1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2020 Joyent, Inc.
24 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2017 Nexenta Systems, Inc. All rights reserved.
27 * Copyright (c) 2011, 2019, Delphix. All rights reserved.
28 * Copyright (c) 2020, George Amanakis. All rights reserved.
29 * Copyright (c) 2020, The FreeBSD Foundation [1]
30 *
31 * [1] Portions of this software were developed by Allan Jude
32 * under sponsorship from the FreeBSD Foundation.
33 */
34
35 /*
36 * DVA-based Adjustable Replacement Cache
37 *
38 * While much of the theory of operation used here is
39 * based on the self-tuning, low overhead replacement cache
40 * presented by Megiddo and Modha at FAST 2003, there are some
41 * significant differences:
42 *
43 * 1. The Megiddo and Modha model assumes any page is evictable.
44 * Pages in its cache cannot be "locked" into memory. This makes
45 * the eviction algorithm simple: evict the last page in the list.
46 * This also make the performance characteristics easy to reason
47 * about. Our cache is not so simple. At any given moment, some
48 * subset of the blocks in the cache are un-evictable because we
49 * have handed out a reference to them. Blocks are only evictable
50 * when there are no external references active. This makes
51 * eviction far more problematic: we choose to evict the evictable
52 * blocks that are the "lowest" in the list.
53 *
54 * There are times when it is not possible to evict the requested
55 * space. In these circumstances we are unable to adjust the cache
56 * size. To prevent the cache growing unbounded at these times we
57 * implement a "cache throttle" that slows the flow of new data
58 * into the cache until we can make space available.
59 *
60 * 2. The Megiddo and Modha model assumes a fixed cache size.
61 * Pages are evicted when the cache is full and there is a cache
62 * miss. Our model has a variable sized cache. It grows with
63 * high use, but also tries to react to memory pressure from the
64 * operating system: decreasing its size when system memory is
65 * tight.
66 *
67 * 3. The Megiddo and Modha model assumes a fixed page size. All
68 * elements of the cache are therefore exactly the same size. So
69 * when adjusting the cache size following a cache miss, its simply
70 * a matter of choosing a single page to evict. In our model, we
71 * have variable sized cache blocks (rangeing from 512 bytes to
72 * 128K bytes). We therefore choose a set of blocks to evict to make
73 * space for a cache miss that approximates as closely as possible
74 * the space used by the new block.
75 *
76 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
77 * by N. Megiddo & D. Modha, FAST 2003
78 */
79
80 /*
81 * The locking model:
82 *
83 * A new reference to a cache buffer can be obtained in two
84 * ways: 1) via a hash table lookup using the DVA as a key,
85 * or 2) via one of the ARC lists. The arc_read() interface
86 * uses method 1, while the internal ARC algorithms for
87 * adjusting the cache use method 2. We therefore provide two
88 * types of locks: 1) the hash table lock array, and 2) the
89 * ARC list locks.
90 *
91 * Buffers do not have their own mutexes, rather they rely on the
92 * hash table mutexes for the bulk of their protection (i.e. most
93 * fields in the arc_buf_hdr_t are protected by these mutexes).
94 *
95 * buf_hash_find() returns the appropriate mutex (held) when it
96 * locates the requested buffer in the hash table. It returns
97 * NULL for the mutex if the buffer was not in the table.
98 *
99 * buf_hash_remove() expects the appropriate hash mutex to be
100 * already held before it is invoked.
101 *
102 * Each ARC state also has a mutex which is used to protect the
103 * buffer list associated with the state. When attempting to
104 * obtain a hash table lock while holding an ARC list lock you
105 * must use: mutex_tryenter() to avoid deadlock. Also note that
106 * the active state mutex must be held before the ghost state mutex.
107 *
108 * Note that the majority of the performance stats are manipulated
109 * with atomic operations.
110 *
111 * The L2ARC uses the l2ad_mtx on each vdev for the following:
112 *
113 * - L2ARC buflist creation
114 * - L2ARC buflist eviction
115 * - L2ARC write completion, which walks L2ARC buflists
116 * - ARC header destruction, as it removes from L2ARC buflists
117 * - ARC header release, as it removes from L2ARC buflists
118 */
119
120 /*
121 * ARC operation:
122 *
123 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
124 * This structure can point either to a block that is still in the cache or to
125 * one that is only accessible in an L2 ARC device, or it can provide
126 * information about a block that was recently evicted. If a block is
127 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
128 * information to retrieve it from the L2ARC device. This information is
129 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
130 * that is in this state cannot access the data directly.
131 *
132 * Blocks that are actively being referenced or have not been evicted
133 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
134 * the arc_buf_hdr_t that will point to the data block in memory. A block can
135 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
136 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
137 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
138 *
139 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
140 * ability to store the physical data (b_pabd) associated with the DVA of the
141 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
142 * it will match its on-disk compression characteristics. This behavior can be
143 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
144 * compressed ARC functionality is disabled, the b_pabd will point to an
145 * uncompressed version of the on-disk data.
146 *
147 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
148 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
149 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
150 * consumer. The ARC will provide references to this data and will keep it
151 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
152 * data block and will evict any arc_buf_t that is no longer referenced. The
153 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
154 * "overhead_size" kstat.
155 *
156 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
157 * compressed form. The typical case is that consumers will want uncompressed
158 * data, and when that happens a new data buffer is allocated where the data is
159 * decompressed for them to use. Currently the only consumer who wants
160 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
161 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
162 * with the arc_buf_hdr_t.
163 *
164 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
165 * first one is owned by a compressed send consumer (and therefore references
166 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
167 * used by any other consumer (and has its own uncompressed copy of the data
168 * buffer).
169 *
170 * arc_buf_hdr_t
171 * +-----------+
172 * | fields |
173 * | common to |
174 * | L1- and |
175 * | L2ARC |
176 * +-----------+
177 * | l2arc_buf_hdr_t
178 * | |
179 * +-----------+
180 * | l1arc_buf_hdr_t
181 * | | arc_buf_t
182 * | b_buf +------------>+-----------+ arc_buf_t
183 * | b_pabd +-+ |b_next +---->+-----------+
184 * +-----------+ | |-----------| |b_next +-->NULL
185 * | |b_comp = T | +-----------+
186 * | |b_data +-+ |b_comp = F |
187 * | +-----------+ | |b_data +-+
188 * +->+------+ | +-----------+ |
189 * compressed | | | |
190 * data | |<--------------+ | uncompressed
191 * +------+ compressed, | data
192 * shared +-->+------+
193 * data | |
194 * | |
195 * +------+
196 *
197 * When a consumer reads a block, the ARC must first look to see if the
198 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
199 * arc_buf_t and either copies uncompressed data into a new data buffer from an
200 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
201 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
202 * hdr is compressed and the desired compression characteristics of the
203 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
204 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
205 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
206 * be anywhere in the hdr's list.
207 *
208 * The diagram below shows an example of an uncompressed ARC hdr that is
209 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
210 * the last element in the buf list):
211 *
212 * arc_buf_hdr_t
213 * +-----------+
214 * | |
215 * | |
216 * | |
217 * +-----------+
218 * l2arc_buf_hdr_t| |
219 * | |
220 * +-----------+
221 * l1arc_buf_hdr_t| |
222 * | | arc_buf_t (shared)
223 * | b_buf +------------>+---------+ arc_buf_t
224 * | | |b_next +---->+---------+
225 * | b_pabd +-+ |---------| |b_next +-->NULL
226 * +-----------+ | | | +---------+
227 * | |b_data +-+ | |
228 * | +---------+ | |b_data +-+
229 * +->+------+ | +---------+ |
230 * | | | |
231 * uncompressed | | | |
232 * data +------+ | |
233 * ^ +->+------+ |
234 * | uncompressed | | |
235 * | data | | |
236 * | +------+ |
237 * +---------------------------------+
238 *
239 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
240 * since the physical block is about to be rewritten. The new data contents
241 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
242 * it may compress the data before writing it to disk. The ARC will be called
243 * with the transformed data and will bcopy the transformed on-disk block into
244 * a newly allocated b_pabd. Writes are always done into buffers which have
245 * either been loaned (and hence are new and don't have other readers) or
246 * buffers which have been released (and hence have their own hdr, if there
247 * were originally other readers of the buf's original hdr). This ensures that
248 * the ARC only needs to update a single buf and its hdr after a write occurs.
249 *
250 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
251 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
252 * that when compressed ARC is enabled that the L2ARC blocks are identical
253 * to the on-disk block in the main data pool. This provides a significant
254 * advantage since the ARC can leverage the bp's checksum when reading from the
255 * L2ARC to determine if the contents are valid. However, if the compressed
256 * ARC is disabled, then the L2ARC's block must be transformed to look
257 * like the physical block in the main data pool before comparing the
258 * checksum and determining its validity.
259 *
260 * The L1ARC has a slightly different system for storing encrypted data.
261 * Raw (encrypted + possibly compressed) data has a few subtle differences from
262 * data that is just compressed. The biggest difference is that it is not
263 * possible to decrypt encrypted data (or visa versa) if the keys aren't loaded.
264 * The other difference is that encryption cannot be treated as a suggestion.
265 * If a caller would prefer compressed data, but they actually wind up with
266 * uncompressed data the worst thing that could happen is there might be a
267 * performance hit. If the caller requests encrypted data, however, we must be
268 * sure they actually get it or else secret information could be leaked. Raw
269 * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore,
270 * may have both an encrypted version and a decrypted version of its data at
271 * once. When a caller needs a raw arc_buf_t, it is allocated and the data is
272 * copied out of this header. To avoid complications with b_pabd, raw buffers
273 * cannot be shared.
274 */
275
276 #include <sys/spa.h>
277 #include <sys/zio.h>
278 #include <sys/spa_impl.h>
279 #include <sys/zio_compress.h>
280 #include <sys/zio_checksum.h>
281 #include <sys/zfs_context.h>
282 #include <sys/arc.h>
283 #include <sys/refcount.h>
284 #include <sys/vdev.h>
285 #include <sys/vdev_impl.h>
286 #include <sys/dsl_pool.h>
287 #include <sys/zfs_zone.h>
288 #include <sys/zio_checksum.h>
289 #include <sys/multilist.h>
290 #include <sys/abd.h>
291 #include <sys/zil.h>
292 #include <sys/fm/fs/zfs.h>
293 #ifdef _KERNEL
294 #include <sys/vmsystm.h>
295 #include <vm/anon.h>
296 #include <sys/fs/swapnode.h>
297 #include <sys/dnlc.h>
298 #endif
299 #include <sys/callb.h>
300 #include <sys/kstat.h>
301 #include <sys/zthr.h>
302 #include <zfs_fletcher.h>
303 #include <sys/arc_impl.h>
304 #include <sys/aggsum.h>
305 #include <sys/cityhash.h>
306 #include <sys/param.h>
307
308 #ifndef _KERNEL
309 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
310 boolean_t arc_watch = B_FALSE;
311 int arc_procfd;
312 #endif
313
314 /*
315 * This thread's job is to keep enough free memory in the system, by
316 * calling arc_kmem_reap_now() plus arc_shrink(), which improves
317 * arc_available_memory().
318 */
319 static zthr_t *arc_reap_zthr;
320
321 /*
322 * This thread's job is to keep arc_size under arc_c, by calling
323 * arc_adjust(), which improves arc_is_overflowing().
324 */
325 static zthr_t *arc_adjust_zthr;
326
327 static kmutex_t arc_adjust_lock;
328 static kcondvar_t arc_adjust_waiters_cv;
329 static boolean_t arc_adjust_needed = B_FALSE;
330
331 uint_t arc_reduce_dnlc_percent = 3;
332
333 /*
334 * The number of headers to evict in arc_evict_state_impl() before
335 * dropping the sublist lock and evicting from another sublist. A lower
336 * value means we're more likely to evict the "correct" header (i.e. the
337 * oldest header in the arc state), but comes with higher overhead
338 * (i.e. more invocations of arc_evict_state_impl()).
339 */
340 int zfs_arc_evict_batch_limit = 10;
341
342 /* number of seconds before growing cache again */
343 int arc_grow_retry = 60;
344
345 /*
346 * Minimum time between calls to arc_kmem_reap_soon(). Note that this will
347 * be converted to ticks, so with the default hz=100, a setting of 15 ms
348 * will actually wait 2 ticks, or 20ms.
349 */
350 int arc_kmem_cache_reap_retry_ms = 1000;
351
352 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
353 int zfs_arc_overflow_shift = 3;
354
355 /* shift of arc_c for calculating both min and max arc_p */
356 int arc_p_min_shift = 4;
357
358 /* log2(fraction of arc to reclaim) */
359 int arc_shrink_shift = 7;
360
361 /*
362 * log2(fraction of ARC which must be free to allow growing).
363 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
364 * when reading a new block into the ARC, we will evict an equal-sized block
365 * from the ARC.
366 *
367 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
368 * we will still not allow it to grow.
369 */
370 int arc_no_grow_shift = 5;
371
372
373 /*
374 * minimum lifespan of a prefetch block in clock ticks
375 * (initialized in arc_init())
376 */
377 static int zfs_arc_min_prefetch_ms = 1;
378 static int zfs_arc_min_prescient_prefetch_ms = 6;
379
380 /*
381 * If this percent of memory is free, don't throttle.
382 */
383 int arc_lotsfree_percent = 10;
384
385 static boolean_t arc_initialized;
386
387 /*
388 * The arc has filled available memory and has now warmed up.
389 */
390 static boolean_t arc_warm;
391
392 /*
393 * log2 fraction of the zio arena to keep free.
394 */
395 int arc_zio_arena_free_shift = 2;
396
397 /*
398 * These tunables are for performance analysis.
399 */
400 uint64_t zfs_arc_max;
401 uint64_t zfs_arc_min;
402 uint64_t zfs_arc_meta_limit = 0;
403 uint64_t zfs_arc_meta_min = 0;
404 int zfs_arc_grow_retry = 0;
405 int zfs_arc_shrink_shift = 0;
406 int zfs_arc_p_min_shift = 0;
407 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
408
409 /*
410 * ARC dirty data constraints for arc_tempreserve_space() throttle
411 */
412 uint_t zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */
413 uint_t zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */
414 uint_t zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */
415
416 boolean_t zfs_compressed_arc_enabled = B_TRUE;
417
418 /* The 6 states: */
419 static arc_state_t ARC_anon;
420 static arc_state_t ARC_mru;
421 static arc_state_t ARC_mru_ghost;
422 static arc_state_t ARC_mfu;
423 static arc_state_t ARC_mfu_ghost;
424 static arc_state_t ARC_l2c_only;
425
426 arc_stats_t arc_stats = {
427 { "hits", KSTAT_DATA_UINT64 },
428 { "misses", KSTAT_DATA_UINT64 },
429 { "demand_data_hits", KSTAT_DATA_UINT64 },
430 { "demand_data_misses", KSTAT_DATA_UINT64 },
431 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
432 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
433 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
434 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
435 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
436 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
437 { "mru_hits", KSTAT_DATA_UINT64 },
438 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
439 { "mfu_hits", KSTAT_DATA_UINT64 },
440 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
441 { "deleted", KSTAT_DATA_UINT64 },
442 { "mutex_miss", KSTAT_DATA_UINT64 },
443 { "access_skip", KSTAT_DATA_UINT64 },
444 { "evict_skip", KSTAT_DATA_UINT64 },
445 { "evict_not_enough", KSTAT_DATA_UINT64 },
446 { "evict_l2_cached", KSTAT_DATA_UINT64 },
447 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
448 { "evict_l2_eligible_mfu", KSTAT_DATA_UINT64 },
449 { "evict_l2_eligible_mru", KSTAT_DATA_UINT64 },
450 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
451 { "evict_l2_skip", KSTAT_DATA_UINT64 },
452 { "hash_elements", KSTAT_DATA_UINT64 },
453 { "hash_elements_max", KSTAT_DATA_UINT64 },
454 { "hash_collisions", KSTAT_DATA_UINT64 },
455 { "hash_chains", KSTAT_DATA_UINT64 },
456 { "hash_chain_max", KSTAT_DATA_UINT64 },
457 { "p", KSTAT_DATA_UINT64 },
458 { "c", KSTAT_DATA_UINT64 },
459 { "c_min", KSTAT_DATA_UINT64 },
460 { "c_max", KSTAT_DATA_UINT64 },
461 { "size", KSTAT_DATA_UINT64 },
462 { "compressed_size", KSTAT_DATA_UINT64 },
463 { "uncompressed_size", KSTAT_DATA_UINT64 },
464 { "overhead_size", KSTAT_DATA_UINT64 },
465 { "hdr_size", KSTAT_DATA_UINT64 },
466 { "data_size", KSTAT_DATA_UINT64 },
467 { "metadata_size", KSTAT_DATA_UINT64 },
468 { "other_size", KSTAT_DATA_UINT64 },
469 { "anon_size", KSTAT_DATA_UINT64 },
470 { "anon_evictable_data", KSTAT_DATA_UINT64 },
471 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
472 { "mru_size", KSTAT_DATA_UINT64 },
473 { "mru_evictable_data", KSTAT_DATA_UINT64 },
474 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
475 { "mru_ghost_size", KSTAT_DATA_UINT64 },
476 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
477 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
478 { "mfu_size", KSTAT_DATA_UINT64 },
479 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
480 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
481 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
482 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
483 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
484 { "l2_hits", KSTAT_DATA_UINT64 },
485 { "l2_misses", KSTAT_DATA_UINT64 },
486 { "l2_prefetch_asize", KSTAT_DATA_UINT64 },
487 { "l2_mru_asize", KSTAT_DATA_UINT64 },
488 { "l2_mfu_asize", KSTAT_DATA_UINT64 },
489 { "l2_bufc_data_asize", KSTAT_DATA_UINT64 },
490 { "l2_bufc_metadata_asize", KSTAT_DATA_UINT64 },
491 { "l2_feeds", KSTAT_DATA_UINT64 },
492 { "l2_rw_clash", KSTAT_DATA_UINT64 },
493 { "l2_read_bytes", KSTAT_DATA_UINT64 },
494 { "l2_write_bytes", KSTAT_DATA_UINT64 },
495 { "l2_writes_sent", KSTAT_DATA_UINT64 },
496 { "l2_writes_done", KSTAT_DATA_UINT64 },
497 { "l2_writes_error", KSTAT_DATA_UINT64 },
498 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
499 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
500 { "l2_evict_reading", KSTAT_DATA_UINT64 },
501 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
502 { "l2_free_on_write", KSTAT_DATA_UINT64 },
503 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
504 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
505 { "l2_io_error", KSTAT_DATA_UINT64 },
506 { "l2_size", KSTAT_DATA_UINT64 },
507 { "l2_asize", KSTAT_DATA_UINT64 },
508 { "l2_hdr_size", KSTAT_DATA_UINT64 },
509 { "l2_log_blk_writes", KSTAT_DATA_UINT64 },
510 { "l2_log_blk_avg_asize", KSTAT_DATA_UINT64 },
511 { "l2_log_blk_asize", KSTAT_DATA_UINT64 },
512 { "l2_log_blk_count", KSTAT_DATA_UINT64 },
513 { "l2_data_to_meta_ratio", KSTAT_DATA_UINT64 },
514 { "l2_rebuild_success", KSTAT_DATA_UINT64 },
515 { "l2_rebuild_unsupported", KSTAT_DATA_UINT64 },
516 { "l2_rebuild_io_errors", KSTAT_DATA_UINT64 },
517 { "l2_rebuild_dh_errors", KSTAT_DATA_UINT64 },
518 { "l2_rebuild_cksum_lb_errors", KSTAT_DATA_UINT64 },
519 { "l2_rebuild_lowmem", KSTAT_DATA_UINT64 },
520 { "l2_rebuild_size", KSTAT_DATA_UINT64 },
521 { "l2_rebuild_asize", KSTAT_DATA_UINT64 },
522 { "l2_rebuild_bufs", KSTAT_DATA_UINT64 },
523 { "l2_rebuild_bufs_precached", KSTAT_DATA_UINT64 },
524 { "l2_rebuild_log_blks", KSTAT_DATA_UINT64 },
525 { "memory_throttle_count", KSTAT_DATA_UINT64 },
526 { "arc_meta_used", KSTAT_DATA_UINT64 },
527 { "arc_meta_limit", KSTAT_DATA_UINT64 },
528 { "arc_meta_max", KSTAT_DATA_UINT64 },
529 { "arc_meta_min", KSTAT_DATA_UINT64 },
530 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
531 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
532 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
533 };
534
535 #define ARCSTAT_MAX(stat, val) { \
536 uint64_t m; \
537 while ((val) > (m = arc_stats.stat.value.ui64) && \
538 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
539 continue; \
540 }
541
542 #define ARCSTAT_MAXSTAT(stat) \
543 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
544
545 /*
546 * We define a macro to allow ARC hits/misses to be easily broken down by
547 * two separate conditions, giving a total of four different subtypes for
548 * each of hits and misses (so eight statistics total).
549 */
550 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
551 if (cond1) { \
552 if (cond2) { \
553 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
554 } else { \
555 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
556 } \
557 } else { \
558 if (cond2) { \
559 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
560 } else { \
561 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
562 } \
563 }
564
565 /*
566 * This macro allows us to use kstats as floating averages. Each time we
567 * update this kstat, we first factor it and the update value by
568 * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall
569 * average. This macro assumes that integer loads and stores are atomic, but
570 * is not safe for multiple writers updating the kstat in parallel (only the
571 * last writer's update will remain).
572 */
573 #define ARCSTAT_F_AVG_FACTOR 3
574 #define ARCSTAT_F_AVG(stat, value) \
575 do { \
576 uint64_t x = ARCSTAT(stat); \
577 x = x - x / ARCSTAT_F_AVG_FACTOR + \
578 (value) / ARCSTAT_F_AVG_FACTOR; \
579 ARCSTAT(stat) = x; \
580 _NOTE(CONSTCOND) \
581 } while (0)
582
583 kstat_t *arc_ksp;
584 static arc_state_t *arc_anon;
585 static arc_state_t *arc_mru;
586 static arc_state_t *arc_mru_ghost;
587 static arc_state_t *arc_mfu;
588 static arc_state_t *arc_mfu_ghost;
589 static arc_state_t *arc_l2c_only;
590
591 /*
592 * There are also some ARC variables that we want to export, but that are
593 * updated so often that having the canonical representation be the statistic
594 * variable causes a performance bottleneck. We want to use aggsum_t's for these
595 * instead, but still be able to export the kstat in the same way as before.
596 * The solution is to always use the aggsum version, except in the kstat update
597 * callback.
598 */
599 aggsum_t arc_size;
600 aggsum_t arc_meta_used;
601 aggsum_t astat_data_size;
602 aggsum_t astat_metadata_size;
603 aggsum_t astat_hdr_size;
604 aggsum_t astat_other_size;
605 aggsum_t astat_l2_hdr_size;
606
607 static int arc_no_grow; /* Don't try to grow cache size */
608 static hrtime_t arc_growtime;
609 static uint64_t arc_tempreserve;
610 static uint64_t arc_loaned_bytes;
611
612 #define GHOST_STATE(state) \
613 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
614 (state) == arc_l2c_only)
615
616 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
617 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
618 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
619 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
620 #define HDR_PRESCIENT_PREFETCH(hdr) \
621 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
622 #define HDR_COMPRESSION_ENABLED(hdr) \
623 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
624
625 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
626 #define HDR_L2_READING(hdr) \
627 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
628 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
629 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
630 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
631 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
632 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
633 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
634 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
635
636 #define HDR_ISTYPE_METADATA(hdr) \
637 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
638 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
639
640 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
641 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
642 #define HDR_HAS_RABD(hdr) \
643 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
644 (hdr)->b_crypt_hdr.b_rabd != NULL)
645 #define HDR_ENCRYPTED(hdr) \
646 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
647 #define HDR_AUTHENTICATED(hdr) \
648 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
649
650 /* For storing compression mode in b_flags */
651 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
652
653 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
654 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
655 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
656 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
657
658 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
659 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
660 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
661 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
662
663 /*
664 * Other sizes
665 */
666
667 #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
668 #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
669 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
670
671 /*
672 * Hash table routines
673 */
674
675 #define HT_LOCK_PAD 64
676
677 struct ht_lock {
678 kmutex_t ht_lock;
679 #ifdef _KERNEL
680 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
681 #endif
682 };
683
684 #define BUF_LOCKS 256
685 typedef struct buf_hash_table {
686 uint64_t ht_mask;
687 arc_buf_hdr_t **ht_table;
688 struct ht_lock ht_locks[BUF_LOCKS];
689 } buf_hash_table_t;
690
691 static buf_hash_table_t buf_hash_table;
692
693 #define BUF_HASH_INDEX(spa, dva, birth) \
694 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
695 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
696 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
697 #define HDR_LOCK(hdr) \
698 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
699
700 uint64_t zfs_crc64_table[256];
701
702 /*
703 * Level 2 ARC
704 */
705
706 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
707 #define L2ARC_HEADROOM 2 /* num of writes */
708 /*
709 * If we discover during ARC scan any buffers to be compressed, we boost
710 * our headroom for the next scanning cycle by this percentage multiple.
711 */
712 #define L2ARC_HEADROOM_BOOST 200
713 #define L2ARC_FEED_SECS 1 /* caching interval secs */
714 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
715
716 /*
717 * We can feed L2ARC from two states of ARC buffers, mru and mfu,
718 * and each of the state has two types: data and metadata.
719 */
720 #define L2ARC_FEED_TYPES 4
721
722
723 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
724 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
725
726 /* L2ARC Performance Tunables */
727 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
728 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
729 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
730 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
731 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
732 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
733 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
734 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
735 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
736 int l2arc_meta_percent = 33; /* limit on headers size */
737
738 /*
739 * L2ARC Internals
740 */
741 static list_t L2ARC_dev_list; /* device list */
742 static list_t *l2arc_dev_list; /* device list pointer */
743 static kmutex_t l2arc_dev_mtx; /* device list mutex */
744 static l2arc_dev_t *l2arc_dev_last; /* last device used */
745 static list_t L2ARC_free_on_write; /* free after write buf list */
746 static list_t *l2arc_free_on_write; /* free after write list ptr */
747 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
748 static uint64_t l2arc_ndev; /* number of devices */
749
750 typedef struct l2arc_read_callback {
751 arc_buf_hdr_t *l2rcb_hdr; /* read header */
752 blkptr_t l2rcb_bp; /* original blkptr */
753 zbookmark_phys_t l2rcb_zb; /* original bookmark */
754 int l2rcb_flags; /* original flags */
755 abd_t *l2rcb_abd; /* temporary buffer */
756 } l2arc_read_callback_t;
757
758 typedef struct l2arc_data_free {
759 /* protected by l2arc_free_on_write_mtx */
760 abd_t *l2df_abd;
761 size_t l2df_size;
762 arc_buf_contents_t l2df_type;
763 list_node_t l2df_list_node;
764 } l2arc_data_free_t;
765
766 static kmutex_t l2arc_feed_thr_lock;
767 static kcondvar_t l2arc_feed_thr_cv;
768 static uint8_t l2arc_thread_exit;
769
770 static kmutex_t l2arc_rebuild_thr_lock;
771 static kcondvar_t l2arc_rebuild_thr_cv;
772
773 enum arc_hdr_alloc_flags {
774 ARC_HDR_ALLOC_RDATA = 0x1,
775 ARC_HDR_DO_ADAPT = 0x2,
776 };
777
778
779 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *, boolean_t);
780 typedef enum arc_fill_flags {
781 ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */
782 ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */
783 ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */
784 ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */
785 ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */
786 } arc_fill_flags_t;
787
788 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
789 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *, boolean_t);
790 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
791 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
792 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
793 static void arc_hdr_free_pabd(arc_buf_hdr_t *, boolean_t);
794 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *, int);
795 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
796 static boolean_t arc_is_overflowing();
797 static void arc_buf_watch(arc_buf_t *);
798 static l2arc_dev_t *l2arc_vdev_get(vdev_t *vd);
799
800 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
801 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
802 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
803 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
804
805 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
806 static void l2arc_read_done(zio_t *);
807 static void l2arc_do_free_on_write(void);
808 static void l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
809 boolean_t state_only);
810
811 #define l2arc_hdr_arcstats_increment(hdr) \
812 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_FALSE)
813 #define l2arc_hdr_arcstats_decrement(hdr) \
814 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_FALSE)
815 #define l2arc_hdr_arcstats_increment_state(hdr) \
816 l2arc_hdr_arcstats_update((hdr), B_TRUE, B_TRUE)
817 #define l2arc_hdr_arcstats_decrement_state(hdr) \
818 l2arc_hdr_arcstats_update((hdr), B_FALSE, B_TRUE)
819
820 /*
821 * The arc_all_memory function is a ZoL enhancement that lives in their OSL
822 * code. In user-space code, which is used primarily for testing, we return
823 * half of all memory.
