10,
6915 (u_longlong_t)reserve >> 10,
6916 (u_longlong_t)rarc_c >> 10);
6917 #endif
6918 DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle);
6919 return (SET_ERROR(ERESTART));
6920 }
6921 atomic_add_64(&arc_tempreserve, reserve);
6922 return (0);
6923 }
6924
6925 static void
arc_kstat_update_state(arc_state_t * state,kstat_named_t * size,kstat_named_t * data,kstat_named_t * metadata,kstat_named_t * evict_data,kstat_named_t * evict_metadata)6926 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6927 kstat_named_t *data, kstat_named_t *metadata,
6928 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6929 {
6930 data->value.ui64 =
6931 zfs_refcount_count(&state->arcs_size[ARC_BUFC_DATA]);
6932 metadata->value.ui64 =
6933 zfs_refcount_count(&state->arcs_size[ARC_BUFC_METADATA]);
6934 size->value.ui64 = data->value.ui64 + metadata->value.ui64;
6935 evict_data->value.ui64 =
6936 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6937 evict_metadata->value.ui64 =
6938 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6939 }
6940
6941 static int
arc_kstat_update(kstat_t * ksp,int rw)6942 arc_kstat_update(kstat_t *ksp, int rw)
6943 {
6944 arc_stats_t *as = ksp->ks_data;
6945
6946 if (rw == KSTAT_WRITE)
6947 return (SET_ERROR(EACCES));
6948
6949 as->arcstat_hits.value.ui64 =
6950 wmsum_value(&arc_sums.arcstat_hits);
6951 as->arcstat_iohits.value.ui64 =
6952 wmsum_value(&arc_sums.arcstat_iohits);
6953 as->arcstat_misses.value.ui64 =
6954 wmsum_value(&arc_sums.arcstat_misses);
6955 as->arcstat_demand_data_hits.value.ui64 =
6956 wmsum_value(&arc_sums.arcstat_demand_data_hits);
6957 as->arcstat_demand_data_iohits.value.ui64 =
6958 wmsum_value(&arc_sums.arcstat_demand_data_iohits);
6959 as->arcstat_demand_data_misses.value.ui64 =
6960 wmsum_value(&arc_sums.arcstat_demand_data_misses);
6961 as->arcstat_demand_metadata_hits.value.ui64 =
6962 wmsum_value(&arc_sums.arcstat_demand_metadata_hits);
6963 as->arcstat_demand_metadata_iohits.value.ui64 =
6964 wmsum_value(&arc_sums.arcstat_demand_metadata_iohits);
6965 as->arcstat_demand_metadata_misses.value.ui64 =
6966 wmsum_value(&arc_sums.arcstat_demand_metadata_misses);
6967 as->arcstat_prefetch_data_hits.value.ui64 =
6968 wmsum_value(&arc_sums.arcstat_prefetch_data_hits);
6969 as->arcstat_prefetch_data_iohits.value.ui64 =
6970 wmsum_value(&arc_sums.arcstat_prefetch_data_iohits);
6971 as->arcstat_prefetch_data_misses.value.ui64 =
6972 wmsum_value(&arc_sums.arcstat_prefetch_data_misses);
6973 as->arcstat_prefetch_metadata_hits.value.ui64 =
6974 wmsum_value(&arc_sums.arcstat_prefetch_metadata_hits);
6975 as->arcstat_prefetch_metadata_iohits.value.ui64 =
6976 wmsum_value(&arc_sums.arcstat_prefetch_metadata_iohits);
6977 as->arcstat_prefetch_metadata_misses.value.ui64 =
6978 wmsum_value(&arc_sums.arcstat_prefetch_metadata_misses);
6979 as->arcstat_mru_hits.value.ui64 =
6980 wmsum_value(&arc_sums.arcstat_mru_hits);
6981 as->arcstat_mru_ghost_hits.value.ui64 =
6982 wmsum_value(&arc_sums.arcstat_mru_ghost_hits);
6983 as->arcstat_mfu_hits.value.ui64 =
6984 wmsum_value(&arc_sums.arcstat_mfu_hits);
6985 as->arcstat_mfu_ghost_hits.value.ui64 =
6986 wmsum_value(&arc_sums.arcstat_mfu_ghost_hits);
6987 as->arcstat_uncached_hits.value.ui64 =
6988 wmsum_value(&arc_sums.arcstat_uncached_hits);
6989 as->arcstat_deleted.value.ui64 =
6990 wmsum_value(&arc_sums.arcstat_deleted);
6991 as->arcstat_mutex_miss.value.ui64 =
6992 wmsum_value(&arc_sums.arcstat_mutex_miss);
6993 as->arcstat_access_skip.value.ui64 =
6994 wmsum_value(&arc_sums.arcstat_access_skip);
6995 as->arcstat_evict_skip.value.ui64 =
6996 wmsum_value(&arc_sums.arcstat_evict_skip);
6997 as->arcstat_evict_not_enough.value.ui64 =
6998 wmsum_value(&arc_sums.arcstat_evict_not_enough);
6999 as->arcstat_evict_l2_cached.value.ui64 =
7000 wmsum_value(&arc_sums.arcstat_evict_l2_cached);
7001 as->arcstat_evict_l2_eligible.value.ui64 =
7002 wmsum_value(&arc_sums.arcstat_evict_l2_eligible);
7003 as->arcstat_evict_l2_eligible_mfu.value.ui64 =
7004 wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mfu);
7005 as->arcstat_evict_l2_eligible_mru.value.ui64 =
7006 wmsum_value(&arc_sums.arcstat_evict_l2_eligible_mru);
7007 as->arcstat_evict_l2_ineligible.value.ui64 =
7008 wmsum_value(&arc_sums.arcstat_evict_l2_ineligible);
7009 as->arcstat_evict_l2_skip.value.ui64 =
7010 wmsum_value(&arc_sums.arcstat_evict_l2_skip);
7011 as->arcstat_hash_collisions.value.ui64 =
7012 wmsum_value(&arc_sums.arcstat_hash_collisions);
7013 as->arcstat_hash_chains.value.ui64 =
7014 wmsum_value(&arc_sums.arcstat_hash_chains);
7015 as->arcstat_size.value.ui64 =
7016 aggsum_value(&arc_sums.arcstat_size);
7017 as->arcstat_compressed_size.value.ui64 =
7018 wmsum_value(&arc_sums.arcstat_compressed_size);
7019 as->arcstat_uncompressed_size.value.ui64 =
7020 wmsum_value(&arc_sums.arcstat_uncompressed_size);
7021 as->arcstat_overhead_size.value.ui64 =
7022 wmsum_value(&arc_sums.arcstat_overhead_size);
7023 as->arcstat_hdr_size.value.ui64 =
7024 wmsum_value(&arc_sums.arcstat_hdr_size);
7025 as->arcstat_data_size.value.ui64 =
7026 wmsum_value(&arc_sums.arcstat_data_size);
7027 as->arcstat_metadata_size.value.ui64 =
7028 wmsum_value(&arc_sums.arcstat_metadata_size);
7029 as->arcstat_dbuf_size.value.ui64 =
7030 wmsum_value(&arc_sums.arcstat_dbuf_size);
7031 #if defined(COMPAT_FREEBSD11)
7032 as->arcstat_other_size.value.ui64 =
7033 wmsum_value(&arc_sums.arcstat_bonus_size) +
7034 wmsum_value(&arc_sums.arcstat_dnode_size) +
7035 wmsum_value(&arc_sums.arcstat_dbuf_size);
7036 #endif
7037
7038 arc_kstat_update_state(arc_anon,
7039 &as->arcstat_anon_size,
7040 &as->arcstat_anon_data,
7041 &as->arcstat_anon_metadata,
7042 &as->arcstat_anon_evictable_data,
7043 &as->arcstat_anon_evictable_metadata);
7044 arc_kstat_update_state(arc_mru,
7045 &as->arcstat_mru_size,
7046 &as->arcstat_mru_data,
7047 &as->arcstat_mru_metadata,
7048 &as->arcstat_mru_evictable_data,
7049 &as->arcstat_mru_evictable_metadata);
7050 arc_kstat_update_state(arc_mru_ghost,
7051 &as->arcstat_mru_ghost_size,
7052 &as->arcstat_mru_ghost_data,
7053 &as->arcstat_mru_ghost_metadata,
7054 &as->arcstat_mru_ghost_evictable_data,
7055 &as->arcstat_mru_ghost_evictable_metadata);
7056 arc_kstat_update_state(arc_mfu,
7057 &as->arcstat_mfu_size,
7058 &as->arcstat_mfu_data,
7059 &as->arcstat_mfu_metadata,
7060 &as->arcstat_mfu_evictable_data,
7061 &as->arcstat_mfu_evictable_metadata);
7062 arc_kstat_update_state(arc_mfu_ghost,
7063 &as->arcstat_mfu_ghost_size,
7064 &as->arcstat_mfu_ghost_data,
7065 &as->arcstat_mfu_ghost_metadata,
7066 &as->arcstat_mfu_ghost_evictable_data,
7067 &as->arcstat_mfu_ghost_evictable_metadata);
7068 arc_kstat_update_state(arc_uncached,
7069 &as->arcstat_uncached_size,
7070 &as->arcstat_uncached_data,
7071 &as->arcstat_uncached_metadata,
7072 &as->arcstat_uncached_evictable_data,
7073 &as->arcstat_uncached_evictable_metadata);
7074
7075 as->arcstat_dnode_size.value.ui64 =
7076 wmsum_value(&arc_sums.arcstat_dnode_size);
7077 as->arcstat_bonus_size.value.ui64 =
7078 wmsum_value(&arc_sums.arcstat_bonus_size);
7079 as->arcstat_l2_hits.value.ui64 =
7080 wmsum_value(&arc_sums.arcstat_l2_hits);
7081 as->arcstat_l2_misses.value.ui64 =
7082 wmsum_value(&arc_sums.arcstat_l2_misses);
7083 as->arcstat_l2_prefetch_asize.value.ui64 =
7084 wmsum_value(&arc_sums.arcstat_l2_prefetch_asize);
7085 as->arcstat_l2_mru_asize.value.ui64 =
7086 wmsum_value(&arc_sums.arcstat_l2_mru_asize);
7087 as->arcstat_l2_mfu_asize.value.ui64 =
7088 wmsum_value(&arc_sums.arcstat_l2_mfu_asize);
7089 as->arcstat_l2_bufc_data_asize.value.ui64 =
7090 wmsum_value(&arc_sums.arcstat_l2_bufc_data_asize);
7091 as->arcstat_l2_bufc_metadata_asize.value.ui64 =
7092 wmsum_value(&arc_sums.arcstat_l2_bufc_metadata_asize);
7093 as->arcstat_l2_feeds.value.ui64 =
7094 wmsum_value(&arc_sums.arcstat_l2_feeds);
7095 as->arcstat_l2_rw_clash.value.ui64 =
7096 wmsum_value(&arc_sums.arcstat_l2_rw_clash);
7097 as->arcstat_l2_read_bytes.value.ui64 =
7098 wmsum_value(&arc_sums.arcstat_l2_read_bytes);
7099 as->arcstat_l2_write_bytes.value.ui64 =
7100 wmsum_value(&arc_sums.arcstat_l2_write_bytes);
7101 as->arcstat_l2_writes_sent.value.ui64 =
7102 wmsum_value(&arc_sums.arcstat_l2_writes_sent);
7103 as->arcstat_l2_writes_done.value.ui64 =
7104 wmsum_value(&arc_sums.arcstat_l2_writes_done);
7105 as->arcstat_l2_writes_error.value.ui64 =
7106 wmsum_value(&arc_sums.arcstat_l2_writes_error);
7107 as->arcstat_l2_writes_lock_retry.value.ui64 =
7108 wmsum_value(&arc_sums.arcstat_l2_writes_lock_retry);
7109 as->arcstat_l2_evict_lock_retry.value.ui64 =
7110 wmsum_value(&arc_sums.arcstat_l2_evict_lock_retry);
7111 as->arcstat_l2_evict_reading.value.ui64 =
7112 wmsum_value(&arc_sums.arcstat_l2_evict_reading);
7113 as->arcstat_l2_evict_l1cached.value.ui64 =
7114 wmsum_value(&arc_sums.arcstat_l2_evict_l1cached);
7115 as->arcstat_l2_free_on_write.value.ui64 =
7116 wmsum_value(&arc_sums.arcstat_l2_free_on_write);
7117 as->arcstat_l2_abort_lowmem.value.ui64 =
7118 wmsum_value(&arc_sums.arcstat_l2_abort_lowmem);
7119 as->arcstat_l2_cksum_bad.value.ui64 =
7120 wmsum_value(&arc_sums.arcstat_l2_cksum_bad);
7121 as->arcstat_l2_io_error.value.ui64 =
7122 wmsum_value(&arc_sums.arcstat_l2_io_error);
7123 as->arcstat_l2_lsize.value.ui64 =
7124 wmsum_value(&arc_sums.arcstat_l2_lsize);
7125 as->arcstat_l2_psize.value.ui64 =
7126 wmsum_value(&arc_sums.arcstat_l2_psize);
7127 as->arcstat_l2_hdr_size.value.ui64 =
7128 aggsum_value(&arc_sums.arcstat_l2_hdr_size);
7129 as->arcstat_l2_log_blk_writes.value.ui64 =
7130 wmsum_value(&arc_sums.arcstat_l2_log_blk_writes);
7131 as->arcstat_l2_log_blk_asize.value.ui64 =
7132 wmsum_value(&arc_sums.arcstat_l2_log_blk_asize);
7133 as->arcstat_l2_log_blk_count.value.ui64 =
7134 wmsum_value(&arc_sums.arcstat_l2_log_blk_count);
7135 as->arcstat_l2_rebuild_success.value.ui64 =
7136 wmsum_value(&arc_sums.arcstat_l2_rebuild_success);
7137 as->arcstat_l2_rebuild_abort_unsupported.value.ui64 =
7138 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_unsupported);
7139 as->arcstat_l2_rebuild_abort_io_errors.value.ui64 =
7140 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_io_errors);
7141 as->arcstat_l2_rebuild_abort_dh_errors.value.ui64 =
7142 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_dh_errors);
7143 as->arcstat_l2_rebuild_abort_cksum_lb_errors.value.ui64 =
7144 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors);
7145 as->arcstat_l2_rebuild_abort_lowmem.value.ui64 =
7146 wmsum_value(&arc_sums.arcstat_l2_rebuild_abort_lowmem);
7147 as->arcstat_l2_rebuild_size.value.ui64 =
7148 wmsum_value(&arc_sums.arcstat_l2_rebuild_size);
7149 as->arcstat_l2_rebuild_asize.value.ui64 =
7150 wmsum_value(&arc_sums.arcstat_l2_rebuild_asize);
7151 as->arcstat_l2_rebuild_bufs.value.ui64 =
7152 wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs);
7153 as->arcstat_l2_rebuild_bufs_precached.value.ui64 =
7154 wmsum_value(&arc_sums.arcstat_l2_rebuild_bufs_precached);
7155 as->arcstat_l2_rebuild_log_blks.value.ui64 =
7156 wmsum_value(&arc_sums.arcstat_l2_rebuild_log_blks);
7157 as->arcstat_memory_throttle_count.value.ui64 =
7158 wmsum_value(&arc_sums.arcstat_memory_throttle_count);
7159 as->arcstat_memory_direct_count.value.ui64 =
7160 wmsum_value(&arc_sums.arcstat_memory_direct_count);
7161 as->arcstat_memory_indirect_count.value.ui64 =
7162 wmsum_value(&arc_sums.arcstat_memory_indirect_count);
7163
7164 as->arcstat_memory_all_bytes.value.ui64 =
7165 arc_all_memory();
7166 as->arcstat_memory_free_bytes.value.ui64 =
7167 arc_free_memory();
7168 as->arcstat_memory_available_bytes.value.i64 =
7169 arc_available_memory();
7170
7171 as->arcstat_prune.value.ui64 =
7172 wmsum_value(&arc_sums.arcstat_prune);
7173 as->arcstat_meta_used.value.ui64 =
7174 wmsum_value(&arc_sums.arcstat_meta_used);
7175 as->arcstat_async_upgrade_sync.value.ui64 =
7176 wmsum_value(&arc_sums.arcstat_async_upgrade_sync);
7177 as->arcstat_predictive_prefetch.value.ui64 =
7178 wmsum_value(&arc_sums.arcstat_predictive_prefetch);
7179 as->arcstat_demand_hit_predictive_prefetch.value.ui64 =
7180 wmsum_value(&arc_sums.arcstat_demand_hit_predictive_prefetch);
7181 as->arcstat_demand_iohit_predictive_prefetch.value.ui64 =
7182 wmsum_value(&arc_sums.arcstat_demand_iohit_predictive_prefetch);
7183 as->arcstat_prescient_prefetch.value.ui64 =
7184 wmsum_value(&arc_sums.arcstat_prescient_prefetch);
7185 as->arcstat_demand_hit_prescient_prefetch.value.ui64 =
7186 wmsum_value(&arc_sums.arcstat_demand_hit_prescient_prefetch);
7187 as->arcstat_demand_iohit_prescient_prefetch.value.ui64 =
7188 wmsum_value(&arc_sums.arcstat_demand_iohit_prescient_prefetch);
7189 as->arcstat_raw_size.value.ui64 =
7190 wmsum_value(&arc_sums.arcstat_raw_size);
7191 as->arcstat_cached_only_in_progress.value.ui64 =
7192 wmsum_value(&arc_sums.arcstat_cached_only_in_progress);
7193 as->arcstat_abd_chunk_waste_size.value.ui64 =
7194 wmsum_value(&arc_sums.arcstat_abd_chunk_waste_size);
7195
7196 return (0);
7197 }
7198
7199 /*
7200 * This function *must* return indices evenly distributed between all
7201 * sublists of the multilist. This is needed due to how the ARC eviction
7202 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7203 * distributed between all sublists and uses this assumption when
7204 * deciding which sublist to evict from and how much to evict from it.
7205 */
7206 static unsigned int
arc_state_multilist_index_func(multilist_t * ml,void * obj)7207 arc_state_multilist_index_func(multilist_t *ml, void *obj)
7208 {
7209 arc_buf_hdr_t *hdr = obj;
7210
7211 /*
7212 * We rely on b_dva to generate evenly distributed index
7213 * numbers using buf_hash below. So, as an added precaution,
7214 * let's make sure we never add empty buffers to the arc lists.
7215 */
7216 ASSERT(!HDR_EMPTY(hdr));
7217
7218 /*
7219 * The assumption here, is the hash value for a given
7220 * arc_buf_hdr_t will remain constant throughout its lifetime
7221 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7222 * Thus, we don't need to store the header's sublist index
7223 * on insertion, as this index can be recalculated on removal.
7224 *
7225 * Also, the low order bits of the hash value are thought to be
7226 * distributed evenly. Otherwise, in the case that the multilist
7227 * has a power of two number of sublists, each sublists' usage
7228 * would not be evenly distributed. In this context full 64bit
7229 * division would be a waste of time, so limit it to 32 bits.
7230 */
7231 return ((unsigned int)buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
7232 multilist_get_num_sublists(ml));
7233 }
7234
7235 static unsigned int
arc_state_l2c_multilist_index_func(multilist_t * ml,void * obj)7236 arc_state_l2c_multilist_index_func(multilist_t *ml, void *obj)
7237 {
7238 panic("Header %p insert into arc_l2c_only %p", obj, ml);
7239 }
7240
7241 #define WARN_IF_TUNING_IGNORED(tuning, value, do_warn) do { \
7242 if ((do_warn) && (tuning) && ((tuning) != (value))) { \
7243 cmn_err(CE_WARN, \
7244 "ignoring tunable %s (using %llu instead)", \
7245 (#tuning), (u_longlong_t)(value)); \
7246 } \
7247 } while (0)
7248
7249 /*
7250 * Called during module initialization and periodically thereafter to
7251 * apply reasonable changes to the exposed performance tunings. Can also be
7252 * called explicitly by param_set_arc_*() functions when ARC tunables are
7253 * updated manually. Non-zero zfs_* values which differ from the currently set
7254 * values will be applied.
