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