824 */
825 uint64_t
826 arc_all_memory(void)
827 {
828 #ifdef _KERNEL
829 return (ptob(physmem));
830 #else
831 return ((sysconf(_SC_PAGESIZE) * sysconf(_SC_PHYS_PAGES)) / 2);
832 #endif
833 }
834
835 /*
836 * We use Cityhash for this. It's fast, and has good hash properties without
837 * requiring any large static buffers.
838 */
839 static uint64_t
840 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
841 {
842 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
843 }
844
845 #define HDR_EMPTY(hdr) \
846 ((hdr)->b_dva.dva_word[0] == 0 && \
847 (hdr)->b_dva.dva_word[1] == 0)
848
849 #define HDR_EMPTY_OR_LOCKED(hdr) \
850 (HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr)))
851
852 #define HDR_EQUAL(spa, dva, birth, hdr) \
853 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
854 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
855 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
856
857 static void
858 buf_discard_identity(arc_buf_hdr_t *hdr)
859 {
860 hdr->b_dva.dva_word[0] = 0;
861 hdr->b_dva.dva_word[1] = 0;
862 hdr->b_birth = 0;
863 }
864
865 static arc_buf_hdr_t *
866 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
867 {
868 const dva_t *dva = BP_IDENTITY(bp);
869 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
870 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
871 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
872 arc_buf_hdr_t *hdr;
873
874 mutex_enter(hash_lock);
875 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
876 hdr = hdr->b_hash_next) {
877 if (HDR_EQUAL(spa, dva, birth, hdr)) {
878 *lockp = hash_lock;
879 return (hdr);
880 }
881 }
882 mutex_exit(hash_lock);
883 *lockp = NULL;
884 return (NULL);
885 }
886
887 /*
888 * Insert an entry into the hash table. If there is already an element
889 * equal to elem in the hash table, then the already existing element
890 * will be returned and the new element will not be inserted.
891 * Otherwise returns NULL.
892 * If lockp == NULL, the caller is assumed to already hold the hash lock.
893 */
894 static arc_buf_hdr_t *
895 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
896 {
897 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
898 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
899 arc_buf_hdr_t *fhdr;
900 uint32_t i;
901
902 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
903 ASSERT(hdr->b_birth != 0);
904 ASSERT(!HDR_IN_HASH_TABLE(hdr));
905
906 if (lockp != NULL) {
907 *lockp = hash_lock;
908 mutex_enter(hash_lock);
909 } else {
910 ASSERT(MUTEX_HELD(hash_lock));
911 }
912
913 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
914 fhdr = fhdr->b_hash_next, i++) {
915 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
916 return (fhdr);
917 }
918
919 hdr->b_hash_next = buf_hash_table.ht_table[idx];
920 buf_hash_table.ht_table[idx] = hdr;
921 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
922
923 /* collect some hash table performance data */
924 if (i > 0) {
925 ARCSTAT_BUMP(arcstat_hash_collisions);
926 if (i == 1)
927 ARCSTAT_BUMP(arcstat_hash_chains);
928
929 ARCSTAT_MAX(arcstat_hash_chain_max, i);
930 }
931
932 ARCSTAT_BUMP(arcstat_hash_elements);
933 ARCSTAT_MAXSTAT(arcstat_hash_elements);
934
935 return (NULL);
936 }
937
938 static void
939 buf_hash_remove(arc_buf_hdr_t *hdr)
940 {
941 arc_buf_hdr_t *fhdr, **hdrp;
942 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
943
944 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
945 ASSERT(HDR_IN_HASH_TABLE(hdr));
946
947 hdrp = &buf_hash_table.ht_table[idx];
948 while ((fhdr = *hdrp) != hdr) {
949 ASSERT3P(fhdr, !=, NULL);
950 hdrp = &fhdr->b_hash_next;
951 }
952 *hdrp = hdr->b_hash_next;
953 hdr->b_hash_next = NULL;
954 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
955
956 /* collect some hash table performance data */
957 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
958
959 if (buf_hash_table.ht_table[idx] &&
960 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
961 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
962 }
963
964 /*
965 * l2arc_mfuonly : A ZFS module parameter that controls whether only MFU
966 * metadata and data are cached from ARC into L2ARC.
967 */
968 int l2arc_mfuonly = 0;
969
970 /*
971 * Global data structures and functions for the buf kmem cache.
972 */
973
974 static kmem_cache_t *hdr_full_cache;
975 static kmem_cache_t *hdr_full_crypt_cache;
976 static kmem_cache_t *hdr_l2only_cache;
977 static kmem_cache_t *buf_cache;
978
979 static void
980 buf_fini(void)
981 {
982 int i;
983
984 kmem_free(buf_hash_table.ht_table,
985 (buf_hash_table.ht_mask + 1) * sizeof (void *));
986 for (i = 0; i < BUF_LOCKS; i++)
987 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
988 kmem_cache_destroy(hdr_full_cache);
989 kmem_cache_destroy(hdr_full_crypt_cache);
990 kmem_cache_destroy(hdr_l2only_cache);
991 kmem_cache_destroy(buf_cache);
992 }
993
994 /*
995 * Constructor callback - called when the cache is empty
996 * and a new buf is requested.
997 */
998 /* ARGSUSED */
999 static int
1000 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1001 {
1002 arc_buf_hdr_t *hdr = vbuf;
1003
1004 bzero(hdr, HDR_FULL_SIZE);
1005 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
1006 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1007 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
1008 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1009 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1010 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1011
1012 return (0);
1013 }
1014
1015 /* ARGSUSED */
1016 static int
1017 hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag)
1018 {
1019 arc_buf_hdr_t *hdr = vbuf;
1020
1021 (void) hdr_full_cons(vbuf, unused, kmflag);
1022 bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr));
1023 arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1024
1025 return (0);
1026 }
1027
1028 /* ARGSUSED */
1029 static int
1030 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1031 {
1032 arc_buf_hdr_t *hdr = vbuf;
1033
1034 bzero(hdr, HDR_L2ONLY_SIZE);
1035 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1036
1037 return (0);
1038 }
1039
1040 /* ARGSUSED */
1041 static int
1042 buf_cons(void *vbuf, void *unused, int kmflag)
1043 {
1044 arc_buf_t *buf = vbuf;
1045
1046 bzero(buf, sizeof (arc_buf_t));
1047 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1048 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1049
1050 return (0);
1051 }
1052
1053 /*
1054 * Destructor callback - called when a cached buf is
1055 * no longer required.
1056 */
1057 /* ARGSUSED */
1058 static void
1059 hdr_full_dest(void *vbuf, void *unused)
1060 {
1061 arc_buf_hdr_t *hdr = vbuf;
1062
1063 ASSERT(HDR_EMPTY(hdr));
1064 cv_destroy(&hdr->b_l1hdr.b_cv);
1065 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1066 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1067 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1068 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1069 }
1070
1071 /* ARGSUSED */
1072 static void
1073 hdr_full_crypt_dest(void *vbuf, void *unused)
1074 {
1075 arc_buf_hdr_t *hdr = vbuf;
1076
1077 hdr_full_dest(hdr, unused);
1078 arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1079 }
1080
1081 /* ARGSUSED */
1082 static void
1083 hdr_l2only_dest(void *vbuf, void *unused)
1084 {
1085 arc_buf_hdr_t *hdr = vbuf;
1086
1087 ASSERT(HDR_EMPTY(hdr));
1088 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1089 }
1090
1091 /* ARGSUSED */
1092 static void
1093 buf_dest(void *vbuf, void *unused)
1094 {
1095 arc_buf_t *buf = vbuf;
1096
1097 mutex_destroy(&buf->b_evict_lock);
1098 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1099 }
1100
1101 /*
1102 * Reclaim callback -- invoked when memory is low.
1103 */
1104 /* ARGSUSED */
1105 static void
1106 hdr_recl(void *unused)
1107 {
1108 dprintf("hdr_recl called\n");
1109 /*
1110 * umem calls the reclaim func when we destroy the buf cache,
1111 * which is after we do arc_fini().
1112 */
1113 if (arc_initialized)
1114 zthr_wakeup(arc_reap_zthr);
1115 }
1116
1117 static void
1118 buf_init(void)
1119 {
1120 uint64_t *ct;
1121 uint64_t hsize = 1ULL << 12;
1122 int i, j;
1123
1124 /*
1125 * The hash table is big enough to fill all of physical memory
1126 * with an average block size of zfs_arc_average_blocksize (default 8K).
1127 * By default, the table will take up
1128 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1129 */
1130 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1131 hsize <<= 1;
1132 retry:
1133 buf_hash_table.ht_mask = hsize - 1;
1134 buf_hash_table.ht_table =
1135 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1136 if (buf_hash_table.ht_table == NULL) {
1137 ASSERT(hsize > (1ULL << 8));
1138 hsize >>= 1;
1139 goto retry;
1140 }
1141
1142 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1143 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1144 hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt",
1145 HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest,
1146 hdr_recl, NULL, NULL, 0);
1147 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1148 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1149 NULL, NULL, 0);
1150 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1151 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1152
1153 for (i = 0; i < 256; i++)
1154 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1155 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1156
1157 for (i = 0; i < BUF_LOCKS; i++) {
1158 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1159 NULL, MUTEX_DEFAULT, NULL);
1160 }
1161 }
1162
1163 /*
1164 * This is the size that the buf occupies in memory. If the buf is compressed,
1165 * it will correspond to the compressed size. You should use this method of
1166 * getting the buf size unless you explicitly need the logical size.
1167 */
1168 int32_t
1169 arc_buf_size(arc_buf_t *buf)
1170 {
1171 return (ARC_BUF_COMPRESSED(buf) ?
1172 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1173 }
1174
1175 int32_t
1176 arc_buf_lsize(arc_buf_t *buf)
1177 {
1178 return (HDR_GET_LSIZE(buf->b_hdr));
1179 }
1180
1181 /*
1182 * This function will return B_TRUE if the buffer is encrypted in memory.
1183 * This buffer can be decrypted by calling arc_untransform().
1184 */
1185 boolean_t
1186 arc_is_encrypted(arc_buf_t *buf)
1187 {
1188 return (ARC_BUF_ENCRYPTED(buf) != 0);
1189 }
1190
1191 /*
1192 * Returns B_TRUE if the buffer represents data that has not had its MAC
1193 * verified yet.
1194 */
1195 boolean_t
1196 arc_is_unauthenticated(arc_buf_t *buf)
1197 {
1198 return (HDR_NOAUTH(buf->b_hdr) != 0);
1199 }
1200
1201 void
1202 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1203 uint8_t *iv, uint8_t *mac)
1204 {
1205 arc_buf_hdr_t *hdr = buf->b_hdr;
1206
1207 ASSERT(HDR_PROTECTED(hdr));
1208
1209 bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
1210 bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
1211 bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
1212 *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1213 /* CONSTCOND */
1214 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1215 }
1216
1217 /*
1218 * Indicates how this buffer is compressed in memory. If it is not compressed
1219 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1220 * arc_untransform() as long as it is also unencrypted.
1221 */
1222 enum zio_compress
1223 arc_get_compression(arc_buf_t *buf)
1224 {
1225 return (ARC_BUF_COMPRESSED(buf) ?
1226 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1227 }
1228
1229 #define ARC_MINTIME (hz>>4) /* 62 ms */
1230
1231 /*
1232 * Return the compression algorithm used to store this data in the ARC. If ARC
1233 * compression is enabled or this is an encrypted block, this will be the same
1234 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1235 */
1236 static inline enum zio_compress
1237 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1238 {
1239 return (HDR_COMPRESSION_ENABLED(hdr) ?
1240 HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1241 }
1242
1243 static inline boolean_t
1244 arc_buf_is_shared(arc_buf_t *buf)
1245 {
1246 boolean_t shared = (buf->b_data != NULL &&
1247 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1248 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1249 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1250 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1251 IMPLY(shared, ARC_BUF_SHARED(buf));
1252 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1253
1254 /*
1255 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1256 * already being shared" requirement prevents us from doing that.
1257 */
1258
1259 return (shared);
1260 }
1261
1262 /*
1263 * Free the checksum associated with this header. If there is no checksum, this
1264 * is a no-op.
1265 */
1266 static inline void
1267 arc_cksum_free(arc_buf_hdr_t *hdr)
1268 {
1269 ASSERT(HDR_HAS_L1HDR(hdr));
1270
1271 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1272 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1273 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1274 hdr->b_l1hdr.b_freeze_cksum = NULL;
1275 }
1276 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1277 }
1278
1279 /*
1280 * Return true iff at least one of the bufs on hdr is not compressed.
1281 * Encrypted buffers count as compressed.
1282 */
1283 static boolean_t
1284 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1285 {
1286 ASSERT(hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY_OR_LOCKED(hdr));
1287
1288 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1289 if (!ARC_BUF_COMPRESSED(b)) {
1290 return (B_TRUE);
1291 }
1292 }
1293 return (B_FALSE);
1294 }
1295
1296 /*
1297 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1298 * matches the checksum that is stored in the hdr. If there is no checksum,
1299 * or if the buf is compressed, this is a no-op.
1300 */
1301 static void
1302 arc_cksum_verify(arc_buf_t *buf)
1303 {
1304 arc_buf_hdr_t *hdr = buf->b_hdr;
1305 zio_cksum_t zc;
1306
1307 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1308 return;
1309
1310 if (ARC_BUF_COMPRESSED(buf))
1311 return;
1312
1313 ASSERT(HDR_HAS_L1HDR(hdr));
1314
1315 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1316
1317 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1318 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1319 return;
1320 }
1321
1322 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1323 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1324 panic("buffer modified while frozen!");
1325 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1326 }
1327
1328 /*
1329 * This function makes the assumption that data stored in the L2ARC
1330 * will be transformed exactly as it is in the main pool. Because of
1331 * this we can verify the checksum against the reading process's bp.
1332 */
1333 static boolean_t
1334 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1335 {
1336 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1337 boolean_t valid_cksum;
1338
1339 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1340 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1341
1342 /*
1343 * We rely on the blkptr's checksum to determine if the block
1344 * is valid or not. When compressed arc is enabled, the l2arc
1345 * writes the block to the l2arc just as it appears in the pool.
1346 * This allows us to use the blkptr's checksum to validate the
1347 * data that we just read off of the l2arc without having to store
1348 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1349 * arc is disabled, then the data written to the l2arc is always
1350 * uncompressed and won't match the block as it exists in the main
1351 * pool. When this is the case, we must first compress it if it is
1352 * compressed on the main pool before we can validate the checksum.
1353 */
1354 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1355 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1356 uint64_t lsize = HDR_GET_LSIZE(hdr);
1357 uint64_t csize;
1358
1359 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1360 csize = zio_compress_data(compress, zio->io_abd,
1361 abd_to_buf(cdata), lsize);
1362
1363 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1364 if (csize < HDR_GET_PSIZE(hdr)) {
1365 /*
1366 * Compressed blocks are always a multiple of the
1367 * smallest ashift in the pool. Ideally, we would
1368 * like to round up the csize to the next
1369 * spa_min_ashift but that value may have changed
1370 * since the block was last written. Instead,
1371 * we rely on the fact that the hdr's psize
1372 * was set to the psize of the block when it was
1373 * last written. We set the csize to that value
1374 * and zero out any part that should not contain
1375 * data.
1376 */
1377 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
1378 csize = HDR_GET_PSIZE(hdr);
1379 }
1380 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
1381 }
1382
1383 /*
1384 * Block pointers always store the checksum for the logical data.
1385 * If the block pointer has the gang bit set, then the checksum
1386 * it represents is for the reconstituted data and not for an
1387 * individual gang member. The zio pipeline, however, must be able to
1388 * determine the checksum of each of the gang constituents so it
1389 * treats the checksum comparison differently than what we need
1390 * for l2arc blocks. This prevents us from using the
1391 * zio_checksum_error() interface directly. Instead we must call the
1392 * zio_checksum_error_impl() so that we can ensure the checksum is
1393 * generated using the correct checksum algorithm and accounts for the
1394 * logical I/O size and not just a gang fragment.
1395 */
1396 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1397 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1398 zio->io_offset, NULL) == 0);
1399 zio_pop_transforms(zio);
1400 return (valid_cksum);
1401 }
1402
1403 /*
1404 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1405 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1406 * isn't modified later on. If buf is compressed or there is already a checksum
1407 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1408 */
1409 static void
1410 arc_cksum_compute(arc_buf_t *buf)
1411 {
1412 arc_buf_hdr_t *hdr = buf->b_hdr;
1413
1414 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1415 return;
1416
1417 ASSERT(HDR_HAS_L1HDR(hdr));
1418
1419 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1420 if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
1421 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1422 return;
1423 }
1424
1425 ASSERT(!ARC_BUF_ENCRYPTED(buf));
1426 ASSERT(!ARC_BUF_COMPRESSED(buf));
1427 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1428 KM_SLEEP);
1429 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1430 hdr->b_l1hdr.b_freeze_cksum);
1431 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1432 arc_buf_watch(buf);
1433 }
1434
1435 #ifndef _KERNEL
1436 typedef struct procctl {
1437 long cmd;
1438 prwatch_t prwatch;
1439 } procctl_t;
1440 #endif
1441
1442 /* ARGSUSED */
1443 static void
1444 arc_buf_unwatch(arc_buf_t *buf)
1445 {
1446 #ifndef _KERNEL
1447 if (arc_watch) {
1448 int result;
1449 procctl_t ctl;
1450 ctl.cmd = PCWATCH;
1451 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1452 ctl.prwatch.pr_size = 0;
1453 ctl.prwatch.pr_wflags = 0;
1454 result = write(arc_procfd, &ctl, sizeof (ctl));
1455 ASSERT3U(result, ==, sizeof (ctl));
1456 }
1457 #endif
1458 }
1459
1460 /* ARGSUSED */
1461 static void
1462 arc_buf_watch(arc_buf_t *buf)
1463 {
1464 #ifndef _KERNEL
1465 if (arc_watch) {
1466 int result;
1467 procctl_t ctl;
1468 ctl.cmd = PCWATCH;
1469 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1470 ctl.prwatch.pr_size = arc_buf_size(buf);
1471 ctl.prwatch.pr_wflags = WA_WRITE;
1472 result = write(arc_procfd, &ctl, sizeof (ctl));
1473 ASSERT3U(result, ==, sizeof (ctl));
1474 }
1475 #endif
1476 }
1477
1478 static arc_buf_contents_t
1479 arc_buf_type(arc_buf_hdr_t *hdr)
1480 {
1481 arc_buf_contents_t type;
1482 if (HDR_ISTYPE_METADATA(hdr)) {
1483 type = ARC_BUFC_METADATA;
1484 } else {
1485 type = ARC_BUFC_DATA;
1486 }
1487 VERIFY3U(hdr->b_type, ==, type);
1488 return (type);
1489 }
1490
1491 boolean_t
1492 arc_is_metadata(arc_buf_t *buf)
1493 {
1494 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1495 }
1496
1497 static uint32_t
1498 arc_bufc_to_flags(arc_buf_contents_t type)
1499 {
1500 switch (type) {
1501 case ARC_BUFC_DATA:
1502 /* metadata field is 0 if buffer contains normal data */
1503 return (0);
1504 case ARC_BUFC_METADATA:
1505 return (ARC_FLAG_BUFC_METADATA);
1506 default:
1507 break;
1508 }
1509 panic("undefined ARC buffer type!");
1510 return ((uint32_t)-1);
1511 }
1512
1513 void
1514 arc_buf_thaw(arc_buf_t *buf)
1515 {
1516 arc_buf_hdr_t *hdr = buf->b_hdr;
1517
1518 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1519 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1520
1521 arc_cksum_verify(buf);
1522
1523 /*
1524 * Compressed buffers do not manipulate the b_freeze_cksum.
1525 */
1526 if (ARC_BUF_COMPRESSED(buf))
1527 return;
1528
1529 ASSERT(HDR_HAS_L1HDR(hdr));
1530 arc_cksum_free(hdr);
1531
1532 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1533 #ifdef ZFS_DEBUG
1534 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1535 if (hdr->b_l1hdr.b_thawed != NULL)
1536 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1537 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1538 }
1539 #endif
1540
1541 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1542
1543 arc_buf_unwatch(buf);
1544 }
1545
1546 void
1547 arc_buf_freeze(arc_buf_t *buf)
1548 {
1549 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1550 return;
1551
1552 if (ARC_BUF_COMPRESSED(buf))
1553 return;
1554
1555 ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
1556 arc_cksum_compute(buf);
1557 }
1558
1559 /*
1560 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1561 * the following functions should be used to ensure that the flags are
1562 * updated in a thread-safe way. When manipulating the flags either
1563 * the hash_lock must be held or the hdr must be undiscoverable. This
1564 * ensures that we're not racing with any other threads when updating
1565 * the flags.
1566 */
1567 static inline void
1568 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1569 {
1570 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1571 hdr->b_flags |= flags;
1572 }
1573
1574 static inline void
1575 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1576 {
1577 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1578 hdr->b_flags &= ~flags;
1579 }
1580
1581 /*
1582 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1583 * done in a special way since we have to clear and set bits
1584 * at the same time. Consumers that wish to set the compression bits
1585 * must use this function to ensure that the flags are updated in
1586 * thread-safe manner.
1587 */
1588 static void
1589 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1590 {
1591 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1592
1593 /*
1594 * Holes and embedded blocks will always have a psize = 0 so
1595 * we ignore the compression of the blkptr and set the
1596 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
1597 * Holes and embedded blocks remain anonymous so we don't
1598 * want to uncompress them. Mark them as uncompressed.
1599 */
1600 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1601 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1602 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1603 } else {
1604 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1605 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1606 }
1607
1608 HDR_SET_COMPRESS(hdr, cmp);
1609 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1610 }
1611
1612 /*
1613 * Looks for another buf on the same hdr which has the data decompressed, copies
1614 * from it, and returns true. If no such buf exists, returns false.
1615 */
1616 static boolean_t
1617 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1618 {
1619 arc_buf_hdr_t *hdr = buf->b_hdr;
1620 boolean_t copied = B_FALSE;
1621
1622 ASSERT(HDR_HAS_L1HDR(hdr));
1623 ASSERT3P(buf->b_data, !=, NULL);
1624 ASSERT(!ARC_BUF_COMPRESSED(buf));
1625
1626 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1627 from = from->b_next) {
1628 /* can't use our own data buffer */
1629 if (from == buf) {
1630 continue;
1631 }
1632
1633 if (!ARC_BUF_COMPRESSED(from)) {
1634 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1635 copied = B_TRUE;
1636 break;
1637 }
1638 }
1639
1640 /*
1641 * Note: With encryption support, the following assertion is no longer
1642 * necessarily valid. If we receive two back to back raw snapshots
1643 * (send -w), the second receive can use a hdr with a cksum already
1644 * calculated. This happens via:
1645 * dmu_recv_stream() -> receive_read_record() -> arc_loan_raw_buf()
1646 * The rsend/send_mixed_raw test case exercises this code path.
1647 *
1648 * There were no decompressed bufs, so there should not be a
1649 * checksum on the hdr either.
1650 * EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1651 */
1652
1653 return (copied);
1654 }
1655
1656 /*
1657 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1658 */
1659 static uint64_t
1660 arc_hdr_size(arc_buf_hdr_t *hdr)
1661 {
1662 uint64_t size;
1663
1664 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1665 HDR_GET_PSIZE(hdr) > 0) {
1666 size = HDR_GET_PSIZE(hdr);
1667 } else {
1668 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1669 size = HDR_GET_LSIZE(hdr);
1670 }
1671 return (size);
1672 }
1673
1674 static int
1675 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1676 {
1677 int ret;
1678 uint64_t csize;
1679 uint64_t lsize = HDR_GET_LSIZE(hdr);
1680 uint64_t psize = HDR_GET_PSIZE(hdr);
1681 void *tmpbuf = NULL;
1682 abd_t *abd = hdr->b_l1hdr.b_pabd;
1683
1684 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1685 ASSERT(HDR_AUTHENTICATED(hdr));
1686 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1687
1688 /*
1689 * The MAC is calculated on the compressed data that is stored on disk.
1690 * However, if compressed arc is disabled we will only have the
1691 * decompressed data available to us now. Compress it into a temporary
1692 * abd so we can verify the MAC. The performance overhead of this will
1693 * be relatively low, since most objects in an encrypted objset will
1694 * be encrypted (instead of authenticated) anyway.
1695 */
1696 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1697 !HDR_COMPRESSION_ENABLED(hdr)) {
1698 tmpbuf = zio_buf_alloc(lsize);
1699 abd = abd_get_from_buf(tmpbuf, lsize);
1700 abd_take_ownership_of_buf(abd, B_TRUE);
1701
1702 csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1703 hdr->b_l1hdr.b_pabd, tmpbuf, lsize);
1704 ASSERT3U(csize, <=, psize);
1705 abd_zero_off(abd, csize, psize - csize);
1706 }
1707
1708 /*
1709 * Authentication is best effort. We authenticate whenever the key is
1710 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1711 */
1712 if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1713 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1714 ASSERT3U(lsize, ==, psize);
1715 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1716 psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1717 } else {
1718 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1719 hdr->b_crypt_hdr.b_mac);
1720 }
1721
1722 if (ret == 0)
1723 arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1724 else if (ret != ENOENT)
1725 goto error;
1726
1727 if (tmpbuf != NULL)
1728 abd_free(abd);
1729
1730 return (0);
1731
1732 error:
1733 if (tmpbuf != NULL)
1734 abd_free(abd);
1735
1736 return (ret);
1737 }
1738
1739 /*
1740 * This function will take a header that only has raw encrypted data in
1741 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1742 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1743 * also decompress the data.