7255 */
7256 void
arc_tuning_update(boolean_t verbose)7257 arc_tuning_update(boolean_t verbose)
7258 {
7259 uint64_t allmem = arc_all_memory();
7260
7261 /* Valid range: 32M - <arc_c_max> */
7262 if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) &&
7263 (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) &&
7264 (zfs_arc_min <= arc_c_max)) {
7265 arc_c_min = zfs_arc_min;
7266 arc_c = MAX(arc_c, arc_c_min);
7267 }
7268 WARN_IF_TUNING_IGNORED(zfs_arc_min, arc_c_min, verbose);
7269
7270 /* Valid range: 64M - <all physical memory> */
7271 if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) &&
7272 (zfs_arc_max >= MIN_ARC_MAX) && (zfs_arc_max < allmem) &&
7273 (zfs_arc_max > arc_c_min)) {
7274 arc_c_max = zfs_arc_max;
7275 arc_c = MIN(arc_c, arc_c_max);
7276 if (arc_dnode_limit > arc_c_max)
7277 arc_dnode_limit = arc_c_max;
7278 }
7279 WARN_IF_TUNING_IGNORED(zfs_arc_max, arc_c_max, verbose);
7280
7281 /* Valid range: 0 - <all physical memory> */
7282 arc_dnode_limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit :
7283 MIN(zfs_arc_dnode_limit_percent, 100) * arc_c_max / 100;
7284 WARN_IF_TUNING_IGNORED(zfs_arc_dnode_limit, arc_dnode_limit, verbose);
7285
7286 /* Valid range: 1 - N */
7287 if (zfs_arc_grow_retry)
7288 arc_grow_retry = zfs_arc_grow_retry;
7289
7290 /* Valid range: 1 - N */
7291 if (zfs_arc_shrink_shift) {
7292 arc_shrink_shift = zfs_arc_shrink_shift;
7293 arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1);
7294 }
7295
7296 /* Valid range: 1 - N ms */
7297 if (zfs_arc_min_prefetch_ms)
7298 arc_min_prefetch_ms = zfs_arc_min_prefetch_ms;
7299
7300 /* Valid range: 1 - N ms */
7301 if (zfs_arc_min_prescient_prefetch_ms) {
7302 arc_min_prescient_prefetch_ms =
7303 zfs_arc_min_prescient_prefetch_ms;
7304 }
7305
7306 /* Valid range: 0 - 100 */
7307 if (zfs_arc_lotsfree_percent <= 100)
7308 arc_lotsfree_percent = zfs_arc_lotsfree_percent;
7309 WARN_IF_TUNING_IGNORED(zfs_arc_lotsfree_percent, arc_lotsfree_percent,
7310 verbose);
7311
7312 /* Valid range: 0 - <all physical memory> */
7313 if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free))
7314 arc_sys_free = MIN(zfs_arc_sys_free, allmem);
7315 WARN_IF_TUNING_IGNORED(zfs_arc_sys_free, arc_sys_free, verbose);
7316 }
7317
7318 static void
arc_state_multilist_init(multilist_t * ml,multilist_sublist_index_func_t * index_func,int * maxcountp)7319 arc_state_multilist_init(multilist_t *ml,
7320 multilist_sublist_index_func_t *index_func, int *maxcountp)
7321 {
7322 multilist_create(ml, sizeof (arc_buf_hdr_t),
7323 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), index_func);
7324 *maxcountp = MAX(*maxcountp, multilist_get_num_sublists(ml));
7325 }
7326
7327 static void
arc_state_init(void)7328 arc_state_init(void)
7329 {
7330 int num_sublists = 0;
7331
7332 arc_state_multilist_init(&arc_mru->arcs_list[ARC_BUFC_METADATA],
7333 arc_state_multilist_index_func, &num_sublists);
7334 arc_state_multilist_init(&arc_mru->arcs_list[ARC_BUFC_DATA],
7335 arc_state_multilist_index_func, &num_sublists);
7336 arc_state_multilist_init(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
7337 arc_state_multilist_index_func, &num_sublists);
7338 arc_state_multilist_init(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
7339 arc_state_multilist_index_func, &num_sublists);
7340 arc_state_multilist_init(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
7341 arc_state_multilist_index_func, &num_sublists);
7342 arc_state_multilist_init(&arc_mfu->arcs_list[ARC_BUFC_DATA],
7343 arc_state_multilist_index_func, &num_sublists);
7344 arc_state_multilist_init(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
7345 arc_state_multilist_index_func, &num_sublists);
7346 arc_state_multilist_init(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
7347 arc_state_multilist_index_func, &num_sublists);
7348 arc_state_multilist_init(&arc_uncached->arcs_list[ARC_BUFC_METADATA],
7349 arc_state_multilist_index_func, &num_sublists);
7350 arc_state_multilist_init(&arc_uncached->arcs_list[ARC_BUFC_DATA],
7351 arc_state_multilist_index_func, &num_sublists);
7352
7353 /*
7354 * L2 headers should never be on the L2 state list since they don't
7355 * have L1 headers allocated. Special index function asserts that.
7356 */
7357 arc_state_multilist_init(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
7358 arc_state_l2c_multilist_index_func, &num_sublists);
7359 arc_state_multilist_init(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
7360 arc_state_l2c_multilist_index_func, &num_sublists);
7361
7362 /*
7363 * Keep track of the number of markers needed to reclaim buffers from
7364 * any ARC state. The markers will be pre-allocated so as to minimize
7365 * the number of memory allocations performed by the eviction thread.
7366 */
7367 arc_state_evict_marker_count = num_sublists;
7368
7369 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7370 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7371 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7372 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7373 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7374 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7375 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7376 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7377 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7378 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7379 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7380 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7381 zfs_refcount_create(&arc_uncached->arcs_esize[ARC_BUFC_METADATA]);
7382 zfs_refcount_create(&arc_uncached->arcs_esize[ARC_BUFC_DATA]);
7383
7384 zfs_refcount_create(&arc_anon->arcs_size[ARC_BUFC_DATA]);
7385 zfs_refcount_create(&arc_anon->arcs_size[ARC_BUFC_METADATA]);
7386 zfs_refcount_create(&arc_mru->arcs_size[ARC_BUFC_DATA]);
7387 zfs_refcount_create(&arc_mru->arcs_size[ARC_BUFC_METADATA]);
7388 zfs_refcount_create(&arc_mru_ghost->arcs_size[ARC_BUFC_DATA]);
7389 zfs_refcount_create(&arc_mru_ghost->arcs_size[ARC_BUFC_METADATA]);
7390 zfs_refcount_create(&arc_mfu->arcs_size[ARC_BUFC_DATA]);
7391 zfs_refcount_create(&arc_mfu->arcs_size[ARC_BUFC_METADATA]);
7392 zfs_refcount_create(&arc_mfu_ghost->arcs_size[ARC_BUFC_DATA]);
7393 zfs_refcount_create(&arc_mfu_ghost->arcs_size[ARC_BUFC_METADATA]);
7394 zfs_refcount_create(&arc_l2c_only->arcs_size[ARC_BUFC_DATA]);
7395 zfs_refcount_create(&arc_l2c_only->arcs_size[ARC_BUFC_METADATA]);
7396 zfs_refcount_create(&arc_uncached->arcs_size[ARC_BUFC_DATA]);
7397 zfs_refcount_create(&arc_uncached->arcs_size[ARC_BUFC_METADATA]);
7398
7399 wmsum_init(&arc_mru_ghost->arcs_hits[ARC_BUFC_DATA], 0);
7400 wmsum_init(&arc_mru_ghost->arcs_hits[ARC_BUFC_METADATA], 0);
7401 wmsum_init(&arc_mfu_ghost->arcs_hits[ARC_BUFC_DATA], 0);
7402 wmsum_init(&arc_mfu_ghost->arcs_hits[ARC_BUFC_METADATA], 0);
7403
7404 wmsum_init(&arc_sums.arcstat_hits, 0);
7405 wmsum_init(&arc_sums.arcstat_iohits, 0);
7406 wmsum_init(&arc_sums.arcstat_misses, 0);
7407 wmsum_init(&arc_sums.arcstat_demand_data_hits, 0);
7408 wmsum_init(&arc_sums.arcstat_demand_data_iohits, 0);
7409 wmsum_init(&arc_sums.arcstat_demand_data_misses, 0);
7410 wmsum_init(&arc_sums.arcstat_demand_metadata_hits, 0);
7411 wmsum_init(&arc_sums.arcstat_demand_metadata_iohits, 0);
7412 wmsum_init(&arc_sums.arcstat_demand_metadata_misses, 0);
7413 wmsum_init(&arc_sums.arcstat_prefetch_data_hits, 0);
7414 wmsum_init(&arc_sums.arcstat_prefetch_data_iohits, 0);
7415 wmsum_init(&arc_sums.arcstat_prefetch_data_misses, 0);
7416 wmsum_init(&arc_sums.arcstat_prefetch_metadata_hits, 0);
7417 wmsum_init(&arc_sums.arcstat_prefetch_metadata_iohits, 0);
7418 wmsum_init(&arc_sums.arcstat_prefetch_metadata_misses, 0);
7419 wmsum_init(&arc_sums.arcstat_mru_hits, 0);
7420 wmsum_init(&arc_sums.arcstat_mru_ghost_hits, 0);
7421 wmsum_init(&arc_sums.arcstat_mfu_hits, 0);
7422 wmsum_init(&arc_sums.arcstat_mfu_ghost_hits, 0);
7423 wmsum_init(&arc_sums.arcstat_uncached_hits, 0);
7424 wmsum_init(&arc_sums.arcstat_deleted, 0);
7425 wmsum_init(&arc_sums.arcstat_mutex_miss, 0);
7426 wmsum_init(&arc_sums.arcstat_access_skip, 0);
7427 wmsum_init(&arc_sums.arcstat_evict_skip, 0);
7428 wmsum_init(&arc_sums.arcstat_evict_not_enough, 0);
7429 wmsum_init(&arc_sums.arcstat_evict_l2_cached, 0);
7430 wmsum_init(&arc_sums.arcstat_evict_l2_eligible, 0);
7431 wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mfu, 0);
7432 wmsum_init(&arc_sums.arcstat_evict_l2_eligible_mru, 0);
7433 wmsum_init(&arc_sums.arcstat_evict_l2_ineligible, 0);
7434 wmsum_init(&arc_sums.arcstat_evict_l2_skip, 0);
7435 wmsum_init(&arc_sums.arcstat_hash_collisions, 0);
7436 wmsum_init(&arc_sums.arcstat_hash_chains, 0);
7437 aggsum_init(&arc_sums.arcstat_size, 0);
7438 wmsum_init(&arc_sums.arcstat_compressed_size, 0);
7439 wmsum_init(&arc_sums.arcstat_uncompressed_size, 0);
7440 wmsum_init(&arc_sums.arcstat_overhead_size, 0);
7441 wmsum_init(&arc_sums.arcstat_hdr_size, 0);
7442 wmsum_init(&arc_sums.arcstat_data_size, 0);
7443 wmsum_init(&arc_sums.arcstat_metadata_size, 0);
7444 wmsum_init(&arc_sums.arcstat_dbuf_size, 0);
7445 wmsum_init(&arc_sums.arcstat_dnode_size, 0);
7446 wmsum_init(&arc_sums.arcstat_bonus_size, 0);
7447 wmsum_init(&arc_sums.arcstat_l2_hits, 0);
7448 wmsum_init(&arc_sums.arcstat_l2_misses, 0);
7449 wmsum_init(&arc_sums.arcstat_l2_prefetch_asize, 0);
7450 wmsum_init(&arc_sums.arcstat_l2_mru_asize, 0);
7451 wmsum_init(&arc_sums.arcstat_l2_mfu_asize, 0);
7452 wmsum_init(&arc_sums.arcstat_l2_bufc_data_asize, 0);
7453 wmsum_init(&arc_sums.arcstat_l2_bufc_metadata_asize, 0);
7454 wmsum_init(&arc_sums.arcstat_l2_feeds, 0);
7455 wmsum_init(&arc_sums.arcstat_l2_rw_clash, 0);
7456 wmsum_init(&arc_sums.arcstat_l2_read_bytes, 0);
7457 wmsum_init(&arc_sums.arcstat_l2_write_bytes, 0);
7458 wmsum_init(&arc_sums.arcstat_l2_writes_sent, 0);
7459 wmsum_init(&arc_sums.arcstat_l2_writes_done, 0);
7460 wmsum_init(&arc_sums.arcstat_l2_writes_error, 0);
7461 wmsum_init(&arc_sums.arcstat_l2_writes_lock_retry, 0);
7462 wmsum_init(&arc_sums.arcstat_l2_evict_lock_retry, 0);
7463 wmsum_init(&arc_sums.arcstat_l2_evict_reading, 0);
7464 wmsum_init(&arc_sums.arcstat_l2_evict_l1cached, 0);
7465 wmsum_init(&arc_sums.arcstat_l2_free_on_write, 0);
7466 wmsum_init(&arc_sums.arcstat_l2_abort_lowmem, 0);
7467 wmsum_init(&arc_sums.arcstat_l2_cksum_bad, 0);
7468 wmsum_init(&arc_sums.arcstat_l2_io_error, 0);
7469 wmsum_init(&arc_sums.arcstat_l2_lsize, 0);
7470 wmsum_init(&arc_sums.arcstat_l2_psize, 0);
7471 aggsum_init(&arc_sums.arcstat_l2_hdr_size, 0);
7472 wmsum_init(&arc_sums.arcstat_l2_log_blk_writes, 0);
7473 wmsum_init(&arc_sums.arcstat_l2_log_blk_asize, 0);
7474 wmsum_init(&arc_sums.arcstat_l2_log_blk_count, 0);
7475 wmsum_init(&arc_sums.arcstat_l2_rebuild_success, 0);
7476 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_unsupported, 0);
7477 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_io_errors, 0);
7478 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_dh_errors, 0);
7479 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors, 0);
7480 wmsum_init(&arc_sums.arcstat_l2_rebuild_abort_lowmem, 0);
7481 wmsum_init(&arc_sums.arcstat_l2_rebuild_size, 0);
7482 wmsum_init(&arc_sums.arcstat_l2_rebuild_asize, 0);
7483 wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs, 0);
7484 wmsum_init(&arc_sums.arcstat_l2_rebuild_bufs_precached, 0);
7485 wmsum_init(&arc_sums.arcstat_l2_rebuild_log_blks, 0);
7486 wmsum_init(&arc_sums.arcstat_memory_throttle_count, 0);
7487 wmsum_init(&arc_sums.arcstat_memory_direct_count, 0);
7488 wmsum_init(&arc_sums.arcstat_memory_indirect_count, 0);
7489 wmsum_init(&arc_sums.arcstat_prune, 0);
7490 wmsum_init(&arc_sums.arcstat_meta_used, 0);
7491 wmsum_init(&arc_sums.arcstat_async_upgrade_sync, 0);
7492 wmsum_init(&arc_sums.arcstat_predictive_prefetch, 0);
7493 wmsum_init(&arc_sums.arcstat_demand_hit_predictive_prefetch, 0);
7494 wmsum_init(&arc_sums.arcstat_demand_iohit_predictive_prefetch, 0);
7495 wmsum_init(&arc_sums.arcstat_prescient_prefetch, 0);
7496 wmsum_init(&arc_sums.arcstat_demand_hit_prescient_prefetch, 0);
7497 wmsum_init(&arc_sums.arcstat_demand_iohit_prescient_prefetch, 0);
7498 wmsum_init(&arc_sums.arcstat_raw_size, 0);
7499 wmsum_init(&arc_sums.arcstat_cached_only_in_progress, 0);
7500 wmsum_init(&arc_sums.arcstat_abd_chunk_waste_size, 0);
7501
7502 arc_anon->arcs_state = ARC_STATE_ANON;
7503 arc_mru->arcs_state = ARC_STATE_MRU;
7504 arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
7505 arc_mfu->arcs_state = ARC_STATE_MFU;
7506 arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
7507 arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
7508 arc_uncached->arcs_state = ARC_STATE_UNCACHED;
7509 }
7510
7511 static void
arc_state_fini(void)7512 arc_state_fini(void)
7513 {
7514 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7515 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7516 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7517 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7518 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7519 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7520 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7521 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7522 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7523 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7524 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7525 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7526 zfs_refcount_destroy(&arc_uncached->arcs_esize[ARC_BUFC_METADATA]);
7527 zfs_refcount_destroy(&arc_uncached->arcs_esize[ARC_BUFC_DATA]);
7528
7529 zfs_refcount_destroy(&arc_anon->arcs_size[ARC_BUFC_DATA]);
7530 zfs_refcount_destroy(&arc_anon->arcs_size[ARC_BUFC_METADATA]);
7531 zfs_refcount_destroy(&arc_mru->arcs_size[ARC_BUFC_DATA]);
7532 zfs_refcount_destroy(&arc_mru->arcs_size[ARC_BUFC_METADATA]);
7533 zfs_refcount_destroy(&arc_mru_ghost->arcs_size[ARC_BUFC_DATA]);
7534 zfs_refcount_destroy(&arc_mru_ghost->arcs_size[ARC_BUFC_METADATA]);
7535 zfs_refcount_destroy(&arc_mfu->arcs_size[ARC_BUFC_DATA]);
7536 zfs_refcount_destroy(&arc_mfu->arcs_size[ARC_BUFC_METADATA]);
7537 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size[ARC_BUFC_DATA]);
7538 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size[ARC_BUFC_METADATA]);
7539 zfs_refcount_destroy(&arc_l2c_only->arcs_size[ARC_BUFC_DATA]);
7540 zfs_refcount_destroy(&arc_l2c_only->arcs_size[ARC_BUFC_METADATA]);
7541 zfs_refcount_destroy(&arc_uncached->arcs_size[ARC_BUFC_DATA]);
7542 zfs_refcount_destroy(&arc_uncached->arcs_size[ARC_BUFC_METADATA]);
7543
7544 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
7545 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7546 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7547 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7548 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
7549 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7550 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
7551 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7552 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
7553 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
7554 multilist_destroy(&arc_uncached->arcs_list[ARC_BUFC_METADATA]);
7555 multilist_destroy(&arc_uncached->arcs_list[ARC_BUFC_DATA]);
7556
7557 wmsum_fini(&arc_mru_ghost->arcs_hits[ARC_BUFC_DATA]);
7558 wmsum_fini(&arc_mru_ghost->arcs_hits[ARC_BUFC_METADATA]);
7559 wmsum_fini(&arc_mfu_ghost->arcs_hits[ARC_BUFC_DATA]);
7560 wmsum_fini(&arc_mfu_ghost->arcs_hits[ARC_BUFC_METADATA]);
7561
7562 wmsum_fini(&arc_sums.arcstat_hits);
7563 wmsum_fini(&arc_sums.arcstat_iohits);
7564 wmsum_fini(&arc_sums.arcstat_misses);
7565 wmsum_fini(&arc_sums.arcstat_demand_data_hits);
7566 wmsum_fini(&arc_sums.arcstat_demand_data_iohits);
7567 wmsum_fini(&arc_sums.arcstat_demand_data_misses);
7568 wmsum_fini(&arc_sums.arcstat_demand_metadata_hits);
7569 wmsum_fini(&arc_sums.arcstat_demand_metadata_iohits);
7570 wmsum_fini(&arc_sums.arcstat_demand_metadata_misses);
7571 wmsum_fini(&arc_sums.arcstat_prefetch_data_hits);
7572 wmsum_fini(&arc_sums.arcstat_prefetch_data_iohits);
7573 wmsum_fini(&arc_sums.arcstat_prefetch_data_misses);
7574 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_hits);
7575 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_iohits);
7576 wmsum_fini(&arc_sums.arcstat_prefetch_metadata_misses);
7577 wmsum_fini(&arc_sums.arcstat_mru_hits);
7578 wmsum_fini(&arc_sums.arcstat_mru_ghost_hits);
7579 wmsum_fini(&arc_sums.arcstat_mfu_hits);
7580 wmsum_fini(&arc_sums.arcstat_mfu_ghost_hits);
7581 wmsum_fini(&arc_sums.arcstat_uncached_hits);
7582 wmsum_fini(&arc_sums.arcstat_deleted);
7583 wmsum_fini(&arc_sums.arcstat_mutex_miss);
7584 wmsum_fini(&arc_sums.arcstat_access_skip);
7585 wmsum_fini(&arc_sums.arcstat_evict_skip);
7586 wmsum_fini(&arc_sums.arcstat_evict_not_enough);
7587 wmsum_fini(&arc_sums.arcstat_evict_l2_cached);
7588 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible);
7589 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mfu);
7590 wmsum_fini(&arc_sums.arcstat_evict_l2_eligible_mru);
7591 wmsum_fini(&arc_sums.arcstat_evict_l2_ineligible);
7592 wmsum_fini(&arc_sums.arcstat_evict_l2_skip);
7593 wmsum_fini(&arc_sums.arcstat_hash_collisions);
7594 wmsum_fini(&arc_sums.arcstat_hash_chains);
7595 aggsum_fini(&arc_sums.arcstat_size);
7596 wmsum_fini(&arc_sums.arcstat_compressed_size);
7597 wmsum_fini(&arc_sums.arcstat_uncompressed_size);
7598 wmsum_fini(&arc_sums.arcstat_overhead_size);
7599 wmsum_fini(&arc_sums.arcstat_hdr_size);
7600 wmsum_fini(&arc_sums.arcstat_data_size);
7601 wmsum_fini(&arc_sums.arcstat_metadata_size);
7602 wmsum_fini(&arc_sums.arcstat_dbuf_size);
7603 wmsum_fini(&arc_sums.arcstat_dnode_size);
7604 wmsum_fini(&arc_sums.arcstat_bonus_size);
7605 wmsum_fini(&arc_sums.arcstat_l2_hits);
7606 wmsum_fini(&arc_sums.arcstat_l2_misses);
7607 wmsum_fini(&arc_sums.arcstat_l2_prefetch_asize);
7608 wmsum_fini(&arc_sums.arcstat_l2_mru_asize);
7609 wmsum_fini(&arc_sums.arcstat_l2_mfu_asize);
7610 wmsum_fini(&arc_sums.arcstat_l2_bufc_data_asize);
7611 wmsum_fini(&arc_sums.arcstat_l2_bufc_metadata_asize);
7612 wmsum_fini(&arc_sums.arcstat_l2_feeds);
7613 wmsum_fini(&arc_sums.arcstat_l2_rw_clash);
7614 wmsum_fini(&arc_sums.arcstat_l2_read_bytes);
7615 wmsum_fini(&arc_sums.arcstat_l2_write_bytes);
7616 wmsum_fini(&arc_sums.arcstat_l2_writes_sent);
7617 wmsum_fini(&arc_sums.arcstat_l2_writes_done);
7618 wmsum_fini(&arc_sums.arcstat_l2_writes_error);
7619 wmsum_fini(&arc_sums.arcstat_l2_writes_lock_retry);
7620 wmsum_fini(&arc_sums.arcstat_l2_evict_lock_retry);
7621 wmsum_fini(&arc_sums.arcstat_l2_evict_reading);
7622 wmsum_fini(&arc_sums.arcstat_l2_evict_l1cached);
7623 wmsum_fini(&arc_sums.arcstat_l2_free_on_write);
7624 wmsum_fini(&arc_sums.arcstat_l2_abort_lowmem);
7625 wmsum_fini(&arc_sums.arcstat_l2_cksum_bad);
7626 wmsum_fini(&arc_sums.arcstat_l2_io_error);
7627 wmsum_fini(&arc_sums.arcstat_l2_lsize);
7628 wmsum_fini(&arc_sums.arcstat_l2_psize);
7629 aggsum_fini(&arc_sums.arcstat_l2_hdr_size);
7630 wmsum_fini(&arc_sums.arcstat_l2_log_blk_writes);
7631 wmsum_fini(&arc_sums.arcstat_l2_log_blk_asize);
7632 wmsum_fini(&arc_sums.arcstat_l2_log_blk_count);
7633 wmsum_fini(&arc_sums.arcstat_l2_rebuild_success);
7634 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_unsupported);
7635 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_io_errors);
7636 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_dh_errors);
7637 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_cksum_lb_errors);
7638 wmsum_fini(&arc_sums.arcstat_l2_rebuild_abort_lowmem);
7639 wmsum_fini(&arc_sums.arcstat_l2_rebuild_size);
7640 wmsum_fini(&arc_sums.arcstat_l2_rebuild_asize);
7641 wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs);
7642 wmsum_fini(&arc_sums.arcstat_l2_rebuild_bufs_precached);
7643 wmsum_fini(&arc_sums.arcstat_l2_rebuild_log_blks);
7644 wmsum_fini(&arc_sums.arcstat_memory_throttle_count);
7645 wmsum_fini(&arc_sums.arcstat_memory_direct_count);
7646 wmsum_fini(&arc_sums.arcstat_memory_indirect_count);
7647 wmsum_fini(&arc_sums.arcstat_prune);
7648 wmsum_fini(&arc_sums.arcstat_meta_used);
7649 wmsum_fini(&arc_sums.arcstat_async_upgrade_sync);
7650 wmsum_fini(&arc_sums.arcstat_predictive_prefetch);
7651 wmsum_fini(&arc_sums.arcstat_demand_hit_predictive_prefetch);
7652 wmsum_fini(&arc_sums.arcstat_demand_iohit_predictive_prefetch);
7653 wmsum_fini(&arc_sums.arcstat_prescient_prefetch);
7654 wmsum_fini(&arc_sums.arcstat_demand_hit_prescient_prefetch);
7655 wmsum_fini(&arc_sums.arcstat_demand_iohit_prescient_prefetch);
7656 wmsum_fini(&arc_sums.arcstat_raw_size);
7657 wmsum_fini(&arc_sums.arcstat_cached_only_in_progress);
7658 wmsum_fini(&arc_sums.arcstat_abd_chunk_waste_size);
7659 }
7660
7661 uint64_t
arc_target_bytes(void)7662 arc_target_bytes(void)
7663 {
7664 return (arc_c);
7665 }
7666
7667 void
arc_set_limits(uint64_t allmem)7668 arc_set_limits(uint64_t allmem)
7669 {
7670 /* Set min cache to 1/32 of all memory, or 32MB, whichever is more. */
7671 arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT);
7672
7673 /* How to set default max varies by platform. */
7674 arc_c_max = arc_default_max(arc_c_min, allmem);
7675 }
7676 void
arc_init(void)7677 arc_init(void)
7678 {
7679 uint64_t percent, allmem = arc_all_memory();
7680 mutex_init(&arc_evict_lock, NULL, MUTEX_DEFAULT, NULL);
7681 list_create(&arc_evict_waiters, sizeof (arc_evict_waiter_t),
7682 offsetof(arc_evict_waiter_t, aew_node));
7683
7684 arc_min_prefetch_ms = 1000;
7685 arc_min_prescient_prefetch_ms = 6000;
7686
7687 #if defined(_KERNEL)
7688 arc_lowmem_init();
7689 #endif
7690
7691 arc_set_limits(allmem);
7692
7693 #ifdef _KERNEL
7694 /*
7695 * If zfs_arc_max is non-zero at init, meaning it was set in the kernel
7696 * environment before the module was loaded, don't block setting the
7697 * maximum because it is less than arc_c_min, instead, reset arc_c_min
7698 * to a lower value.
