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