1744 */
1745 static int
1746 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
1747 {
1748 int ret;
1749 abd_t *cabd = NULL;
1750 void *tmp = NULL;
1751 boolean_t no_crypt = B_FALSE;
1752 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1753
1754 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1755 ASSERT(HDR_ENCRYPTED(hdr));
1756
1757 arc_hdr_alloc_pabd(hdr, ARC_HDR_DO_ADAPT);
1758
1759 ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
1760 B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
1761 hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
1762 hdr->b_crypt_hdr.b_rabd, &no_crypt);
1763 if (ret != 0)
1764 goto error;
1765
1766 if (no_crypt) {
1767 abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1768 HDR_GET_PSIZE(hdr));
1769 }
1770
1771 /*
1772 * If this header has disabled arc compression but the b_pabd is
1773 * compressed after decrypting it, we need to decompress the newly
1774 * decrypted data.
1775 */
1776 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1777 !HDR_COMPRESSION_ENABLED(hdr)) {
1778 /*
1779 * We want to make sure that we are correctly honoring the
1780 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1781 * and then loan a buffer from it, rather than allocating a
1782 * linear buffer and wrapping it in an abd later.
1783 */
1784 cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr, B_TRUE);
1785 tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
1786
1787 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1788 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
1789 HDR_GET_LSIZE(hdr));
1790 if (ret != 0) {
1791 abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
1792 goto error;
1793 }
1794
1795 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
1796 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
1797 arc_hdr_size(hdr), hdr);
1798 hdr->b_l1hdr.b_pabd = cabd;
1799 }
1800
1801 return (0);
1802
1803 error:
1804 arc_hdr_free_pabd(hdr, B_FALSE);
1805 if (cabd != NULL)
1806 arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
1807
1808 return (ret);
1809 }
1810
1811 /*
1812 * This function is called during arc_buf_fill() to prepare the header's
1813 * abd plaintext pointer for use. This involves authenticated protected
1814 * data and decrypting encrypted data into the plaintext abd.
1815 */
1816 static int
1817 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
1818 const zbookmark_phys_t *zb, boolean_t noauth)
1819 {
1820 int ret;
1821
1822 ASSERT(HDR_PROTECTED(hdr));
1823
1824 if (hash_lock != NULL)
1825 mutex_enter(hash_lock);
1826
1827 if (HDR_NOAUTH(hdr) && !noauth) {
1828 /*
1829 * The caller requested authenticated data but our data has
1830 * not been authenticated yet. Verify the MAC now if we can.
1831 */
1832 ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
1833 if (ret != 0)
1834 goto error;
1835 } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
1836 /*
1837 * If we only have the encrypted version of the data, but the
1838 * unencrypted version was requested we take this opportunity
1839 * to store the decrypted version in the header for future use.
1840 */
1841 ret = arc_hdr_decrypt(hdr, spa, zb);
1842 if (ret != 0)
1843 goto error;
1844 }
1845
1846 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1847
1848 if (hash_lock != NULL)
1849 mutex_exit(hash_lock);
1850
1851 return (0);
1852
1853 error:
1854 if (hash_lock != NULL)
1855 mutex_exit(hash_lock);
1856
1857 return (ret);
1858 }
1859
1860 /*
1861 * This function is used by the dbuf code to decrypt bonus buffers in place.
1862 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
1863 * block, so we use the hash lock here to protect against concurrent calls to
1864 * arc_buf_fill().
1865 */
1866 /* ARGSUSED */
1867 static void
1868 arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock)
1869 {
1870 arc_buf_hdr_t *hdr = buf->b_hdr;
1871
1872 ASSERT(HDR_ENCRYPTED(hdr));
1873 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1874 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1875 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1876
1877 zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
1878 arc_buf_size(buf));
1879 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
1880 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1881 hdr->b_crypt_hdr.b_ebufcnt -= 1;
1882 }
1883
1884 /*
1885 * Given a buf that has a data buffer attached to it, this function will
1886 * efficiently fill the buf with data of the specified compression setting from
1887 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1888 * are already sharing a data buf, no copy is performed.
1889 *
1890 * If the buf is marked as compressed but uncompressed data was requested, this
1891 * will allocate a new data buffer for the buf, remove that flag, and fill the
1892 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1893 * uncompressed data, and (since we haven't added support for it yet) if you
1894 * want compressed data your buf must already be marked as compressed and have
1895 * the correct-sized data buffer.
1896 */
1897 static int
1898 arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
1899 arc_fill_flags_t flags)
1900 {
1901 int error = 0;
1902 arc_buf_hdr_t *hdr = buf->b_hdr;
1903 boolean_t hdr_compressed =
1904 (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
1905 boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
1906 boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
1907 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
1908 kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
1909
1910 ASSERT3P(buf->b_data, !=, NULL);
1911 IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
1912 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
1913 IMPLY(encrypted, HDR_ENCRYPTED(hdr));
1914 IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
1915 IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
1916 IMPLY(encrypted, !ARC_BUF_SHARED(buf));
1917
1918 /*
1919 * If the caller wanted encrypted data we just need to copy it from
1920 * b_rabd and potentially byteswap it. We won't be able to do any
1921 * further transforms on it.
1922 */
1923 if (encrypted) {
1924 ASSERT(HDR_HAS_RABD(hdr));
1925 abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
1926 HDR_GET_PSIZE(hdr));
1927 goto byteswap;
1928 }
1929
1930 /*
1931 * Adjust encrypted and authenticated headers to accomodate
1932 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
1933 * allowed to fail decryption due to keys not being loaded
1934 * without being marked as an IO error.
1935 */
1936 if (HDR_PROTECTED(hdr)) {
1937 error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
1938 zb, !!(flags & ARC_FILL_NOAUTH));
1939 if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
1940 return (error);
1941 } else if (error != 0) {
1942 if (hash_lock != NULL)
1943 mutex_enter(hash_lock);
1944 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
1945 if (hash_lock != NULL)
1946 mutex_exit(hash_lock);
1947 return (error);
1948 }
1949 }
1950
1951 /*
1952 * There is a special case here for dnode blocks which are
1953 * decrypting their bonus buffers. These blocks may request to
1954 * be decrypted in-place. This is necessary because there may
1955 * be many dnodes pointing into this buffer and there is
1956 * currently no method to synchronize replacing the backing
1957 * b_data buffer and updating all of the pointers. Here we use
1958 * the hash lock to ensure there are no races. If the need
1959 * arises for other types to be decrypted in-place, they must
1960 * add handling here as well.
1961 */
1962 if ((flags & ARC_FILL_IN_PLACE) != 0) {
1963 ASSERT(!hdr_compressed);
1964 ASSERT(!compressed);
1965 ASSERT(!encrypted);
1966
1967 if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
1968 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1969
1970 if (hash_lock != NULL)
1971 mutex_enter(hash_lock);
1972 arc_buf_untransform_in_place(buf, hash_lock);
1973 if (hash_lock != NULL)
1974 mutex_exit(hash_lock);
1975
1976 /* Compute the hdr's checksum if necessary */
1977 arc_cksum_compute(buf);
1978 }
1979
1980 return (0);
1981 }
1982
1983 if (hdr_compressed == compressed) {
1984 if (!arc_buf_is_shared(buf)) {
1985 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
1986 arc_buf_size(buf));
1987 }
1988 } else {
1989 ASSERT(hdr_compressed);
1990 ASSERT(!compressed);
1991 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
1992
1993 /*
1994 * If the buf is sharing its data with the hdr, unlink it and
1995 * allocate a new data buffer for the buf.
1996 */
1997 if (arc_buf_is_shared(buf)) {
1998 ASSERT(ARC_BUF_COMPRESSED(buf));
1999
2000 /* We need to give the buf its own b_data */
2001 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2002 buf->b_data =
2003 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2004 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2005
2006 /* Previously overhead was 0; just add new overhead */
2007 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
2008 } else if (ARC_BUF_COMPRESSED(buf)) {
2009 /* We need to reallocate the buf's b_data */
2010 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
2011 buf);
2012 buf->b_data =
2013 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2014
2015 /* We increased the size of b_data; update overhead */
2016 ARCSTAT_INCR(arcstat_overhead_size,
2017 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2018 }
2019
2020 /*
2021 * Regardless of the buf's previous compression settings, it
2022 * should not be compressed at the end of this function.
2023 */
2024 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2025
2026 /*
2027 * Try copying the data from another buf which already has a
2028 * decompressed version. If that's not possible, it's time to
2029 * bite the bullet and decompress the data from the hdr.
2030 */
2031 if (arc_buf_try_copy_decompressed_data(buf)) {
2032 /* Skip byteswapping and checksumming (already done) */
2033 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
2034 return (0);
2035 } else {
2036 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2037 hdr->b_l1hdr.b_pabd, buf->b_data,
2038 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2039
2040 /*
2041 * Absent hardware errors or software bugs, this should
2042 * be impossible, but log it anyway so we can debug it.
2043 */
2044 if (error != 0) {
2045 zfs_dbgmsg(
2046 "hdr %p, compress %d, psize %d, lsize %d",
2047 hdr, arc_hdr_get_compress(hdr),
2048 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2049 if (hash_lock != NULL)
2050 mutex_enter(hash_lock);
2051 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2052 if (hash_lock != NULL)
2053 mutex_exit(hash_lock);
2054 return (SET_ERROR(EIO));
2055 }
2056 }
2057 }
2058
2059 byteswap:
2060 /* Byteswap the buf's data if necessary */
2061 if (bswap != DMU_BSWAP_NUMFUNCS) {
2062 ASSERT(!HDR_SHARED_DATA(hdr));
2063 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2064 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2065 }
2066
2067 /* Compute the hdr's checksum if necessary */
2068 arc_cksum_compute(buf);
2069
2070 return (0);
2071 }
2072
2073 /*
2074 * If this function is being called to decrypt an encrypted buffer or verify an
2075 * authenticated one, the key must be loaded and a mapping must be made
2076 * available in the keystore via spa_keystore_create_mapping() or one of its
2077 * callers.
2078 */
2079 int
2080 arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2081 boolean_t in_place)
2082 {
2083 int ret;
2084 arc_fill_flags_t flags = 0;
2085
2086 if (in_place)
2087 flags |= ARC_FILL_IN_PLACE;
2088
2089 ret = arc_buf_fill(buf, spa, zb, flags);
2090 if (ret == ECKSUM) {
2091 /*
2092 * Convert authentication and decryption errors to EIO
2093 * (and generate an ereport) before leaving the ARC.
2094 */
2095 ret = SET_ERROR(EIO);
2096 spa_log_error(spa, zb);
2097 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
2098 spa, NULL, zb, NULL, 0, 0);
2099 }
2100
2101 return (ret);
2102 }
2103
2104 /*
2105 * Increment the amount of evictable space in the arc_state_t's refcount.
2106 * We account for the space used by the hdr and the arc buf individually
2107 * so that we can add and remove them from the refcount individually.
2108 */
2109 static void
2110 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2111 {
2112 arc_buf_contents_t type = arc_buf_type(hdr);
2113
2114 ASSERT(HDR_HAS_L1HDR(hdr));
2115
2116 if (GHOST_STATE(state)) {
2117 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2118 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2119 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2120 ASSERT(!HDR_HAS_RABD(hdr));
2121 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2122 HDR_GET_LSIZE(hdr), hdr);
2123 return;
2124 }
2125
2126 ASSERT(!GHOST_STATE(state));
2127 if (hdr->b_l1hdr.b_pabd != NULL) {
2128 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2129 arc_hdr_size(hdr), hdr);
2130 }
2131 if (HDR_HAS_RABD(hdr)) {
2132 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2133 HDR_GET_PSIZE(hdr), hdr);
2134 }
2135 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2136 buf = buf->b_next) {
2137 if (arc_buf_is_shared(buf))
2138 continue;
2139 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2140 arc_buf_size(buf), buf);
2141 }
2142 }
2143
2144 /*
2145 * Decrement the amount of evictable space in the arc_state_t's refcount.
2146 * We account for the space used by the hdr and the arc buf individually
2147 * so that we can add and remove them from the refcount individually.
2148 */
2149 static void
2150 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2151 {
2152 arc_buf_contents_t type = arc_buf_type(hdr);
2153
2154 ASSERT(HDR_HAS_L1HDR(hdr));
2155
2156 if (GHOST_STATE(state)) {
2157 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2158 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2159 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2160 ASSERT(!HDR_HAS_RABD(hdr));
2161 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2162 HDR_GET_LSIZE(hdr), hdr);
2163 return;
2164 }
2165
2166 ASSERT(!GHOST_STATE(state));
2167 if (hdr->b_l1hdr.b_pabd != NULL) {
2168 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2169 arc_hdr_size(hdr), hdr);
2170 }
2171 if (HDR_HAS_RABD(hdr)) {
2172 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2173 HDR_GET_PSIZE(hdr), hdr);
2174 }
2175 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2176 buf = buf->b_next) {
2177 if (arc_buf_is_shared(buf))
2178 continue;
2179 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2180 arc_buf_size(buf), buf);
2181 }
2182 }
2183
2184 /*
2185 * Add a reference to this hdr indicating that someone is actively
2186 * referencing that memory. When the refcount transitions from 0 to 1,
2187 * we remove it from the respective arc_state_t list to indicate that
2188 * it is not evictable.
2189 */
2190 static void
2191 add_reference(arc_buf_hdr_t *hdr, void *tag)
2192 {
2193 ASSERT(HDR_HAS_L1HDR(hdr));
2194 if (!HDR_EMPTY(hdr) && !MUTEX_HELD(HDR_LOCK(hdr))) {
2195 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2196 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2197 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2198 }
2199
2200 arc_state_t *state = hdr->b_l1hdr.b_state;
2201
2202 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2203 (state != arc_anon)) {
2204 /* We don't use the L2-only state list. */
2205 if (state != arc_l2c_only) {
2206 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2207 hdr);
2208 arc_evictable_space_decrement(hdr, state);
2209 }
2210 /* remove the prefetch flag if we get a reference */
2211 if (HDR_HAS_L2HDR(hdr))
2212 l2arc_hdr_arcstats_decrement_state(hdr);
2213 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2214 if (HDR_HAS_L2HDR(hdr))
2215 l2arc_hdr_arcstats_increment_state(hdr);
2216 }
2217 }
2218
2219 /*
2220 * Remove a reference from this hdr. When the reference transitions from
2221 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2222 * list making it eligible for eviction.
2223 */
2224 static int
2225 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2226 {
2227 int cnt;
2228 arc_state_t *state = hdr->b_l1hdr.b_state;
2229
2230 ASSERT(HDR_HAS_L1HDR(hdr));
2231 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2232 ASSERT(!GHOST_STATE(state));
2233
2234 /*
2235 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2236 * check to prevent usage of the arc_l2c_only list.
2237 */
2238 if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2239 (state != arc_anon)) {
2240 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2241 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2242 arc_evictable_space_increment(hdr, state);
2243 }
2244 return (cnt);
2245 }
2246
2247 /*
2248 * Move the supplied buffer to the indicated state. The hash lock
2249 * for the buffer must be held by the caller.
2250 */
2251 static void
2252 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2253 kmutex_t *hash_lock)
2254 {
2255 arc_state_t *old_state;
2256 int64_t refcnt;
2257 uint32_t bufcnt;
2258 boolean_t update_old, update_new;
2259 arc_buf_contents_t buftype = arc_buf_type(hdr);
2260
2261 /*
2262 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2263 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2264 * L1 hdr doesn't always exist when we change state to arc_anon before
2265 * destroying a header, in which case reallocating to add the L1 hdr is
2266 * pointless.
2267 */
2268 if (HDR_HAS_L1HDR(hdr)) {
2269 old_state = hdr->b_l1hdr.b_state;
2270 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2271 bufcnt = hdr->b_l1hdr.b_bufcnt;
2272
2273 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL ||
2274 HDR_HAS_RABD(hdr));
2275 } else {
2276 old_state = arc_l2c_only;
2277 refcnt = 0;
2278 bufcnt = 0;
2279 update_old = B_FALSE;
2280 }
2281 update_new = update_old;
2282
2283 ASSERT(MUTEX_HELD(hash_lock));
2284 ASSERT3P(new_state, !=, old_state);
2285 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2286 ASSERT(old_state != arc_anon || bufcnt <= 1);
2287
2288 /*
2289 * If this buffer is evictable, transfer it from the
2290 * old state list to the new state list.
2291 */
2292 if (refcnt == 0) {
2293 if (old_state != arc_anon && old_state != arc_l2c_only) {
2294 ASSERT(HDR_HAS_L1HDR(hdr));
2295 multilist_remove(old_state->arcs_list[buftype], hdr);
2296
2297 if (GHOST_STATE(old_state)) {
2298 ASSERT0(bufcnt);
2299 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2300 update_old = B_TRUE;
2301 }
2302 arc_evictable_space_decrement(hdr, old_state);
2303 }
2304 if (new_state != arc_anon && new_state != arc_l2c_only) {
2305
2306 /*
2307 * An L1 header always exists here, since if we're
2308 * moving to some L1-cached state (i.e. not l2c_only or
2309 * anonymous), we realloc the header to add an L1hdr
2310 * beforehand.
2311 */
2312 ASSERT(HDR_HAS_L1HDR(hdr));
2313 multilist_insert(new_state->arcs_list[buftype], hdr);
2314
2315 if (GHOST_STATE(new_state)) {
2316 ASSERT0(bufcnt);
2317 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2318 update_new = B_TRUE;
2319 }
2320 arc_evictable_space_increment(hdr, new_state);
2321 }
2322 }
2323
2324 ASSERT(!HDR_EMPTY(hdr));
2325 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2326 buf_hash_remove(hdr);
2327
2328 /* adjust state sizes (ignore arc_l2c_only) */
2329
2330 if (update_new && new_state != arc_l2c_only) {
2331 ASSERT(HDR_HAS_L1HDR(hdr));
2332 if (GHOST_STATE(new_state)) {
2333 ASSERT0(bufcnt);
2334
2335 /*
2336 * When moving a header to a ghost state, we first
2337 * remove all arc buffers. Thus, we'll have a
2338 * bufcnt of zero, and no arc buffer to use for
2339 * the reference. As a result, we use the arc
2340 * header pointer for the reference.
2341 */
2342 (void) zfs_refcount_add_many(&new_state->arcs_size,
2343 HDR_GET_LSIZE(hdr), hdr);
2344 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2345 ASSERT(!HDR_HAS_RABD(hdr));
2346 } else {
2347 uint32_t buffers = 0;
2348
2349 /*
2350 * Each individual buffer holds a unique reference,
2351 * thus we must remove each of these references one
2352 * at a time.
2353 */
2354 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2355 buf = buf->b_next) {
2356 ASSERT3U(bufcnt, !=, 0);
2357 buffers++;
2358
2359 /*
2360 * When the arc_buf_t is sharing the data
2361 * block with the hdr, the owner of the
2362 * reference belongs to the hdr. Only
2363 * add to the refcount if the arc_buf_t is
2364 * not shared.
2365 */
2366 if (arc_buf_is_shared(buf))
2367 continue;
2368
2369 (void) zfs_refcount_add_many(
2370 &new_state->arcs_size,
2371 arc_buf_size(buf), buf);
2372 }
2373 ASSERT3U(bufcnt, ==, buffers);
2374
2375 if (hdr->b_l1hdr.b_pabd != NULL) {
2376 (void) zfs_refcount_add_many(
2377 &new_state->arcs_size,
2378 arc_hdr_size(hdr), hdr);
2379 }
2380
2381 if (HDR_HAS_RABD(hdr)) {
2382 (void) zfs_refcount_add_many(
2383 &new_state->arcs_size,
2384 HDR_GET_PSIZE(hdr), hdr);
2385 }
2386 }
2387 }
2388
2389 if (update_old && old_state != arc_l2c_only) {
2390 ASSERT(HDR_HAS_L1HDR(hdr));
2391 if (GHOST_STATE(old_state)) {
2392 ASSERT0(bufcnt);
2393 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2394 ASSERT(!HDR_HAS_RABD(hdr));
2395
2396 /*
2397 * When moving a header off of a ghost state,
2398 * the header will not contain any arc buffers.
2399 * We use the arc header pointer for the reference
2400 * which is exactly what we did when we put the
2401 * header on the ghost state.
2402 */
2403
2404 (void) zfs_refcount_remove_many(&old_state->arcs_size,
2405 HDR_GET_LSIZE(hdr), hdr);
2406 } else {
2407 uint32_t buffers = 0;
2408
2409 /*
2410 * Each individual buffer holds a unique reference,
2411 * thus we must remove each of these references one
2412 * at a time.
2413 */
2414 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2415 buf = buf->b_next) {
2416 ASSERT3U(bufcnt, !=, 0);
2417 buffers++;
2418
2419 /*
2420 * When the arc_buf_t is sharing the data
2421 * block with the hdr, the owner of the
2422 * reference belongs to the hdr. Only
2423 * add to the refcount if the arc_buf_t is
2424 * not shared.
2425 */
2426 if (arc_buf_is_shared(buf))
2427 continue;
2428
2429 (void) zfs_refcount_remove_many(
2430 &old_state->arcs_size, arc_buf_size(buf),
2431 buf);
2432 }
2433 ASSERT3U(bufcnt, ==, buffers);
2434 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2435 HDR_HAS_RABD(hdr));
2436
2437 if (hdr->b_l1hdr.b_pabd != NULL) {
2438 (void) zfs_refcount_remove_many(
2439 &old_state->arcs_size, arc_hdr_size(hdr),
2440 hdr);
2441 }
2442
2443 if (HDR_HAS_RABD(hdr)) {
2444 (void) zfs_refcount_remove_many(
2445 &old_state->arcs_size, HDR_GET_PSIZE(hdr),
2446 hdr);
2447 }
2448 }
2449 }
2450
2451 if (HDR_HAS_L1HDR(hdr)) {
2452 hdr->b_l1hdr.b_state = new_state;
2453
2454 if (HDR_HAS_L2HDR(hdr) && new_state != arc_l2c_only) {
2455 l2arc_hdr_arcstats_decrement_state(hdr);
2456 hdr->b_l2hdr.b_arcs_state = new_state->arcs_state;
2457 l2arc_hdr_arcstats_increment_state(hdr);
2458 }
2459 }
2460
2461 /*
2462 * L2 headers should never be on the L2 state list since they don't
2463 * have L1 headers allocated.
2464 */
2465 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2466 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2467 }
2468
2469 void
2470 arc_space_consume(uint64_t space, arc_space_type_t type)
2471 {
2472 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2473
2474 switch (type) {
2475 case ARC_SPACE_DATA:
2476 aggsum_add(&astat_data_size, space);
2477 break;
2478 case ARC_SPACE_META:
2479 aggsum_add(&astat_metadata_size, space);
2480 break;
2481 case ARC_SPACE_OTHER:
2482 aggsum_add(&astat_other_size, space);
2483 break;
2484 case ARC_SPACE_HDRS:
2485 aggsum_add(&astat_hdr_size, space);
2486 break;
2487 case ARC_SPACE_L2HDRS:
2488 aggsum_add(&astat_l2_hdr_size, space);
2489 break;
2490 }
2491
2492 if (type != ARC_SPACE_DATA)
2493 aggsum_add(&arc_meta_used, space);
2494
2495 aggsum_add(&arc_size, space);
2496 }
2497
2498 void
2499 arc_space_return(uint64_t space, arc_space_type_t type)
2500 {
2501 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2502
2503 switch (type) {
2504 case ARC_SPACE_DATA:
2505 aggsum_add(&astat_data_size, -space);
2506 break;
2507 case ARC_SPACE_META:
2508 aggsum_add(&astat_metadata_size, -space);
2509 break;
2510 case ARC_SPACE_OTHER:
2511 aggsum_add(&astat_other_size, -space);
2512 break;
2513 case ARC_SPACE_HDRS:
2514 aggsum_add(&astat_hdr_size, -space);
2515 break;
2516 case ARC_SPACE_L2HDRS:
2517 aggsum_add(&astat_l2_hdr_size, -space);
2518 break;
2519 }
2520
2521 if (type != ARC_SPACE_DATA) {
2522 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2523 /*
2524 * We use the upper bound here rather than the precise value
2525 * because the arc_meta_max value doesn't need to be
2526 * precise. It's only consumed by humans via arcstats.
2527 */
2528 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2529 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2530 aggsum_add(&arc_meta_used, -space);
2531 }
2532
2533 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2534 aggsum_add(&arc_size, -space);
2535 }
2536
2537 /*
2538 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2539 * with the hdr's b_pabd.
2540 */
2541 static boolean_t
2542 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2543 {
2544 /*
2545 * The criteria for sharing a hdr's data are:
2546 * 1. the buffer is not encrypted
2547 * 2. the hdr's compression matches the buf's compression
2548 * 3. the hdr doesn't need to be byteswapped
2549 * 4. the hdr isn't already being shared
2550 * 5. the buf is either compressed or it is the last buf in the hdr list
2551 *
2552 * Criterion #5 maintains the invariant that shared uncompressed
2553 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2554 * might ask, "if a compressed buf is allocated first, won't that be the
2555 * last thing in the list?", but in that case it's impossible to create
2556 * a shared uncompressed buf anyway (because the hdr must be compressed
2557 * to have the compressed buf). You might also think that #3 is
2558 * sufficient to make this guarantee, however it's possible
2559 * (specifically in the rare L2ARC write race mentioned in
2560 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2561 * is sharable, but wasn't at the time of its allocation. Rather than
2562 * allow a new shared uncompressed buf to be created and then shuffle
2563 * the list around to make it the last element, this simply disallows
2564 * sharing if the new buf isn't the first to be added.
2565 */
2566 ASSERT3P(buf->b_hdr, ==, hdr);
2567 boolean_t hdr_compressed = arc_hdr_get_compress(hdr) !=
2568 ZIO_COMPRESS_OFF;
2569 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2570 return (!ARC_BUF_ENCRYPTED(buf) &&
2571 buf_compressed == hdr_compressed &&
2572 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2573 !HDR_SHARED_DATA(hdr) &&
2574 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2575 }
2576
2577 /*
2578 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2579 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2580 * copy was made successfully, or an error code otherwise.
2581 */
2582 static int
2583 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
2584 void *tag, boolean_t encrypted, boolean_t compressed, boolean_t noauth,
2585 boolean_t fill, arc_buf_t **ret)
2586 {
2587 arc_buf_t *buf;
2588 arc_fill_flags_t flags = ARC_FILL_LOCKED;
2589
2590 ASSERT(HDR_HAS_L1HDR(hdr));
2591 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2592 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2593 hdr->b_type == ARC_BUFC_METADATA);
2594 ASSERT3P(ret, !=, NULL);
2595 ASSERT3P(*ret, ==, NULL);
2596 IMPLY(encrypted, compressed);
2597
2598 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2599 buf->b_hdr = hdr;
2600 buf->b_data = NULL;
2601 buf->b_next = hdr->b_l1hdr.b_buf;
2602 buf->b_flags = 0;
2603
2604 add_reference(hdr, tag);
2605
2606 /*
2607 * We're about to change the hdr's b_flags. We must either
2608 * hold the hash_lock or be undiscoverable.
2609 */
2610 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2611
2612 /*
2613 * Only honor requests for compressed bufs if the hdr is actually
2614 * compressed. This must be overriden if the buffer is encrypted since
2615 * encrypted buffers cannot be decompressed.