7699 * zfs_arc_min will be handled by arc_tuning_update().
7700 */
7701 if (zfs_arc_max != 0 && zfs_arc_max >= MIN_ARC_MAX &&
7702 zfs_arc_max < allmem) {
7703 arc_c_max = zfs_arc_max;
7704 if (arc_c_min >= arc_c_max) {
7705 arc_c_min = MAX(zfs_arc_max / 2,
7706 2ULL << SPA_MAXBLOCKSHIFT);
7707 }
7708 }
7709 #else
7710 /*
7711 * In userland, there's only the memory pressure that we artificially
7712 * create (see arc_available_memory()). Don't let arc_c get too
7713 * small, because it can cause transactions to be larger than
7714 * arc_c, causing arc_tempreserve_space() to fail.
7715 */
7716 arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT);
7717 #endif
7718
7719 arc_c = arc_c_min;
7720 /*
7721 * 32-bit fixed point fractions of metadata from total ARC size,
7722 * MRU data from all data and MRU metadata from all metadata.
7723 */
7724 arc_meta = (1ULL << 32) / 4; /* Metadata is 25% of arc_c. */
7725 arc_pd = (1ULL << 32) / 2; /* Data MRU is 50% of data. */
7726 arc_pm = (1ULL << 32) / 2; /* Metadata MRU is 50% of metadata. */
7727
7728 percent = MIN(zfs_arc_dnode_limit_percent, 100);
7729 arc_dnode_limit = arc_c_max * percent / 100;
7730
7731 /* Apply user specified tunings */
7732 arc_tuning_update(B_TRUE);
7733
7734 /* if kmem_flags are set, lets try to use less memory */
7735 if (kmem_debugging())
7736 arc_c = arc_c / 2;
7737 if (arc_c < arc_c_min)
7738 arc_c = arc_c_min;
7739
7740 arc_register_hotplug();
7741
7742 arc_state_init();
7743
7744 buf_init();
7745
7746 list_create(&arc_prune_list, sizeof (arc_prune_t),
7747 offsetof(arc_prune_t, p_node));
7748 mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL);
7749
7750 arc_prune_taskq = taskq_create("arc_prune", zfs_arc_prune_task_threads,
7751 defclsyspri, 100, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
7752
7753 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7754 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7755
7756 if (arc_ksp != NULL) {
7757 arc_ksp->ks_data = &arc_stats;
7758 arc_ksp->ks_update = arc_kstat_update;
7759 kstat_install(arc_ksp);
7760 }
7761
7762 arc_state_evict_markers =
7763 arc_state_alloc_markers(arc_state_evict_marker_count);
7764 arc_evict_zthr = zthr_create_timer("arc_evict",
7765 arc_evict_cb_check, arc_evict_cb, NULL, SEC2NSEC(1), defclsyspri);
7766 arc_reap_zthr = zthr_create_timer("arc_reap",
7767 arc_reap_cb_check, arc_reap_cb, NULL, SEC2NSEC(1), minclsyspri);
7768
7769 arc_warm = B_FALSE;
7770
7771 /*
7772 * Calculate maximum amount of dirty data per pool.
7773 *
7774 * If it has been set by a module parameter, take that.
7775 * Otherwise, use a percentage of physical memory defined by
7776 * zfs_dirty_data_max_percent (default 10%) with a cap at
7777 * zfs_dirty_data_max_max (default 4G or 25% of physical memory).
7778 */
7779 #ifdef __LP64__
7780 if (zfs_dirty_data_max_max == 0)
7781 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024,
7782 allmem * zfs_dirty_data_max_max_percent / 100);
7783 #else
7784 if (zfs_dirty_data_max_max == 0)
7785 zfs_dirty_data_max_max = MIN(1ULL * 1024 * 1024 * 1024,
7786 allmem * zfs_dirty_data_max_max_percent / 100);
7787 #endif
7788
7789 if (zfs_dirty_data_max == 0) {
7790 zfs_dirty_data_max = allmem *
7791 zfs_dirty_data_max_percent / 100;
7792 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7793 zfs_dirty_data_max_max);
7794 }
7795
7796 if (zfs_wrlog_data_max == 0) {
7797
7798 /*
7799 * dp_wrlog_total is reduced for each txg at the end of
7800 * spa_sync(). However, dp_dirty_total is reduced every time
7801 * a block is written out. Thus under normal operation,
7802 * dp_wrlog_total could grow 2 times as big as
7803 * zfs_dirty_data_max.
7804 */
7805 zfs_wrlog_data_max = zfs_dirty_data_max * 2;
7806 }
7807 }
7808
7809 void
arc_fini(void)7810 arc_fini(void)
7811 {
7812 arc_prune_t *p;
7813
7814 #ifdef _KERNEL
7815 arc_lowmem_fini();
7816 #endif /* _KERNEL */
7817
7818 /* Use B_TRUE to ensure *all* buffers are evicted */
7819 arc_flush(NULL, B_TRUE);
7820
7821 if (arc_ksp != NULL) {
7822 kstat_delete(arc_ksp);
7823 arc_ksp = NULL;
7824 }
7825
7826 taskq_wait(arc_prune_taskq);
7827 taskq_destroy(arc_prune_taskq);
7828
7829 mutex_enter(&arc_prune_mtx);
7830 while ((p = list_remove_head(&arc_prune_list)) != NULL) {
7831 (void) zfs_refcount_remove(&p->p_refcnt, &arc_prune_list);
7832 zfs_refcount_destroy(&p->p_refcnt);
7833 kmem_free(p, sizeof (*p));
7834 }
7835 mutex_exit(&arc_prune_mtx);
7836
7837 list_destroy(&arc_prune_list);
7838 mutex_destroy(&arc_prune_mtx);
7839
7840 (void) zthr_cancel(arc_evict_zthr);
7841 (void) zthr_cancel(arc_reap_zthr);
7842 arc_state_free_markers(arc_state_evict_markers,
7843 arc_state_evict_marker_count);
7844
7845 mutex_destroy(&arc_evict_lock);
7846 list_destroy(&arc_evict_waiters);
7847
7848 /*
7849 * Free any buffers that were tagged for destruction. This needs
7850 * to occur before arc_state_fini() runs and destroys the aggsum
7851 * values which are updated when freeing scatter ABDs.
7852 */
7853 l2arc_do_free_on_write();
7854
7855 /*
7856 * buf_fini() must proceed arc_state_fini() because buf_fin() may
7857 * trigger the release of kmem magazines, which can callback to
7858 * arc_space_return() which accesses aggsums freed in act_state_fini().
7859 */
7860 buf_fini();
7861 arc_state_fini();
7862
7863 arc_unregister_hotplug();
7864
7865 /*
7866 * We destroy the zthrs after all the ARC state has been
7867 * torn down to avoid the case of them receiving any
7868 * wakeup() signals after they are destroyed.
7869 */
7870 zthr_destroy(arc_evict_zthr);
7871 zthr_destroy(arc_reap_zthr);
7872
7873 ASSERT0(arc_loaned_bytes);
7874 }
7875
7876 /*
7877 * Level 2 ARC
7878 *
7879 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7880 * It uses dedicated storage devices to hold cached data, which are populated
7881 * using large infrequent writes. The main role of this cache is to boost
7882 * the performance of random read workloads. The intended L2ARC devices
7883 * include short-stroked disks, solid state disks, and other media with
7884 * substantially faster read latency than disk.
7885 *
7886 * +-----------------------+
7887 * | ARC |
7888 * +-----------------------+
7889 * | ^ ^
7890 * | | |
7891 * l2arc_feed_thread() arc_read()
7892 * | | |
7893 * | l2arc read |
7894 * V | |
7895 * +---------------+ |
7896 * | L2ARC | |
7897 * +---------------+ |
7898 * | ^ |
7899 * l2arc_write() | |
7900 * | | |
7901 * V | |
7902 * +-------+ +-------+
7903 * | vdev | | vdev |
7904 * | cache | | cache |
7905 * +-------+ +-------+
7906 * +=========+ .-----.
7907 * : L2ARC : |-_____-|
7908 * : devices : | Disks |
7909 * +=========+ `-_____-'
7910 *
7911 * Read requests are satisfied from the following sources, in order:
7912 *
7913 * 1) ARC
7914 * 2) vdev cache of L2ARC devices
7915 * 3) L2ARC devices
7916 * 4) vdev cache of disks
7917 * 5) disks
7918 *
7919 * Some L2ARC device types exhibit extremely slow write performance.
7920 * To accommodate for this there are some significant differences between
7921 * the L2ARC and traditional cache design:
7922 *
7923 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7924 * the ARC behave as usual, freeing buffers and placing headers on ghost
7925 * lists. The ARC does not send buffers to the L2ARC during eviction as
7926 * this would add inflated write latencies for all ARC memory pressure.
7927 *
7928 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7929 * It does this by periodically scanning buffers from the eviction-end of
7930 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7931 * not already there. It scans until a headroom of buffers is satisfied,
7932 * which itself is a buffer for ARC eviction. If a compressible buffer is
7933 * found during scanning and selected for writing to an L2ARC device, we
7934 * temporarily boost scanning headroom during the next scan cycle to make
7935 * sure we adapt to compression effects (which might significantly reduce
7936 * the data volume we write to L2ARC). The thread that does this is
7937 * l2arc_feed_thread(), illustrated below; example sizes are included to
7938 * provide a better sense of ratio than this diagram:
7939 *
7940 * head --> tail
7941 * +---------------------+----------+
7942 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7943 * +---------------------+----------+ | o L2ARC eligible
7944 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7945 * +---------------------+----------+ |
7946 * 15.9 Gbytes ^ 32 Mbytes |
7947 * headroom |
7948 * l2arc_feed_thread()
7949 * |
7950 * l2arc write hand <--[oooo]--'
7951 * | 8 Mbyte
7952 * | write max
7953 * V
7954 * +==============================+
7955 * L2ARC dev |####|#|###|###| |####| ... |
7956 * +==============================+
7957 * 32 Gbytes
7958 *
7959 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7960 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7961 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7962 * safe to say that this is an uncommon case, since buffers at the end of
7963 * the ARC lists have moved there due to inactivity.
7964 *
7965 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7966 * then the L2ARC simply misses copying some buffers. This serves as a
7967 * pressure valve to prevent heavy read workloads from both stalling the ARC
7968 * with waits and clogging the L2ARC with writes. This also helps prevent
7969 * the potential for the L2ARC to churn if it attempts to cache content too
7970 * quickly, such as during backups of the entire pool.
7971 *
7972 * 5. After system boot and before the ARC has filled main memory, there are
7973 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7974 * lists can remain mostly static. Instead of searching from tail of these
7975 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7976 * for eligible buffers, greatly increasing its chance of finding them.
7977 *
7978 * The L2ARC device write speed is also boosted during this time so that
7979 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7980 * there are no L2ARC reads, and no fear of degrading read performance
7981 * through increased writes.
7982 *
7983 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7984 * the vdev queue can aggregate them into larger and fewer writes. Each
7985 * device is written to in a rotor fashion, sweeping writes through
7986 * available space then repeating.
7987 *
7988 * 7. The L2ARC does not store dirty content. It never needs to flush
7989 * write buffers back to disk based storage.
7990 *
7991 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7992 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7993 *
7994 * The performance of the L2ARC can be tweaked by a number of tunables, which
7995 * may be necessary for different workloads:
7996 *
7997 * l2arc_write_max max write bytes per interval
7998 * l2arc_write_boost extra write bytes during device warmup
7999 * l2arc_noprefetch skip caching prefetched buffers
8000 * l2arc_headroom number of max device writes to precache
8001 * l2arc_headroom_boost when we find compressed buffers during ARC
8002 * scanning, we multiply headroom by this
8003 * percentage factor for the next scan cycle,
8004 * since more compressed buffers are likely to
8005 * be present
8006 * l2arc_feed_secs seconds between L2ARC writing
8007 *
8008 * Tunables may be removed or added as future performance improvements are
8009 * integrated, and also may become zpool properties.
8010 *
8011 * There are three key functions that control how the L2ARC warms up:
8012 *
8013 * l2arc_write_eligible() check if a buffer is eligible to cache
8014 * l2arc_write_size() calculate how much to write
8015 * l2arc_write_interval() calculate sleep delay between writes
8016 *
8017 * These three functions determine what to write, how much, and how quickly
8018 * to send writes.
8019 *
8020 * L2ARC persistence:
8021 *
8022 * When writing buffers to L2ARC, we periodically add some metadata to
8023 * make sure we can pick them up after reboot, thus dramatically reducing
8024 * the impact that any downtime has on the performance of storage systems
8025 * with large caches.
8026 *
8027 * The implementation works fairly simply by integrating the following two
8028 * modifications:
8029 *
8030 * *) When writing to the L2ARC, we occasionally write a "l2arc log block",
8031 * which is an additional piece of metadata which describes what's been
8032 * written. This allows us to rebuild the arc_buf_hdr_t structures of the
8033 * main ARC buffers. There are 2 linked-lists of log blocks headed by
8034 * dh_start_lbps[2]. We alternate which chain we append to, so they are
8035 * time-wise and offset-wise interleaved, but that is an optimization rather
8036 * than for correctness. The log block also includes a pointer to the
8037 * previous block in its chain.
8038 *
8039 * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device
8040 * for our header bookkeeping purposes. This contains a device header,
8041 * which contains our top-level reference structures. We update it each
8042 * time we write a new log block, so that we're able to locate it in the
8043 * L2ARC device. If this write results in an inconsistent device header
8044 * (e.g. due to power failure), we detect this by verifying the header's
8045 * checksum and simply fail to reconstruct the L2ARC after reboot.
8046 *
8047 * Implementation diagram:
8048 *
8049 * +=== L2ARC device (not to scale) ======================================+
8050 * | ___two newest log block pointers__.__________ |
8051 * | / \dh_start_lbps[1] |
8052 * | / \ \dh_start_lbps[0]|
8053 * |.___/__. V V |
8054 * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---|
8055 * || hdr| ^ /^ /^ / / |
8056 * |+------+ ...--\-------/ \-----/--\------/ / |
8057 * | \--------------/ \--------------/ |
8058 * +======================================================================+
8059 *
8060 * As can be seen on the diagram, rather than using a simple linked list,
8061 * we use a pair of linked lists with alternating elements. This is a
8062 * performance enhancement due to the fact that we only find out the
8063 * address of the next log block access once the current block has been
8064 * completely read in. Obviously, this hurts performance, because we'd be
8065 * keeping the device's I/O queue at only a 1 operation deep, thus
8066 * incurring a large amount of I/O round-trip latency. Having two lists
8067 * allows us to fetch two log blocks ahead of where we are currently
8068 * rebuilding L2ARC buffers.
8069 *
8070 * On-device data structures:
8071 *
8072 * L2ARC device header: l2arc_dev_hdr_phys_t
8073 * L2ARC log block: l2arc_log_blk_phys_t
8074 *
8075 * L2ARC reconstruction:
8076 *
8077 * When writing data, we simply write in the standard rotary fashion,
8078 * evicting buffers as we go and simply writing new data over them (writing
8079 * a new log block every now and then). This obviously means that once we
8080 * loop around the end of the device, we will start cutting into an already
8081 * committed log block (and its referenced data buffers), like so:
8082 *
8083 * current write head__ __old tail
8084 * \ /
8085 * V V
8086 * <--|bufs |lb |bufs |lb | |bufs |lb |bufs |lb |-->
8087 * ^ ^^^^^^^^^___________________________________
8088 * | \
8089 * <<nextwrite>> may overwrite this blk and/or its bufs --'
8090 *
8091 * When importing the pool, we detect this situation and use it to stop
8092 * our scanning process (see l2arc_rebuild).
8093 *
8094 * There is one significant caveat to consider when rebuilding ARC contents
8095 * from an L2ARC device: what about invalidated buffers? Given the above
8096 * construction, we cannot update blocks which we've already written to amend
8097 * them to remove buffers which were invalidated. Thus, during reconstruction,
8098 * we might be populating the cache with buffers for data that's not on the
8099 * main pool anymore, or may have been overwritten!
8100 *
8101 * As it turns out, this isn't a problem. Every arc_read request includes
8102 * both the DVA and, crucially, the birth TXG of the BP the caller is
8103 * looking for. So even if the cache were populated by completely rotten
8104 * blocks for data that had been long deleted and/or overwritten, we'll
8105 * never actually return bad data from the cache, since the DVA with the
8106 * birth TXG uniquely identify a block in space and time - once created,
8107 * a block is immutable on disk. The worst thing we have done is wasted
8108 * some time and memory at l2arc rebuild to reconstruct outdated ARC
8109 * entries that will get dropped from the l2arc as it is being updated
8110 * with new blocks.
8111 *
8112 * L2ARC buffers that have been evicted by l2arc_evict() ahead of the write
8113 * hand are not restored. This is done by saving the offset (in bytes)
8114 * l2arc_evict() has evicted to in the L2ARC device header and taking it
8115 * into account when restoring buffers.
8116 */
8117
8118 static boolean_t
l2arc_write_eligible(uint64_t spa_guid,arc_buf_hdr_t * hdr)8119 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
8120 {
8121 /*
8122 * A buffer is *not* eligible for the L2ARC if it:
8123 * 1. belongs to a different spa.
8124 * 2. is already cached on the L2ARC.
8125 * 3. has an I/O in progress (it may be an incomplete read).
8126 * 4. is flagged not eligible (zfs property).
8127 */
8128 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
8129 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
8130 return (B_FALSE);
8131
8132 return (B_TRUE);
8133 }
8134
8135 static uint64_t
l2arc_write_size(l2arc_dev_t * dev)8136 l2arc_write_size(l2arc_dev_t *dev)
8137 {
8138 uint64_t size;
8139
8140 /*
8141 * Make sure our globals have meaningful values in case the user
8142 * altered them.
8143 */
8144 size = l2arc_write_max;
8145 if (size == 0) {
8146 cmn_err(CE_NOTE, "l2arc_write_max must be greater than zero, "
8147 "resetting it to the default (%d)", L2ARC_WRITE_SIZE);
8148 size = l2arc_write_max = L2ARC_WRITE_SIZE;
8149 }
8150
8151 if (arc_warm == B_FALSE)
8152 size += l2arc_write_boost;
8153
8154 /* We need to add in the worst case scenario of log block overhead. */
8155 size += l2arc_log_blk_overhead(size, dev);
8156 if (dev->l2ad_vdev->vdev_has_trim && l2arc_trim_ahead > 0) {
8157 /*
8158 * Trim ahead of the write size 64MB or (l2arc_trim_ahead/100)
8159 * times the writesize, whichever is greater.
8160 */
8161 size += MAX(64 * 1024 * 1024,
8162 (size * l2arc_trim_ahead) / 100);
8163 }
8164
8165 /*
8166 * Make sure the write size does not exceed the size of the cache
8167 * device. This is important in l2arc_evict(), otherwise infinite
8168 * iteration can occur.
8169 */
8170 size = MIN(size, (dev->l2ad_end - dev->l2ad_start) / 4);
8171
8172 size = P2ROUNDUP(size, 1ULL << dev->l2ad_vdev->vdev_ashift);
8173
8174 return (size);
8175
8176 }
8177
8178 static clock_t
l2arc_write_interval(clock_t began,uint64_t wanted,uint64_t wrote)8179 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
8180 {
8181 clock_t interval, next, now;
8182
8183 /*
8184 * If the ARC lists are busy, increase our write rate; if the
8185 * lists are stale, idle back. This is achieved by checking
8186 * how much we previously wrote - if it was more than half of
8187 * what we wanted, schedule the next write much sooner.
8188 */
8189 if (l2arc_feed_again && wrote > (wanted / 2))
8190 interval = (hz * l2arc_feed_min_ms) / 1000;
8191 else
8192 interval = hz * l2arc_feed_secs;
8193
8194 now = ddi_get_lbolt();
8195 next = MAX(now, MIN(now + interval, began + interval));
8196
8197 return (next);
8198 }
8199
8200 /*
8201 * Cycle through L2ARC devices. This is how L2ARC load balances.
8202 * If a device is returned, this also returns holding the spa config lock.
8203 */
8204 static l2arc_dev_t *
l2arc_dev_get_next(void)8205 l2arc_dev_get_next(void)
8206 {
8207 l2arc_dev_t *first, *next = NULL;
8208
8209 /*
8210 * Lock out the removal of spas (spa_namespace_lock), then removal
8211 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
8212 * both locks will be dropped and a spa config lock held instead.
8213 */
8214 mutex_enter(&spa_namespace_lock);
8215 mutex_enter(&l2arc_dev_mtx);
8216
8217 /* if there are no vdevs, there is nothing to do */
8218 if (l2arc_ndev == 0)
8219 goto out;
8220
8221 first = NULL;
8222 next = l2arc_dev_last;
8223 do {
8224 /* loop around the list looking for a non-faulted vdev */
8225 if (next == NULL) {
8226 next = list_head(l2arc_dev_list);
8227 } else {
8228 next = list_next(l2arc_dev_list, next);
8229 if (next == NULL)
8230 next = list_head(l2arc_dev_list);
8231 }
8232
8233 /* if we have come back to the start, bail out */
8234 if (first == NULL)
8235 first = next;
8236 else if (next == first)
8237 break;
8238
8239 ASSERT3P(next, !=, NULL);
8240 } while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
8241 next->l2ad_trim_all || next->l2ad_spa->spa_is_exporting);
8242
8243 /* if we were unable to find any usable vdevs, return NULL */
8244 if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild ||
8245 next->l2ad_trim_all || next->l2ad_spa->spa_is_exporting)
8246 next = NULL;
8247
8248 l2arc_dev_last = next;
8249
8250 out:
8251 mutex_exit(&l2arc_dev_mtx);
8252
8253 /*
8254 * Grab the config lock to prevent the 'next' device from being
8255 * removed while we are writing to it.
8256 */
8257 if (next != NULL)
8258 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
8259 mutex_exit(&spa_namespace_lock);
8260
8261 return (next);
8262 }
8263
8264 /*
8265 * Free buffers that were tagged for destruction.
8266 */
8267 static void
l2arc_do_free_on_write(void)8268 l2arc_do_free_on_write(void)
8269 {
8270 l2arc_data_free_t *df;
8271
8272 mutex_enter(&l2arc_free_on_write_mtx);
8273 while ((df = list_remove_head(l2arc_free_on_write)) != NULL) {
8274 ASSERT3P(df->l2df_abd, !=, NULL);
8275 abd_free(df->l2df_abd);
8276 kmem_free(df, sizeof (l2arc_data_free_t));
8277 }
8278 mutex_exit(&l2arc_free_on_write_mtx);
8279 }
8280
8281 /*
8282 * A write to a cache device has completed. Update all headers to allow
8283 * reads from these buffers to begin.
8284 */
8285 static void
l2arc_write_done(zio_t * zio)8286 l2arc_write_done(zio_t *zio)
8287 {
8288 l2arc_write_callback_t *cb;
8289 l2arc_lb_abd_buf_t *abd_buf;
8290 l2arc_lb_ptr_buf_t *lb_ptr_buf;
8291 l2arc_dev_t *dev;
8292 l2arc_dev_hdr_phys_t *l2dhdr;
8293 list_t *buflist;
8294 arc_buf_hdr_t *head, *hdr, *hdr_prev;
8295 kmutex_t *hash_lock;
8296 int64_t bytes_dropped = 0;
8297
8298 cb = zio->io_private;
8299 ASSERT3P(cb, !=, NULL);
8300 dev = cb->l2wcb_dev;
8301 l2dhdr = dev->l2ad_dev_hdr;
8302 ASSERT3P(dev, !=, NULL);
8303 head = cb->l2wcb_head;
8304 ASSERT3P(head, !=, NULL);
8305 buflist = &dev->l2ad_buflist;
8306 ASSERT3P(buflist, !=, NULL);
8307 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
8308 l2arc_write_callback_t *, cb);
8309
8310 /*
8311 * All writes completed, or an error was hit.
8312 */
8313 top:
8314 mutex_enter(&dev->l2ad_mtx);
8315 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
8316 hdr_prev = list_prev(buflist, hdr);
8317
8318 hash_lock = HDR_LOCK(hdr);
8319
8320 /*
8321 * We cannot use mutex_enter or else we can deadlock
8322 * with l2arc_write_buffers (due to swapping the order
8323 * the hash lock and l2ad_mtx are taken).
8324 */
8325 if (!mutex_tryenter(hash_lock)) {
8326 /*
8327 * Missed the hash lock. We must retry so we
8328 * don't leave the ARC_FLAG_L2_WRITING bit set.
8329 */
8330 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
8331
8332 /*
8333 * We don't want to rescan the headers we've
8334 * already marked as having been written out, so
8335 * we reinsert the head node so we can pick up
8336 * where we left off.
8337 */
8338 list_remove(buflist, head);
8339 list_insert_after(buflist, hdr, head);
8340
8341 mutex_exit(&dev->l2ad_mtx);
8342
8343 /*
8344 * We wait for the hash lock to become available
8345 * to try and prevent busy waiting, and increase
8346 * the chance we'll be able to acquire the lock
8347 * the next time around.
8348 */
8349 mutex_enter(hash_lock);
8350 mutex_exit(hash_lock);
8351 goto top;
8352 }
8353
8354 /*
8355 * We could not have been moved into the arc_l2c_only
8356 * state while in-flight due to our ARC_FLAG_L2_WRITING
8357 * bit being set. Let's just ensure that's being enforced.
8358 */
8359 ASSERT(HDR_HAS_L1HDR(hdr));
8360
8361 /*
8362 * Skipped - drop L2ARC entry and mark the header as no
8363 * longer L2 eligibile.
8364 */
8365 if (zio->io_error != 0) {
8366 /*
8367 * Error - drop L2ARC entry.
8368 */
8369 list_remove(buflist, hdr);
8370 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
8371
8372 uint64_t psize = HDR_GET_PSIZE(hdr);
8373 l2arc_hdr_arcstats_decrement(hdr);
8374
8375 bytes_dropped +=
8376 vdev_psize_to_asize(dev->l2ad_vdev, psize);
8377 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
8378 arc_hdr_size(hdr), hdr);
8379 }
8380
8381 /*
8382 * Allow ARC to begin reads and ghost list evictions to
8383 * this L2ARC entry.
8384 */
8385 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
8386
8387 mutex_exit(hash_lock);
8388 }
8389
8390 /*
8391 * Free the allocated abd buffers for writing the log blocks.
8392 * If the zio failed reclaim the allocated space and remove the
8393 * pointers to these log blocks from the log block pointer list
8394 * of the L2ARC device.
8395 */
8396 while ((abd_buf = list_remove_tail(&cb->l2wcb_abd_list)) != NULL) {
8397 abd_free(abd_buf->abd);
8398 zio_buf_free(abd_buf, sizeof (*abd_buf));
8399 if (zio->io_error != 0) {
8400 lb_ptr_buf = list_remove_head(&dev->l2ad_lbptr_list);
8401 /*
8402 * L2BLK_GET_PSIZE returns aligned size for log
8403 * blocks.
8404 */
8405 uint64_t asize =
8406 L2BLK_GET_PSIZE((lb_ptr_buf->lb_ptr)->lbp_prop);
8407 bytes_dropped += asize;
8408 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8409 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8410 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8411 lb_ptr_buf);
8412 (void) zfs_refcount_remove(&dev->l2ad_lb_count,
8413 lb_ptr_buf);
8414 kmem_free(lb_ptr_buf->lb_ptr,
8415 sizeof (l2arc_log_blkptr_t));
8416 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8417 }
8418 }
8419 list_destroy(&cb->l2wcb_abd_list);
8420
8421 if (zio->io_error != 0) {
8422 ARCSTAT_BUMP(arcstat_l2_writes_error);
8423
8424 /*
8425 * Restore the lbps array in the header to its previous state.
8426 * If the list of log block pointers is empty, zero out the
8427 * log block pointers in the device header.
8428 */
8429 lb_ptr_buf = list_head(&dev->l2ad_lbptr_list);
8430 for (int i = 0; i < 2; i++) {
8431 if (lb_ptr_buf == NULL) {
8432 /*
8433 * If the list is empty zero out the device
8434 * header. Otherwise zero out the second log
8435 * block pointer in the header.
8436 */
8437 if (i == 0) {
8438 memset(l2dhdr, 0,
8439 dev->l2ad_dev_hdr_asize);
8440 } else {
8441 memset(&l2dhdr->dh_start_lbps[i], 0,
8442 sizeof (l2arc_log_blkptr_t));
8443 }
8444 break;
8445 }
8446 memcpy(&l2dhdr->dh_start_lbps[i], lb_ptr_buf->lb_ptr,
8447 sizeof (l2arc_log_blkptr_t));
8448 lb_ptr_buf = list_next(&dev->l2ad_lbptr_list,
8449 lb_ptr_buf);
8450 }
8451 }
8452
8453 ARCSTAT_BUMP(arcstat_l2_writes_done);
8454 list_remove(buflist, head);
8455 ASSERT(!HDR_HAS_L1HDR(head));
8456 kmem_cache_free(hdr_l2only_cache, head);
8457 mutex_exit(&dev->l2ad_mtx);
8458
8459 ASSERT(dev->l2ad_vdev != NULL);
8460 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
8461
8462 l2arc_do_free_on_write();
8463
8464 kmem_free(cb, sizeof (l2arc_write_callback_t));
8465 }
8466
8467 static int
l2arc_untransform(zio_t * zio,l2arc_read_callback_t * cb)8468 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
8469 {
8470 int ret;
8471 spa_t *spa = zio->io_spa;
8472 arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
8473 blkptr_t *bp = zio->io_bp;
8474 uint8_t salt[ZIO_DATA_SALT_LEN];
8475 uint8_t iv[ZIO_DATA_IV_LEN];
8476 uint8_t mac[ZIO_DATA_MAC_LEN];
8477 boolean_t no_crypt = B_FALSE;
8478
8479 /*
8480 * ZIL data is never be written to the L2ARC, so we don't need
8481 * special handling for its unique MAC storage.
8482 */
8483 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
8484 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
8485 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8486
8487 /*
8488 * If the data was encrypted, decrypt it now. Note that
8489 * we must check the bp here and not the hdr, since the
8490 * hdr does not have its encryption parameters updated
8491 * until arc_read_done().
8492 */
8493 if (BP_IS_ENCRYPTED(bp)) {
8494 abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8495 ARC_HDR_USE_RESERVE);
8496
8497 zio_crypt_decode_params_bp(bp, salt, iv);
8498 zio_crypt_decode_mac_bp(bp, mac);
8499
8500 ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb,
8501 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
8502 salt, iv, mac, HDR_GET_PSIZE(hdr), eabd,
8503 hdr->b_l1hdr.b_pabd, &no_crypt);
8504 if (ret != 0) {
8505 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8506 goto error;
8507 }
8508
8509 /*
8510 * If we actually performed decryption, replace b_pabd
8511 * with the decrypted data. Otherwise we can just throw
8512 * our decryption buffer away.
8513 */
8514 if (!no_crypt) {
8515 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8516 arc_hdr_size(hdr), hdr);
8517 hdr->b_l1hdr.b_pabd = eabd;
8518 zio->io_abd = eabd;
8519 } else {
8520 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8521 }
8522 }
8523
8524 /*
8525 * If the L2ARC block was compressed, but ARC compression
8526 * is disabled we decompress the data into a new buffer and
8527 * replace the existing data.
8528 */
8529 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8530 !HDR_COMPRESSION_ENABLED(hdr)) {
8531 abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8532 ARC_HDR_USE_RESERVE);
8533
8534 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
8535 hdr->b_l1hdr.b_pabd, cabd, HDR_GET_PSIZE(hdr),
8536 HDR_GET_LSIZE(hdr), &hdr->b_complevel);
8537 if (ret != 0) {
8538 arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
8539 goto error;
8540 }
8541
8542 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8543 arc_hdr_size(hdr), hdr);
8544 hdr->b_l1hdr.b_pabd = cabd;
8545 zio->io_abd = cabd;
8546 zio->io_size = HDR_GET_LSIZE(hdr);
8547 }
8548
8549 return (0);
8550
8551 error:
8552 return (ret);
8553 }
8554
8555
8556 /*
8557 * A read to a cache device completed. Validate buffer contents before
8558 * handing over to the regular ARC routines.
8559 */
8560 static void
l2arc_read_done(zio_t * zio)8561 l2arc_read_done(zio_t *zio)
8562 {
8563 int tfm_error = 0;
8564 l2arc_read_callback_t *cb = zio->io_private;
8565 arc_buf_hdr_t *hdr;
8566 kmutex_t *hash_lock;
8567 boolean_t valid_cksum;
8568 boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
8569 (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT));
8570
8571 ASSERT3P(zio->io_vd, !=, NULL);
8572 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
8573
8574 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
8575
8576 ASSERT3P(cb, !=, NULL);
8577 hdr = cb->l2rcb_hdr;
8578 ASSERT3P(hdr, !=, NULL);
8579
8580 hash_lock = HDR_LOCK(hdr);
8581 mutex_enter(hash_lock);
8582 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
8583
8584 /*
8585 * If the data was read into a temporary buffer,
8586 * move it and free the buffer.
8587 */
8588 if (cb->l2rcb_abd != NULL) {
8589 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
8590 if (zio->io_error == 0) {
8591 if (using_rdata) {
8592 abd_copy(hdr->b_crypt_hdr.b_rabd,
8593 cb->l2rcb_abd, arc_hdr_size(hdr));
8594 } else {
8595 abd_copy(hdr->b_l1hdr.b_pabd,
8596 cb->l2rcb_abd, arc_hdr_size(hdr));
8597 }
8598 }
8599
8600 /*
8601 * The following must be done regardless of whether
8602 * there was an error:
8603 * - free the temporary buffer
8604 * - point zio to the real ARC buffer
8605 * - set zio size accordingly
8606 * These are required because zio is either re-used for
8607 * an I/O of the block in the case of the error
8608 * or the zio is passed to arc_read_done() and it
8609 * needs real data.
8610 */
8611 abd_free(cb->l2rcb_abd);
8612 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
8613
8614 if (using_rdata) {
8615 ASSERT(HDR_HAS_RABD(hdr));
8616 zio->io_abd = zio->io_orig_abd =
8617 hdr->b_crypt_hdr.b_rabd;
8618 } else {
8619 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8620 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
8621 }
8622 }
8623
8624 ASSERT3P(zio->io_abd, !=, NULL);
8625
8626 /*
8627 * Check this survived the L2ARC journey.
8628 */
8629 ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
8630 (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
8631 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
8632 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
8633 zio->io_prop.zp_complevel = hdr->b_complevel;
8634
8635 valid_cksum = arc_cksum_is_equal(hdr, zio);
8636
8637 /*
8638 * b_rabd will always match the data as it exists on disk if it is
8639 * being used. Therefore if we are reading into b_rabd we do not
8640 * attempt to untransform the data.
8641 */
8642 if (valid_cksum && !using_rdata)
8643 tfm_error = l2arc_untransform(zio, cb);
8644
8645 if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
8646 !HDR_L2_EVICTED(hdr)) {
8647 mutex_exit(hash_lock);
8648 zio->io_private = hdr;
8649 arc_read_done(zio);
8650 } else {
8651 /*
8652 * Buffer didn't survive caching. Increment stats and
8653 * reissue to the original storage device.
8654 */
8655 if (zio->io_error != 0) {
8656 ARCSTAT_BUMP(arcstat_l2_io_error);
8657 } else {
8658 zio->io_error = SET_ERROR(EIO);
8659 }
8660 if (!valid_cksum || tfm_error != 0)
8661 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
8662
8663 /*
8664 * If there's no waiter, issue an async i/o to the primary
8665 * storage now. If there *is* a waiter, the caller must
8666 * issue the i/o in a context where it's OK to block.
8667 */
8668 if (zio->io_waiter == NULL) {
8669 zio_t *pio = zio_unique_parent(zio);
8670 void *abd = (using_rdata) ?
8671 hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
8672
8673 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
8674
8675 zio = zio_read(pio, zio->io_spa, zio->io_bp,
8676 abd, zio->io_size, arc_read_done,
8677 hdr, zio->io_priority, cb->l2rcb_flags,
8678 &cb->l2rcb_zb);
8679
8680 /*
8681 * Original ZIO will be freed, so we need to update
8682 * ARC header with the new ZIO pointer to be used
8683 * by zio_change_priority() in arc_read().