2616 */
2617 if (encrypted) {
2618 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2619 buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2620 flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2621 } else if (compressed &&
2622 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2623 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2624 flags |= ARC_FILL_COMPRESSED;
2625 }
2626
2627 if (noauth) {
2628 ASSERT0(encrypted);
2629 flags |= ARC_FILL_NOAUTH;
2630 }
2631
2632 /*
2633 * If the hdr's data can be shared then we share the data buffer and
2634 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2635 * allocate a new buffer to store the buf's data.
2636 *
2637 * There are two additional restrictions here because we're sharing
2638 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2639 * actively involved in an L2ARC write, because if this buf is used by
2640 * an arc_write() then the hdr's data buffer will be released when the
2641 * write completes, even though the L2ARC write might still be using it.
2642 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2643 * need to be ABD-aware.
2644 */
2645 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2646 hdr->b_l1hdr.b_pabd != NULL && abd_is_linear(hdr->b_l1hdr.b_pabd);
2647
2648 /* Set up b_data and sharing */
2649 if (can_share) {
2650 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2651 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2652 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2653 } else {
2654 buf->b_data =
2655 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2656 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2657 }
2658 VERIFY3P(buf->b_data, !=, NULL);
2659
2660 hdr->b_l1hdr.b_buf = buf;
2661 hdr->b_l1hdr.b_bufcnt += 1;
2662 if (encrypted)
2663 hdr->b_crypt_hdr.b_ebufcnt += 1;
2664
2665 /*
2666 * If the user wants the data from the hdr, we need to either copy or
2667 * decompress the data.
2668 */
2669 if (fill) {
2670 ASSERT3P(zb, !=, NULL);
2671 return (arc_buf_fill(buf, spa, zb, flags));
2672 }
2673
2674 return (0);
2675 }
2676
2677 static char *arc_onloan_tag = "onloan";
2678
2679 static inline void
2680 arc_loaned_bytes_update(int64_t delta)
2681 {
2682 atomic_add_64(&arc_loaned_bytes, delta);
2683
2684 /* assert that it did not wrap around */
2685 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2686 }
2687
2688 /*
2689 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2690 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2691 * buffers must be returned to the arc before they can be used by the DMU or
2692 * freed.
2693 */
2694 arc_buf_t *
2695 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2696 {
2697 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2698 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2699
2700 arc_loaned_bytes_update(arc_buf_size(buf));
2701
2702 return (buf);
2703 }
2704
2705 arc_buf_t *
2706 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2707 enum zio_compress compression_type)
2708 {
2709 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2710 psize, lsize, compression_type);
2711
2712 arc_loaned_bytes_update(arc_buf_size(buf));
2713
2714 return (buf);
2715 }
2716
2717 arc_buf_t *
2718 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
2719 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
2720 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
2721 enum zio_compress compression_type)
2722 {
2723 arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
2724 byteorder, salt, iv, mac, ot, psize, lsize, compression_type);
2725
2726 atomic_add_64(&arc_loaned_bytes, psize);
2727 return (buf);
2728 }
2729
2730 /*
2731 * Performance tuning of L2ARC persistence:
2732 *
2733 * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding
2734 * an L2ARC device (either at pool import or later) will attempt
2735 * to rebuild L2ARC buffer contents.
2736 * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls
2737 * whether log blocks are written to the L2ARC device. If the L2ARC
2738 * device is less than 1GB, the amount of data l2arc_evict()
2739 * evicts is significant compared to the amount of restored L2ARC
2740 * data. In this case do not write log blocks in L2ARC in order
2741 * not to waste space.
2742 */
2743 int l2arc_rebuild_enabled = B_TRUE;
2744 unsigned long l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024;
2745
2746 /* L2ARC persistence rebuild control routines. */
2747 void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen);
2748 static void l2arc_dev_rebuild_start(l2arc_dev_t *dev);
2749 static int l2arc_rebuild(l2arc_dev_t *dev);
2750
2751 /* L2ARC persistence read I/O routines. */
2752 static int l2arc_dev_hdr_read(l2arc_dev_t *dev);
2753 static int l2arc_log_blk_read(l2arc_dev_t *dev,
2754 const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp,
2755 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
2756 zio_t *this_io, zio_t **next_io);
2757 static zio_t *l2arc_log_blk_fetch(vdev_t *vd,
2758 const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb);
2759 static void l2arc_log_blk_fetch_abort(zio_t *zio);
2760
2761 /* L2ARC persistence block restoration routines. */
2762 static void l2arc_log_blk_restore(l2arc_dev_t *dev,
2763 const l2arc_log_blk_phys_t *lb, uint64_t lb_asize);
2764 static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le,
2765 l2arc_dev_t *dev);
2766
2767 /* L2ARC persistence write I/O routines. */
2768 static void l2arc_dev_hdr_update(l2arc_dev_t *dev);
2769 static void l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
2770 l2arc_write_callback_t *cb);
2771
2772 /* L2ARC persistence auxilliary routines. */
2773 boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev,
2774 const l2arc_log_blkptr_t *lbp);
2775 static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev,
2776 const arc_buf_hdr_t *ab);
2777 boolean_t l2arc_range_check_overlap(uint64_t bottom,
2778 uint64_t top, uint64_t check);
2779 static void l2arc_blk_fetch_done(zio_t *zio);
2780 static inline uint64_t
2781 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev);
2782
2783 /*
2784 * Return a loaned arc buffer to the arc.
2785 */
2786 void
2787 arc_return_buf(arc_buf_t *buf, void *tag)
2788 {
2789 arc_buf_hdr_t *hdr = buf->b_hdr;
2790
2791 ASSERT3P(buf->b_data, !=, NULL);
2792 ASSERT(HDR_HAS_L1HDR(hdr));
2793 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2794 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2795
2796 arc_loaned_bytes_update(-arc_buf_size(buf));
2797 }
2798
2799 /* Detach an arc_buf from a dbuf (tag) */
2800 void
2801 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2802 {
2803 arc_buf_hdr_t *hdr = buf->b_hdr;
2804
2805 ASSERT3P(buf->b_data, !=, NULL);
2806 ASSERT(HDR_HAS_L1HDR(hdr));
2807 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2808 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2809
2810 arc_loaned_bytes_update(arc_buf_size(buf));
2811 }
2812
2813 static void
2814 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2815 {
2816 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2817
2818 df->l2df_abd = abd;
2819 df->l2df_size = size;
2820 df->l2df_type = type;
2821 mutex_enter(&l2arc_free_on_write_mtx);
2822 list_insert_head(l2arc_free_on_write, df);
2823 mutex_exit(&l2arc_free_on_write_mtx);
2824 }
2825
2826 static void
2827 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
2828 {
2829 arc_state_t *state = hdr->b_l1hdr.b_state;
2830 arc_buf_contents_t type = arc_buf_type(hdr);
2831 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
2832
2833 /* protected by hash lock, if in the hash table */
2834 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2835 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2836 ASSERT(state != arc_anon && state != arc_l2c_only);
2837
2838 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2839 size, hdr);
2840 }
2841 (void) zfs_refcount_remove_many(&state->arcs_size, size, hdr);
2842 if (type == ARC_BUFC_METADATA) {
2843 arc_space_return(size, ARC_SPACE_META);
2844 } else {
2845 ASSERT(type == ARC_BUFC_DATA);
2846 arc_space_return(size, ARC_SPACE_DATA);
2847 }
2848
2849 if (free_rdata) {
2850 l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
2851 } else {
2852 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2853 }
2854 }
2855
2856 /*
2857 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2858 * data buffer, we transfer the refcount ownership to the hdr and update
2859 * the appropriate kstats.
2860 */
2861 static void
2862 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2863 {
2864 /* LINTED */
2865 arc_state_t *state = hdr->b_l1hdr.b_state;
2866
2867 ASSERT(arc_can_share(hdr, buf));
2868 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2869 ASSERT(!ARC_BUF_ENCRYPTED(buf));
2870 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2871
2872 /*
2873 * Start sharing the data buffer. We transfer the
2874 * refcount ownership to the hdr since it always owns
2875 * the refcount whenever an arc_buf_t is shared.
2876 */
2877 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
2878 arc_hdr_size(hdr), buf, hdr);
2879 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2880 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2881 HDR_ISTYPE_METADATA(hdr));
2882 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2883 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2884
2885 /*
2886 * Since we've transferred ownership to the hdr we need
2887 * to increment its compressed and uncompressed kstats and
2888 * decrement the overhead size.
2889 */
2890 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2891 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2892 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2893 }
2894
2895 static void
2896 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2897 {
2898 /* LINTED */
2899 arc_state_t *state = hdr->b_l1hdr.b_state;
2900
2901 ASSERT(arc_buf_is_shared(buf));
2902 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2903 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2904
2905 /*
2906 * We are no longer sharing this buffer so we need
2907 * to transfer its ownership to the rightful owner.
2908 */
2909 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
2910 arc_hdr_size(hdr), hdr, buf);
2911 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2912 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
2913 abd_put(hdr->b_l1hdr.b_pabd);
2914 hdr->b_l1hdr.b_pabd = NULL;
2915 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2916
2917 /*
2918 * Since the buffer is no longer shared between
2919 * the arc buf and the hdr, count it as overhead.
2920 */
2921 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2922 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2923 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2924 }
2925
2926 /*
2927 * Remove an arc_buf_t from the hdr's buf list and return the last
2928 * arc_buf_t on the list. If no buffers remain on the list then return
2929 * NULL.
2930 */
2931 static arc_buf_t *
2932 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2933 {
2934 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
2935 arc_buf_t *lastbuf = NULL;
2936
2937 ASSERT(HDR_HAS_L1HDR(hdr));
2938 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2939
2940 /*
2941 * Remove the buf from the hdr list and locate the last
2942 * remaining buffer on the list.
2943 */
2944 while (*bufp != NULL) {
2945 if (*bufp == buf)
2946 *bufp = buf->b_next;
2947
2948 /*
2949 * If we've removed a buffer in the middle of
2950 * the list then update the lastbuf and update
2951 * bufp.
2952 */
2953 if (*bufp != NULL) {
2954 lastbuf = *bufp;
2955 bufp = &(*bufp)->b_next;
2956 }
2957 }
2958 buf->b_next = NULL;
2959 ASSERT3P(lastbuf, !=, buf);
2960 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
2961 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
2962 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
2963
2964 return (lastbuf);
2965 }
2966
2967 /*
2968 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
2969 * list and free it.
2970 */
2971 static void
2972 arc_buf_destroy_impl(arc_buf_t *buf)
2973 {
2974 arc_buf_hdr_t *hdr = buf->b_hdr;
2975
2976 /*
2977 * Free up the data associated with the buf but only if we're not
2978 * sharing this with the hdr. If we are sharing it with the hdr, the
2979 * hdr is responsible for doing the free.
2980 */
2981 if (buf->b_data != NULL) {
2982 /*
2983 * We're about to change the hdr's b_flags. We must either
2984 * hold the hash_lock or be undiscoverable.
2985 */
2986 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2987
2988 arc_cksum_verify(buf);
2989 arc_buf_unwatch(buf);
2990
2991 if (arc_buf_is_shared(buf)) {
2992 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2993 } else {
2994 uint64_t size = arc_buf_size(buf);
2995 arc_free_data_buf(hdr, buf->b_data, size, buf);
2996 ARCSTAT_INCR(arcstat_overhead_size, -size);
2997 }
2998 buf->b_data = NULL;
2999
3000 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3001 hdr->b_l1hdr.b_bufcnt -= 1;
3002
3003 if (ARC_BUF_ENCRYPTED(buf)) {
3004 hdr->b_crypt_hdr.b_ebufcnt -= 1;
3005
3006 /*
3007 * If we have no more encrypted buffers and we've
3008 * already gotten a copy of the decrypted data we can
3009 * free b_rabd to save some space.
3010 */
3011 if (hdr->b_crypt_hdr.b_ebufcnt == 0 &&
3012 HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL &&
3013 !HDR_IO_IN_PROGRESS(hdr)) {
3014 arc_hdr_free_pabd(hdr, B_TRUE);
3015 }
3016 }
3017 }
3018
3019 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
3020
3021 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
3022 /*
3023 * If the current arc_buf_t is sharing its data buffer with the
3024 * hdr, then reassign the hdr's b_pabd to share it with the new
3025 * buffer at the end of the list. The shared buffer is always
3026 * the last one on the hdr's buffer list.
3027 *
3028 * There is an equivalent case for compressed bufs, but since
3029 * they aren't guaranteed to be the last buf in the list and
3030 * that is an exceedingly rare case, we just allow that space be
3031 * wasted temporarily. We must also be careful not to share
3032 * encrypted buffers, since they cannot be shared.
3033 */
3034 if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
3035 /* Only one buf can be shared at once */
3036 VERIFY(!arc_buf_is_shared(lastbuf));
3037 /* hdr is uncompressed so can't have compressed buf */
3038 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
3039
3040 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3041 arc_hdr_free_pabd(hdr, B_FALSE);
3042
3043 /*
3044 * We must setup a new shared block between the
3045 * last buffer and the hdr. The data would have
3046 * been allocated by the arc buf so we need to transfer
3047 * ownership to the hdr since it's now being shared.
3048 */
3049 arc_share_buf(hdr, lastbuf);
3050 }
3051 } else if (HDR_SHARED_DATA(hdr)) {
3052 /*
3053 * Uncompressed shared buffers are always at the end
3054 * of the list. Compressed buffers don't have the
3055 * same requirements. This makes it hard to
3056 * simply assert that the lastbuf is shared so
3057 * we rely on the hdr's compression flags to determine
3058 * if we have a compressed, shared buffer.
3059 */
3060 ASSERT3P(lastbuf, !=, NULL);
3061 ASSERT(arc_buf_is_shared(lastbuf) ||
3062 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3063 }
3064
3065 /*
3066 * Free the checksum if we're removing the last uncompressed buf from
3067 * this hdr.
3068 */
3069 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3070 arc_cksum_free(hdr);
3071 }
3072
3073 /* clean up the buf */
3074 buf->b_hdr = NULL;
3075 kmem_cache_free(buf_cache, buf);
3076 }
3077
3078 static void
3079 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr, int alloc_flags)
3080 {
3081 uint64_t size;
3082 boolean_t alloc_rdata = ((alloc_flags & ARC_HDR_ALLOC_RDATA) != 0);
3083 boolean_t do_adapt = ((alloc_flags & ARC_HDR_DO_ADAPT) != 0);
3084
3085 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3086 ASSERT(HDR_HAS_L1HDR(hdr));
3087 ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3088 IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3089
3090 if (alloc_rdata) {
3091 size = HDR_GET_PSIZE(hdr);
3092 ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3093 hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr,
3094 do_adapt);
3095 ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3096 } else {
3097 size = arc_hdr_size(hdr);
3098 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3099 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr,
3100 do_adapt);
3101 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3102 }
3103
3104 ARCSTAT_INCR(arcstat_compressed_size, size);
3105 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3106 }
3107
3108 static void
3109 arc_hdr_free_pabd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3110 {
3111 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3112
3113 ASSERT(HDR_HAS_L1HDR(hdr));
3114 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3115 IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3116
3117
3118 /*
3119 * If the hdr is currently being written to the l2arc then
3120 * we defer freeing the data by adding it to the l2arc_free_on_write
3121 * list. The l2arc will free the data once it's finished
3122 * writing it to the l2arc device.
3123 */
3124 if (HDR_L2_WRITING(hdr)) {
3125 arc_hdr_free_on_write(hdr, free_rdata);
3126 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3127 } else if (free_rdata) {
3128 arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3129 } else {
3130 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3131 size, hdr);
3132 }
3133
3134 if (free_rdata) {
3135 hdr->b_crypt_hdr.b_rabd = NULL;
3136 } else {
3137 hdr->b_l1hdr.b_pabd = NULL;
3138 }
3139
3140 if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3141 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3142
3143 ARCSTAT_INCR(arcstat_compressed_size, -size);
3144 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3145 }
3146
3147 static arc_buf_hdr_t *
3148 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3149 boolean_t protected, enum zio_compress compression_type,
3150 arc_buf_contents_t type, boolean_t alloc_rdata)
3151 {
3152 arc_buf_hdr_t *hdr;
3153 int flags = ARC_HDR_DO_ADAPT;
3154
3155 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3156 if (protected) {
3157 hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE);
3158 } else {
3159 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3160 }
3161 flags |= alloc_rdata ? ARC_HDR_ALLOC_RDATA : 0;
3162 ASSERT(HDR_EMPTY(hdr));
3163 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3164 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3165 HDR_SET_PSIZE(hdr, psize);
3166 HDR_SET_LSIZE(hdr, lsize);
3167 hdr->b_spa = spa;
3168 hdr->b_type = type;
3169 hdr->b_flags = 0;
3170 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3171 arc_hdr_set_compress(hdr, compression_type);
3172 if (protected)
3173 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3174
3175 hdr->b_l1hdr.b_state = arc_anon;
3176 hdr->b_l1hdr.b_arc_access = 0;
3177 hdr->b_l1hdr.b_bufcnt = 0;
3178 hdr->b_l1hdr.b_buf = NULL;
3179
3180 /*
3181 * Allocate the hdr's buffer. This will contain either
3182 * the compressed or uncompressed data depending on the block
3183 * it references and compressed arc enablement.
3184 */
3185 arc_hdr_alloc_pabd(hdr, flags);
3186 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3187
3188 return (hdr);
3189 }
3190
3191 /*
3192 * Transition between the two allocation states for the arc_buf_hdr struct.
3193 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3194 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3195 * version is used when a cache buffer is only in the L2ARC in order to reduce
3196 * memory usage.
3197 */
3198 static arc_buf_hdr_t *
3199 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3200 {
3201 ASSERT(HDR_HAS_L2HDR(hdr));
3202
3203 arc_buf_hdr_t *nhdr;
3204 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3205
3206 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3207 (old == hdr_l2only_cache && new == hdr_full_cache));
3208
3209 /*
3210 * if the caller wanted a new full header and the header is to be
3211 * encrypted we will actually allocate the header from the full crypt
3212 * cache instead. The same applies to freeing from the old cache.
3213 */
3214 if (HDR_PROTECTED(hdr) && new == hdr_full_cache)
3215 new = hdr_full_crypt_cache;
3216 if (HDR_PROTECTED(hdr) && old == hdr_full_cache)
3217 old = hdr_full_crypt_cache;
3218
3219 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3220
3221 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3222 buf_hash_remove(hdr);
3223
3224 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3225
3226 if (new == hdr_full_cache || new == hdr_full_crypt_cache) {
3227 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3228 /*
3229 * arc_access and arc_change_state need to be aware that a
3230 * header has just come out of L2ARC, so we set its state to
3231 * l2c_only even though it's about to change.
3232 */
3233 nhdr->b_l1hdr.b_state = arc_l2c_only;
3234
3235 /* Verify previous threads set to NULL before freeing */
3236 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3237 ASSERT(!HDR_HAS_RABD(hdr));
3238 } else {
3239 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3240 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3241 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3242
3243 /*
3244 * If we've reached here, We must have been called from
3245 * arc_evict_hdr(), as such we should have already been
3246 * removed from any ghost list we were previously on
3247 * (which protects us from racing with arc_evict_state),
3248 * thus no locking is needed during this check.
3249 */
3250 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3251
3252 /*
3253 * A buffer must not be moved into the arc_l2c_only
3254 * state if it's not finished being written out to the
3255 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3256 * might try to be accessed, even though it was removed.
3257 */
3258 VERIFY(!HDR_L2_WRITING(hdr));
3259 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3260 ASSERT(!HDR_HAS_RABD(hdr));
3261
3262 #ifdef ZFS_DEBUG
3263 if (hdr->b_l1hdr.b_thawed != NULL) {
3264 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3265 hdr->b_l1hdr.b_thawed = NULL;
3266 }
3267 #endif
3268
3269 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3270 }
3271 /*
3272 * The header has been reallocated so we need to re-insert it into any
3273 * lists it was on.
3274 */
3275 (void) buf_hash_insert(nhdr, NULL);
3276
3277 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3278
3279 mutex_enter(&dev->l2ad_mtx);
3280
3281 /*
3282 * We must place the realloc'ed header back into the list at
3283 * the same spot. Otherwise, if it's placed earlier in the list,
3284 * l2arc_write_buffers() could find it during the function's
3285 * write phase, and try to write it out to the l2arc.
3286 */
3287 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3288 list_remove(&dev->l2ad_buflist, hdr);
3289
3290 mutex_exit(&dev->l2ad_mtx);
3291
3292 /*
3293 * Since we're using the pointer address as the tag when
3294 * incrementing and decrementing the l2ad_alloc refcount, we
3295 * must remove the old pointer (that we're about to destroy) and
3296 * add the new pointer to the refcount. Otherwise we'd remove
3297 * the wrong pointer address when calling arc_hdr_destroy() later.
3298 */
3299
3300 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3301 hdr);
3302 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr),
3303 nhdr);
3304
3305 buf_discard_identity(hdr);
3306 kmem_cache_free(old, hdr);
3307
3308 return (nhdr);
3309 }
3310
3311 /*
3312 * This function allows an L1 header to be reallocated as a crypt
3313 * header and vice versa. If we are going to a crypt header, the
3314 * new fields will be zeroed out.
3315 */
3316 static arc_buf_hdr_t *
3317 arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt)
3318 {
3319 arc_buf_hdr_t *nhdr;
3320 arc_buf_t *buf;
3321 kmem_cache_t *ncache, *ocache;
3322
3323 ASSERT(HDR_HAS_L1HDR(hdr));
3324 ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt);
3325 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3326 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3327 ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node));
3328 ASSERT3P(hdr->b_hash_next, ==, NULL);
3329
3330 if (need_crypt) {
3331 ncache = hdr_full_crypt_cache;
3332 ocache = hdr_full_cache;
3333 } else {
3334 ncache = hdr_full_cache;
3335 ocache = hdr_full_crypt_cache;
3336 }
3337
3338 nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE);
3339
3340 /*
3341 * Copy all members that aren't locks or condvars to the new header.
3342 * No lists are pointing to us (as we asserted above), so we don't
3343 * need to worry about the list nodes.
3344 */
3345 nhdr->b_dva = hdr->b_dva;
3346 nhdr->b_birth = hdr->b_birth;
3347 nhdr->b_type = hdr->b_type;
3348 nhdr->b_flags = hdr->b_flags;
3349 nhdr->b_psize = hdr->b_psize;
3350 nhdr->b_lsize = hdr->b_lsize;
3351 nhdr->b_spa = hdr->b_spa;
3352 nhdr->b_l2hdr.b_dev = hdr->b_l2hdr.b_dev;
3353 nhdr->b_l2hdr.b_daddr = hdr->b_l2hdr.b_daddr;
3354 nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum;
3355 nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt;
3356 nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap;
3357 nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state;
3358 nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access;
3359 nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb;
3360 nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd;
3361 #ifdef ZFS_DEBUG
3362 if (hdr->b_l1hdr.b_thawed != NULL) {
3363 nhdr->b_l1hdr.b_thawed = hdr->b_l1hdr.b_thawed;
3364 hdr->b_l1hdr.b_thawed = NULL;
3365 }
3366 #endif
3367
3368 /*
3369 * This refcount_add() exists only to ensure that the individual
3370 * arc buffers always point to a header that is referenced, avoiding
3371 * a small race condition that could trigger ASSERTs.
3372 */
3373 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG);
3374 nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf;
3375 for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) {
3376 mutex_enter(&buf->b_evict_lock);
3377 buf->b_hdr = nhdr;
3378 mutex_exit(&buf->b_evict_lock);
3379 }
3380 zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt);
3381 (void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG);
3382 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3383
3384 if (need_crypt) {
3385 arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED);
3386 } else {
3387 arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED);
3388 }
3389
3390 /* unset all members of the original hdr */
3391 bzero(&hdr->b_dva, sizeof (dva_t));
3392 hdr->b_birth = 0;
3393 hdr->b_type = ARC_BUFC_INVALID;
3394 hdr->b_flags = 0;
3395 hdr->b_psize = 0;
3396 hdr->b_lsize = 0;
3397 hdr->b_spa = 0;
3398 hdr->b_l2hdr.b_dev = NULL;
3399 hdr->b_l2hdr.b_daddr = 0;
3400 hdr->b_l1hdr.b_freeze_cksum = NULL;
3401 hdr->b_l1hdr.b_buf = NULL;
3402 hdr->b_l1hdr.b_bufcnt = 0;
3403 hdr->b_l1hdr.b_byteswap = 0;
3404 hdr->b_l1hdr.b_state = NULL;
3405 hdr->b_l1hdr.b_arc_access = 0;
3406 hdr->b_l1hdr.b_acb = NULL;
3407 hdr->b_l1hdr.b_pabd = NULL;
3408
3409 if (ocache == hdr_full_crypt_cache) {
3410 ASSERT(!HDR_HAS_RABD(hdr));
3411 hdr->b_crypt_hdr.b_ot = DMU_OT_NONE;
3412 hdr->b_crypt_hdr.b_ebufcnt = 0;
3413 hdr->b_crypt_hdr.b_dsobj = 0;
3414 bzero(hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3415 bzero(hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3416 bzero(hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3417 }
3418
3419 buf_discard_identity(hdr);
3420 kmem_cache_free(ocache, hdr);
3421
3422 return (nhdr);
3423 }
3424
3425 /*
3426 * This function is used by the send / receive code to convert a newly
3427 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3428 * is also used to allow the root objset block to be uupdated without altering
3429 * its embedded MACs. Both block types will always be uncompressed so we do not
3430 * have to worry about compression type or psize.
3431 */
3432 void
3433 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3434 dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3435 const uint8_t *mac)
3436 {
3437 arc_buf_hdr_t *hdr = buf->b_hdr;
3438
3439 ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3440 ASSERT(HDR_HAS_L1HDR(hdr));
3441 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3442
3443 buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3444 if (!HDR_PROTECTED(hdr))
3445 hdr = arc_hdr_realloc_crypt(hdr, B_TRUE);
3446 hdr->b_crypt_hdr.b_dsobj = dsobj;
3447 hdr->b_crypt_hdr.b_ot = ot;
3448 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3449 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3450 if (!arc_hdr_has_uncompressed_buf(hdr))
3451 arc_cksum_free(hdr);
3452
3453 if (salt != NULL)
3454 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3455 if (iv != NULL)
3456 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3457 if (mac != NULL)
3458 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3459 }
3460
3461 /*
3462 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3463 * The buf is returned thawed since we expect the consumer to modify it.
3464 */
3465 arc_buf_t *
3466 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3467 {
3468 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3469 B_FALSE, ZIO_COMPRESS_OFF, type, B_FALSE);
3470
3471 arc_buf_t *buf = NULL;
3472 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
3473 B_FALSE, B_FALSE, &buf));
3474 arc_buf_thaw(buf);
3475
3476 return (buf);
3477 }
3478
3479 /*
3480 * Allocates an ARC buf header that's in an evicted & L2-cached state.
3481 * This is used during l2arc reconstruction to make empty ARC buffers
3482 * which circumvent the regular disk->arc->l2arc path and instead come
3483 * into being in the reverse order, i.e. l2arc->arc.