8684 */
8685 for (struct arc_callback *acb = hdr->b_l1hdr.b_acb;
8686 acb != NULL; acb = acb->acb_next)
8687 acb->acb_zio_head = zio;
8688
8689 mutex_exit(hash_lock);
8690 zio_nowait(zio);
8691 } else {
8692 mutex_exit(hash_lock);
8693 }
8694 }
8695
8696 kmem_free(cb, sizeof (l2arc_read_callback_t));
8697 }
8698
8699 /*
8700 * This is the list priority from which the L2ARC will search for pages to
8701 * cache. This is used within loops (0..3) to cycle through lists in the
8702 * desired order. This order can have a significant effect on cache
8703 * performance.
8704 *
8705 * Currently the metadata lists are hit first, MFU then MRU, followed by
8706 * the data lists. This function returns a locked list, and also returns
8707 * the lock pointer.
8708 */
8709 static multilist_sublist_t *
l2arc_sublist_lock(int list_num)8710 l2arc_sublist_lock(int list_num)
8711 {
8712 multilist_t *ml = NULL;
8713 unsigned int idx;
8714
8715 ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES);
8716
8717 switch (list_num) {
8718 case 0:
8719 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
8720 break;
8721 case 1:
8722 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
8723 break;
8724 case 2:
8725 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
8726 break;
8727 case 3:
8728 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
8729 break;
8730 default:
8731 return (NULL);
8732 }
8733
8734 /*
8735 * Return a randomly-selected sublist. This is acceptable
8736 * because the caller feeds only a little bit of data for each
8737 * call (8MB). Subsequent calls will result in different
8738 * sublists being selected.
8739 */
8740 idx = multilist_get_random_index(ml);
8741 return (multilist_sublist_lock_idx(ml, idx));
8742 }
8743
8744 /*
8745 * Calculates the maximum overhead of L2ARC metadata log blocks for a given
8746 * L2ARC write size. l2arc_evict and l2arc_write_size need to include this
8747 * overhead in processing to make sure there is enough headroom available
8748 * when writing buffers.
8749 */
8750 static inline uint64_t
l2arc_log_blk_overhead(uint64_t write_sz,l2arc_dev_t * dev)8751 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev)
8752 {
8753 if (dev->l2ad_log_entries == 0) {
8754 return (0);
8755 } else {
8756 uint64_t log_entries = write_sz >> SPA_MINBLOCKSHIFT;
8757
8758 uint64_t log_blocks = (log_entries +
8759 dev->l2ad_log_entries - 1) /
8760 dev->l2ad_log_entries;
8761
8762 return (vdev_psize_to_asize(dev->l2ad_vdev,
8763 sizeof (l2arc_log_blk_phys_t)) * log_blocks);
8764 }
8765 }
8766
8767 /*
8768 * Evict buffers from the device write hand to the distance specified in
8769 * bytes. This distance may span populated buffers, it may span nothing.
8770 * This is clearing a region on the L2ARC device ready for writing.
8771 * If the 'all' boolean is set, every buffer is evicted.
8772 */
8773 static void
l2arc_evict(l2arc_dev_t * dev,uint64_t distance,boolean_t all)8774 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
8775 {
8776 list_t *buflist;
8777 arc_buf_hdr_t *hdr, *hdr_prev;
8778 kmutex_t *hash_lock;
8779 uint64_t taddr;
8780 l2arc_lb_ptr_buf_t *lb_ptr_buf, *lb_ptr_buf_prev;
8781 vdev_t *vd = dev->l2ad_vdev;
8782 boolean_t rerun;
8783
8784 buflist = &dev->l2ad_buflist;
8785
8786 top:
8787 rerun = B_FALSE;
8788 if (dev->l2ad_hand + distance > dev->l2ad_end) {
8789 /*
8790 * When there is no space to accommodate upcoming writes,
8791 * evict to the end. Then bump the write and evict hands
8792 * to the start and iterate. This iteration does not
8793 * happen indefinitely as we make sure in
8794 * l2arc_write_size() that when the write hand is reset,
8795 * the write size does not exceed the end of the device.
8796 */
8797 rerun = B_TRUE;
8798 taddr = dev->l2ad_end;
8799 } else {
8800 taddr = dev->l2ad_hand + distance;
8801 }
8802 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
8803 uint64_t, taddr, boolean_t, all);
8804
8805 if (!all) {
8806 /*
8807 * This check has to be placed after deciding whether to
8808 * iterate (rerun).
8809 */
8810 if (dev->l2ad_first) {
8811 /*
8812 * This is the first sweep through the device. There is
8813 * nothing to evict. We have already trimmmed the
8814 * whole device.
8815 */
8816 goto out;
8817 } else {
8818 /*
8819 * Trim the space to be evicted.
8820 */
8821 if (vd->vdev_has_trim && dev->l2ad_evict < taddr &&
8822 l2arc_trim_ahead > 0) {
8823 /*
8824 * We have to drop the spa_config lock because
8825 * vdev_trim_range() will acquire it.
8826 * l2ad_evict already accounts for the label
8827 * size. To prevent vdev_trim_ranges() from
8828 * adding it again, we subtract it from
8829 * l2ad_evict.
8830 */
8831 spa_config_exit(dev->l2ad_spa, SCL_L2ARC, dev);
8832 vdev_trim_simple(vd,
8833 dev->l2ad_evict - VDEV_LABEL_START_SIZE,
8834 taddr - dev->l2ad_evict);
8835 spa_config_enter(dev->l2ad_spa, SCL_L2ARC, dev,
8836 RW_READER);
8837 }
8838
8839 /*
8840 * When rebuilding L2ARC we retrieve the evict hand
8841 * from the header of the device. Of note, l2arc_evict()
8842 * does not actually delete buffers from the cache
8843 * device, but trimming may do so depending on the
8844 * hardware implementation. Thus keeping track of the
8845 * evict hand is useful.
8846 */
8847 dev->l2ad_evict = MAX(dev->l2ad_evict, taddr);
8848 }
8849 }
8850
8851 retry:
8852 mutex_enter(&dev->l2ad_mtx);
8853 /*
8854 * We have to account for evicted log blocks. Run vdev_space_update()
8855 * on log blocks whose offset (in bytes) is before the evicted offset
8856 * (in bytes) by searching in the list of pointers to log blocks
8857 * present in the L2ARC device.
8858 */
8859 for (lb_ptr_buf = list_tail(&dev->l2ad_lbptr_list); lb_ptr_buf;
8860 lb_ptr_buf = lb_ptr_buf_prev) {
8861
8862 lb_ptr_buf_prev = list_prev(&dev->l2ad_lbptr_list, lb_ptr_buf);
8863
8864 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
8865 uint64_t asize = L2BLK_GET_PSIZE(
8866 (lb_ptr_buf->lb_ptr)->lbp_prop);
8867
8868 /*
8869 * We don't worry about log blocks left behind (ie
8870 * lbp_payload_start < l2ad_hand) because l2arc_write_buffers()
8871 * will never write more than l2arc_evict() evicts.
8872 */
8873 if (!all && l2arc_log_blkptr_valid(dev, lb_ptr_buf->lb_ptr)) {
8874 break;
8875 } else {
8876 vdev_space_update(vd, -asize, 0, 0);
8877 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8878 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8879 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8880 lb_ptr_buf);
8881 (void) zfs_refcount_remove(&dev->l2ad_lb_count,
8882 lb_ptr_buf);
8883 list_remove(&dev->l2ad_lbptr_list, lb_ptr_buf);
8884 kmem_free(lb_ptr_buf->lb_ptr,
8885 sizeof (l2arc_log_blkptr_t));
8886 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8887 }
8888 }
8889
8890 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
8891 hdr_prev = list_prev(buflist, hdr);
8892
8893 ASSERT(!HDR_EMPTY(hdr));
8894 hash_lock = HDR_LOCK(hdr);
8895
8896 /*
8897 * We cannot use mutex_enter or else we can deadlock
8898 * with l2arc_write_buffers (due to swapping the order
8899 * the hash lock and l2ad_mtx are taken).
8900 */
8901 if (!mutex_tryenter(hash_lock)) {
8902 /*
8903 * Missed the hash lock. Retry.
8904 */
8905 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
8906 mutex_exit(&dev->l2ad_mtx);
8907 mutex_enter(hash_lock);
8908 mutex_exit(hash_lock);
8909 goto retry;
8910 }
8911
8912 /*
8913 * A header can't be on this list if it doesn't have L2 header.
8914 */
8915 ASSERT(HDR_HAS_L2HDR(hdr));
8916
8917 /* Ensure this header has finished being written. */
8918 ASSERT(!HDR_L2_WRITING(hdr));
8919 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
8920
8921 if (!all && (hdr->b_l2hdr.b_daddr >= dev->l2ad_evict ||
8922 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
8923 /*
8924 * We've evicted to the target address,
8925 * or the end of the device.
8926 */
8927 mutex_exit(hash_lock);
8928 break;
8929 }
8930
8931 if (!HDR_HAS_L1HDR(hdr)) {
8932 ASSERT(!HDR_L2_READING(hdr));
8933 /*
8934 * This doesn't exist in the ARC. Destroy.
8935 * arc_hdr_destroy() will call list_remove()
8936 * and decrement arcstat_l2_lsize.
8937 */
8938 arc_change_state(arc_anon, hdr);
8939 arc_hdr_destroy(hdr);
8940 } else {
8941 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
8942 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
8943 /*
8944 * Invalidate issued or about to be issued
8945 * reads, since we may be about to write
8946 * over this location.
8947 */
8948 if (HDR_L2_READING(hdr)) {
8949 ARCSTAT_BUMP(arcstat_l2_evict_reading);
8950 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
8951 }
8952
8953 arc_hdr_l2hdr_destroy(hdr);
8954 }
8955 mutex_exit(hash_lock);
8956 }
8957 mutex_exit(&dev->l2ad_mtx);
8958
8959 out:
8960 /*
8961 * We need to check if we evict all buffers, otherwise we may iterate
8962 * unnecessarily.
8963 */
8964 if (!all && rerun) {
8965 /*
8966 * Bump device hand to the device start if it is approaching the
8967 * end. l2arc_evict() has already evicted ahead for this case.
8968 */
8969 dev->l2ad_hand = dev->l2ad_start;
8970 dev->l2ad_evict = dev->l2ad_start;
8971 dev->l2ad_first = B_FALSE;
8972 goto top;
8973 }
8974
8975 if (!all) {
8976 /*
8977 * In case of cache device removal (all) the following
8978 * assertions may be violated without functional consequences
8979 * as the device is about to be removed.
8980 */
8981 ASSERT3U(dev->l2ad_hand + distance, <=, dev->l2ad_end);
8982 if (!dev->l2ad_first)
8983 ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
8984 }
8985 }
8986
8987 /*
8988 * Handle any abd transforms that might be required for writing to the L2ARC.
8989 * If successful, this function will always return an abd with the data
8990 * transformed as it is on disk in a new abd of asize bytes.
8991 */
8992 static int
l2arc_apply_transforms(spa_t * spa,arc_buf_hdr_t * hdr,uint64_t asize,abd_t ** abd_out)8993 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
8994 abd_t **abd_out)
8995 {
8996 int ret;
8997 abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
8998 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
8999 uint64_t psize = HDR_GET_PSIZE(hdr);
9000 uint64_t size = arc_hdr_size(hdr);
9001 boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
9002 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
9003 dsl_crypto_key_t *dck = NULL;
9004 uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
9005 boolean_t no_crypt = B_FALSE;
9006
9007 ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
9008 !HDR_COMPRESSION_ENABLED(hdr)) ||
9009 HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
9010 ASSERT3U(psize, <=, asize);
9011
9012 /*
9013 * If this data simply needs its own buffer, we simply allocate it
9014 * and copy the data. This may be done to eliminate a dependency on a
9015 * shared buffer or to reallocate the buffer to match asize.
9016 */
9017 if (HDR_HAS_RABD(hdr)) {
9018 ASSERT3U(asize, >, psize);
9019 to_write = abd_alloc_for_io(asize, ismd);
9020 abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize);
9021 abd_zero_off(to_write, psize, asize - psize);
9022 goto out;
9023 }
9024
9025 if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
9026 !HDR_ENCRYPTED(hdr)) {
9027 ASSERT3U(size, ==, psize);
9028 to_write = abd_alloc_for_io(asize, ismd);
9029 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
9030 if (asize > size)
9031 abd_zero_off(to_write, size, asize - size);
9032 goto out;
9033 }
9034
9035 if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
9036 cabd = abd_alloc_for_io(MAX(size, asize), ismd);
9037 uint64_t csize = zio_compress_data(compress, to_write, &cabd,
9038 size, MIN(size, psize), hdr->b_complevel);
9039 if (csize >= size || csize > psize) {
9040 /*
9041 * We can't re-compress the block into the original
9042 * psize. Even if it fits into asize, it does not
9043 * matter, since checksum will never match on read.
9044 */
9045 abd_free(cabd);
9046 return (SET_ERROR(EIO));
9047 }
9048 if (asize > csize)
9049 abd_zero_off(cabd, csize, asize - csize);
9050 to_write = cabd;
9051 }
9052
9053 if (HDR_ENCRYPTED(hdr)) {
9054 eabd = abd_alloc_for_io(asize, ismd);
9055
9056 /*
9057 * If the dataset was disowned before the buffer
9058 * made it to this point, the key to re-encrypt
9059 * it won't be available. In this case we simply
9060 * won't write the buffer to the L2ARC.
9061 */
9062 ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
9063 FTAG, &dck);
9064 if (ret != 0)
9065 goto error;
9066
9067 ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
9068 hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
9069 hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
9070 &no_crypt);
9071 if (ret != 0)
9072 goto error;
9073
9074 if (no_crypt)
9075 abd_copy(eabd, to_write, psize);
9076
9077 if (psize != asize)
9078 abd_zero_off(eabd, psize, asize - psize);
9079
9080 /* assert that the MAC we got here matches the one we saved */
9081 ASSERT0(memcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
9082 spa_keystore_dsl_key_rele(spa, dck, FTAG);
9083
9084 if (to_write == cabd)
9085 abd_free(cabd);
9086
9087 to_write = eabd;
9088 }
9089
9090 out:
9091 ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
9092 *abd_out = to_write;
9093 return (0);
9094
9095 error:
9096 if (dck != NULL)
9097 spa_keystore_dsl_key_rele(spa, dck, FTAG);
9098 if (cabd != NULL)
9099 abd_free(cabd);
9100 if (eabd != NULL)
9101 abd_free(eabd);
9102
9103 *abd_out = NULL;
9104 return (ret);
9105 }
9106
9107 static void
l2arc_blk_fetch_done(zio_t * zio)9108 l2arc_blk_fetch_done(zio_t *zio)
9109 {
9110 l2arc_read_callback_t *cb;
9111
9112 cb = zio->io_private;
9113 if (cb->l2rcb_abd != NULL)
9114 abd_free(cb->l2rcb_abd);
9115 kmem_free(cb, sizeof (l2arc_read_callback_t));
9116 }
9117
9118 /*
9119 * Find and write ARC buffers to the L2ARC device.
9120 *
9121 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
9122 * for reading until they have completed writing.
9123 * The headroom_boost is an in-out parameter used to maintain headroom boost
9124 * state between calls to this function.
9125 *
9126 * Returns the number of bytes actually written (which may be smaller than
9127 * the delta by which the device hand has changed due to alignment and the
9128 * writing of log blocks).
9129 */
9130 static uint64_t
l2arc_write_buffers(spa_t * spa,l2arc_dev_t * dev,uint64_t target_sz)9131 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
9132 {
9133 arc_buf_hdr_t *hdr, *head, *marker;
9134 uint64_t write_asize, write_psize, headroom;
9135 boolean_t full, from_head = !arc_warm;
9136 l2arc_write_callback_t *cb = NULL;
9137 zio_t *pio, *wzio;
9138 uint64_t guid = spa_load_guid(spa);
9139 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9140
9141 ASSERT3P(dev->l2ad_vdev, !=, NULL);
9142
9143 pio = NULL;
9144 write_asize = write_psize = 0;
9145 full = B_FALSE;
9146 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
9147 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
9148 marker = arc_state_alloc_marker();
9149
9150 /*
9151 * Copy buffers for L2ARC writing.
9152 */
9153 for (int pass = 0; pass < L2ARC_FEED_TYPES; pass++) {
9154 /*
9155 * pass == 0: MFU meta
9156 * pass == 1: MRU meta
9157 * pass == 2: MFU data
9158 * pass == 3: MRU data
9159 */
9160 if (l2arc_mfuonly == 1) {
9161 if (pass == 1 || pass == 3)
9162 continue;
9163 } else if (l2arc_mfuonly > 1) {
9164 if (pass == 3)
9165 continue;
9166 }
9167
9168 uint64_t passed_sz = 0;
9169 headroom = target_sz * l2arc_headroom;
9170 if (zfs_compressed_arc_enabled)
9171 headroom = (headroom * l2arc_headroom_boost) / 100;
9172
9173 /*
9174 * Until the ARC is warm and starts to evict, read from the
9175 * head of the ARC lists rather than the tail.
9176 */
9177 multilist_sublist_t *mls = l2arc_sublist_lock(pass);
9178 ASSERT3P(mls, !=, NULL);
9179 if (from_head)
9180 hdr = multilist_sublist_head(mls);
9181 else
9182 hdr = multilist_sublist_tail(mls);
9183
9184 while (hdr != NULL) {
9185 kmutex_t *hash_lock;
9186 abd_t *to_write = NULL;
9187
9188 hash_lock = HDR_LOCK(hdr);
9189 if (!mutex_tryenter(hash_lock)) {
9190 skip:
9191 /* Skip this buffer rather than waiting. */
9192 if (from_head)
9193 hdr = multilist_sublist_next(mls, hdr);
9194 else
9195 hdr = multilist_sublist_prev(mls, hdr);
9196 continue;
9197 }
9198
9199 passed_sz += HDR_GET_LSIZE(hdr);
9200 if (l2arc_headroom != 0 && passed_sz > headroom) {
9201 /*
9202 * Searched too far.
9203 */
9204 mutex_exit(hash_lock);
9205 break;
9206 }
9207
9208 if (!l2arc_write_eligible(guid, hdr)) {
9209 mutex_exit(hash_lock);
9210 goto skip;
9211 }
9212
9213 ASSERT(HDR_HAS_L1HDR(hdr));
9214 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
9215 ASSERT3U(arc_hdr_size(hdr), >, 0);
9216 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
9217 HDR_HAS_RABD(hdr));
9218 uint64_t psize = HDR_GET_PSIZE(hdr);
9219 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
9220 psize);
9221
9222 /*
9223 * If the allocated size of this buffer plus the max
9224 * size for the pending log block exceeds the evicted
9225 * target size, terminate writing buffers for this run.
9226 */
9227 if (write_asize + asize +
9228 sizeof (l2arc_log_blk_phys_t) > target_sz) {
9229 full = B_TRUE;
9230 mutex_exit(hash_lock);
9231 break;
9232 }
9233
9234 /*
9235 * We should not sleep with sublist lock held or it
9236 * may block ARC eviction. Insert a marker to save
9237 * the position and drop the lock.
9238 */
9239 if (from_head) {
9240 multilist_sublist_insert_after(mls, hdr,
9241 marker);
9242 } else {
9243 multilist_sublist_insert_before(mls, hdr,
9244 marker);
9245 }
9246 multilist_sublist_unlock(mls);
9247
9248 /*
9249 * If this header has b_rabd, we can use this since it
9250 * must always match the data exactly as it exists on
9251 * disk. Otherwise, the L2ARC can normally use the
9252 * hdr's data, but if we're sharing data between the
9253 * hdr and one of its bufs, L2ARC needs its own copy of
9254 * the data so that the ZIO below can't race with the
9255 * buf consumer. To ensure that this copy will be
9256 * available for the lifetime of the ZIO and be cleaned
9257 * up afterwards, we add it to the l2arc_free_on_write
9258 * queue. If we need to apply any transforms to the
9259 * data (compression, encryption) we will also need the
9260 * extra buffer.
9261 */
9262 if (HDR_HAS_RABD(hdr) && psize == asize) {
9263 to_write = hdr->b_crypt_hdr.b_rabd;
9264 } else if ((HDR_COMPRESSION_ENABLED(hdr) ||
9265 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
9266 !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
9267 psize == asize) {
9268 to_write = hdr->b_l1hdr.b_pabd;
9269 } else {
9270 int ret;
9271 arc_buf_contents_t type = arc_buf_type(hdr);
9272
9273 ret = l2arc_apply_transforms(spa, hdr, asize,
9274 &to_write);
9275 if (ret != 0) {
9276 arc_hdr_clear_flags(hdr,
9277 ARC_FLAG_L2CACHE);
9278 mutex_exit(hash_lock);
9279 goto next;
9280 }
9281
9282 l2arc_free_abd_on_write(to_write, asize, type);
9283 }
9284
9285 hdr->b_l2hdr.b_dev = dev;
9286 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
9287 hdr->b_l2hdr.b_hits = 0;
9288 hdr->b_l2hdr.b_arcs_state =
9289 hdr->b_l1hdr.b_state->arcs_state;
9290 mutex_enter(&dev->l2ad_mtx);
9291 if (pio == NULL) {
9292 /*
9293 * Insert a dummy header on the buflist so
9294 * l2arc_write_done() can find where the
9295 * write buffers begin without searching.