3484 */
3485 arc_buf_hdr_t *
3486 arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev,
3487 dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth,
3488 enum zio_compress compress, boolean_t protected,
3489 boolean_t prefetch, arc_state_type_t arcs_state)
3490 {
3491 arc_buf_hdr_t *hdr;
3492
3493 ASSERT(size != 0);
3494 hdr = kmem_cache_alloc(hdr_l2only_cache, KM_SLEEP);
3495 hdr->b_birth = birth;
3496 hdr->b_type = type;
3497 hdr->b_flags = 0;
3498 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR);
3499 HDR_SET_LSIZE(hdr, size);
3500 HDR_SET_PSIZE(hdr, psize);
3501 arc_hdr_set_compress(hdr, compress);
3502 if (protected)
3503 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3504 if (prefetch)
3505 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
3506 hdr->b_spa = spa_load_guid(dev->l2ad_vdev->vdev_spa);
3507
3508 hdr->b_dva = dva;
3509
3510 hdr->b_l2hdr.b_dev = dev;
3511 hdr->b_l2hdr.b_daddr = daddr;
3512 hdr->b_l2hdr.b_arcs_state = arcs_state;
3513
3514 return (hdr);
3515 }
3516
3517 /*
3518 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3519 * for bufs containing metadata.
3520 */
3521 arc_buf_t *
3522 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3523 enum zio_compress compression_type)
3524 {
3525 ASSERT3U(lsize, >, 0);
3526 ASSERT3U(lsize, >=, psize);
3527 ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3528 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3529
3530 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3531 B_FALSE, compression_type, ARC_BUFC_DATA, B_FALSE);
3532
3533 arc_buf_t *buf = NULL;
3534 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
3535 B_TRUE, B_FALSE, B_FALSE, &buf));
3536 arc_buf_thaw(buf);
3537 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3538
3539 if (!arc_buf_is_shared(buf)) {
3540 /*
3541 * To ensure that the hdr has the correct data in it if we call
3542 * arc_untransform() on this buf before it's been written to
3543 * disk, it's easiest if we just set up sharing between the
3544 * buf and the hdr.
3545 */
3546 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3547 arc_hdr_free_pabd(hdr, B_FALSE);
3548 arc_share_buf(hdr, buf);
3549 }
3550
3551 return (buf);
3552 }
3553
3554 arc_buf_t *
3555 arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder,
3556 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
3557 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3558 enum zio_compress compression_type)
3559 {
3560 arc_buf_hdr_t *hdr;
3561 arc_buf_t *buf;
3562 arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3563 ARC_BUFC_METADATA : ARC_BUFC_DATA;
3564
3565 ASSERT3U(lsize, >, 0);
3566 ASSERT3U(lsize, >=, psize);
3567 ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3568 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3569
3570 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3571 compression_type, type, B_TRUE);
3572
3573 hdr->b_crypt_hdr.b_dsobj = dsobj;
3574 hdr->b_crypt_hdr.b_ot = ot;
3575 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3576 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3577 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3578 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3579 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3580
3581 /*
3582 * This buffer will be considered encrypted even if the ot is not an
3583 * encrypted type. It will become authenticated instead in
3584 * arc_write_ready().
3585 */
3586 buf = NULL;
3587 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
3588 B_FALSE, B_FALSE, &buf));
3589 arc_buf_thaw(buf);
3590 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3591
3592 return (buf);
3593 }
3594
3595 static void
3596 l2arc_hdr_arcstats_update(arc_buf_hdr_t *hdr, boolean_t incr,
3597 boolean_t state_only)
3598 {
3599 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3600 l2arc_dev_t *dev = l2hdr->b_dev;
3601 uint64_t lsize = HDR_GET_LSIZE(hdr);
3602 uint64_t psize = HDR_GET_PSIZE(hdr);
3603 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3604 arc_buf_contents_t type = hdr->b_type;
3605 int64_t lsize_s;
3606 int64_t psize_s;
3607 int64_t asize_s;
3608
3609 if (incr) {
3610 lsize_s = lsize;
3611 psize_s = psize;
3612 asize_s = asize;
3613 } else {
3614 lsize_s = -lsize;
3615 psize_s = -psize;
3616 asize_s = -asize;
3617 }
3618
3619 /* If the buffer is a prefetch, count it as such. */
3620 if (HDR_PREFETCH(hdr)) {
3621 ARCSTAT_INCR(arcstat_l2_prefetch_asize, asize_s);
3622 } else {
3623 /*
3624 * We use the value stored in the L2 header upon initial
3625 * caching in L2ARC. This value will be updated in case
3626 * an MRU/MRU_ghost buffer transitions to MFU but the L2ARC
3627 * metadata (log entry) cannot currently be updated. Having
3628 * the ARC state in the L2 header solves the problem of a
3629 * possibly absent L1 header (apparent in buffers restored
3630 * from persistent L2ARC).
3631 */
3632 switch (hdr->b_l2hdr.b_arcs_state) {
3633 case ARC_STATE_MRU_GHOST:
3634 case ARC_STATE_MRU:
3635 ARCSTAT_INCR(arcstat_l2_mru_asize, asize_s);
3636 break;
3637 case ARC_STATE_MFU_GHOST:
3638 case ARC_STATE_MFU:
3639 ARCSTAT_INCR(arcstat_l2_mfu_asize, asize_s);
3640 break;
3641 default:
3642 break;
3643 }
3644 }
3645
3646 if (state_only)
3647 return;
3648
3649 ARCSTAT_INCR(arcstat_l2_psize, psize_s);
3650 ARCSTAT_INCR(arcstat_l2_lsize, lsize_s);
3651
3652 switch (type) {
3653 case ARC_BUFC_DATA:
3654 ARCSTAT_INCR(arcstat_l2_bufc_data_asize, asize_s);
3655 break;
3656 case ARC_BUFC_METADATA:
3657 ARCSTAT_INCR(arcstat_l2_bufc_metadata_asize, asize_s);
3658 break;
3659 default:
3660 break;
3661 }
3662 }
3663
3664
3665 static void
3666 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3667 {
3668 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3669 l2arc_dev_t *dev = l2hdr->b_dev;
3670 uint64_t psize = HDR_GET_PSIZE(hdr);
3671 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3672
3673 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3674 ASSERT(HDR_HAS_L2HDR(hdr));
3675
3676 list_remove(&dev->l2ad_buflist, hdr);
3677
3678 l2arc_hdr_arcstats_decrement(hdr);
3679 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3680
3681 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3682 hdr);
3683 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3684 }
3685
3686 static void
3687 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3688 {
3689 if (HDR_HAS_L1HDR(hdr)) {
3690 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3691 hdr->b_l1hdr.b_bufcnt > 0);
3692 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3693 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3694 }
3695 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3696 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3697
3698 if (HDR_HAS_L2HDR(hdr)) {
3699 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3700 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3701
3702 if (!buflist_held)
3703 mutex_enter(&dev->l2ad_mtx);
3704
3705 /*
3706 * Even though we checked this conditional above, we
3707 * need to check this again now that we have the
3708 * l2ad_mtx. This is because we could be racing with
3709 * another thread calling l2arc_evict() which might have
3710 * destroyed this header's L2 portion as we were waiting
3711 * to acquire the l2ad_mtx. If that happens, we don't
3712 * want to re-destroy the header's L2 portion.
3713 */
3714 if (HDR_HAS_L2HDR(hdr))
3715 arc_hdr_l2hdr_destroy(hdr);
3716
3717 if (!buflist_held)
3718 mutex_exit(&dev->l2ad_mtx);
3719 }
3720
3721 /*
3722 * The header's identity can only be safely discarded once it is no
3723 * longer discoverable. This requires removing it from the hash table
3724 * and the l2arc header list. After this point the hash lock can not
3725 * be used to protect the header.
3726 */
3727 if (!HDR_EMPTY(hdr))
3728 buf_discard_identity(hdr);
3729
3730 if (HDR_HAS_L1HDR(hdr)) {
3731 arc_cksum_free(hdr);
3732
3733 while (hdr->b_l1hdr.b_buf != NULL)
3734 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3735
3736 #ifdef ZFS_DEBUG
3737 if (hdr->b_l1hdr.b_thawed != NULL) {
3738 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3739 hdr->b_l1hdr.b_thawed = NULL;
3740 }
3741 #endif
3742
3743 if (hdr->b_l1hdr.b_pabd != NULL)
3744 arc_hdr_free_pabd(hdr, B_FALSE);
3745
3746 if (HDR_HAS_RABD(hdr))
3747 arc_hdr_free_pabd(hdr, B_TRUE);
3748 }
3749
3750 ASSERT3P(hdr->b_hash_next, ==, NULL);
3751 if (HDR_HAS_L1HDR(hdr)) {
3752 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3753 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3754
3755 if (!HDR_PROTECTED(hdr)) {
3756 kmem_cache_free(hdr_full_cache, hdr);
3757 } else {
3758 kmem_cache_free(hdr_full_crypt_cache, hdr);
3759 }
3760 } else {
3761 kmem_cache_free(hdr_l2only_cache, hdr);
3762 }
3763 }
3764
3765 void
3766 arc_buf_destroy(arc_buf_t *buf, void* tag)
3767 {
3768 arc_buf_hdr_t *hdr = buf->b_hdr;
3769
3770 if (hdr->b_l1hdr.b_state == arc_anon) {
3771 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3772 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3773 VERIFY0(remove_reference(hdr, NULL, tag));
3774 arc_hdr_destroy(hdr);
3775 return;
3776 }
3777
3778 kmutex_t *hash_lock = HDR_LOCK(hdr);
3779 mutex_enter(hash_lock);
3780
3781 ASSERT3P(hdr, ==, buf->b_hdr);
3782 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3783 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3784 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3785 ASSERT3P(buf->b_data, !=, NULL);
3786
3787 (void) remove_reference(hdr, hash_lock, tag);
3788 arc_buf_destroy_impl(buf);
3789 mutex_exit(hash_lock);
3790 }
3791
3792 /*
3793 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3794 * state of the header is dependent on its state prior to entering this
3795 * function. The following transitions are possible:
3796 *
3797 * - arc_mru -> arc_mru_ghost
3798 * - arc_mfu -> arc_mfu_ghost
3799 * - arc_mru_ghost -> arc_l2c_only
3800 * - arc_mru_ghost -> deleted
3801 * - arc_mfu_ghost -> arc_l2c_only
3802 * - arc_mfu_ghost -> deleted
3803 */
3804 static int64_t
3805 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3806 {
3807 arc_state_t *evicted_state, *state;
3808 int64_t bytes_evicted = 0;
3809 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3810 zfs_arc_min_prescient_prefetch_ms : zfs_arc_min_prefetch_ms;
3811
3812 ASSERT(MUTEX_HELD(hash_lock));
3813 ASSERT(HDR_HAS_L1HDR(hdr));
3814
3815 state = hdr->b_l1hdr.b_state;
3816 if (GHOST_STATE(state)) {
3817 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3818 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3819
3820 /*
3821 * l2arc_write_buffers() relies on a header's L1 portion
3822 * (i.e. its b_pabd field) during its write phase.
3823 * Thus, we cannot push a header onto the arc_l2c_only
3824 * state (removing its L1 piece) until the header is
3825 * done being written to the l2arc.
3826 */
3827 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3828 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3829 return (bytes_evicted);
3830 }
3831
3832 ARCSTAT_BUMP(arcstat_deleted);
3833 bytes_evicted += HDR_GET_LSIZE(hdr);
3834
3835 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3836
3837 if (HDR_HAS_L2HDR(hdr)) {
3838 ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3839 ASSERT(!HDR_HAS_RABD(hdr));
3840 /*
3841 * This buffer is cached on the 2nd Level ARC;
3842 * don't destroy the header.
3843 */
3844 arc_change_state(arc_l2c_only, hdr, hash_lock);
3845 /*
3846 * dropping from L1+L2 cached to L2-only,
3847 * realloc to remove the L1 header.
3848 */
3849 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3850 hdr_l2only_cache);
3851 } else {
3852 arc_change_state(arc_anon, hdr, hash_lock);
3853 arc_hdr_destroy(hdr);
3854 }
3855 return (bytes_evicted);
3856 }
3857
3858 ASSERT(state == arc_mru || state == arc_mfu);
3859 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3860
3861 /* prefetch buffers have a minimum lifespan */
3862 if (HDR_IO_IN_PROGRESS(hdr) ||
3863 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3864 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < min_lifetime * hz)) {
3865 ARCSTAT_BUMP(arcstat_evict_skip);
3866 return (bytes_evicted);
3867 }
3868
3869 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3870 while (hdr->b_l1hdr.b_buf) {
3871 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3872 if (!mutex_tryenter(&buf->b_evict_lock)) {
3873 ARCSTAT_BUMP(arcstat_mutex_miss);
3874 break;
3875 }
3876 if (buf->b_data != NULL)
3877 bytes_evicted += HDR_GET_LSIZE(hdr);
3878 mutex_exit(&buf->b_evict_lock);
3879 arc_buf_destroy_impl(buf);
3880 }
3881
3882 if (HDR_HAS_L2HDR(hdr)) {
3883 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3884 } else {
3885 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3886 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3887 HDR_GET_LSIZE(hdr));
3888
3889 switch (state->arcs_state) {
3890 case ARC_STATE_MRU:
3891 ARCSTAT_INCR(
3892 arcstat_evict_l2_eligible_mru,
3893 HDR_GET_LSIZE(hdr));
3894 break;
3895 case ARC_STATE_MFU:
3896 ARCSTAT_INCR(
3897 arcstat_evict_l2_eligible_mfu,
3898 HDR_GET_LSIZE(hdr));
3899 break;
3900 default:
3901 break;
3902 }
3903 } else {
3904 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3905 HDR_GET_LSIZE(hdr));
3906 }
3907 }
3908
3909 if (hdr->b_l1hdr.b_bufcnt == 0) {
3910 arc_cksum_free(hdr);
3911
3912 bytes_evicted += arc_hdr_size(hdr);
3913
3914 /*
3915 * If this hdr is being evicted and has a compressed
3916 * buffer then we discard it here before we change states.
3917 * This ensures that the accounting is updated correctly
3918 * in arc_free_data_impl().
3919 */
3920 if (hdr->b_l1hdr.b_pabd != NULL)
3921 arc_hdr_free_pabd(hdr, B_FALSE);
3922
3923 if (HDR_HAS_RABD(hdr))
3924 arc_hdr_free_pabd(hdr, B_TRUE);
3925
3926 arc_change_state(evicted_state, hdr, hash_lock);
3927 ASSERT(HDR_IN_HASH_TABLE(hdr));
3928 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3929 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3930 }
3931
3932 return (bytes_evicted);
3933 }
3934
3935 static uint64_t
3936 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3937 uint64_t spa, int64_t bytes)
3938 {
3939 multilist_sublist_t *mls;
3940 uint64_t bytes_evicted = 0;
3941 arc_buf_hdr_t *hdr;
3942 kmutex_t *hash_lock;
3943 int evict_count = 0;
3944
3945 ASSERT3P(marker, !=, NULL);
3946 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3947
3948 mls = multilist_sublist_lock(ml, idx);
3949
3950 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3951 hdr = multilist_sublist_prev(mls, marker)) {
3952 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3953 (evict_count >= zfs_arc_evict_batch_limit))
3954 break;
3955
3956 /*
3957 * To keep our iteration location, move the marker
3958 * forward. Since we're not holding hdr's hash lock, we
3959 * must be very careful and not remove 'hdr' from the
3960 * sublist. Otherwise, other consumers might mistake the
3961 * 'hdr' as not being on a sublist when they call the
3962 * multilist_link_active() function (they all rely on
3963 * the hash lock protecting concurrent insertions and
3964 * removals). multilist_sublist_move_forward() was
3965 * specifically implemented to ensure this is the case
3966 * (only 'marker' will be removed and re-inserted).
3967 */
3968 multilist_sublist_move_forward(mls, marker);
3969
3970 /*
3971 * The only case where the b_spa field should ever be
3972 * zero, is the marker headers inserted by
3973 * arc_evict_state(). It's possible for multiple threads
3974 * to be calling arc_evict_state() concurrently (e.g.
3975 * dsl_pool_close() and zio_inject_fault()), so we must
3976 * skip any markers we see from these other threads.
3977 */
3978 if (hdr->b_spa == 0)
3979 continue;
3980
3981 /* we're only interested in evicting buffers of a certain spa */
3982 if (spa != 0 && hdr->b_spa != spa) {
3983 ARCSTAT_BUMP(arcstat_evict_skip);
3984 continue;
3985 }
3986
3987 hash_lock = HDR_LOCK(hdr);
3988
3989 /*
3990 * We aren't calling this function from any code path
3991 * that would already be holding a hash lock, so we're
3992 * asserting on this assumption to be defensive in case
3993 * this ever changes. Without this check, it would be
3994 * possible to incorrectly increment arcstat_mutex_miss
3995 * below (e.g. if the code changed such that we called
3996 * this function with a hash lock held).
3997 */
3998 ASSERT(!MUTEX_HELD(hash_lock));
3999
4000 if (mutex_tryenter(hash_lock)) {
4001 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
4002 mutex_exit(hash_lock);
4003
4004 bytes_evicted += evicted;
4005
4006 /*
4007 * If evicted is zero, arc_evict_hdr() must have
4008 * decided to skip this header, don't increment
4009 * evict_count in this case.
4010 */
4011 if (evicted != 0)
4012 evict_count++;
4013
4014 /*
4015 * If arc_size isn't overflowing, signal any
4016 * threads that might happen to be waiting.
4017 *
4018 * For each header evicted, we wake up a single
4019 * thread. If we used cv_broadcast, we could
4020 * wake up "too many" threads causing arc_size
4021 * to significantly overflow arc_c; since
4022 * arc_get_data_impl() doesn't check for overflow
4023 * when it's woken up (it doesn't because it's
4024 * possible for the ARC to be overflowing while
4025 * full of un-evictable buffers, and the
4026 * function should proceed in this case).
4027 *
4028 * If threads are left sleeping, due to not
4029 * using cv_broadcast here, they will be woken
4030 * up via cv_broadcast in arc_adjust_cb() just
4031 * before arc_adjust_zthr sleeps.
4032 */
4033 mutex_enter(&arc_adjust_lock);
4034 if (!arc_is_overflowing())
4035 cv_signal(&arc_adjust_waiters_cv);
4036 mutex_exit(&arc_adjust_lock);
4037 } else {
4038 ARCSTAT_BUMP(arcstat_mutex_miss);
4039 }
4040 }
4041
4042 multilist_sublist_unlock(mls);
4043
4044 return (bytes_evicted);
4045 }
4046
4047 /*
4048 * Evict buffers from the given arc state, until we've removed the
4049 * specified number of bytes. Move the removed buffers to the
4050 * appropriate evict state.
4051 *
4052 * This function makes a "best effort". It skips over any buffers
4053 * it can't get a hash_lock on, and so, may not catch all candidates.
4054 * It may also return without evicting as much space as requested.
4055 *
4056 * If bytes is specified using the special value ARC_EVICT_ALL, this
4057 * will evict all available (i.e. unlocked and evictable) buffers from
4058 * the given arc state; which is used by arc_flush().
4059 */
4060 static uint64_t
4061 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
4062 arc_buf_contents_t type)
4063 {
4064 uint64_t total_evicted = 0;
4065 multilist_t *ml = state->arcs_list[type];
4066 int num_sublists;
4067 arc_buf_hdr_t **markers;
4068
4069 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
4070
4071 num_sublists = multilist_get_num_sublists(ml);
4072
4073 /*
4074 * If we've tried to evict from each sublist, made some
4075 * progress, but still have not hit the target number of bytes
4076 * to evict, we want to keep trying. The markers allow us to
4077 * pick up where we left off for each individual sublist, rather
4078 * than starting from the tail each time.
4079 */
4080 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
4081 for (int i = 0; i < num_sublists; i++) {
4082 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
4083
4084 /*
4085 * A b_spa of 0 is used to indicate that this header is
4086 * a marker. This fact is used in arc_adjust_type() and
4087 * arc_evict_state_impl().
4088 */
4089 markers[i]->b_spa = 0;
4090
4091 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4092 multilist_sublist_insert_tail(mls, markers[i]);
4093 multilist_sublist_unlock(mls);
4094 }
4095
4096 /*
4097 * While we haven't hit our target number of bytes to evict, or
4098 * we're evicting all available buffers.
4099 */
4100 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
4101 /*
4102 * Start eviction using a randomly selected sublist,
4103 * this is to try and evenly balance eviction across all
4104 * sublists. Always starting at the same sublist
4105 * (e.g. index 0) would cause evictions to favor certain
4106 * sublists over others.
4107 */
4108 int sublist_idx = multilist_get_random_index(ml);
4109 uint64_t scan_evicted = 0;
4110
4111 for (int i = 0; i < num_sublists; i++) {
4112 uint64_t bytes_remaining;
4113 uint64_t bytes_evicted;
4114
4115 if (bytes == ARC_EVICT_ALL)
4116 bytes_remaining = ARC_EVICT_ALL;
4117 else if (total_evicted < bytes)
4118 bytes_remaining = bytes - total_evicted;
4119 else
4120 break;
4121
4122 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
4123 markers[sublist_idx], spa, bytes_remaining);
4124
4125 scan_evicted += bytes_evicted;
4126 total_evicted += bytes_evicted;
4127
4128 /* we've reached the end, wrap to the beginning */
4129 if (++sublist_idx >= num_sublists)
4130 sublist_idx = 0;
4131 }
4132
4133 /*
4134 * If we didn't evict anything during this scan, we have
4135 * no reason to believe we'll evict more during another
4136 * scan, so break the loop.
4137 */
4138 if (scan_evicted == 0) {
4139 /* This isn't possible, let's make that obvious */
4140 ASSERT3S(bytes, !=, 0);
4141
4142 /*
4143 * When bytes is ARC_EVICT_ALL, the only way to
4144 * break the loop is when scan_evicted is zero.
4145 * In that case, we actually have evicted enough,
4146 * so we don't want to increment the kstat.
4147 */
4148 if (bytes != ARC_EVICT_ALL) {
4149 ASSERT3S(total_evicted, <, bytes);
4150 ARCSTAT_BUMP(arcstat_evict_not_enough);
4151 }
4152
4153 break;
4154 }
4155 }
4156
4157 for (int i = 0; i < num_sublists; i++) {
4158 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4159 multilist_sublist_remove(mls, markers[i]);
4160 multilist_sublist_unlock(mls);
4161
4162 kmem_cache_free(hdr_full_cache, markers[i]);
4163 }
4164 kmem_free(markers, sizeof (*markers) * num_sublists);
4165
4166 return (total_evicted);
4167 }
4168
4169 /*
4170 * Flush all "evictable" data of the given type from the arc state
4171 * specified. This will not evict any "active" buffers (i.e. referenced).
4172 *
4173 * When 'retry' is set to B_FALSE, the function will make a single pass
4174 * over the state and evict any buffers that it can. Since it doesn't
4175 * continually retry the eviction, it might end up leaving some buffers
4176 * in the ARC due to lock misses.
4177 *
4178 * When 'retry' is set to B_TRUE, the function will continually retry the
4179 * eviction until *all* evictable buffers have been removed from the
4180 * state. As a result, if concurrent insertions into the state are
4181 * allowed (e.g. if the ARC isn't shutting down), this function might
4182 * wind up in an infinite loop, continually trying to evict buffers.
4183 */
4184 static uint64_t
4185 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4186 boolean_t retry)
4187 {
4188 uint64_t evicted = 0;
4189
4190 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
4191 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4192
4193 if (!retry)
4194 break;
4195 }
4196
4197 return (evicted);
4198 }
4199
4200 /*
4201 * Evict the specified number of bytes from the state specified,
4202 * restricting eviction to the spa and type given. This function
4203 * prevents us from trying to evict more from a state's list than
4204 * is "evictable", and to skip evicting altogether when passed a
4205 * negative value for "bytes". In contrast, arc_evict_state() will
4206 * evict everything it can, when passed a negative value for "bytes".
4207 */
4208 static uint64_t
4209 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4210 arc_buf_contents_t type)
4211 {
4212 int64_t delta;
4213
4214 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
4215 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
4216 bytes);
4217 return (arc_evict_state(state, spa, delta, type));
4218 }
4219
4220 return (0);
4221 }
4222
4223 /*
4224 * Evict metadata buffers from the cache, such that arc_meta_used is
4225 * capped by the arc_meta_limit tunable.
4226 */
4227 static uint64_t
4228 arc_adjust_meta(uint64_t meta_used)
4229 {
4230 uint64_t total_evicted = 0;
4231 int64_t target;
4232
4233 /*
4234 * If we're over the meta limit, we want to evict enough
4235 * metadata to get back under the meta limit. We don't want to
4236 * evict so much that we drop the MRU below arc_p, though. If
4237 * we're over the meta limit more than we're over arc_p, we
4238 * evict some from the MRU here, and some from the MFU below.
4239 */
4240 target = MIN((int64_t)(meta_used - arc_meta_limit),
4241 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4242 zfs_refcount_count(&arc_mru->arcs_size) - arc_p));
4243
4244 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4245
4246 /*
4247 * Similar to the above, we want to evict enough bytes to get us
4248 * below the meta limit, but not so much as to drop us below the
4249 * space allotted to the MFU (which is defined as arc_c - arc_p).
4250 */
4251 target = MIN((int64_t)(meta_used - arc_meta_limit),
4252 (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) -
4253 (arc_c - arc_p)));
4254
4255 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4256
4257 return (total_evicted);
4258 }
4259
4260 /*
4261 * Return the type of the oldest buffer in the given arc state
4262 *
4263 * This function will select a random sublist of type ARC_BUFC_DATA and
4264 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4265 * is compared, and the type which contains the "older" buffer will be
4266 * returned.
4267 */
4268 static arc_buf_contents_t
4269 arc_adjust_type(arc_state_t *state)
4270 {
4271 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4272 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4273 int data_idx = multilist_get_random_index(data_ml);
4274 int meta_idx = multilist_get_random_index(meta_ml);
4275 multilist_sublist_t *data_mls;
4276 multilist_sublist_t *meta_mls;
4277 arc_buf_contents_t type;
4278 arc_buf_hdr_t *data_hdr;
4279 arc_buf_hdr_t *meta_hdr;
4280
4281 /*
4282 * We keep the sublist lock until we're finished, to prevent
4283 * the headers from being destroyed via arc_evict_state().
4284 */
4285 data_mls = multilist_sublist_lock(data_ml, data_idx);
4286 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4287
4288 /*
4289 * These two loops are to ensure we skip any markers that
4290 * might be at the tail of the lists due to arc_evict_state().
4291 */
4292
4293 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4294 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4295 if (data_hdr->b_spa != 0)
4296 break;
4297 }
4298
4299 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4300 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4301 if (meta_hdr->b_spa != 0)
4302 break;
4303 }
4304
4305 if (data_hdr == NULL && meta_hdr == NULL) {
4306 type = ARC_BUFC_DATA;
4307 } else if (data_hdr == NULL) {
4308 ASSERT3P(meta_hdr, !=, NULL);
4309 type = ARC_BUFC_METADATA;
4310 } else if (meta_hdr == NULL) {
4311 ASSERT3P(data_hdr, !=, NULL);
4312 type = ARC_BUFC_DATA;
4313 } else {
4314 ASSERT3P(data_hdr, !=, NULL);
4315 ASSERT3P(meta_hdr, !=, NULL);
4316
4317 /* The headers can't be on the sublist without an L1 header */
4318 ASSERT(HDR_HAS_L1HDR(data_hdr));
4319 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4320
4321 if (data_hdr->b_l1hdr.b_arc_access <
4322 meta_hdr->b_l1hdr.b_arc_access) {
4323 type = ARC_BUFC_DATA;
4324 } else {
4325 type = ARC_BUFC_METADATA;
4326 }
4327 }
4328
4329 multilist_sublist_unlock(meta_mls);
4330 multilist_sublist_unlock(data_mls);
4331
4332 return (type);
4333 }
4334
4335 /*
4336 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4337 */
4338 static uint64_t
4339 arc_adjust(void)
4340 {
4341 uint64_t total_evicted = 0;
4342 uint64_t bytes;
4343 int64_t target;
4344 uint64_t asize = aggsum_value(&arc_size);
4345 uint64_t ameta = aggsum_value(&arc_meta_used);
4346
4347 /*
4348 * If we're over arc_meta_limit, we want to correct that before
4349 * potentially evicting data buffers below.
4350 */
4351 total_evicted += arc_adjust_meta(ameta);
4352
4353 /*
4354 * Adjust MRU size
4355 *
4356 * If we're over the target cache size, we want to evict enough
4357 * from the list to get back to our target size. We don't want
4358 * to evict too much from the MRU, such that it drops below
4359 * arc_p. So, if we're over our target cache size more than
4360 * the MRU is over arc_p, we'll evict enough to get back to
4361 * arc_p here, and then evict more from the MFU below.
4362 */
4363 target = MIN((int64_t)(asize - arc_c),
4364 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4365 zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4366
4367 /*
4368 * If we're below arc_meta_min, always prefer to evict data.
4369 * Otherwise, try to satisfy the requested number of bytes to
4370 * evict from the type which contains older buffers; in an
4371 * effort to keep newer buffers in the cache regardless of their
4372 * type. If we cannot satisfy the number of bytes from this
4373 * type, spill over into the next type.
4374 */
4375 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4376 ameta > arc_meta_min) {
4377 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4378 total_evicted += bytes;
4379
4380 /*
4381 * If we couldn't evict our target number of bytes from
4382 * metadata, we try to get the rest from data.
4383 */
4384 target -= bytes;
4385
4386 total_evicted +=
4387 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4388 } else {
4389 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4390 total_evicted += bytes;
4391
4392 /*
4393 * If we couldn't evict our target number of bytes from
4394 * data, we try to get the rest from metadata.
4395 */
4396 target -= bytes;
4397
4398 total_evicted +=
4399 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4400 }
4401
4402 /*
4403 * Adjust MFU size
4404 *
4405 * Now that we've tried to evict enough from the MRU to get its
4406 * size back to arc_p, if we're still above the target cache
4407 * size, we evict the rest from the MFU.
4408 */
4409 target = asize - arc_c;
4410
4411 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4412 ameta > arc_meta_min) {
4413 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4414 total_evicted += bytes;
4415
4416 /*
4417 * If we couldn't evict our target number of bytes from
4418 * metadata, we try to get the rest from data.
4419 */
4420 target -= bytes;
4421
4422 total_evicted +=
4423 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4424 } else {
4425 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4426 total_evicted += bytes;
4427
4428 /*
4429 * If we couldn't evict our target number of bytes from
4430 * data, we try to get the rest from data.
4431 */
4432 target -= bytes;
4433
4434 total_evicted +=
4435 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4436 }
4437
4438 /*
4439 * Adjust ghost lists
4440 *
4441 * In addition to the above, the ARC also defines target values
4442 * for the ghost lists. The sum of the mru list and mru ghost
4443 * list should never exceed the target size of the cache, and
4444 * the sum of the mru list, mfu list, mru ghost list, and mfu
4445 * ghost list should never exceed twice the target size of the
4446 * cache. The following logic enforces these limits on the ghost
4447 * caches, and evicts from them as needed.
4448 */
4449 target = zfs_refcount_count(&arc_mru->arcs_size) +
4450 zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4451
4452 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4453 total_evicted += bytes;
4454
4455 target -= bytes;
4456
4457 total_evicted +=
4458 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4459
4460 /*
4461 * We assume the sum of the mru list and mfu list is less than
4462 * or equal to arc_c (we enforced this above), which means we
4463 * can use the simpler of the two equations below:
4464 *
4465 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4466 * mru ghost + mfu ghost <= arc_c
4467 */
4468 target = zfs_refcount_count(&arc_mru_ghost->arcs_size) +
4469 zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4470
4471 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4472 total_evicted += bytes;
4473
4474 target -= bytes;
4475
4476 total_evicted +=
4477 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4478
4479 return (total_evicted);
4480 }
4481
4482 void
4483 arc_flush(spa_t *spa, boolean_t retry)
4484 {
4485 uint64_t guid = 0;
4486
4487 /*
4488 * If retry is B_TRUE, a spa must not be specified since we have
4489 * no good way to determine if all of a spa's buffers have been
4490 * evicted from an arc state.
4491 */
4492 ASSERT(!retry || spa == 0);
4493
4494 if (spa != NULL)
4495 guid = spa_load_guid(spa);
4496
4497 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4498 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4499
4500 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4501 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4502
4503 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4504 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4505
4506 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4507 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4508 }
4509
4510 static void
4511 arc_reduce_target_size(int64_t to_free)
4512 {
4513 uint64_t asize = aggsum_value(&arc_size);
4514 if (arc_c > arc_c_min) {
4515
4516 if (arc_c > arc_c_min + to_free)
4517 atomic_add_64(&arc_c, -to_free);
4518 else
4519 arc_c = arc_c_min;
4520
4521 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4522 if (asize < arc_c)
4523 arc_c = MAX(asize, arc_c_min);
4524 if (arc_p > arc_c)
4525 arc_p = (arc_c >> 1);
4526 ASSERT(arc_c >= arc_c_min);
4527 ASSERT((int64_t)arc_p >= 0);
4528 }
4529
4530 if (asize > arc_c) {
4531 /* See comment in arc_adjust_cb_check() on why lock+flag */
4532 mutex_enter(&arc_adjust_lock);
4533 arc_adjust_needed = B_TRUE;
4534 mutex_exit(&arc_adjust_lock);
4535 zthr_wakeup(arc_adjust_zthr);
4536 }
4537 }
4538
4539 typedef enum free_memory_reason_t {
4540 FMR_UNKNOWN,
4541 FMR_NEEDFREE,
4542 FMR_LOTSFREE,
4543 FMR_SWAPFS_MINFREE,
4544 FMR_PAGES_PP_MAXIMUM,
4545 FMR_HEAP_ARENA,
4546 FMR_ZIO_ARENA,
4547 FMR_VIRT_MACHINE, /* 'VM' seems ambiguous in this context */
4548 } free_memory_reason_t;
4549
4550 int64_t last_free_memory;
4551 free_memory_reason_t last_free_reason;
4552
4553 /*
4554 * Additional reserve of pages for pp_reserve.
4555 */
4556 int64_t arc_pages_pp_reserve = 64;
4557
4558 /*
4559 * Additional reserve of pages for swapfs.
4560 */
4561 int64_t arc_swapfs_reserve = 64;
4562
4563 static volatile uint64_t arc_virt_machine_reserved;
4564
4565 /*
4566 * XXX: A possible concern is that we allow arc_virt_machine_reserved to
4567 * get so large that we cause the arc to perform a lot of additional
4568 * work to keep the arc extremely small. We may want to set limits to
4569 * the size of arc_virt_machine_reserved and disallow reservations
4570 * beyond that limit.
4571 */
4572 int
4573 arc_virt_machine_reserve(size_t pages)
4574 {
4575 uint64_t newv;
4576
4577 newv = atomic_add_64_nv(&arc_virt_machine_reserved, pages);
4578
4579 /*
4580 * Since arc_virt_machine_reserved effectively lowers arc_c_max
4581 * as needed for vmm memory, if this request would put the arc
4582 * under arc_c_min, we reject it. arc_c_min should be a value that
4583 * ensures reasonable performance for non-VMM stuff, as well as keep
4584 * us from dipping below lotsfree, which could trigger the pager
4585 * (and send the system toa grinding halt while it pages).
4586 *
4587 * XXX: This is a bit hacky and might be better done w/ a mutex
4588 * instead of atomic ops.
4589 */
4590 if (newv + arc_c_min > arc_c_max) {
4591 atomic_add_64(&arc_virt_machine_reserved, -(int64_t)pages);
4592 return (ENOMEM);
4593 }
4594
4595 zthr_wakeup(arc_reap_zthr);
4596 return (0);
4597 }
4598
4599 void
4600 arc_virt_machine_release(size_t pages)
4601 {
4602 atomic_add_64(&arc_virt_machine_reserved, -(int64_t)pages);
4603 }
4604
4605 /*
4606 * Return the amount of memory that can be consumed before reclaim will be
4607 * needed. Positive if there is sufficient free memory, negative indicates
4608 * the amount of memory that needs to be freed up.
4609 */
4610 static int64_t
4611 arc_available_memory(void)
4612 {
4613 int64_t lowest = INT64_MAX;
4614 int64_t n;
4615 free_memory_reason_t r = FMR_UNKNOWN;
4616
4617 #ifdef _KERNEL
4618 if (needfree > 0) {
4619 n = PAGESIZE * (-needfree);
4620 if (n < lowest) {
4621 lowest = n;
4622 r = FMR_NEEDFREE;
4623 }
4624 }
4625
4626 /*
4627 * check that we're out of range of the pageout scanner. It starts to
4628 * schedule paging if freemem is less than lotsfree and needfree.
4629 * lotsfree is the high-water mark for pageout, and needfree is the
4630 * number of needed free pages. We add extra pages here to make sure
4631 * the scanner doesn't start up while we're freeing memory.
4632 */
4633 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4634 if (n < lowest) {
4635 lowest = n;
4636 r = FMR_LOTSFREE;
4637 }
4638
4639 /*
4640 * check to make sure that swapfs has enough space so that anon
4641 * reservations can still succeed. anon_resvmem() checks that the
4642 * availrmem is greater than swapfs_minfree, and the number of reserved
4643 * swap pages. We also add a bit of extra here just to prevent
4644 * circumstances from getting really dire.
4645 */
4646 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4647 desfree - arc_swapfs_reserve);
4648 if (n < lowest) {
4649 lowest = n;
4650 r = FMR_SWAPFS_MINFREE;
4651 }
4652
4653
4654 /*
4655 * Check that we have enough availrmem that memory locking (e.g., via
4656 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4657 * stores the number of pages that cannot be locked; when availrmem
4658 * drops below pages_pp_maximum, page locking mechanisms such as
4659 * page_pp_lock() will fail.)
4660 */
4661 n = PAGESIZE * (availrmem - pages_pp_maximum -
4662 arc_pages_pp_reserve);
4663 if (n < lowest) {
4664 lowest = n;
4665 r = FMR_PAGES_PP_MAXIMUM;
4666 }
4667
4668 /*
4669 * Check that we have enough memory for any virtual machines that
4670 * are running or starting. We add desfree to keep us out of
4671 * particularly dire circumstances.
4672 */
4673 n = PAGESIZE * (availrmem - arc_virt_machine_reserved - desfree);
4674 if (n < lowest) {
4675 lowest = n;
4676 r = FMR_VIRT_MACHINE;
4677 }
4678
4679 #if defined(__i386)
4680 /*
4681 * If we're on an i386 platform, it's possible that we'll exhaust the
4682 * kernel heap space before we ever run out of available physical
4683 * memory. Most checks of the size of the heap_area compare against
4684 * tune.t_minarmem, which is the minimum available real memory that we
4685 * can have in the system. However, this is generally fixed at 25 pages
4686 * which is so low that it's useless. In this comparison, we seek to
4687 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4688 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4689 * free)
4690 */
4691 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
4692 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
4693 if (n < lowest) {
4694 lowest = n;
4695 r = FMR_HEAP_ARENA;
4696 }
4697 #endif
4698
4699 /*
4700 * If zio data pages are being allocated out of a separate heap segment,
4701 * then enforce that the size of available vmem for this arena remains
4702 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4703 *
4704 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4705 * memory (in the zio_arena) free, which can avoid memory
4706 * fragmentation issues.
4707 */
4708 if (zio_arena != NULL) {
4709 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4710 (vmem_size(zio_arena, VMEM_ALLOC) >>
4711 arc_zio_arena_free_shift);
4712 if (n < lowest) {
4713 lowest = n;
4714 r = FMR_ZIO_ARENA;
4715 }
4716 }
4717 #else
4718 /* Every 100 calls, free a small amount */
4719 if (spa_get_random(100) == 0)
4720 lowest = -1024;
4721 #endif
4722
4723 last_free_memory = lowest;
4724 last_free_reason = r;
4725
4726 return (lowest);
4727 }
4728
4729
4730 /*
4731 * Determine if the system is under memory pressure and is asking
4732 * to reclaim memory. A return value of B_TRUE indicates that the system
4733 * is under memory pressure and that the arc should adjust accordingly.
4734 */
4735 static boolean_t
4736 arc_reclaim_needed(void)
4737 {
4738 return (arc_available_memory() < 0);
4739 }
4740
4741 static void
4742 arc_kmem_reap_soon(void)
4743 {
4744 size_t i;
4745 kmem_cache_t *prev_cache = NULL;
4746 kmem_cache_t *prev_data_cache = NULL;
4747 extern kmem_cache_t *zio_buf_cache[];
4748 extern kmem_cache_t *zio_data_buf_cache[];
4749 extern kmem_cache_t *zfs_btree_leaf_cache;
4750 extern kmem_cache_t *abd_chunk_cache;
4751
4752 #ifdef _KERNEL
4753 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4754 /*
4755 * We are exceeding our meta-data cache limit.
4756 * Purge some DNLC entries to release holds on meta-data.
4757 */
4758 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4759 }
4760 #if defined(__i386)
4761 /*
4762 * Reclaim unused memory from all kmem caches.
4763 */
4764 kmem_reap();
4765 #endif
4766 #endif
4767
4768 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4769 if (zio_buf_cache[i] != prev_cache) {
4770 prev_cache = zio_buf_cache[i];
4771 kmem_cache_reap_soon(zio_buf_cache[i]);
4772 }
4773 if (zio_data_buf_cache[i] != prev_data_cache) {
4774 prev_data_cache = zio_data_buf_cache[i];
4775 kmem_cache_reap_soon(zio_data_buf_cache[i]);
4776 }
4777 }
4778 kmem_cache_reap_soon(abd_chunk_cache);
4779 kmem_cache_reap_soon(buf_cache);
4780 kmem_cache_reap_soon(hdr_full_cache);
4781 kmem_cache_reap_soon(hdr_l2only_cache);
4782 kmem_cache_reap_soon(zfs_btree_leaf_cache);
4783
4784 if (zio_arena != NULL) {
4785 /*
4786 * Ask the vmem arena to reclaim unused memory from its
4787 * quantum caches.
4788 */
4789 vmem_qcache_reap(zio_arena);
4790 }
4791 }
4792
4793 /* ARGSUSED */
4794 static boolean_t
4795 arc_adjust_cb_check(void *arg, zthr_t *zthr)
4796 {
4797 /*
4798 * This is necessary in order for the mdb ::arc dcmd to
4799 * show up to date information. Since the ::arc command
4800 * does not call the kstat's update function, without
4801 * this call, the command may show stale stats for the
4802 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4803 * with this change, the data might be up to 1 second
4804 * out of date(the arc_adjust_zthr has a maximum sleep
4805 * time of 1 second); but that should suffice. The
4806 * arc_state_t structures can be queried directly if more
4807 * accurate information is needed.
4808 */
4809 if (arc_ksp != NULL)
4810 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4811
4812 /*
4813 * We have to rely on arc_get_data_impl() to tell us when to adjust,
4814 * rather than checking if we are overflowing here, so that we are
4815 * sure to not leave arc_get_data_impl() waiting on
4816 * arc_adjust_waiters_cv. If we have become "not overflowing" since
4817 * arc_get_data_impl() checked, we need to wake it up. We could
4818 * broadcast the CV here, but arc_get_data_impl() may have not yet
4819 * gone to sleep. We would need to use a mutex to ensure that this
4820 * function doesn't broadcast until arc_get_data_impl() has gone to
4821 * sleep (e.g. the arc_adjust_lock). However, the lock ordering of
4822 * such a lock would necessarily be incorrect with respect to the
4823 * zthr_lock, which is held before this function is called, and is
4824 * held by arc_get_data_impl() when it calls zthr_wakeup().
4825 */
4826 return (arc_adjust_needed);
4827 }
4828
4829 /*
4830 * Keep arc_size under arc_c by running arc_adjust which evicts data
4831 * from the ARC.
4832 */
4833 /* ARGSUSED */
4834 static void
4835 arc_adjust_cb(void *arg, zthr_t *zthr)
4836 {
4837 uint64_t evicted = 0;
4838
4839 /* Evict from cache */
4840 evicted = arc_adjust();
4841
4842 /*
4843 * If evicted is zero, we couldn't evict anything
4844 * via arc_adjust(). This could be due to hash lock
4845 * collisions, but more likely due to the majority of
4846 * arc buffers being unevictable. Therefore, even if
4847 * arc_size is above arc_c, another pass is unlikely to
4848 * be helpful and could potentially cause us to enter an
4849 * infinite loop. Additionally, zthr_iscancelled() is
4850 * checked here so that if the arc is shutting down, the
4851 * broadcast will wake any remaining arc adjust waiters.
4852 */
4853 mutex_enter(&arc_adjust_lock);
4854 arc_adjust_needed = !zthr_iscancelled(arc_adjust_zthr) &&
4855 evicted > 0 && aggsum_compare(&arc_size, arc_c) > 0;
4856 if (!arc_adjust_needed) {
4857 /*
4858 * We're either no longer overflowing, or we
4859 * can't evict anything more, so we should wake
4860 * up any waiters.
4861 */
4862 cv_broadcast(&arc_adjust_waiters_cv);
4863 }
4864 mutex_exit(&arc_adjust_lock);
4865 }
4866
4867 /* ARGSUSED */
4868 static boolean_t
4869 arc_reap_cb_check(void *arg, zthr_t *zthr)
4870 {
4871 int64_t free_memory = arc_available_memory();
4872
4873 /*
4874 * If a kmem reap is already active, don't schedule more. We must
4875 * check for this because kmem_cache_reap_soon() won't actually
4876 * block on the cache being reaped (this is to prevent callers from
4877 * becoming implicitly blocked by a system-wide kmem reap -- which,
4878 * on a system with many, many full magazines, can take minutes).
4879 */
4880 if (!kmem_cache_reap_active() &&
4881 free_memory < 0) {
4882 arc_no_grow = B_TRUE;
4883 arc_warm = B_TRUE;
4884 /*
4885 * Wait at least zfs_grow_retry (default 60) seconds
4886 * before considering growing.
4887 */
4888 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4889 return (B_TRUE);
4890 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4891 arc_no_grow = B_TRUE;
4892 } else if (gethrtime() >= arc_growtime) {
4893 arc_no_grow = B_FALSE;
4894 }
4895
4896 return (B_FALSE);
4897 }
4898
4899 /*
4900 * Keep enough free memory in the system by reaping the ARC's kmem
4901 * caches. To cause more slabs to be reapable, we may reduce the
4902 * target size of the cache (arc_c), causing the arc_adjust_cb()
4903 * to free more buffers.
4904 */
4905 /* ARGSUSED */
4906 static void
4907 arc_reap_cb(void *arg, zthr_t *zthr)
4908 {
4909 int64_t free_memory;
4910
4911 /*
4912 * Kick off asynchronous kmem_reap()'s of all our caches.
4913 */
4914 arc_kmem_reap_soon();
4915
4916 /*
4917 * Wait at least arc_kmem_cache_reap_retry_ms between
4918 * arc_kmem_reap_soon() calls. Without this check it is possible to
4919 * end up in a situation where we spend lots of time reaping
4920 * caches, while we're near arc_c_min. Waiting here also gives the
4921 * subsequent free memory check a chance of finding that the
4922 * asynchronous reap has already freed enough memory, and we don't
4923 * need to call arc_reduce_target_size().
4924 */
4925 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
4926
4927 /*
4928 * Reduce the target size as needed to maintain the amount of free
4929 * memory in the system at a fraction of the arc_size (1/128th by
4930 * default). If oversubscribed (free_memory < 0) then reduce the
4931 * target arc_size by the deficit amount plus the fractional
4932 * amount. If free memory is positive but less then the fractional
4933 * amount, reduce by what is needed to hit the fractional amount.
4934 */
4935 free_memory = arc_available_memory();
4936
4937 int64_t to_free =
4938 (arc_c >> arc_shrink_shift) - free_memory;
4939 if (to_free > 0) {
4940 #ifdef _KERNEL
4941 to_free = MAX(to_free, ptob(needfree));
4942 #endif
4943 arc_reduce_target_size(to_free);
4944 }
4945 }
4946
4947 /*
4948 * Adapt arc info given the number of bytes we are trying to add and
4949 * the state that we are coming from. This function is only called
4950 * when we are adding new content to the cache.
4951 */
4952 static void
4953 arc_adapt(int bytes, arc_state_t *state)
4954 {
4955 int mult;
4956 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4957 int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size);
4958 int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size);
4959
4960 ASSERT(bytes > 0);
4961 /*
4962 * Adapt the target size of the MRU list:
4963 * - if we just hit in the MRU ghost list, then increase
4964 * the target size of the MRU list.
4965 * - if we just hit in the MFU ghost list, then increase
4966 * the target size of the MFU list by decreasing the
4967 * target size of the MRU list.
4968 */
4969 if (state == arc_mru_ghost) {
4970 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4971 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4972
4973 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4974 } else if (state == arc_mfu_ghost) {
4975 uint64_t delta;
4976
4977 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4978 mult = MIN(mult, 10);
4979
4980 delta = MIN(bytes * mult, arc_p);
4981 arc_p = MAX(arc_p_min, arc_p - delta);
4982 }
4983 ASSERT((int64_t)arc_p >= 0);
4984
4985 /*
4986 * Wake reap thread if we do not have any available memory
4987 */
4988 if (arc_reclaim_needed()) {
4989 zthr_wakeup(arc_reap_zthr);
4990 return;
4991 }
4992
4993
4994 if (arc_no_grow)
4995 return;
4996
4997 if (arc_c >= arc_c_max)
4998 return;
4999
5000 /*
5001 * If we're within (2 * maxblocksize) bytes of the target
5002 * cache size, increment the target cache size
5003 */
5004 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
5005 0) {
5006 atomic_add_64(&arc_c, (int64_t)bytes);
5007 if (arc_c > arc_c_max)
5008 arc_c = arc_c_max;
5009 else if (state == arc_anon)
5010 atomic_add_64(&arc_p, (int64_t)bytes);
5011 if (arc_p > arc_c)
5012 arc_p = arc_c;
5013 }
5014 ASSERT((int64_t)arc_p >= 0);
5015 }
5016
5017 /*
5018 * Check if arc_size has grown past our upper threshold, determined by
5019 * zfs_arc_overflow_shift.
5020 */
5021 static boolean_t
5022 arc_is_overflowing(void)
5023 {
5024 /* Always allow at least one block of overflow */
5025 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5026 arc_c >> zfs_arc_overflow_shift);
5027
5028 /*
5029 * We just compare the lower bound here for performance reasons. Our
5030 * primary goals are to make sure that the arc never grows without
5031 * bound, and that it can reach its maximum size. This check
5032 * accomplishes both goals. The maximum amount we could run over by is
5033 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
5034 * in the ARC. In practice, that's in the tens of MB, which is low
5035 * enough to be safe.
5036 */
5037 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
5038 }
5039
5040 static abd_t *
5041 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag,
5042 boolean_t do_adapt)
5043 {
5044 arc_buf_contents_t type = arc_buf_type(hdr);
5045
5046 arc_get_data_impl(hdr, size, tag, do_adapt);
5047 if (type == ARC_BUFC_METADATA) {
5048 return (abd_alloc(size, B_TRUE));
5049 } else {
5050 ASSERT(type == ARC_BUFC_DATA);
5051 return (abd_alloc(size, B_FALSE));
5052 }
5053 }
5054
5055 static void *
5056 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5057 {
5058 arc_buf_contents_t type = arc_buf_type(hdr);
5059
5060 arc_get_data_impl(hdr, size, tag, B_TRUE);
5061 if (type == ARC_BUFC_METADATA) {
5062 return (zio_buf_alloc(size));
5063 } else {
5064 ASSERT(type == ARC_BUFC_DATA);
5065 return (zio_data_buf_alloc(size));
5066 }
5067 }
5068
5069 /*
5070 * Allocate a block and return it to the caller. If we are hitting the
5071 * hard limit for the cache size, we must sleep, waiting for the eviction
5072 * thread to catch up. If we're past the target size but below the hard
5073 * limit, we'll only signal the reclaim thread and continue on.
5074 */
5075 static void
5076 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag,
5077 boolean_t do_adapt)
5078 {
5079 arc_state_t *state = hdr->b_l1hdr.b_state;
5080 arc_buf_contents_t type = arc_buf_type(hdr);
5081
5082 if (do_adapt)
5083 arc_adapt(size, state);
5084
5085 /*
5086 * If arc_size is currently overflowing, and has grown past our
5087 * upper limit, we must be adding data faster than the evict
5088 * thread can evict. Thus, to ensure we don't compound the
5089 * problem by adding more data and forcing arc_size to grow even
5090 * further past its target size, we halt and wait for the
5091 * eviction thread to catch up.
5092 *
5093 * It's also possible that the reclaim thread is unable to evict
5094 * enough buffers to get arc_size below the overflow limit (e.g.
5095 * due to buffers being un-evictable, or hash lock collisions).
5096 * In this case, we want to proceed regardless if we're
5097 * overflowing; thus we don't use a while loop here.
5098 */
5099 if (arc_is_overflowing()) {
5100 mutex_enter(&arc_adjust_lock);
5101
5102 /*
5103 * Now that we've acquired the lock, we may no longer be
5104 * over the overflow limit, lets check.
5105 *
5106 * We're ignoring the case of spurious wake ups. If that
5107 * were to happen, it'd let this thread consume an ARC
5108 * buffer before it should have (i.e. before we're under
5109 * the overflow limit and were signalled by the reclaim
5110 * thread). As long as that is a rare occurrence, it
5111 * shouldn't cause any harm.
5112 */
5113 if (arc_is_overflowing()) {
5114 arc_adjust_needed = B_TRUE;
5115 zthr_wakeup(arc_adjust_zthr);
5116 (void) cv_wait(&arc_adjust_waiters_cv,
5117 &arc_adjust_lock);
5118 }
5119 mutex_exit(&arc_adjust_lock);
5120 }
5121
5122 VERIFY3U(hdr->b_type, ==, type);
5123 if (type == ARC_BUFC_METADATA) {
5124 arc_space_consume(size, ARC_SPACE_META);
5125 } else {
5126 arc_space_consume(size, ARC_SPACE_DATA);
5127 }
5128
5129 /*
5130 * Update the state size. Note that ghost states have a
5131 * "ghost size" and so don't need to be updated.