9296 */
9297 list_insert_head(&dev->l2ad_buflist, head);
9298 }
9299 list_insert_head(&dev->l2ad_buflist, hdr);
9300 mutex_exit(&dev->l2ad_mtx);
9301 arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR |
9302 ARC_FLAG_L2_WRITING);
9303
9304 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
9305 arc_hdr_size(hdr), hdr);
9306 l2arc_hdr_arcstats_increment(hdr);
9307
9308 boolean_t commit = l2arc_log_blk_insert(dev, hdr);
9309 mutex_exit(hash_lock);
9310
9311 if (pio == NULL) {
9312 cb = kmem_alloc(
9313 sizeof (l2arc_write_callback_t), KM_SLEEP);
9314 cb->l2wcb_dev = dev;
9315 cb->l2wcb_head = head;
9316 list_create(&cb->l2wcb_abd_list,
9317 sizeof (l2arc_lb_abd_buf_t),
9318 offsetof(l2arc_lb_abd_buf_t, node));
9319 pio = zio_root(spa, l2arc_write_done, cb,
9320 ZIO_FLAG_CANFAIL);
9321 }
9322
9323 wzio = zio_write_phys(pio, dev->l2ad_vdev,
9324 dev->l2ad_hand, asize, to_write,
9325 ZIO_CHECKSUM_OFF, NULL, hdr,
9326 ZIO_PRIORITY_ASYNC_WRITE,
9327 ZIO_FLAG_CANFAIL, B_FALSE);
9328
9329 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
9330 zio_t *, wzio);
9331 zio_nowait(wzio);
9332
9333 write_psize += psize;
9334 write_asize += asize;
9335 dev->l2ad_hand += asize;
9336 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9337
9338 if (commit) {
9339 /* l2ad_hand will be adjusted inside. */
9340 write_asize +=
9341 l2arc_log_blk_commit(dev, pio, cb);
9342 }
9343
9344 next:
9345 multilist_sublist_lock(mls);
9346 if (from_head)
9347 hdr = multilist_sublist_next(mls, marker);
9348 else
9349 hdr = multilist_sublist_prev(mls, marker);
9350 multilist_sublist_remove(mls, marker);
9351 }
9352
9353 multilist_sublist_unlock(mls);
9354
9355 if (full == B_TRUE)
9356 break;
9357 }
9358
9359 arc_state_free_marker(marker);
9360
9361 /* No buffers selected for writing? */
9362 if (pio == NULL) {
9363 ASSERT0(write_psize);
9364 ASSERT(!HDR_HAS_L1HDR(head));
9365 kmem_cache_free(hdr_l2only_cache, head);
9366
9367 /*
9368 * Although we did not write any buffers l2ad_evict may
9369 * have advanced.
9370 */
9371 if (dev->l2ad_evict != l2dhdr->dh_evict)
9372 l2arc_dev_hdr_update(dev);
9373
9374 return (0);
9375 }
9376
9377 if (!dev->l2ad_first)
9378 ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
9379
9380 ASSERT3U(write_asize, <=, target_sz);
9381 ARCSTAT_BUMP(arcstat_l2_writes_sent);
9382 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
9383
9384 dev->l2ad_writing = B_TRUE;
9385 (void) zio_wait(pio);
9386 dev->l2ad_writing = B_FALSE;
9387
9388 /*
9389 * Update the device header after the zio completes as
9390 * l2arc_write_done() may have updated the memory holding the log block
9391 * pointers in the device header.
9392 */
9393 l2arc_dev_hdr_update(dev);
9394
9395 return (write_asize);
9396 }
9397
9398 static boolean_t
l2arc_hdr_limit_reached(void)9399 l2arc_hdr_limit_reached(void)
9400 {
9401 int64_t s = aggsum_upper_bound(&arc_sums.arcstat_l2_hdr_size);
9402
9403 return (arc_reclaim_needed() ||
9404 (s > (arc_warm ? arc_c : arc_c_max) * l2arc_meta_percent / 100));
9405 }
9406
9407 /*
9408 * This thread feeds the L2ARC at regular intervals. This is the beating
9409 * heart of the L2ARC.
9410 */
9411 static __attribute__((noreturn)) void
l2arc_feed_thread(void * unused)9412 l2arc_feed_thread(void *unused)
9413 {
9414 (void) unused;
9415 callb_cpr_t cpr;
9416 l2arc_dev_t *dev;
9417 spa_t *spa;
9418 uint64_t size, wrote;
9419 clock_t begin, next = ddi_get_lbolt();
9420 fstrans_cookie_t cookie;
9421
9422 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
9423
9424 mutex_enter(&l2arc_feed_thr_lock);
9425
9426 cookie = spl_fstrans_mark();
9427 while (l2arc_thread_exit == 0) {
9428 CALLB_CPR_SAFE_BEGIN(&cpr);
9429 (void) cv_timedwait_idle(&l2arc_feed_thr_cv,
9430 &l2arc_feed_thr_lock, next);
9431 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
9432 next = ddi_get_lbolt() + hz;
9433
9434 /*
9435 * Quick check for L2ARC devices.
9436 */
9437 mutex_enter(&l2arc_dev_mtx);
9438 if (l2arc_ndev == 0) {
9439 mutex_exit(&l2arc_dev_mtx);
9440 continue;
9441 }
9442 mutex_exit(&l2arc_dev_mtx);
9443 begin = ddi_get_lbolt();
9444
9445 /*
9446 * This selects the next l2arc device to write to, and in
9447 * doing so the next spa to feed from: dev->l2ad_spa. This
9448 * will return NULL if there are now no l2arc devices or if
9449 * they are all faulted.
9450 *
9451 * If a device is returned, its spa's config lock is also
9452 * held to prevent device removal. l2arc_dev_get_next()
9453 * will grab and release l2arc_dev_mtx.
9454 */
9455 if ((dev = l2arc_dev_get_next()) == NULL)
9456 continue;
9457
9458 spa = dev->l2ad_spa;
9459 ASSERT3P(spa, !=, NULL);
9460
9461 /*
9462 * If the pool is read-only then force the feed thread to
9463 * sleep a little longer.
9464 */
9465 if (!spa_writeable(spa)) {
9466 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
9467 spa_config_exit(spa, SCL_L2ARC, dev);
9468 continue;
9469 }
9470
9471 /*
9472 * Avoid contributing to memory pressure.
9473 */
9474 if (l2arc_hdr_limit_reached()) {
9475 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
9476 spa_config_exit(spa, SCL_L2ARC, dev);
9477 continue;
9478 }
9479
9480 ARCSTAT_BUMP(arcstat_l2_feeds);
9481
9482 size = l2arc_write_size(dev);
9483
9484 /*
9485 * Evict L2ARC buffers that will be overwritten.
9486 */
9487 l2arc_evict(dev, size, B_FALSE);
9488
9489 /*
9490 * Write ARC buffers.
9491 */
9492 wrote = l2arc_write_buffers(spa, dev, size);
9493
9494 /*
9495 * Calculate interval between writes.
9496 */
9497 next = l2arc_write_interval(begin, size, wrote);
9498 spa_config_exit(spa, SCL_L2ARC, dev);
9499 }
9500 spl_fstrans_unmark(cookie);
9501
9502 l2arc_thread_exit = 0;
9503 cv_broadcast(&l2arc_feed_thr_cv);
9504 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
9505 thread_exit();
9506 }
9507
9508 boolean_t
l2arc_vdev_present(vdev_t * vd)9509 l2arc_vdev_present(vdev_t *vd)
9510 {
9511 return (l2arc_vdev_get(vd) != NULL);
9512 }
9513
9514 /*
9515 * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if
9516 * the vdev_t isn't an L2ARC device.
9517 */
9518 l2arc_dev_t *
l2arc_vdev_get(vdev_t * vd)9519 l2arc_vdev_get(vdev_t *vd)
9520 {
9521 l2arc_dev_t *dev;
9522
9523 mutex_enter(&l2arc_dev_mtx);
9524 for (dev = list_head(l2arc_dev_list); dev != NULL;
9525 dev = list_next(l2arc_dev_list, dev)) {
9526 if (dev->l2ad_vdev == vd)
9527 break;
9528 }
9529 mutex_exit(&l2arc_dev_mtx);
9530
9531 return (dev);
9532 }
9533
9534 static void
l2arc_rebuild_dev(l2arc_dev_t * dev,boolean_t reopen)9535 l2arc_rebuild_dev(l2arc_dev_t *dev, boolean_t reopen)
9536 {
9537 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9538 uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9539 spa_t *spa = dev->l2ad_spa;
9540
9541 /*
9542 * The L2ARC has to hold at least the payload of one log block for
9543 * them to be restored (persistent L2ARC). The payload of a log block
9544 * depends on the amount of its log entries. We always write log blocks
9545 * with 1022 entries. How many of them are committed or restored depends
9546 * on the size of the L2ARC device. Thus the maximum payload of
9547 * one log block is 1022 * SPA_MAXBLOCKSIZE = 16GB. If the L2ARC device
9548 * is less than that, we reduce the amount of committed and restored
9549 * log entries per block so as to enable persistence.
9550 */
9551 if (dev->l2ad_end < l2arc_rebuild_blocks_min_l2size) {
9552 dev->l2ad_log_entries = 0;
9553 } else {
9554 dev->l2ad_log_entries = MIN((dev->l2ad_end -
9555 dev->l2ad_start) >> SPA_MAXBLOCKSHIFT,
9556 L2ARC_LOG_BLK_MAX_ENTRIES);
9557 }
9558
9559 /*
9560 * Read the device header, if an error is returned do not rebuild L2ARC.
9561 */
9562 if (l2arc_dev_hdr_read(dev) == 0 && dev->l2ad_log_entries > 0) {
9563 /*
9564 * If we are onlining a cache device (vdev_reopen) that was
9565 * still present (l2arc_vdev_present()) and rebuild is enabled,
9566 * we should evict all ARC buffers and pointers to log blocks
9567 * and reclaim their space before restoring its contents to
9568 * L2ARC.
9569 */
9570 if (reopen) {
9571 if (!l2arc_rebuild_enabled) {
9572 return;
9573 } else {
9574 l2arc_evict(dev, 0, B_TRUE);
9575 /* start a new log block */
9576 dev->l2ad_log_ent_idx = 0;
9577 dev->l2ad_log_blk_payload_asize = 0;
9578 dev->l2ad_log_blk_payload_start = 0;
9579 }
9580 }
9581 /*
9582 * Just mark the device as pending for a rebuild. We won't
9583 * be starting a rebuild in line here as it would block pool
9584 * import. Instead spa_load_impl will hand that off to an
9585 * async task which will call l2arc_spa_rebuild_start.
9586 */
9587 dev->l2ad_rebuild = B_TRUE;
9588 } else if (spa_writeable(spa)) {
9589 /*
9590 * In this case TRIM the whole device if l2arc_trim_ahead > 0,
9591 * otherwise create a new header. We zero out the memory holding
9592 * the header to reset dh_start_lbps. If we TRIM the whole
9593 * device the new header will be written by
9594 * vdev_trim_l2arc_thread() at the end of the TRIM to update the
9595 * trim_state in the header too. When reading the header, if
9596 * trim_state is not VDEV_TRIM_COMPLETE and l2arc_trim_ahead > 0
9597 * we opt to TRIM the whole device again.
9598 */
9599 if (l2arc_trim_ahead > 0) {
9600 dev->l2ad_trim_all = B_TRUE;
9601 } else {
9602 memset(l2dhdr, 0, l2dhdr_asize);
9603 l2arc_dev_hdr_update(dev);
9604 }
9605 }
9606 }
9607
9608 /*
9609 * Add a vdev for use by the L2ARC. By this point the spa has already
9610 * validated the vdev and opened it.
9611 */
9612 void
l2arc_add_vdev(spa_t * spa,vdev_t * vd)9613 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
9614 {
9615 l2arc_dev_t *adddev;
9616 uint64_t l2dhdr_asize;
9617
9618 ASSERT(!l2arc_vdev_present(vd));
9619
9620 /*
9621 * Create a new l2arc device entry.
9622 */
9623 adddev = vmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
9624 adddev->l2ad_spa = spa;
9625 adddev->l2ad_vdev = vd;
9626 /* leave extra size for an l2arc device header */
9627 l2dhdr_asize = adddev->l2ad_dev_hdr_asize =
9628 MAX(sizeof (*adddev->l2ad_dev_hdr), 1 << vd->vdev_ashift);
9629 adddev->l2ad_start = VDEV_LABEL_START_SIZE + l2dhdr_asize;
9630 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
9631 ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end);
9632 adddev->l2ad_hand = adddev->l2ad_start;
9633 adddev->l2ad_evict = adddev->l2ad_start;
9634 adddev->l2ad_first = B_TRUE;
9635 adddev->l2ad_writing = B_FALSE;
9636 adddev->l2ad_trim_all = B_FALSE;
9637 list_link_init(&adddev->l2ad_node);
9638 adddev->l2ad_dev_hdr = kmem_zalloc(l2dhdr_asize, KM_SLEEP);
9639
9640 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
9641 /*
9642 * This is a list of all ARC buffers that are still valid on the
9643 * device.
9644 */
9645 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
9646 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
9647
9648 /*
9649 * This is a list of pointers to log blocks that are still present
9650 * on the device.
9651 */
9652 list_create(&adddev->l2ad_lbptr_list, sizeof (l2arc_lb_ptr_buf_t),
9653 offsetof(l2arc_lb_ptr_buf_t, node));
9654
9655 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
9656 zfs_refcount_create(&adddev->l2ad_alloc);
9657 zfs_refcount_create(&adddev->l2ad_lb_asize);
9658 zfs_refcount_create(&adddev->l2ad_lb_count);
9659
9660 /*
9661 * Decide if dev is eligible for L2ARC rebuild or whole device
9662 * trimming. This has to happen before the device is added in the
9663 * cache device list and l2arc_dev_mtx is released. Otherwise
9664 * l2arc_feed_thread() might already start writing on the
9665 * device.
9666 */
9667 l2arc_rebuild_dev(adddev, B_FALSE);
9668
9669 /*
9670 * Add device to global list
9671 */
9672 mutex_enter(&l2arc_dev_mtx);
9673 list_insert_head(l2arc_dev_list, adddev);
9674 atomic_inc_64(&l2arc_ndev);
9675 mutex_exit(&l2arc_dev_mtx);
9676 }
9677
9678 /*
9679 * Decide if a vdev is eligible for L2ARC rebuild, called from vdev_reopen()
9680 * in case of onlining a cache device.
9681 */
9682 void
l2arc_rebuild_vdev(vdev_t * vd,boolean_t reopen)9683 l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen)
9684 {
9685 l2arc_dev_t *dev = NULL;
9686
9687 dev = l2arc_vdev_get(vd);
9688 ASSERT3P(dev, !=, NULL);
9689
9690 /*
9691 * In contrast to l2arc_add_vdev() we do not have to worry about
9692 * l2arc_feed_thread() invalidating previous content when onlining a
9693 * cache device. The device parameters (l2ad*) are not cleared when
9694 * offlining the device and writing new buffers will not invalidate
9695 * all previous content. In worst case only buffers that have not had
9696 * their log block written to the device will be lost.
9697 * When onlining the cache device (ie offline->online without exporting
9698 * the pool in between) this happens:
9699 * vdev_reopen() -> vdev_open() -> l2arc_rebuild_vdev()
9700 * | |
9701 * vdev_is_dead() = B_FALSE l2ad_rebuild = B_TRUE
9702 * During the time where vdev_is_dead = B_FALSE and until l2ad_rebuild
9703 * is set to B_TRUE we might write additional buffers to the device.
9704 */
9705 l2arc_rebuild_dev(dev, reopen);
9706 }
9707
9708 /*
9709 * Remove a vdev from the L2ARC.
9710 */
9711 void
l2arc_remove_vdev(vdev_t * vd)9712 l2arc_remove_vdev(vdev_t *vd)
9713 {
9714 l2arc_dev_t *remdev = NULL;
9715
9716 /*
9717 * Find the device by vdev
9718 */
9719 remdev = l2arc_vdev_get(vd);
9720 ASSERT3P(remdev, !=, NULL);
9721
9722 /*
9723 * Cancel any ongoing or scheduled rebuild.
9724 */
9725 mutex_enter(&l2arc_rebuild_thr_lock);
9726 if (remdev->l2ad_rebuild_began == B_TRUE) {
9727 remdev->l2ad_rebuild_cancel = B_TRUE;
9728 while (remdev->l2ad_rebuild == B_TRUE)
9729 cv_wait(&l2arc_rebuild_thr_cv, &l2arc_rebuild_thr_lock);
9730 }
9731 mutex_exit(&l2arc_rebuild_thr_lock);
9732
9733 /*
9734 * Remove device from global list
9735 */
9736 mutex_enter(&l2arc_dev_mtx);
9737 list_remove(l2arc_dev_list, remdev);
9738 l2arc_dev_last = NULL; /* may have been invalidated */
9739 atomic_dec_64(&l2arc_ndev);
9740 mutex_exit(&l2arc_dev_mtx);
9741
9742 /*
9743 * Clear all buflists and ARC references. L2ARC device flush.
9744 */
9745 l2arc_evict(remdev, 0, B_TRUE);
9746 list_destroy(&remdev->l2ad_buflist);
9747 ASSERT(list_is_empty(&remdev->l2ad_lbptr_list));
9748 list_destroy(&remdev->l2ad_lbptr_list);
9749 mutex_destroy(&remdev->l2ad_mtx);
9750 zfs_refcount_destroy(&remdev->l2ad_alloc);
9751 zfs_refcount_destroy(&remdev->l2ad_lb_asize);
9752 zfs_refcount_destroy(&remdev->l2ad_lb_count);
9753 kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize);
9754 vmem_free(remdev, sizeof (l2arc_dev_t));
9755 }
9756
9757 void
l2arc_init(void)9758 l2arc_init(void)
9759 {
9760 l2arc_thread_exit = 0;
9761 l2arc_ndev = 0;
9762
9763 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9764 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
9765 mutex_init(&l2arc_rebuild_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9766 cv_init(&l2arc_rebuild_thr_cv, NULL, CV_DEFAULT, NULL);
9767 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
9768 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
9769
9770 l2arc_dev_list = &L2ARC_dev_list;
9771 l2arc_free_on_write = &L2ARC_free_on_write;
9772 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
9773 offsetof(l2arc_dev_t, l2ad_node));
9774 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
9775 offsetof(l2arc_data_free_t, l2df_list_node));
9776 }
9777
9778 void
l2arc_fini(void)9779 l2arc_fini(void)
9780 {
9781 mutex_destroy(&l2arc_feed_thr_lock);
9782 cv_destroy(&l2arc_feed_thr_cv);
9783 mutex_destroy(&l2arc_rebuild_thr_lock);
9784 cv_destroy(&l2arc_rebuild_thr_cv);
9785 mutex_destroy(&l2arc_dev_mtx);
9786 mutex_destroy(&l2arc_free_on_write_mtx);
9787
9788 list_destroy(l2arc_dev_list);
9789 list_destroy(l2arc_free_on_write);
9790 }
9791
9792 void
l2arc_start(void)9793 l2arc_start(void)
9794 {
9795 if (!(spa_mode_global & SPA_MODE_WRITE))
9796 return;
9797
9798 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
9799 TS_RUN, defclsyspri);
9800 }
9801
9802 void
l2arc_stop(void)9803 l2arc_stop(void)
9804 {
9805 if (!(spa_mode_global & SPA_MODE_WRITE))
9806 return;
9807
9808 mutex_enter(&l2arc_feed_thr_lock);
9809 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
9810 l2arc_thread_exit = 1;
9811 while (l2arc_thread_exit != 0)
9812 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
9813 mutex_exit(&l2arc_feed_thr_lock);
9814 }
9815
9816 /*
9817 * Punches out rebuild threads for the L2ARC devices in a spa. This should
9818 * be called after pool import from the spa async thread, since starting
9819 * these threads directly from spa_import() will make them part of the
9820 * "zpool import" context and delay process exit (and thus pool import).
9821 */
9822 void
l2arc_spa_rebuild_start(spa_t * spa)9823 l2arc_spa_rebuild_start(spa_t *spa)
9824 {
9825 ASSERT(MUTEX_HELD(&spa_namespace_lock));
9826
9827 /*
9828 * Locate the spa's l2arc devices and kick off rebuild threads.
9829 */
9830 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
9831 l2arc_dev_t *dev =
9832 l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]);
9833 if (dev == NULL) {
9834 /* Don't attempt a rebuild if the vdev is UNAVAIL */
9835 continue;
9836 }
9837 mutex_enter(&l2arc_rebuild_thr_lock);
9838 if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) {
9839 dev->l2ad_rebuild_began = B_TRUE;
9840 (void) thread_create(NULL, 0, l2arc_dev_rebuild_thread,
9841 dev, 0, &p0, TS_RUN, minclsyspri);
9842 }
9843 mutex_exit(&l2arc_rebuild_thr_lock);
9844 }
9845 }
9846
9847 /*
9848 * Main entry point for L2ARC rebuilding.