5132 */
5133 if (!GHOST_STATE(state)) {
5134
5135 (void) zfs_refcount_add_many(&state->arcs_size, size, tag);
5136
5137 /*
5138 * If this is reached via arc_read, the link is
5139 * protected by the hash lock. If reached via
5140 * arc_buf_alloc, the header should not be accessed by
5141 * any other thread. And, if reached via arc_read_done,
5142 * the hash lock will protect it if it's found in the
5143 * hash table; otherwise no other thread should be
5144 * trying to [add|remove]_reference it.
5145 */
5146 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5147 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5148 (void) zfs_refcount_add_many(&state->arcs_esize[type],
5149 size, tag);
5150 }
5151
5152 /*
5153 * If we are growing the cache, and we are adding anonymous
5154 * data, and we have outgrown arc_p, update arc_p
5155 */
5156 if (aggsum_compare(&arc_size, arc_c) < 0 &&
5157 hdr->b_l1hdr.b_state == arc_anon &&
5158 (zfs_refcount_count(&arc_anon->arcs_size) +
5159 zfs_refcount_count(&arc_mru->arcs_size) > arc_p))
5160 arc_p = MIN(arc_c, arc_p + size);
5161 }
5162 }
5163
5164 static void
5165 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5166 {
5167 arc_free_data_impl(hdr, size, tag);
5168 abd_free(abd);
5169 }
5170
5171 static void
5172 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5173 {
5174 arc_buf_contents_t type = arc_buf_type(hdr);
5175
5176 arc_free_data_impl(hdr, size, tag);
5177 if (type == ARC_BUFC_METADATA) {
5178 zio_buf_free(buf, size);
5179 } else {
5180 ASSERT(type == ARC_BUFC_DATA);
5181 zio_data_buf_free(buf, size);
5182 }
5183 }
5184
5185 /*
5186 * Free the arc data buffer.
5187 */
5188 static void
5189 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5190 {
5191 arc_state_t *state = hdr->b_l1hdr.b_state;
5192 arc_buf_contents_t type = arc_buf_type(hdr);
5193
5194 /* protected by hash lock, if in the hash table */
5195 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5196 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5197 ASSERT(state != arc_anon && state != arc_l2c_only);
5198
5199 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
5200 size, tag);
5201 }
5202 (void) zfs_refcount_remove_many(&state->arcs_size, size, tag);
5203
5204 VERIFY3U(hdr->b_type, ==, type);
5205 if (type == ARC_BUFC_METADATA) {
5206 arc_space_return(size, ARC_SPACE_META);
5207 } else {
5208 ASSERT(type == ARC_BUFC_DATA);
5209 arc_space_return(size, ARC_SPACE_DATA);
5210 }
5211 }
5212
5213 /*
5214 * This routine is called whenever a buffer is accessed.
5215 * NOTE: the hash lock is dropped in this function.
5216 */
5217 static void
5218 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5219 {
5220 clock_t now;
5221
5222 ASSERT(MUTEX_HELD(hash_lock));
5223 ASSERT(HDR_HAS_L1HDR(hdr));
5224
5225 if (hdr->b_l1hdr.b_state == arc_anon) {
5226 /*
5227 * This buffer is not in the cache, and does not
5228 * appear in our "ghost" list. Add the new buffer
5229 * to the MRU state.
5230 */
5231
5232 ASSERT0(hdr->b_l1hdr.b_arc_access);
5233 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5234 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5235 arc_change_state(arc_mru, hdr, hash_lock);
5236
5237 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5238 now = ddi_get_lbolt();
5239
5240 /*
5241 * If this buffer is here because of a prefetch, then either:
5242 * - clear the flag if this is a "referencing" read
5243 * (any subsequent access will bump this into the MFU state).
5244 * or
5245 * - move the buffer to the head of the list if this is
5246 * another prefetch (to make it less likely to be evicted).
5247 */
5248 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5249 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5250 /* link protected by hash lock */
5251 ASSERT(multilist_link_active(
5252 &hdr->b_l1hdr.b_arc_node));
5253 } else {
5254 if (HDR_HAS_L2HDR(hdr))
5255 l2arc_hdr_arcstats_decrement_state(hdr);
5256 arc_hdr_clear_flags(hdr,
5257 ARC_FLAG_PREFETCH |
5258 ARC_FLAG_PRESCIENT_PREFETCH);
5259 ARCSTAT_BUMP(arcstat_mru_hits);
5260 if (HDR_HAS_L2HDR(hdr))
5261 l2arc_hdr_arcstats_increment_state(hdr);
5262 }
5263 hdr->b_l1hdr.b_arc_access = now;
5264 return;
5265 }
5266
5267 /*
5268 * This buffer has been "accessed" only once so far,
5269 * but it is still in the cache. Move it to the MFU
5270 * state.
5271 */
5272 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
5273 /*
5274 * More than 125ms have passed since we
5275 * instantiated this buffer. Move it to the
5276 * most frequently used state.
5277 */
5278 hdr->b_l1hdr.b_arc_access = now;
5279 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5280 arc_change_state(arc_mfu, hdr, hash_lock);
5281 }
5282 ARCSTAT_BUMP(arcstat_mru_hits);
5283 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5284 arc_state_t *new_state;
5285 /*
5286 * This buffer has been "accessed" recently, but
5287 * was evicted from the cache. Move it to the
5288 * MFU state.
5289 */
5290 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5291 new_state = arc_mru;
5292 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5293 if (HDR_HAS_L2HDR(hdr))
5294 l2arc_hdr_arcstats_decrement_state(hdr);
5295 arc_hdr_clear_flags(hdr,
5296 ARC_FLAG_PREFETCH |
5297 ARC_FLAG_PRESCIENT_PREFETCH);
5298 if (HDR_HAS_L2HDR(hdr))
5299 l2arc_hdr_arcstats_increment_state(hdr);
5300 }
5301 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5302 } else {
5303 new_state = arc_mfu;
5304 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5305 }
5306
5307 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5308 arc_change_state(new_state, hdr, hash_lock);
5309
5310 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5311 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5312 /*
5313 * This buffer has been accessed more than once and is
5314 * still in the cache. Keep it in the MFU state.
5315 *
5316 * NOTE: an add_reference() that occurred when we did
5317 * the arc_read() will have kicked this off the list.
5318 * If it was a prefetch, we will explicitly move it to
5319 * the head of the list now.
5320 */
5321 ARCSTAT_BUMP(arcstat_mfu_hits);
5322 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5323 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5324 arc_state_t *new_state = arc_mfu;
5325 /*
5326 * This buffer has been accessed more than once but has
5327 * been evicted from the cache. Move it back to the
5328 * MFU state.
5329 */
5330
5331 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5332 /*
5333 * This is a prefetch access...
5334 * move this block back to the MRU state.
5335 */
5336 new_state = arc_mru;
5337 }
5338
5339 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5340 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5341 arc_change_state(new_state, hdr, hash_lock);
5342
5343 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5344 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5345 /*
5346 * This buffer is on the 2nd Level ARC.
5347 */
5348
5349 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5350 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5351 arc_change_state(arc_mfu, hdr, hash_lock);
5352 } else {
5353 ASSERT(!"invalid arc state");
5354 }
5355 }
5356
5357 /*
5358 * This routine is called by dbuf_hold() to update the arc_access() state
5359 * which otherwise would be skipped for entries in the dbuf cache.
5360 */
5361 void
5362 arc_buf_access(arc_buf_t *buf)
5363 {
5364 mutex_enter(&buf->b_evict_lock);
5365 arc_buf_hdr_t *hdr = buf->b_hdr;
5366
5367 /*
5368 * Avoid taking the hash_lock when possible as an optimization.
5369 * The header must be checked again under the hash_lock in order
5370 * to handle the case where it is concurrently being released.
5371 */
5372 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5373 mutex_exit(&buf->b_evict_lock);
5374 return;
5375 }
5376
5377 kmutex_t *hash_lock = HDR_LOCK(hdr);
5378 mutex_enter(hash_lock);
5379
5380 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5381 mutex_exit(hash_lock);
5382 mutex_exit(&buf->b_evict_lock);
5383 ARCSTAT_BUMP(arcstat_access_skip);
5384 return;
5385 }
5386
5387 mutex_exit(&buf->b_evict_lock);
5388
5389 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5390 hdr->b_l1hdr.b_state == arc_mfu);
5391
5392 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5393 arc_access(hdr, hash_lock);
5394 mutex_exit(hash_lock);
5395
5396 ARCSTAT_BUMP(arcstat_hits);
5397 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5398 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5399 }
5400
5401 /* a generic arc_read_done_func_t which you can use */
5402 /* ARGSUSED */
5403 void
5404 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5405 arc_buf_t *buf, void *arg)
5406 {
5407 if (buf == NULL)
5408 return;
5409
5410 bcopy(buf->b_data, arg, arc_buf_size(buf));
5411 arc_buf_destroy(buf, arg);
5412 }
5413
5414 /* a generic arc_read_done_func_t */
5415 void
5416 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5417 arc_buf_t *buf, void *arg)
5418 {
5419 arc_buf_t **bufp = arg;
5420
5421 if (buf == NULL) {
5422 ASSERT(zio == NULL || zio->io_error != 0);
5423 *bufp = NULL;
5424 } else {
5425 ASSERT(zio == NULL || zio->io_error == 0);
5426 *bufp = buf;
5427 ASSERT(buf->b_data != NULL);
5428 }
5429 }
5430
5431 static void
5432 arc_hdr_verify(arc_buf_hdr_t *hdr, const blkptr_t *bp)
5433 {
5434 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5435 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5436 ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5437 } else {
5438 if (HDR_COMPRESSION_ENABLED(hdr)) {
5439 ASSERT3U(arc_hdr_get_compress(hdr), ==,
5440 BP_GET_COMPRESS(bp));
5441 }
5442 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5443 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5444 ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5445 }
5446 }
5447
5448 /*
5449 * XXX this should be changed to return an error, and callers
5450 * re-read from disk on failure (on nondebug bits).
5451 */
5452 static void
5453 arc_hdr_verify_checksum(spa_t *spa, arc_buf_hdr_t *hdr, const blkptr_t *bp)
5454 {
5455 arc_hdr_verify(hdr, bp);
5456 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
5457 return;
5458 int err = 0;
5459 abd_t *abd = NULL;
5460 if (BP_IS_ENCRYPTED(bp)) {
5461 if (HDR_HAS_RABD(hdr)) {
5462 abd = hdr->b_crypt_hdr.b_rabd;
5463 }
5464 } else if (HDR_COMPRESSION_ENABLED(hdr)) {
5465 abd = hdr->b_l1hdr.b_pabd;
5466 }
5467 if (abd != NULL) {
5468 /*
5469 * The offset is only used for labels, which are not
5470 * cached in the ARC, so it doesn't matter what we
5471 * pass for the offset parameter.
5472 */
5473 int psize = HDR_GET_PSIZE(hdr);
5474 err = zio_checksum_error_impl(spa, bp,
5475 BP_GET_CHECKSUM(bp), abd, psize, 0, NULL);
5476 if (err != 0) {
5477 /*
5478 * Use abd_copy_to_buf() rather than
5479 * abd_borrow_buf_copy() so that we are sure to
5480 * include the buf in crash dumps.
5481 */
5482 void *buf = kmem_alloc(psize, KM_SLEEP);
5483 abd_copy_to_buf(buf, abd, psize);
5484 panic("checksum of cached data doesn't match BP "
5485 "err=%u hdr=%p bp=%p abd=%p buf=%p",
5486 err, (void *)hdr, (void *)bp, (void *)abd, buf);
5487 }
5488 }
5489 }
5490
5491 static void
5492 arc_read_done(zio_t *zio)
5493 {
5494 blkptr_t *bp = zio->io_bp;
5495 arc_buf_hdr_t *hdr = zio->io_private;
5496 kmutex_t *hash_lock = NULL;
5497 arc_callback_t *callback_list;
5498 arc_callback_t *acb;
5499 boolean_t freeable = B_FALSE;
5500
5501 /*
5502 * The hdr was inserted into hash-table and removed from lists
5503 * prior to starting I/O. We should find this header, since
5504 * it's in the hash table, and it should be legit since it's
5505 * not possible to evict it during the I/O. The only possible
5506 * reason for it not to be found is if we were freed during the
5507 * read.
5508 */
5509 if (HDR_IN_HASH_TABLE(hdr)) {
5510 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5511 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5512 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5513 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5514 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5515
5516 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5517 &hash_lock);
5518
5519 ASSERT((found == hdr &&
5520 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5521 (found == hdr && HDR_L2_READING(hdr)));
5522 ASSERT3P(hash_lock, !=, NULL);
5523 }
5524
5525 if (BP_IS_PROTECTED(bp)) {
5526 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5527 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5528 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5529 hdr->b_crypt_hdr.b_iv);
5530
5531 if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5532 void *tmpbuf;
5533
5534 tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5535 sizeof (zil_chain_t));
5536 zio_crypt_decode_mac_zil(tmpbuf,
5537 hdr->b_crypt_hdr.b_mac);
5538 abd_return_buf(zio->io_abd, tmpbuf,
5539 sizeof (zil_chain_t));
5540 } else {
5541 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
5542 }
5543 }
5544
5545 if (zio->io_error == 0) {
5546 /* byteswap if necessary */
5547 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5548 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5549 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5550 } else {
5551 hdr->b_l1hdr.b_byteswap =
5552 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5553 }
5554 } else {
5555 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5556 }
5557 }
5558
5559 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5560
5561 callback_list = hdr->b_l1hdr.b_acb;
5562 ASSERT3P(callback_list, !=, NULL);
5563
5564 if (hash_lock && zio->io_error == 0 &&
5565 hdr->b_l1hdr.b_state == arc_anon) {
5566 /*
5567 * Only call arc_access on anonymous buffers. This is because
5568 * if we've issued an I/O for an evicted buffer, we've already
5569 * called arc_access (to prevent any simultaneous readers from
5570 * getting confused).
5571 */
5572 arc_access(hdr, hash_lock);
5573 }
5574
5575 /*
5576 * If a read request has a callback (i.e. acb_done is not NULL), then we
5577 * make a buf containing the data according to the parameters which were
5578 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5579 * aren't needlessly decompressing the data multiple times.
5580 */
5581 int callback_cnt = 0;
5582 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5583 if (!acb->acb_done)
5584 continue;
5585
5586 callback_cnt++;
5587
5588 if (zio->io_error != 0)
5589 continue;
5590
5591 int error = arc_buf_alloc_impl(hdr, zio->io_spa,
5592 &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
5593 acb->acb_compressed, acb->acb_noauth, B_TRUE,
5594 &acb->acb_buf);
5595
5596 /*
5597 * Assert non-speculative zios didn't fail because an
5598 * encryption key wasn't loaded
5599 */
5600 ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
5601 error != EACCES);
5602
5603 /*
5604 * If we failed to decrypt, report an error now (as the zio
5605 * layer would have done if it had done the transforms).
5606 */
5607 if (error == ECKSUM) {
5608 ASSERT(BP_IS_PROTECTED(bp));
5609 error = SET_ERROR(EIO);
5610 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5611 spa_log_error(zio->io_spa, &acb->acb_zb);
5612 (void) zfs_ereport_post(
5613 FM_EREPORT_ZFS_AUTHENTICATION,
5614 zio->io_spa, NULL, &acb->acb_zb, zio, 0, 0);
5615 }
5616 }
5617
5618 if (error != 0) {
5619 /*
5620 * Decompression failed. Set io_error
5621 * so that when we call acb_done (below),
5622 * we will indicate that the read failed.
5623 * Note that in the unusual case where one
5624 * callback is compressed and another
5625 * uncompressed, we will mark all of them
5626 * as failed, even though the uncompressed
5627 * one can't actually fail. In this case,
5628 * the hdr will not be anonymous, because
5629 * if there are multiple callbacks, it's
5630 * because multiple threads found the same
5631 * arc buf in the hash table.
5632 */
5633 zio->io_error = error;
5634 }
5635 }
5636
5637 /*
5638 * If there are multiple callbacks, we must have the hash lock,
5639 * because the only way for multiple threads to find this hdr is
5640 * in the hash table. This ensures that if there are multiple
5641 * callbacks, the hdr is not anonymous. If it were anonymous,
5642 * we couldn't use arc_buf_destroy() in the error case below.
5643 */
5644 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5645
5646 hdr->b_l1hdr.b_acb = NULL;
5647 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5648 if (callback_cnt == 0)
5649 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
5650
5651 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5652 callback_list != NULL);
5653
5654 if (zio->io_error == 0) {
5655 arc_hdr_verify(hdr, zio->io_bp);
5656 } else {
5657 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5658 if (hdr->b_l1hdr.b_state != arc_anon)
5659 arc_change_state(arc_anon, hdr, hash_lock);
5660 if (HDR_IN_HASH_TABLE(hdr))
5661 buf_hash_remove(hdr);
5662 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5663 }
5664
5665 /*
5666 * Broadcast before we drop the hash_lock to avoid the possibility
5667 * that the hdr (and hence the cv) might be freed before we get to
5668 * the cv_broadcast().
5669 */
5670 cv_broadcast(&hdr->b_l1hdr.b_cv);
5671
5672 if (hash_lock != NULL) {
5673 mutex_exit(hash_lock);
5674 } else {
5675 /*
5676 * This block was freed while we waited for the read to
5677 * complete. It has been removed from the hash table and
5678 * moved to the anonymous state (so that it won't show up
5679 * in the cache).
5680 */
5681 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5682 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5683 }
5684
5685 /* execute each callback and free its structure */
5686 while ((acb = callback_list) != NULL) {
5687
5688 if (acb->acb_done != NULL) {
5689 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5690 /*
5691 * If arc_buf_alloc_impl() fails during
5692 * decompression, the buf will still be
5693 * allocated, and needs to be freed here.
5694 */
5695 arc_buf_destroy(acb->acb_buf, acb->acb_private);
5696 acb->acb_buf = NULL;
5697 }
5698 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5699 acb->acb_buf, acb->acb_private);
5700 }
5701
5702 if (acb->acb_zio_dummy != NULL) {
5703 acb->acb_zio_dummy->io_error = zio->io_error;
5704 zio_nowait(acb->acb_zio_dummy);
5705 }
5706
5707 callback_list = acb->acb_next;
5708 kmem_free(acb, sizeof (arc_callback_t));
5709 }
5710
5711 if (freeable)
5712 arc_hdr_destroy(hdr);
5713 }
5714
5715 /*
5716 * "Read" the block at the specified DVA (in bp) via the
5717 * cache. If the block is found in the cache, invoke the provided
5718 * callback immediately and return. Note that the `zio' parameter
5719 * in the callback will be NULL in this case, since no IO was
5720 * required. If the block is not in the cache pass the read request
5721 * on to the spa with a substitute callback function, so that the
5722 * requested block will be added to the cache.
5723 *
5724 * If a read request arrives for a block that has a read in-progress,
5725 * either wait for the in-progress read to complete (and return the
5726 * results); or, if this is a read with a "done" func, add a record
5727 * to the read to invoke the "done" func when the read completes,
5728 * and return; or just return.
5729 *
5730 * arc_read_done() will invoke all the requested "done" functions
5731 * for readers of this block.
5732 */
5733 int
5734 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_read_done_func_t *done,
5735 void *private, zio_priority_t priority, int zio_flags,
5736 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5737 {
5738 arc_buf_hdr_t *hdr = NULL;
5739 kmutex_t *hash_lock = NULL;
5740 zio_t *rzio;
5741 uint64_t guid = spa_load_guid(spa);
5742 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
5743 boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
5744 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5745 boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
5746 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5747 int rc = 0;
5748
5749 ASSERT(!BP_IS_EMBEDDED(bp) ||
5750 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5751
5752 top:
5753 if (!BP_IS_EMBEDDED(bp)) {
5754 /*
5755 * Embedded BP's have no DVA and require no I/O to "read".
5756 * Create an anonymous arc buf to back it.
5757 */
5758 hdr = buf_hash_find(guid, bp, &hash_lock);
5759 }
5760
5761 /*
5762 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5763 * we maintain encrypted data seperately from compressed / uncompressed
5764 * data. If the user is requesting raw encrypted data and we don't have
5765 * that in the header we will read from disk to guarantee that we can
5766 * get it even if the encryption keys aren't loaded.
5767 */
5768 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
5769 (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
5770 arc_buf_t *buf = NULL;
5771 *arc_flags |= ARC_FLAG_CACHED;
5772
5773 if (HDR_IO_IN_PROGRESS(hdr)) {
5774 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5775
5776 ASSERT3P(head_zio, !=, NULL);
5777 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5778 priority == ZIO_PRIORITY_SYNC_READ) {
5779 /*
5780 * This is a sync read that needs to wait for
5781 * an in-flight async read. Request that the
5782 * zio have its priority upgraded.
5783 */
5784 zio_change_priority(head_zio, priority);
5785 DTRACE_PROBE1(arc__async__upgrade__sync,
5786 arc_buf_hdr_t *, hdr);
5787 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5788 }
5789 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5790 arc_hdr_clear_flags(hdr,
5791 ARC_FLAG_PREDICTIVE_PREFETCH);
5792 }
5793
5794 if (*arc_flags & ARC_FLAG_WAIT) {
5795 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5796 mutex_exit(hash_lock);
5797 goto top;
5798 }
5799 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5800
5801 if (done) {
5802 arc_callback_t *acb = NULL;
5803
5804 acb = kmem_zalloc(sizeof (arc_callback_t),
5805 KM_SLEEP);
5806 acb->acb_done = done;
5807 acb->acb_private = private;
5808 acb->acb_compressed = compressed_read;
5809 acb->acb_encrypted = encrypted_read;
5810 acb->acb_noauth = noauth_read;
5811 acb->acb_zb = *zb;
5812 if (pio != NULL)
5813 acb->acb_zio_dummy = zio_null(pio,
5814 spa, NULL, NULL, NULL, zio_flags);
5815
5816 ASSERT3P(acb->acb_done, !=, NULL);
5817 acb->acb_zio_head = head_zio;
5818 acb->acb_next = hdr->b_l1hdr.b_acb;
5819 hdr->b_l1hdr.b_acb = acb;
5820 mutex_exit(hash_lock);
5821 return (0);
5822 }
5823 mutex_exit(hash_lock);
5824 return (0);
5825 }
5826
5827 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5828 hdr->b_l1hdr.b_state == arc_mfu);
5829
5830 if (done) {
5831 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5832 /*
5833 * This is a demand read which does not have to
5834 * wait for i/o because we did a predictive
5835 * prefetch i/o for it, which has completed.
5836 */
5837 DTRACE_PROBE1(
5838 arc__demand__hit__predictive__prefetch,
5839 arc_buf_hdr_t *, hdr);
5840 ARCSTAT_BUMP(
5841 arcstat_demand_hit_predictive_prefetch);
5842 arc_hdr_clear_flags(hdr,
5843 ARC_FLAG_PREDICTIVE_PREFETCH);
5844 }
5845
5846 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
5847 ARCSTAT_BUMP(
5848 arcstat_demand_hit_prescient_prefetch);
5849 arc_hdr_clear_flags(hdr,
5850 ARC_FLAG_PRESCIENT_PREFETCH);
5851 }
5852
5853 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5854
5855 arc_hdr_verify_checksum(spa, hdr, bp);
5856
5857 /* Get a buf with the desired data in it. */
5858 rc = arc_buf_alloc_impl(hdr, spa, zb, private,
5859 encrypted_read, compressed_read, noauth_read,
5860 B_TRUE, &buf);
5861 if (rc == ECKSUM) {
5862 /*
5863 * Convert authentication and decryption errors
5864 * to EIO (and generate an ereport if needed)
5865 * before leaving the ARC.
5866 */
5867 rc = SET_ERROR(EIO);
5868 if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5869 spa_log_error(spa, zb);
5870 (void) zfs_ereport_post(
5871 FM_EREPORT_ZFS_AUTHENTICATION,
5872 spa, NULL, zb, NULL, 0, 0);
5873 }
5874 }
5875 if (rc != 0) {
5876 (void) remove_reference(hdr, hash_lock,
5877 private);
5878 arc_buf_destroy_impl(buf);
5879 buf = NULL;
5880 }
5881 /* assert any errors weren't due to unloaded keys */
5882 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
5883 rc != EACCES);
5884 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5885 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
5886 if (HDR_HAS_L2HDR(hdr))
5887 l2arc_hdr_arcstats_decrement_state(hdr);
5888 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5889 if (HDR_HAS_L2HDR(hdr))
5890 l2arc_hdr_arcstats_increment_state(hdr);
5891 }
5892 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5893 arc_access(hdr, hash_lock);
5894 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5895 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5896 if (*arc_flags & ARC_FLAG_L2CACHE)
5897 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5898 mutex_exit(hash_lock);
5899 ARCSTAT_BUMP(arcstat_hits);
5900 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5901 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5902 data, metadata, hits);
5903
5904 if (done)
5905 done(NULL, zb, bp, buf, private);
5906 } else {
5907 uint64_t lsize = BP_GET_LSIZE(bp);
5908 uint64_t psize = BP_GET_PSIZE(bp);
5909 arc_callback_t *acb;
5910 vdev_t *vd = NULL;
5911 uint64_t addr = 0;
5912 boolean_t devw = B_FALSE;
5913 uint64_t size;
5914 abd_t *hdr_abd;
5915 int alloc_flags = encrypted_read ? ARC_HDR_ALLOC_RDATA : 0;
5916
5917 if (hdr == NULL) {
5918 /* this block is not in the cache */
5919 arc_buf_hdr_t *exists = NULL;
5920 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5921 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5922 BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), type,
5923 encrypted_read);
5924
5925 if (!BP_IS_EMBEDDED(bp)) {
5926 hdr->b_dva = *BP_IDENTITY(bp);
5927 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5928 exists = buf_hash_insert(hdr, &hash_lock);
5929 }
5930 if (exists != NULL) {
5931 /* somebody beat us to the hash insert */
5932 mutex_exit(hash_lock);
5933 buf_discard_identity(hdr);
5934 arc_hdr_destroy(hdr);
5935 goto top; /* restart the IO request */
5936 }
5937 } else {
5938 /*
5939 * This block is in the ghost cache or encrypted data
5940 * was requested and we didn't have it. If it was
5941 * L2-only (and thus didn't have an L1 hdr),
5942 * we realloc the header to add an L1 hdr.
5943 */
5944 if (!HDR_HAS_L1HDR(hdr)) {
5945 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5946 hdr_full_cache);
5947 }
5948
5949 if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
5950 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5951 ASSERT(!HDR_HAS_RABD(hdr));
5952 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5953 ASSERT0(zfs_refcount_count(
5954 &hdr->b_l1hdr.b_refcnt));
5955 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5956 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5957 } else if (HDR_IO_IN_PROGRESS(hdr)) {
5958 /*
5959 * If this header already had an IO in progress
5960 * and we are performing another IO to fetch
5961 * encrypted data we must wait until the first
5962 * IO completes so as not to confuse
5963 * arc_read_done(). This should be very rare
5964 * and so the performance impact shouldn't
5965 * matter.
5966 */
5967 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5968 mutex_exit(hash_lock);
5969 goto top;
5970 }
5971
5972 /*
5973 * This is a delicate dance that we play here.
5974 * This hdr might be in the ghost list so we access
5975 * it to move it out of the ghost list before we
5976 * initiate the read. If it's a prefetch then
5977 * it won't have a callback so we'll remove the
5978 * reference that arc_buf_alloc_impl() created. We
5979 * do this after we've called arc_access() to
5980 * avoid hitting an assert in remove_reference().