9849 */
9850 static __attribute__((noreturn)) void
l2arc_dev_rebuild_thread(void * arg)9851 l2arc_dev_rebuild_thread(void *arg)
9852 {
9853 l2arc_dev_t *dev = arg;
9854
9855 VERIFY(!dev->l2ad_rebuild_cancel);
9856 VERIFY(dev->l2ad_rebuild);
9857 (void) l2arc_rebuild(dev);
9858 mutex_enter(&l2arc_rebuild_thr_lock);
9859 dev->l2ad_rebuild_began = B_FALSE;
9860 dev->l2ad_rebuild = B_FALSE;
9861 mutex_exit(&l2arc_rebuild_thr_lock);
9862
9863 thread_exit();
9864 }
9865
9866 /*
9867 * This function implements the actual L2ARC metadata rebuild. It:
9868 * starts reading the log block chain and restores each block's contents
9869 * to memory (reconstructing arc_buf_hdr_t's).
9870 *
9871 * Operation stops under any of the following conditions:
9872 *
9873 * 1) We reach the end of the log block chain.
9874 * 2) We encounter *any* error condition (cksum errors, io errors)
9875 */
9876 static int
l2arc_rebuild(l2arc_dev_t * dev)9877 l2arc_rebuild(l2arc_dev_t *dev)
9878 {
9879 vdev_t *vd = dev->l2ad_vdev;
9880 spa_t *spa = vd->vdev_spa;
9881 int err = 0;
9882 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9883 l2arc_log_blk_phys_t *this_lb, *next_lb;
9884 zio_t *this_io = NULL, *next_io = NULL;
9885 l2arc_log_blkptr_t lbps[2];
9886 l2arc_lb_ptr_buf_t *lb_ptr_buf;
9887 boolean_t lock_held;
9888
9889 this_lb = vmem_zalloc(sizeof (*this_lb), KM_SLEEP);
9890 next_lb = vmem_zalloc(sizeof (*next_lb), KM_SLEEP);
9891
9892 /*
9893 * We prevent device removal while issuing reads to the device,
9894 * then during the rebuilding phases we drop this lock again so
9895 * that a spa_unload or device remove can be initiated - this is
9896 * safe, because the spa will signal us to stop before removing
9897 * our device and wait for us to stop.
9898 */
9899 spa_config_enter(spa, SCL_L2ARC, vd, RW_READER);
9900 lock_held = B_TRUE;
9901
9902 /*
9903 * Retrieve the persistent L2ARC device state.
9904 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9905 */
9906 dev->l2ad_evict = MAX(l2dhdr->dh_evict, dev->l2ad_start);
9907 dev->l2ad_hand = MAX(l2dhdr->dh_start_lbps[0].lbp_daddr +
9908 L2BLK_GET_PSIZE((&l2dhdr->dh_start_lbps[0])->lbp_prop),
9909 dev->l2ad_start);
9910 dev->l2ad_first = !!(l2dhdr->dh_flags & L2ARC_DEV_HDR_EVICT_FIRST);
9911
9912 vd->vdev_trim_action_time = l2dhdr->dh_trim_action_time;
9913 vd->vdev_trim_state = l2dhdr->dh_trim_state;
9914
9915 /*
9916 * In case the zfs module parameter l2arc_rebuild_enabled is false
9917 * we do not start the rebuild process.
9918 */
9919 if (!l2arc_rebuild_enabled)
9920 goto out;
9921
9922 /* Prepare the rebuild process */
9923 memcpy(lbps, l2dhdr->dh_start_lbps, sizeof (lbps));
9924
9925 /* Start the rebuild process */
9926 for (;;) {
9927 if (!l2arc_log_blkptr_valid(dev, &lbps[0]))
9928 break;
9929
9930 if ((err = l2arc_log_blk_read(dev, &lbps[0], &lbps[1],
9931 this_lb, next_lb, this_io, &next_io)) != 0)
9932 goto out;
9933
9934 /*
9935 * Our memory pressure valve. If the system is running low
9936 * on memory, rather than swamping memory with new ARC buf
9937 * hdrs, we opt not to rebuild the L2ARC. At this point,
9938 * however, we have already set up our L2ARC dev to chain in
9939 * new metadata log blocks, so the user may choose to offline/
9940 * online the L2ARC dev at a later time (or re-import the pool)
9941 * to reconstruct it (when there's less memory pressure).
9942 */
9943 if (l2arc_hdr_limit_reached()) {
9944 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem);
9945 cmn_err(CE_NOTE, "System running low on memory, "
9946 "aborting L2ARC rebuild.");
9947 err = SET_ERROR(ENOMEM);
9948 goto out;
9949 }
9950
9951 spa_config_exit(spa, SCL_L2ARC, vd);
9952 lock_held = B_FALSE;
9953
9954 /*
9955 * Now that we know that the next_lb checks out alright, we
9956 * can start reconstruction from this log block.
9957 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9958 */
9959 uint64_t asize = L2BLK_GET_PSIZE((&lbps[0])->lbp_prop);
9960 l2arc_log_blk_restore(dev, this_lb, asize);
9961
9962 /*
9963 * log block restored, include its pointer in the list of
9964 * pointers to log blocks present in the L2ARC device.
9965 */
9966 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
9967 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t),
9968 KM_SLEEP);
9969 memcpy(lb_ptr_buf->lb_ptr, &lbps[0],
9970 sizeof (l2arc_log_blkptr_t));
9971 mutex_enter(&dev->l2ad_mtx);
9972 list_insert_tail(&dev->l2ad_lbptr_list, lb_ptr_buf);
9973 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
9974 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
9975 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
9976 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
9977 mutex_exit(&dev->l2ad_mtx);
9978 vdev_space_update(vd, asize, 0, 0);
9979
9980 /*
9981 * Protection against loops of log blocks:
9982 *
9983 * l2ad_hand l2ad_evict
9984 * V V
9985 * l2ad_start |=======================================| l2ad_end
9986 * -----|||----|||---|||----|||
9987 * (3) (2) (1) (0)
9988 * ---|||---|||----|||---|||
9989 * (7) (6) (5) (4)
9990 *
9991 * In this situation the pointer of log block (4) passes
9992 * l2arc_log_blkptr_valid() but the log block should not be
9993 * restored as it is overwritten by the payload of log block
9994 * (0). Only log blocks (0)-(3) should be restored. We check
9995 * whether l2ad_evict lies in between the payload starting
9996 * offset of the next log block (lbps[1].lbp_payload_start)
9997 * and the payload starting offset of the present log block
9998 * (lbps[0].lbp_payload_start). If true and this isn't the
9999 * first pass, we are looping from the beginning and we should
10000 * stop.
10001 */
10002 if (l2arc_range_check_overlap(lbps[1].lbp_payload_start,
10003 lbps[0].lbp_payload_start, dev->l2ad_evict) &&
10004 !dev->l2ad_first)
10005 goto out;
10006
10007 kpreempt(KPREEMPT_SYNC);
10008 for (;;) {
10009 mutex_enter(&l2arc_rebuild_thr_lock);
10010 if (dev->l2ad_rebuild_cancel) {
10011 dev->l2ad_rebuild = B_FALSE;
10012 cv_signal(&l2arc_rebuild_thr_cv);
10013 mutex_exit(&l2arc_rebuild_thr_lock);
10014 err = SET_ERROR(ECANCELED);
10015 goto out;
10016 }
10017 mutex_exit(&l2arc_rebuild_thr_lock);
10018 if (spa_config_tryenter(spa, SCL_L2ARC, vd,
10019 RW_READER)) {
10020 lock_held = B_TRUE;
10021 break;
10022 }
10023 /*
10024 * L2ARC config lock held by somebody in writer,
10025 * possibly due to them trying to remove us. They'll
10026 * likely to want us to shut down, so after a little
10027 * delay, we check l2ad_rebuild_cancel and retry
10028 * the lock again.
10029 */
10030 delay(1);
10031 }
10032
10033 /*
10034 * Continue with the next log block.
10035 */
10036 lbps[0] = lbps[1];
10037 lbps[1] = this_lb->lb_prev_lbp;
10038 PTR_SWAP(this_lb, next_lb);
10039 this_io = next_io;
10040 next_io = NULL;
10041 }
10042
10043 if (this_io != NULL)
10044 l2arc_log_blk_fetch_abort(this_io);
10045 out:
10046 if (next_io != NULL)
10047 l2arc_log_blk_fetch_abort(next_io);
10048 vmem_free(this_lb, sizeof (*this_lb));
10049 vmem_free(next_lb, sizeof (*next_lb));
10050
10051 if (!l2arc_rebuild_enabled) {
10052 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10053 "disabled");
10054 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) > 0) {
10055 ARCSTAT_BUMP(arcstat_l2_rebuild_success);
10056 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10057 "successful, restored %llu blocks",
10058 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10059 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) == 0) {
10060 /*
10061 * No error but also nothing restored, meaning the lbps array
10062 * in the device header points to invalid/non-present log
10063 * blocks. Reset the header.
10064 */
10065 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10066 "no valid log blocks");
10067 memset(l2dhdr, 0, dev->l2ad_dev_hdr_asize);
10068 l2arc_dev_hdr_update(dev);
10069 } else if (err == ECANCELED) {
10070 /*
10071 * In case the rebuild was canceled do not log to spa history
10072 * log as the pool may be in the process of being removed.
10073 */
10074 zfs_dbgmsg("L2ARC rebuild aborted, restored %llu blocks",
10075 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10076 } else if (err != 0) {
10077 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
10078 "aborted, restored %llu blocks",
10079 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
10080 }
10081
10082 if (lock_held)
10083 spa_config_exit(spa, SCL_L2ARC, vd);
10084
10085 return (err);
10086 }
10087
10088 /*
10089 * Attempts to read the device header on the provided L2ARC device and writes
10090 * it to `hdr'. On success, this function returns 0, otherwise the appropriate
10091 * error code is returned.
10092 */
10093 static int
l2arc_dev_hdr_read(l2arc_dev_t * dev)10094 l2arc_dev_hdr_read(l2arc_dev_t *dev)
10095 {
10096 int err;
10097 uint64_t guid;
10098 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10099 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10100 abd_t *abd;
10101
10102 guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10103
10104 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10105
10106 err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev,
10107 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd,
10108 ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
10109 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY |
10110 ZIO_FLAG_SPECULATIVE, B_FALSE));
10111
10112 abd_free(abd);
10113
10114 if (err != 0) {
10115 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_dh_errors);
10116 zfs_dbgmsg("L2ARC IO error (%d) while reading device header, "
10117 "vdev guid: %llu", err,
10118 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10119 return (err);
10120 }
10121
10122 if (l2dhdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC))
10123 byteswap_uint64_array(l2dhdr, sizeof (*l2dhdr));
10124
10125 if (l2dhdr->dh_magic != L2ARC_DEV_HDR_MAGIC ||
10126 l2dhdr->dh_spa_guid != guid ||
10127 l2dhdr->dh_vdev_guid != dev->l2ad_vdev->vdev_guid ||
10128 l2dhdr->dh_version != L2ARC_PERSISTENT_VERSION ||
10129 l2dhdr->dh_log_entries != dev->l2ad_log_entries ||
10130 l2dhdr->dh_end != dev->l2ad_end ||
10131 !l2arc_range_check_overlap(dev->l2ad_start, dev->l2ad_end,
10132 l2dhdr->dh_evict) ||
10133 (l2dhdr->dh_trim_state != VDEV_TRIM_COMPLETE &&
10134 l2arc_trim_ahead > 0)) {
10135 /*
10136 * Attempt to rebuild a device containing no actual dev hdr
10137 * or containing a header from some other pool or from another
10138 * version of persistent L2ARC.
10139 */
10140 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported);
10141 return (SET_ERROR(ENOTSUP));
10142 }
10143
10144 return (0);
10145 }
10146
10147 /*
10148 * Reads L2ARC log blocks from storage and validates their contents.
10149 *
10150 * This function implements a simple fetcher to make sure that while
10151 * we're processing one buffer the L2ARC is already fetching the next
10152 * one in the chain.
10153 *
10154 * The arguments this_lp and next_lp point to the current and next log block
10155 * address in the block chain. Similarly, this_lb and next_lb hold the
10156 * l2arc_log_blk_phys_t's of the current and next L2ARC blk.
10157 *
10158 * The `this_io' and `next_io' arguments are used for block fetching.
10159 * When issuing the first blk IO during rebuild, you should pass NULL for
10160 * `this_io'. This function will then issue a sync IO to read the block and
10161 * also issue an async IO to fetch the next block in the block chain. The
10162 * fetched IO is returned in `next_io'. On subsequent calls to this
10163 * function, pass the value returned in `next_io' from the previous call
10164 * as `this_io' and a fresh `next_io' pointer to hold the next fetch IO.
10165 * Prior to the call, you should initialize your `next_io' pointer to be
10166 * NULL. If no fetch IO was issued, the pointer is left set at NULL.
10167 *
10168 * On success, this function returns 0, otherwise it returns an appropriate
10169 * error code. On error the fetching IO is aborted and cleared before
10170 * returning from this function. Therefore, if we return `success', the
10171 * caller can assume that we have taken care of cleanup of fetch IOs.
10172 */
10173 static int
l2arc_log_blk_read(l2arc_dev_t * dev,const l2arc_log_blkptr_t * this_lbp,const l2arc_log_blkptr_t * next_lbp,l2arc_log_blk_phys_t * this_lb,l2arc_log_blk_phys_t * next_lb,zio_t * this_io,zio_t ** next_io)10174 l2arc_log_blk_read(l2arc_dev_t *dev,
10175 const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp,
10176 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
10177 zio_t *this_io, zio_t **next_io)
10178 {
10179 int err = 0;
10180 zio_cksum_t cksum;
10181 uint64_t asize;
10182
10183 ASSERT(this_lbp != NULL && next_lbp != NULL);
10184 ASSERT(this_lb != NULL && next_lb != NULL);
10185 ASSERT(next_io != NULL && *next_io == NULL);
10186 ASSERT(l2arc_log_blkptr_valid(dev, this_lbp));
10187
10188 /*
10189 * Check to see if we have issued the IO for this log block in a
10190 * previous run. If not, this is the first call, so issue it now.
10191 */
10192 if (this_io == NULL) {
10193 this_io = l2arc_log_blk_fetch(dev->l2ad_vdev, this_lbp,
10194 this_lb);
10195 }
10196
10197 /*
10198 * Peek to see if we can start issuing the next IO immediately.
10199 */
10200 if (l2arc_log_blkptr_valid(dev, next_lbp)) {
10201 /*
10202 * Start issuing IO for the next log block early - this
10203 * should help keep the L2ARC device busy while we
10204 * decompress and restore this log block.
10205 */
10206 *next_io = l2arc_log_blk_fetch(dev->l2ad_vdev, next_lbp,
10207 next_lb);
10208 }
10209
10210 /* Wait for the IO to read this log block to complete */
10211 if ((err = zio_wait(this_io)) != 0) {
10212 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors);
10213 zfs_dbgmsg("L2ARC IO error (%d) while reading log block, "
10214 "offset: %llu, vdev guid: %llu", err,
10215 (u_longlong_t)this_lbp->lbp_daddr,
10216 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10217 goto cleanup;
10218 }
10219
10220 /*
10221 * Make sure the buffer checks out.
10222 * L2BLK_GET_PSIZE returns aligned size for log blocks.
10223 */
10224 asize = L2BLK_GET_PSIZE((this_lbp)->lbp_prop);
10225 fletcher_4_native(this_lb, asize, NULL, &cksum);
10226 if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) {
10227 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_lb_errors);
10228 zfs_dbgmsg("L2ARC log block cksum failed, offset: %llu, "
10229 "vdev guid: %llu, l2ad_hand: %llu, l2ad_evict: %llu",
10230 (u_longlong_t)this_lbp->lbp_daddr,
10231 (u_longlong_t)dev->l2ad_vdev->vdev_guid,
10232 (u_longlong_t)dev->l2ad_hand,
10233 (u_longlong_t)dev->l2ad_evict);
10234 err = SET_ERROR(ECKSUM);
10235 goto cleanup;
10236 }
10237
10238 /* Now we can take our time decoding this buffer */
10239 switch (L2BLK_GET_COMPRESS((this_lbp)->lbp_prop)) {
10240 case ZIO_COMPRESS_OFF:
10241 break;
10242 case ZIO_COMPRESS_LZ4: {
10243 abd_t *abd = abd_alloc_linear(asize, B_TRUE);
10244 abd_copy_from_buf_off(abd, this_lb, 0, asize);
10245 abd_t dabd;
10246 abd_get_from_buf_struct(&dabd, this_lb, sizeof (*this_lb));
10247 err = zio_decompress_data(
10248 L2BLK_GET_COMPRESS((this_lbp)->lbp_prop),
10249 abd, &dabd, asize, sizeof (*this_lb), NULL);
10250 abd_free(&dabd);
10251 abd_free(abd);
10252 if (err != 0) {
10253 err = SET_ERROR(EINVAL);
10254 goto cleanup;
10255 }
10256 break;
10257 }
10258 default:
10259 err = SET_ERROR(EINVAL);
10260 goto cleanup;
10261 }
10262 if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC))
10263 byteswap_uint64_array(this_lb, sizeof (*this_lb));
10264 if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) {
10265 err = SET_ERROR(EINVAL);
10266 goto cleanup;
10267 }
10268 cleanup:
10269 /* Abort an in-flight fetch I/O in case of error */
10270 if (err != 0 && *next_io != NULL) {
10271 l2arc_log_blk_fetch_abort(*next_io);
10272 *next_io = NULL;
10273 }
10274 return (err);
10275 }
10276
10277 /*
10278 * Restores the payload of a log block to ARC. This creates empty ARC hdr
10279 * entries which only contain an l2arc hdr, essentially restoring the
10280 * buffers to their L2ARC evicted state. This function also updates space
10281 * usage on the L2ARC vdev to make sure it tracks restored buffers.
10282 */
10283 static void
l2arc_log_blk_restore(l2arc_dev_t * dev,const l2arc_log_blk_phys_t * lb,uint64_t lb_asize)10284 l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb,
10285 uint64_t lb_asize)
10286 {
10287 uint64_t size = 0, asize = 0;
10288 uint64_t log_entries = dev->l2ad_log_entries;
10289
10290 /*
10291 * Usually arc_adapt() is called only for data, not headers, but
10292 * since we may allocate significant amount of memory here, let ARC
10293 * grow its arc_c.
10294 */
10295 arc_adapt(log_entries * HDR_L2ONLY_SIZE);
10296
10297 for (int i = log_entries - 1; i >= 0; i--) {
10298 /*
10299 * Restore goes in the reverse temporal direction to preserve
10300 * correct temporal ordering of buffers in the l2ad_buflist.
10301 * l2arc_hdr_restore also does a list_insert_tail instead of
10302 * list_insert_head on the l2ad_buflist:
10303 *
10304 * LIST l2ad_buflist LIST
10305 * HEAD <------ (time) ------ TAIL
10306 * direction +-----+-----+-----+-----+-----+ direction
10307 * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild
10308 * fill +-----+-----+-----+-----+-----+
10309 * ^ ^
10310 * | |
10311 * | |
10312 * l2arc_feed_thread l2arc_rebuild
10313 * will place new bufs here restores bufs here
10314 *
10315 * During l2arc_rebuild() the device is not used by
10316 * l2arc_feed_thread() as dev->l2ad_rebuild is set to true.
10317 */
10318 size += L2BLK_GET_LSIZE((&lb->lb_entries[i])->le_prop);
10319 asize += vdev_psize_to_asize(dev->l2ad_vdev,
10320 L2BLK_GET_PSIZE((&lb->lb_entries[i])->le_prop));
10321 l2arc_hdr_restore(&lb->lb_entries[i], dev);
10322 }
10323
10324 /*
10325 * Record rebuild stats:
10326 * size Logical size of restored buffers in the L2ARC
10327 * asize Aligned size of restored buffers in the L2ARC
10328 */
10329 ARCSTAT_INCR(arcstat_l2_rebuild_size, size);
10330 ARCSTAT_INCR(arcstat_l2_rebuild_asize, asize);
10331 ARCSTAT_INCR(arcstat_l2_rebuild_bufs, log_entries);
10332 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, lb_asize);
10333 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, asize / lb_asize);
10334 ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks);
10335 }
10336
10337 /*
10338 * Restores a single ARC buf hdr from a log entry. The ARC buffer is put
10339 * into a state indicating that it has been evicted to L2ARC.
10340 */
10341 static void
l2arc_hdr_restore(const l2arc_log_ent_phys_t * le,l2arc_dev_t * dev)10342 l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev)
10343 {
10344 arc_buf_hdr_t *hdr, *exists;
10345 kmutex_t *hash_lock;
10346 arc_buf_contents_t type = L2BLK_GET_TYPE((le)->le_prop);
10347 uint64_t asize;
10348
10349 /*
10350 * Do all the allocation before grabbing any locks, this lets us
10351 * sleep if memory is full and we don't have to deal with failed
10352 * allocations.
10353 */
10354 hdr = arc_buf_alloc_l2only(L2BLK_GET_LSIZE((le)->le_prop), type,
10355 dev, le->le_dva, le->le_daddr,
10356 L2BLK_GET_PSIZE((le)->le_prop), le->le_birth,
10357 L2BLK_GET_COMPRESS((le)->le_prop), le->le_complevel,
10358 L2BLK_GET_PROTECTED((le)->le_prop),
10359 L2BLK_GET_PREFETCH((le)->le_prop),
10360 L2BLK_GET_STATE((le)->le_prop));
10361 asize = vdev_psize_to_asize(dev->l2ad_vdev,
10362 L2BLK_GET_PSIZE((le)->le_prop));
10363
10364 /*
10365 * vdev_space_update() has to be called before arc_hdr_destroy() to
10366 * avoid underflow since the latter also calls vdev_space_update().
10367 */
10368 l2arc_hdr_arcstats_increment(hdr);
10369 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10370
10371 mutex_enter(&dev->l2ad_mtx);
10372 list_insert_tail(&dev->l2ad_buflist, hdr);
10373 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
10374 mutex_exit(&dev->l2ad_mtx);
10375
10376 exists = buf_hash_insert(hdr, &hash_lock);
10377 if (exists) {
10378 /* Buffer was already cached, no need to restore it. */
10379 arc_hdr_destroy(hdr);
10380 /*
10381 * If the buffer is already cached, check whether it has
10382 * L2ARC metadata. If not, enter them and update the flag.
10383 * This is important is case of onlining a cache device, since
10384 * we previously evicted all L2ARC metadata from ARC.
10385 */
10386 if (!HDR_HAS_L2HDR(exists)) {
10387 arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR);
10388 exists->b_l2hdr.b_dev = dev;
10389 exists->b_l2hdr.b_daddr = le->le_daddr;
10390 exists->b_l2hdr.b_arcs_state =
10391 L2BLK_GET_STATE((le)->le_prop);
10392 mutex_enter(&dev->l2ad_mtx);
10393 list_insert_tail(&dev->l2ad_buflist, exists);
10394 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
10395 arc_hdr_size(exists), exists);
10396 mutex_exit(&dev->l2ad_mtx);
10397 l2arc_hdr_arcstats_increment(exists);
10398 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10399 }
10400 ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached);
10401 }
10402
10403 mutex_exit(hash_lock);
10404 }
10405
10406 /*
10407 * Starts an asynchronous read IO to read a log block. This is used in log
10408 * block reconstruction to start reading the next block before we are done
10409 * decoding and reconstructing the current block, to keep the l2arc device
10410 * nice and hot with read IO to process.
10411 * The returned zio will contain a newly allocated memory buffers for the IO
10412 * data which should then be freed by the caller once the zio is no longer
10413 * needed (i.e. due to it having completed). If you wish to abort this
10414 * zio, you should do so using l2arc_log_blk_fetch_abort, which takes
10415 * care of disposing of the allocated buffers correctly.
10416 */
10417 static zio_t *
l2arc_log_blk_fetch(vdev_t * vd,const l2arc_log_blkptr_t * lbp,l2arc_log_blk_phys_t * lb)10418 l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp,
10419 l2arc_log_blk_phys_t *lb)
10420 {
10421 uint32_t asize;
10422 zio_t *pio;
10423 l2arc_read_callback_t *cb;
10424
10425 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10426 asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10427 ASSERT(asize <= sizeof (l2arc_log_blk_phys_t));
10428
10429 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP);
10430 cb->l2rcb_abd = abd_get_from_buf(lb, asize);
10431 pio = zio_root(vd->vdev_spa, l2arc_blk_fetch_done, cb,
10432 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY);
10433 (void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, asize,
10434 cb->l2rcb_abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10435 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL |
10436 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE));
10437
10438 return (pio);
10439 }
10440
10441 /*
10442 * Aborts a zio returned from l2arc_log_blk_fetch and frees the data
10443 * buffers allocated for it.
10444 */
10445 static void
l2arc_log_blk_fetch_abort(zio_t * zio)10446 l2arc_log_blk_fetch_abort(zio_t *zio)
10447 {
10448 (void) zio_wait(zio);
10449 }
10450
10451 /*
10452 * Creates a zio to update the device header on an l2arc device.
10453 */
10454 void
l2arc_dev_hdr_update(l2arc_dev_t * dev)10455 l2arc_dev_hdr_update(l2arc_dev_t *dev)
10456 {
10457 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10458 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
10459 abd_t *abd;
10460 int err;
10461
10462 VERIFY(spa_config_held(dev->l2ad_spa, SCL_STATE_ALL, RW_READER));
10463
10464 l2dhdr->dh_magic = L2ARC_DEV_HDR_MAGIC;
10465 l2dhdr->dh_version = L2ARC_PERSISTENT_VERSION;
10466 l2dhdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa);
10467 l2dhdr->dh_vdev_guid = dev->l2ad_vdev->vdev_guid;
10468 l2dhdr->dh_log_entries = dev->l2ad_log_entries;
10469 l2dhdr->dh_evict = dev->l2ad_evict;
10470 l2dhdr->dh_start = dev->l2ad_start;
10471 l2dhdr->dh_end = dev->l2ad_end;
10472 l2dhdr->dh_lb_asize = zfs_refcount_count(&dev->l2ad_lb_asize);
10473 l2dhdr->dh_lb_count = zfs_refcount_count(&dev->l2ad_lb_count);
10474 l2dhdr->dh_flags = 0;
10475 l2dhdr->dh_trim_action_time = dev->l2ad_vdev->vdev_trim_action_time;
10476 l2dhdr->dh_trim_state = dev->l2ad_vdev->vdev_trim_state;
10477 if (dev->l2ad_first)
10478 l2dhdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST;
10479
10480 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
10481
10482 err = zio_wait(zio_write_phys(NULL, dev->l2ad_vdev,
10483 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL,
10484 NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE));
10485
10486 abd_free(abd);
10487
10488 if (err != 0) {
10489 zfs_dbgmsg("L2ARC IO error (%d) while writing device header, "
10490 "vdev guid: %llu", err,
10491 (u_longlong_t)dev->l2ad_vdev->vdev_guid);
10492 }
10493 }
10494
10495 /*
10496 * Commits a log block to the L2ARC device. This routine is invoked from
10497 * l2arc_write_buffers when the log block fills up.
10498 * This function allocates some memory to temporarily hold the serialized
10499 * buffer to be written. This is then released in l2arc_write_done.
10500 */
10501 static uint64_t
l2arc_log_blk_commit(l2arc_dev_t * dev,zio_t * pio,l2arc_write_callback_t * cb)10502 l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb)
10503 {
10504 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10505 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
10506 uint64_t psize, asize;
10507 zio_t *wzio;
10508 l2arc_lb_abd_buf_t *abd_buf;
10509 abd_t *abd = NULL;
10510 l2arc_lb_ptr_buf_t *lb_ptr_buf;
10511
10512 VERIFY3S(dev->l2ad_log_ent_idx, ==, dev->l2ad_log_entries);
10513
10514 abd_buf = zio_buf_alloc(sizeof (*abd_buf));
10515 abd_buf->abd = abd_get_from_buf(lb, sizeof (*lb));
10516 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
10517 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP);
10518
10519 /* link the buffer into the block chain */
10520 lb->lb_prev_lbp = l2dhdr->dh_start_lbps[1];
10521 lb->lb_magic = L2ARC_LOG_BLK_MAGIC;
10522
10523 /*
10524 * l2arc_log_blk_commit() may be called multiple times during a single
10525 * l2arc_write_buffers() call. Save the allocated abd buffers in a list
10526 * so we can free them in l2arc_write_done() later on.
10527 */
10528 list_insert_tail(&cb->l2wcb_abd_list, abd_buf);
10529
10530 /* try to compress the buffer, at least one sector to save */
10531 psize = zio_compress_data(ZIO_COMPRESS_LZ4,
10532 abd_buf->abd, &abd, sizeof (*lb),
10533 zio_get_compression_max_size(ZIO_COMPRESS_LZ4,
10534 dev->l2ad_vdev->vdev_ashift,
10535 dev->l2ad_vdev->vdev_ashift, sizeof (*lb)), 0);
10536
10537 /* a log block is never entirely zero */
10538 ASSERT(psize != 0);
10539 asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
10540 ASSERT(asize <= sizeof (*lb));
10541
10542 /*
10543 * Update the start log block pointer in the device header to point
10544 * to the log block we're about to write.
10545 */
10546 l2dhdr->dh_start_lbps[1] = l2dhdr->dh_start_lbps[0];
10547 l2dhdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand;
10548 l2dhdr->dh_start_lbps[0].lbp_payload_asize =
10549 dev->l2ad_log_blk_payload_asize;
10550 l2dhdr->dh_start_lbps[0].lbp_payload_start =
10551 dev->l2ad_log_blk_payload_start;
10552 L2BLK_SET_LSIZE(
10553 (&l2dhdr->dh_start_lbps[0])->lbp_prop, sizeof (*lb));
10554 L2BLK_SET_PSIZE(
10555 (&l2dhdr->dh_start_lbps[0])->lbp_prop, asize);
10556 L2BLK_SET_CHECKSUM(
10557 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10558 ZIO_CHECKSUM_FLETCHER_4);
10559 if (asize < sizeof (*lb)) {
10560 /* compression succeeded */
10561 abd_zero_off(abd, psize, asize - psize);
10562 L2BLK_SET_COMPRESS(
10563 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10564 ZIO_COMPRESS_LZ4);
10565 } else {
10566 /* compression failed */
10567 abd_copy_from_buf_off(abd, lb, 0, sizeof (*lb));
10568 L2BLK_SET_COMPRESS(
10569 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
10570 ZIO_COMPRESS_OFF);
10571 }
10572
10573 /* checksum what we're about to write */
10574 abd_fletcher_4_native(abd, asize, NULL,
10575 &l2dhdr->dh_start_lbps[0].lbp_cksum);
10576
10577 abd_free(abd_buf->abd);
10578
10579 /* perform the write itself */
10580 abd_buf->abd = abd;
10581 wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand,
10582 asize, abd_buf->abd, ZIO_CHECKSUM_OFF, NULL, NULL,
10583 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE);
10584 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio);
10585 (void) zio_nowait(wzio);
10586
10587 dev->l2ad_hand += asize;
10588 /*
10589 * Include the committed log block's pointer in the list of pointers
10590 * to log blocks present in the L2ARC device.
10591 */
10592 memcpy(lb_ptr_buf->lb_ptr, &l2dhdr->dh_start_lbps[0],
10593 sizeof (l2arc_log_blkptr_t));
10594 mutex_enter(&dev->l2ad_mtx);
10595 list_insert_head(&dev->l2ad_lbptr_list, lb_ptr_buf);
10596 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
10597 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
10598 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
10599 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
10600 mutex_exit(&dev->l2ad_mtx);
10601 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10602
10603 /* bump the kstats */
10604 ARCSTAT_INCR(arcstat_l2_write_bytes, asize);
10605 ARCSTAT_BUMP(arcstat_l2_log_blk_writes);
10606 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, asize);
10607 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio,
10608 dev->l2ad_log_blk_payload_asize / asize);
10609
10610 /* start a new log block */
10611 dev->l2ad_log_ent_idx = 0;
10612 dev->l2ad_log_blk_payload_asize = 0;
10613 dev->l2ad_log_blk_payload_start = 0;
10614
10615 return (asize);
10616 }
10617
10618 /*
10619 * Validates an L2ARC log block address to make sure that it can be read
10620 * from the provided L2ARC device.
10621 */
10622 boolean_t
l2arc_log_blkptr_valid(l2arc_dev_t * dev,const l2arc_log_blkptr_t * lbp)10623 l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp)
10624 {
10625 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10626 uint64_t asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10627 uint64_t end = lbp->lbp_daddr + asize - 1;
10628 uint64_t start = lbp->lbp_payload_start;
10629 boolean_t evicted = B_FALSE;
10630
10631 /*
10632 * A log block is valid if all of the following conditions are true:
10633 * - it fits entirely (including its payload) between l2ad_start and
10634 * l2ad_end
10635 * - it has a valid size
10636 * - neither the log block itself nor part of its payload was evicted
10637 * by l2arc_evict():
10638 *
10639 * l2ad_hand l2ad_evict
10640 * | | lbp_daddr
10641 * | start | | end
10642 * | | | | |
10643 * V V V V V
10644 * l2ad_start ============================================ l2ad_end
10645 * --------------------------||||
10646 * ^ ^
10647 * | log block
10648 * payload
10649 */
10650
10651 evicted =
10652 l2arc_range_check_overlap(start, end, dev->l2ad_hand) ||
10653 l2arc_range_check_overlap(start, end, dev->l2ad_evict) ||
10654 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, start) ||
10655 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, end);
10656
10657 return (start >= dev->l2ad_start && end <= dev->l2ad_end &&
10658 asize > 0 && asize <= sizeof (l2arc_log_blk_phys_t) &&
10659 (!evicted || dev->l2ad_first));
10660 }
10661
10662 /*
10663 * Inserts ARC buffer header `hdr' into the current L2ARC log block on
10664 * the device. The buffer being inserted must be present in L2ARC.
10665 * Returns B_TRUE if the L2ARC log block is full and needs to be committed
10666 * to L2ARC, or B_FALSE if it still has room for more ARC buffers.
10667 */
10668 static boolean_t
l2arc_log_blk_insert(l2arc_dev_t * dev,const arc_buf_hdr_t * hdr)10669 l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr)
10670 {
10671 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10672 l2arc_log_ent_phys_t *le;
10673
10674 if (dev->l2ad_log_entries == 0)
10675 return (B_FALSE);
10676
10677 int index = dev->l2ad_log_ent_idx++;
10678
10679 ASSERT3S(index, <, dev->l2ad_log_entries);
10680 ASSERT(HDR_HAS_L2HDR(hdr));
10681
10682 le = &lb->lb_entries[index];
10683 memset(le, 0, sizeof (*le));
10684 le->le_dva = hdr->b_dva;
10685 le->le_birth = hdr->b_birth;
10686 le->le_daddr = hdr->b_l2hdr.b_daddr;
10687 if (index == 0)
10688 dev->l2ad_log_blk_payload_start = le->le_daddr;
10689 L2BLK_SET_LSIZE((le)->le_prop, HDR_GET_LSIZE(hdr));
10690 L2BLK_SET_PSIZE((le)->le_prop, HDR_GET_PSIZE(hdr));
10691 L2BLK_SET_COMPRESS((le)->le_prop, HDR_GET_COMPRESS(hdr));
10692 le->le_complevel = hdr->b_complevel;
10693 L2BLK_SET_TYPE((le)->le_prop, hdr->b_type);
10694 L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr)));
10695 L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr)));
10696 L2BLK_SET_STATE((le)->le_prop, hdr->b_l2hdr.b_arcs_state);
10697
10698 dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev,
10699 HDR_GET_PSIZE(hdr));
10700
10701 return (dev->l2ad_log_ent_idx == dev->l2ad_log_entries);
10702 }
10703
10704 /*
10705 * Checks whether a given L2ARC device address sits in a time-sequential
10706 * range. The trick here is that the L2ARC is a rotary buffer, so we can't
10707 * just do a range comparison, we need to handle the situation in which the
10708 * range wraps around the end of the L2ARC device. Arguments:
10709 * bottom -- Lower end of the range to check (written to earlier).
10710 * top -- Upper end of the range to check (written to later).
10711 * check -- The address for which we want to determine if it sits in
10712 * between the top and bottom.
10713 *
10714 * The 3-way conditional below represents the following cases:
10715 *
10716 * bottom < top : Sequentially ordered case:
10717 * <check>--------+-------------------+
10718 * | (overlap here?) |
10719 * L2ARC dev V V
10720 * |---------------<bottom>============<top>--------------|
10721 *
10722 * bottom > top: Looped-around case:
10723 * <check>--------+------------------+
10724 * | (overlap here?) |
10725 * L2ARC dev V V
10726 * |===============<top>---------------<bottom>===========|
10727 * ^ ^
10728 * | (or here?) |
10729 * +---------------+---------<check>
10730 *
10731 * top == bottom : Just a single address comparison.
10732 */
10733 boolean_t
l2arc_range_check_overlap(uint64_t bottom,uint64_t top,uint64_t check)10734 l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check)
10735 {
10736 if (bottom < top)
10737 return (bottom <= check && check <= top);
10738 else if (bottom > top)
10739 return (check <= top || bottom <= check);
10740 else
10741 return (check == top);
10742 }
10743
10744 EXPORT_SYMBOL(arc_buf_size);
10745 EXPORT_SYMBOL(arc_write);
10746 EXPORT_SYMBOL(arc_read);
10747 EXPORT_SYMBOL(arc_buf_info);
10748 EXPORT_SYMBOL(arc_getbuf_func);
10749 EXPORT_SYMBOL(arc_add_prune_callback);
10750 EXPORT_SYMBOL(arc_remove_prune_callback);
10751
10752 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min, param_set_arc_min,
10753 spl_param_get_u64, ZMOD_RW, "Minimum ARC size in bytes");
10754
10755 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, max, param_set_arc_max,
10756 spl_param_get_u64, ZMOD_RW, "Maximum ARC size in bytes");
10757
10758 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_balance, UINT, ZMOD_RW,
10759 "Balance between metadata and data on ghost hits.");
10760
10761 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, grow_retry, param_set_arc_int,
10762 param_get_uint, ZMOD_RW, "Seconds before growing ARC size");
10763
10764 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, shrink_shift, param_set_arc_int,
10765 param_get_uint, ZMOD_RW, "log2(fraction of ARC to reclaim)");
10766
10767 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, pc_percent, UINT, ZMOD_RW,
10768 "Percent of pagecache to reclaim ARC to");
10769
10770 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, average_blocksize, UINT, ZMOD_RD,
10771 "Target average block size");
10772
10773 ZFS_MODULE_PARAM(zfs, zfs_, compressed_arc_enabled, INT, ZMOD_RW,
10774 "Disable compressed ARC buffers");
10775
10776 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prefetch_ms, param_set_arc_int,
10777 param_get_uint, ZMOD_RW, "Min life of prefetch block in ms");
10778
10779 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prescient_prefetch_ms,
10780 param_set_arc_int, param_get_uint, ZMOD_RW,
10781 "Min life of prescient prefetched block in ms");
10782
10783 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_max, U64, ZMOD_RW,
10784 "Max write bytes per interval");
10785
10786 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_boost, U64, ZMOD_RW,
10787 "Extra write bytes during device warmup");
10788
10789 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom, U64, ZMOD_RW,
10790 "Number of max device writes to precache");
10791
10792 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom_boost, U64, ZMOD_RW,
10793 "Compressed l2arc_headroom multiplier");
10794
10795 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, trim_ahead, U64, ZMOD_RW,
10796 "TRIM ahead L2ARC write size multiplier");
10797
10798 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_secs, U64, ZMOD_RW,
10799 "Seconds between L2ARC writing");
10800
10801 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_min_ms, U64, ZMOD_RW,
10802 "Min feed interval in milliseconds");
10803
10804 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, noprefetch, INT, ZMOD_RW,
10805 "Skip caching prefetched buffers");
10806
10807 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_again, INT, ZMOD_RW,
10808 "Turbo L2ARC warmup");
10809
10810 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, norw, INT, ZMOD_RW,
10811 "No reads during writes");
10812
10813 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, meta_percent, UINT, ZMOD_RW,
10814 "Percent of ARC size allowed for L2ARC-only headers");
10815
10816 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_enabled, INT, ZMOD_RW,
10817 "Rebuild the L2ARC when importing a pool");
10818
10819 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_blocks_min_l2size, U64, ZMOD_RW,
10820 "Min size in bytes to write rebuild log blocks in L2ARC");
10821
10822 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, mfuonly, INT, ZMOD_RW,
10823 "Cache only MFU data from ARC into L2ARC");
10824
10825 ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, exclude_special, INT, ZMOD_RW,
10826 "Exclude dbufs on special vdevs from being cached to L2ARC if set.");
10827
10828 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, lotsfree_percent, param_set_arc_int,
10829 param_get_uint, ZMOD_RW, "System free memory I/O throttle in bytes");
10830
10831 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, sys_free, param_set_arc_u64,
10832 spl_param_get_u64, ZMOD_RW, "System free memory target size in bytes");
10833
10834 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit, param_set_arc_u64,
10835 spl_param_get_u64, ZMOD_RW, "Minimum bytes of dnodes in ARC");
10836
10837 ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit_percent,
10838 param_set_arc_int, param_get_uint, ZMOD_RW,
10839 "Percent of ARC meta buffers for dnodes");
10840
10841 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, dnode_reduce_percent, UINT, ZMOD_RW,
10842 "Percentage of excess dnodes to try to unpin");
10843
10844 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, eviction_pct, UINT, ZMOD_RW,
10845 "When full, ARC allocation waits for eviction of this % of alloc size");
10846
10847 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, evict_batch_limit, UINT, ZMOD_RW,
10848 "The number of headers to evict per sublist before moving to the next");
10849
10850 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, prune_task_threads, INT, ZMOD_RW,
10851 "Number of arc_prune threads");
10852