5981 */
5982 arc_adapt(arc_hdr_size(hdr), hdr->b_l1hdr.b_state);
5983 arc_access(hdr, hash_lock);
5984 arc_hdr_alloc_pabd(hdr, alloc_flags);
5985 }
5986
5987 if (encrypted_read) {
5988 ASSERT(HDR_HAS_RABD(hdr));
5989 size = HDR_GET_PSIZE(hdr);
5990 hdr_abd = hdr->b_crypt_hdr.b_rabd;
5991 zio_flags |= ZIO_FLAG_RAW;
5992 } else {
5993 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5994 size = arc_hdr_size(hdr);
5995 hdr_abd = hdr->b_l1hdr.b_pabd;
5996
5997 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
5998 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
5999 }
6000
6001 /*
6002 * For authenticated bp's, we do not ask the ZIO layer
6003 * to authenticate them since this will cause the entire
6004 * IO to fail if the key isn't loaded. Instead, we
6005 * defer authentication until arc_buf_fill(), which will
6006 * verify the data when the key is available.
6007 */
6008 if (BP_IS_AUTHENTICATED(bp))
6009 zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
6010 }
6011
6012 if (*arc_flags & ARC_FLAG_PREFETCH &&
6013 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6014 if (HDR_HAS_L2HDR(hdr))
6015 l2arc_hdr_arcstats_decrement_state(hdr);
6016 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
6017 if (HDR_HAS_L2HDR(hdr))
6018 l2arc_hdr_arcstats_increment_state(hdr);
6019 }
6020 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
6021 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
6022
6023 if (*arc_flags & ARC_FLAG_L2CACHE)
6024 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6025 if (BP_IS_AUTHENTICATED(bp))
6026 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6027 if (BP_GET_LEVEL(bp) > 0)
6028 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
6029 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
6030 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
6031 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
6032
6033 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
6034 acb->acb_done = done;
6035 acb->acb_private = private;
6036 acb->acb_compressed = compressed_read;
6037 acb->acb_encrypted = encrypted_read;
6038 acb->acb_noauth = noauth_read;
6039 acb->acb_zb = *zb;
6040
6041 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6042 hdr->b_l1hdr.b_acb = acb;
6043 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6044
6045 if (HDR_HAS_L2HDR(hdr) &&
6046 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
6047 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
6048 addr = hdr->b_l2hdr.b_daddr;
6049 /*
6050 * Lock out L2ARC device removal.
6051 */
6052 if (vdev_is_dead(vd) ||
6053 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
6054 vd = NULL;
6055 }
6056
6057 /*
6058 * We count both async reads and scrub IOs as asynchronous so
6059 * that both can be upgraded in the event of a cache hit while
6060 * the read IO is still in-flight.
6061 */
6062 if (priority == ZIO_PRIORITY_ASYNC_READ ||
6063 priority == ZIO_PRIORITY_SCRUB)
6064 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6065 else
6066 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
6067
6068 /*
6069 * At this point, we have a level 1 cache miss. Try again in
6070 * L2ARC if possible.
6071 */
6072 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
6073
6074 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
6075 uint64_t, lsize, zbookmark_phys_t *, zb);
6076 ARCSTAT_BUMP(arcstat_misses);
6077 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
6078 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
6079 data, metadata, misses);
6080
6081 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
6082 /*
6083 * Read from the L2ARC if the following are true:
6084 * 1. The L2ARC vdev was previously cached.
6085 * 2. This buffer still has L2ARC metadata.
6086 * 3. This buffer isn't currently writing to the L2ARC.
6087 * 4. The L2ARC entry wasn't evicted, which may
6088 * also have invalidated the vdev.
6089 * 5. This isn't prefetch or l2arc_noprefetch is 0.
6090 */
6091 if (HDR_HAS_L2HDR(hdr) &&
6092 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
6093 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
6094 l2arc_read_callback_t *cb;
6095 abd_t *abd;
6096 uint64_t asize;
6097
6098 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
6099 ARCSTAT_BUMP(arcstat_l2_hits);
6100
6101 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
6102 KM_SLEEP);
6103 cb->l2rcb_hdr = hdr;
6104 cb->l2rcb_bp = *bp;
6105 cb->l2rcb_zb = *zb;
6106 cb->l2rcb_flags = zio_flags;
6107
6108 /*
6109 * When Compressed ARC is disabled, but the
6110 * L2ARC block is compressed, arc_hdr_size()
6111 * will have returned LSIZE rather than PSIZE.
6112 */
6113 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
6114 !HDR_COMPRESSION_ENABLED(hdr) &&
6115 HDR_GET_PSIZE(hdr) != 0) {
6116 size = HDR_GET_PSIZE(hdr);
6117 }
6118
6119 asize = vdev_psize_to_asize(vd, size);
6120 if (asize != size) {
6121 abd = abd_alloc_for_io(asize,
6122 HDR_ISTYPE_METADATA(hdr));
6123 cb->l2rcb_abd = abd;
6124 } else {
6125 abd = hdr_abd;
6126 }
6127
6128 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
6129 addr + asize <= vd->vdev_psize -
6130 VDEV_LABEL_END_SIZE);
6131
6132 /*
6133 * l2arc read. The SCL_L2ARC lock will be
6134 * released by l2arc_read_done().
6135 * Issue a null zio if the underlying buffer
6136 * was squashed to zero size by compression.
6137 */
6138 ASSERT3U(arc_hdr_get_compress(hdr), !=,
6139 ZIO_COMPRESS_EMPTY);
6140 rzio = zio_read_phys(pio, vd, addr,
6141 asize, abd,
6142 ZIO_CHECKSUM_OFF,
6143 l2arc_read_done, cb, priority,
6144 zio_flags | ZIO_FLAG_DONT_CACHE |
6145 ZIO_FLAG_CANFAIL |
6146 ZIO_FLAG_DONT_PROPAGATE |
6147 ZIO_FLAG_DONT_RETRY, B_FALSE);
6148 acb->acb_zio_head = rzio;
6149
6150 if (hash_lock != NULL)
6151 mutex_exit(hash_lock);
6152
6153 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
6154 zio_t *, rzio);
6155 ARCSTAT_INCR(arcstat_l2_read_bytes,
6156 HDR_GET_PSIZE(hdr));
6157
6158 if (*arc_flags & ARC_FLAG_NOWAIT) {
6159 zio_nowait(rzio);
6160 return (0);
6161 }
6162
6163 ASSERT(*arc_flags & ARC_FLAG_WAIT);
6164 if (zio_wait(rzio) == 0)
6165 return (0);
6166
6167 /* l2arc read error; goto zio_read() */
6168 if (hash_lock != NULL)
6169 mutex_enter(hash_lock);
6170 } else {
6171 DTRACE_PROBE1(l2arc__miss,
6172 arc_buf_hdr_t *, hdr);
6173 ARCSTAT_BUMP(arcstat_l2_misses);
6174 if (HDR_L2_WRITING(hdr))
6175 ARCSTAT_BUMP(arcstat_l2_rw_clash);
6176 spa_config_exit(spa, SCL_L2ARC, vd);
6177 }
6178 } else {
6179 if (vd != NULL)
6180 spa_config_exit(spa, SCL_L2ARC, vd);
6181 if (l2arc_ndev != 0) {
6182 DTRACE_PROBE1(l2arc__miss,
6183 arc_buf_hdr_t *, hdr);
6184 ARCSTAT_BUMP(arcstat_l2_misses);
6185 }
6186 }
6187
6188 rzio = zio_read(pio, spa, bp, hdr_abd, size,
6189 arc_read_done, hdr, priority, zio_flags, zb);
6190 acb->acb_zio_head = rzio;
6191
6192 if (hash_lock != NULL)
6193 mutex_exit(hash_lock);
6194
6195 /*
6196 * At this point, this read I/O has already missed in the ARC
6197 * and will be going through to the disk. The I/O throttle
6198 * should delay this I/O if this zone is using more than its I/O
6199 * priority allows.
6200 */
6201 zfs_zone_io_throttle(ZFS_ZONE_IOP_READ);
6202
6203 if (*arc_flags & ARC_FLAG_WAIT)
6204 return (zio_wait(rzio));
6205
6206 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
6207 zio_nowait(rzio);
6208 }
6209 return (rc);
6210 }
6211
6212 /*
6213 * Notify the arc that a block was freed, and thus will never be used again.
6214 */
6215 void
6216 arc_freed(spa_t *spa, const blkptr_t *bp)
6217 {
6218 arc_buf_hdr_t *hdr;
6219 kmutex_t *hash_lock;
6220 uint64_t guid = spa_load_guid(spa);
6221
6222 ASSERT(!BP_IS_EMBEDDED(bp));
6223
6224 hdr = buf_hash_find(guid, bp, &hash_lock);
6225 if (hdr == NULL)
6226 return;
6227
6228 /*
6229 * We might be trying to free a block that is still doing I/O
6230 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
6231 * dmu_sync-ed block). If this block is being prefetched, then it
6232 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
6233 * until the I/O completes. A block may also have a reference if it is
6234 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6235 * have written the new block to its final resting place on disk but
6236 * without the dedup flag set. This would have left the hdr in the MRU
6237 * state and discoverable. When the txg finally syncs it detects that
6238 * the block was overridden in open context and issues an override I/O.
6239 * Since this is a dedup block, the override I/O will determine if the
6240 * block is already in the DDT. If so, then it will replace the io_bp
6241 * with the bp from the DDT and allow the I/O to finish. When the I/O
6242 * reaches the done callback, dbuf_write_override_done, it will
6243 * check to see if the io_bp and io_bp_override are identical.
6244 * If they are not, then it indicates that the bp was replaced with
6245 * the bp in the DDT and the override bp is freed. This allows
6246 * us to arrive here with a reference on a block that is being
6247 * freed. So if we have an I/O in progress, or a reference to
6248 * this hdr, then we don't destroy the hdr.
6249 */
6250 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6251 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6252 arc_change_state(arc_anon, hdr, hash_lock);
6253 arc_hdr_destroy(hdr);
6254 mutex_exit(hash_lock);
6255 } else {
6256 mutex_exit(hash_lock);
6257 }
6258
6259 }
6260
6261 /*
6262 * Release this buffer from the cache, making it an anonymous buffer. This
6263 * must be done after a read and prior to modifying the buffer contents.
6264 * If the buffer has more than one reference, we must make
6265 * a new hdr for the buffer.
6266 */
6267 void
6268 arc_release(arc_buf_t *buf, void *tag)
6269 {
6270 arc_buf_hdr_t *hdr = buf->b_hdr;
6271
6272 /*
6273 * It would be nice to assert that if its DMU metadata (level >
6274 * 0 || it's the dnode file), then it must be syncing context.
6275 * But we don't know that information at this level.
6276 */
6277
6278 mutex_enter(&buf->b_evict_lock);
6279
6280 ASSERT(HDR_HAS_L1HDR(hdr));
6281
6282 /*
6283 * We don't grab the hash lock prior to this check, because if
6284 * the buffer's header is in the arc_anon state, it won't be
6285 * linked into the hash table.
6286 */
6287 if (hdr->b_l1hdr.b_state == arc_anon) {
6288 mutex_exit(&buf->b_evict_lock);
6289 /*
6290 * If we are called from dmu_convert_mdn_block_to_raw(),
6291 * a write might be in progress. This is OK because
6292 * the caller won't change the content of this buffer,
6293 * only the flags (via arc_convert_to_raw()).
6294 */
6295 /* ASSERT(!HDR_IO_IN_PROGRESS(hdr)); */
6296 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6297 ASSERT(!HDR_HAS_L2HDR(hdr));
6298 ASSERT(HDR_EMPTY(hdr));
6299
6300 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6301 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6302 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6303
6304 hdr->b_l1hdr.b_arc_access = 0;
6305
6306 /*
6307 * If the buf is being overridden then it may already
6308 * have a hdr that is not empty.
6309 */
6310 buf_discard_identity(hdr);
6311 arc_buf_thaw(buf);
6312
6313 return;
6314 }
6315
6316 kmutex_t *hash_lock = HDR_LOCK(hdr);
6317 mutex_enter(hash_lock);
6318
6319 /*
6320 * This assignment is only valid as long as the hash_lock is
6321 * held, we must be careful not to reference state or the
6322 * b_state field after dropping the lock.
6323 */
6324 arc_state_t *state = hdr->b_l1hdr.b_state;
6325 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6326 ASSERT3P(state, !=, arc_anon);
6327
6328 /* this buffer is not on any list */
6329 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6330
6331 if (HDR_HAS_L2HDR(hdr)) {
6332 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6333
6334 /*
6335 * We have to recheck this conditional again now that
6336 * we're holding the l2ad_mtx to prevent a race with
6337 * another thread which might be concurrently calling
6338 * l2arc_evict(). In that case, l2arc_evict() might have
6339 * destroyed the header's L2 portion as we were waiting
6340 * to acquire the l2ad_mtx.
6341 */
6342 if (HDR_HAS_L2HDR(hdr))
6343 arc_hdr_l2hdr_destroy(hdr);
6344
6345 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6346 }
6347
6348 /*
6349 * Do we have more than one buf?
6350 */
6351 if (hdr->b_l1hdr.b_bufcnt > 1) {
6352 arc_buf_hdr_t *nhdr;
6353 uint64_t spa = hdr->b_spa;
6354 uint64_t psize = HDR_GET_PSIZE(hdr);
6355 uint64_t lsize = HDR_GET_LSIZE(hdr);
6356 boolean_t protected = HDR_PROTECTED(hdr);
6357 enum zio_compress compress = arc_hdr_get_compress(hdr);
6358 arc_buf_contents_t type = arc_buf_type(hdr);
6359 VERIFY3U(hdr->b_type, ==, type);
6360
6361 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6362 (void) remove_reference(hdr, hash_lock, tag);
6363
6364 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6365 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6366 ASSERT(ARC_BUF_LAST(buf));
6367 }
6368
6369 /*
6370 * Pull the data off of this hdr and attach it to
6371 * a new anonymous hdr. Also find the last buffer
6372 * in the hdr's buffer list.
6373 */
6374 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6375 ASSERT3P(lastbuf, !=, NULL);
6376
6377 /*
6378 * If the current arc_buf_t and the hdr are sharing their data
6379 * buffer, then we must stop sharing that block.
6380 */
6381 if (arc_buf_is_shared(buf)) {
6382 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6383 VERIFY(!arc_buf_is_shared(lastbuf));
6384
6385 /*
6386 * First, sever the block sharing relationship between
6387 * buf and the arc_buf_hdr_t.
6388 */
6389 arc_unshare_buf(hdr, buf);
6390
6391 /*
6392 * Now we need to recreate the hdr's b_pabd. Since we
6393 * have lastbuf handy, we try to share with it, but if
6394 * we can't then we allocate a new b_pabd and copy the
6395 * data from buf into it.
6396 */
6397 if (arc_can_share(hdr, lastbuf)) {
6398 arc_share_buf(hdr, lastbuf);
6399 } else {
6400 arc_hdr_alloc_pabd(hdr, ARC_HDR_DO_ADAPT);
6401 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6402 buf->b_data, psize);
6403 }
6404 VERIFY3P(lastbuf->b_data, !=, NULL);
6405 } else if (HDR_SHARED_DATA(hdr)) {
6406 /*
6407 * Uncompressed shared buffers are always at the end
6408 * of the list. Compressed buffers don't have the
6409 * same requirements. This makes it hard to
6410 * simply assert that the lastbuf is shared so
6411 * we rely on the hdr's compression flags to determine
6412 * if we have a compressed, shared buffer.
6413 */
6414 ASSERT(arc_buf_is_shared(lastbuf) ||
6415 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6416 ASSERT(!ARC_BUF_SHARED(buf));
6417 }
6418 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6419 ASSERT3P(state, !=, arc_l2c_only);
6420
6421 (void) zfs_refcount_remove_many(&state->arcs_size,
6422 arc_buf_size(buf), buf);
6423
6424 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6425 ASSERT3P(state, !=, arc_l2c_only);
6426 (void) zfs_refcount_remove_many(
6427 &state->arcs_esize[type],
6428 arc_buf_size(buf), buf);
6429 }
6430
6431 hdr->b_l1hdr.b_bufcnt -= 1;
6432 if (ARC_BUF_ENCRYPTED(buf))
6433 hdr->b_crypt_hdr.b_ebufcnt -= 1;
6434
6435 arc_cksum_verify(buf);
6436 arc_buf_unwatch(buf);
6437
6438 /* if this is the last uncompressed buf free the checksum */
6439 if (!arc_hdr_has_uncompressed_buf(hdr))
6440 arc_cksum_free(hdr);
6441
6442 mutex_exit(hash_lock);
6443
6444 /*
6445 * Allocate a new hdr. The new hdr will contain a b_pabd
6446 * buffer which will be freed in arc_write().
6447 */
6448 nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6449 compress, type, HDR_HAS_RABD(hdr));
6450 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6451 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6452 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
6453 VERIFY3U(nhdr->b_type, ==, type);
6454 ASSERT(!HDR_SHARED_DATA(nhdr));
6455
6456 nhdr->b_l1hdr.b_buf = buf;
6457 nhdr->b_l1hdr.b_bufcnt = 1;
6458 if (ARC_BUF_ENCRYPTED(buf))
6459 nhdr->b_crypt_hdr.b_ebufcnt = 1;
6460 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6461 buf->b_hdr = nhdr;
6462
6463 mutex_exit(&buf->b_evict_lock);
6464 (void) zfs_refcount_add_many(&arc_anon->arcs_size,
6465 arc_buf_size(buf), buf);
6466 } else {
6467 mutex_exit(&buf->b_evict_lock);
6468 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6469 /* protected by hash lock, or hdr is on arc_anon */
6470 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6471 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6472 arc_change_state(arc_anon, hdr, hash_lock);
6473 hdr->b_l1hdr.b_arc_access = 0;
6474
6475 mutex_exit(hash_lock);
6476 buf_discard_identity(hdr);
6477 arc_buf_thaw(buf);
6478 }
6479 }
6480
6481 int
6482 arc_released(arc_buf_t *buf)
6483 {
6484 int released;
6485
6486 mutex_enter(&buf->b_evict_lock);
6487 released = (buf->b_data != NULL &&
6488 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6489 mutex_exit(&buf->b_evict_lock);
6490 return (released);
6491 }
6492
6493 #ifdef ZFS_DEBUG
6494 int
6495 arc_referenced(arc_buf_t *buf)
6496 {
6497 int referenced;
6498
6499 mutex_enter(&buf->b_evict_lock);
6500 referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6501 mutex_exit(&buf->b_evict_lock);
6502 return (referenced);
6503 }
6504 #endif
6505
6506 static void
6507 arc_write_ready(zio_t *zio)
6508 {
6509 arc_write_callback_t *callback = zio->io_private;
6510 arc_buf_t *buf = callback->awcb_buf;
6511 arc_buf_hdr_t *hdr = buf->b_hdr;
6512 blkptr_t *bp = zio->io_bp;
6513 uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6514
6515 ASSERT(HDR_HAS_L1HDR(hdr));
6516 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6517 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6518
6519 /*
6520 * If we're reexecuting this zio because the pool suspended, then
6521 * cleanup any state that was previously set the first time the
6522 * callback was invoked.
6523 */
6524 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6525 arc_cksum_free(hdr);
6526 arc_buf_unwatch(buf);
6527 if (hdr->b_l1hdr.b_pabd != NULL) {
6528 if (arc_buf_is_shared(buf)) {
6529 arc_unshare_buf(hdr, buf);
6530 } else {
6531 arc_hdr_free_pabd(hdr, B_FALSE);
6532 }
6533 }
6534
6535 if (HDR_HAS_RABD(hdr))
6536 arc_hdr_free_pabd(hdr, B_TRUE);
6537 }
6538 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6539 ASSERT(!HDR_HAS_RABD(hdr));
6540 ASSERT(!HDR_SHARED_DATA(hdr));
6541 ASSERT(!arc_buf_is_shared(buf));
6542
6543 callback->awcb_ready(zio, buf, callback->awcb_private);
6544
6545 if (HDR_IO_IN_PROGRESS(hdr))
6546 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6547
6548 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6549
6550 if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr))
6551 hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp));
6552
6553 if (BP_IS_PROTECTED(bp)) {
6554 /* ZIL blocks are written through zio_rewrite */
6555 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
6556 ASSERT(HDR_PROTECTED(hdr));
6557
6558 if (BP_SHOULD_BYTESWAP(bp)) {
6559 if (BP_GET_LEVEL(bp) > 0) {
6560 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
6561 } else {
6562 hdr->b_l1hdr.b_byteswap =
6563 DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
6564 }
6565 } else {
6566 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
6567 }
6568
6569 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
6570 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
6571 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
6572 hdr->b_crypt_hdr.b_iv);
6573 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
6574 }
6575
6576 /*
6577 * If this block was written for raw encryption but the zio layer
6578 * ended up only authenticating it, adjust the buffer flags now.
6579 */
6580 if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
6581 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6582 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6583 if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
6584 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6585 } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
6586 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6587 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6588 }
6589
6590 /* this must be done after the buffer flags are adjusted */
6591 arc_cksum_compute(buf);
6592
6593 enum zio_compress compress;
6594 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
6595 compress = ZIO_COMPRESS_OFF;
6596 } else {
6597 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
6598 compress = BP_GET_COMPRESS(bp);
6599 }
6600 HDR_SET_PSIZE(hdr, psize);
6601 arc_hdr_set_compress(hdr, compress);
6602
6603 if (zio->io_error != 0 || psize == 0)
6604 goto out;
6605
6606 /*
6607 * Fill the hdr with data. If the buffer is encrypted we have no choice
6608 * but to copy the data into b_rabd. If the hdr is compressed, the data
6609 * we want is available from the zio, otherwise we can take it from
6610 * the buf.
6611 *
6612 * We might be able to share the buf's data with the hdr here. However,
6613 * doing so would cause the ARC to be full of linear ABDs if we write a
6614 * lot of shareable data. As a compromise, we check whether scattered
6615 * ABDs are allowed, and assume that if they are then the user wants
6616 * the ARC to be primarily filled with them regardless of the data being
6617 * written. Therefore, if they're allowed then we allocate one and copy
6618 * the data into it; otherwise, we share the data directly if we can.
6619 */
6620 if (ARC_BUF_ENCRYPTED(buf)) {
6621 ASSERT3U(psize, >, 0);
6622 ASSERT(ARC_BUF_COMPRESSED(buf));
6623 arc_hdr_alloc_pabd(hdr, ARC_HDR_DO_ADAPT|ARC_HDR_ALLOC_RDATA);
6624 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6625 } else if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6626 /*
6627 * Ideally, we would always copy the io_abd into b_pabd, but the
6628 * user may have disabled compressed ARC, thus we must check the
6629 * hdr's compression setting rather than the io_bp's.
6630 */
6631 if (BP_IS_ENCRYPTED(bp)) {
6632 ASSERT3U(psize, >, 0);
6633 arc_hdr_alloc_pabd(hdr,
6634 ARC_HDR_DO_ADAPT|ARC_HDR_ALLOC_RDATA);
6635 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6636 } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
6637 !ARC_BUF_COMPRESSED(buf)) {
6638 ASSERT3U(psize, >, 0);
6639 arc_hdr_alloc_pabd(hdr, ARC_HDR_DO_ADAPT);
6640 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6641 } else {
6642 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6643 arc_hdr_alloc_pabd(hdr, ARC_HDR_DO_ADAPT);
6644 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6645 arc_buf_size(buf));
6646 }
6647 } else {
6648 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6649 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6650 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6651 arc_share_buf(hdr, buf);
6652 }
6653
6654 out:
6655 arc_hdr_verify(hdr, bp);
6656 }
6657
6658 static void
6659 arc_write_children_ready(zio_t *zio)
6660 {
6661 arc_write_callback_t *callback = zio->io_private;
6662 arc_buf_t *buf = callback->awcb_buf;
6663
6664 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6665 }
6666
6667 /*
6668 * The SPA calls this callback for each physical write that happens on behalf
6669 * of a logical write. See the comment in dbuf_write_physdone() for details.
6670 */
6671 static void
6672 arc_write_physdone(zio_t *zio)
6673 {
6674 arc_write_callback_t *cb = zio->io_private;
6675 if (cb->awcb_physdone != NULL)
6676 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6677 }
6678
6679 static void
6680 arc_write_done(zio_t *zio)
6681 {
6682 arc_write_callback_t *callback = zio->io_private;
6683 arc_buf_t *buf = callback->awcb_buf;
6684 arc_buf_hdr_t *hdr = buf->b_hdr;
6685
6686 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6687
6688 if (zio->io_error == 0) {
6689 arc_hdr_verify(hdr, zio->io_bp);
6690
6691 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6692 buf_discard_identity(hdr);
6693 } else {
6694 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6695 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6696 }
6697 } else {
6698 ASSERT(HDR_EMPTY(hdr));
6699 }
6700
6701 /*
6702 * If the block to be written was all-zero or compressed enough to be
6703 * embedded in the BP, no write was performed so there will be no
6704 * dva/birth/checksum. The buffer must therefore remain anonymous
6705 * (and uncached).
6706 */
6707 if (!HDR_EMPTY(hdr)) {
6708 arc_buf_hdr_t *exists;
6709 kmutex_t *hash_lock;
6710
6711 ASSERT3U(zio->io_error, ==, 0);
6712
6713 arc_cksum_verify(buf);
6714
6715 exists = buf_hash_insert(hdr, &hash_lock);
6716 if (exists != NULL) {
6717 /*
6718 * This can only happen if we overwrite for
6719 * sync-to-convergence, because we remove
6720 * buffers from the hash table when we arc_free().
6721 */
6722 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6723 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6724 panic("bad overwrite, hdr=%p exists=%p",
6725 (void *)hdr, (void *)exists);
6726 ASSERT(zfs_refcount_is_zero(
6727 &exists->b_l1hdr.b_refcnt));
6728 arc_change_state(arc_anon, exists, hash_lock);
6729 arc_hdr_destroy(exists);
6730 mutex_exit(hash_lock);
6731 exists = buf_hash_insert(hdr, &hash_lock);
6732 ASSERT3P(exists, ==, NULL);
6733 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6734 /* nopwrite */
6735 ASSERT(zio->io_prop.zp_nopwrite);
6736 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6737 panic("bad nopwrite, hdr=%p exists=%p",
6738 (void *)hdr, (void *)exists);
6739 } else {
6740 /* Dedup */
6741 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6742 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6743 ASSERT(BP_GET_DEDUP(zio->io_bp));
6744 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6745 }
6746 }
6747 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6748 /* if it's not anon, we are doing a scrub */
6749 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6750 arc_access(hdr, hash_lock);
6751 mutex_exit(hash_lock);
6752 } else {
6753 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6754 }
6755
6756 ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6757 callback->awcb_done(zio, buf, callback->awcb_private);
6758
6759 abd_put(zio->io_abd);
6760 kmem_free(callback, sizeof (arc_write_callback_t));
6761 }
6762
6763 zio_t *
6764 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6765 boolean_t l2arc, const zio_prop_t *zp, arc_write_done_func_t *ready,
6766 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
6767 arc_write_done_func_t *done, void *private, zio_priority_t priority,
6768 int zio_flags, const zbookmark_phys_t *zb)
6769 {
6770 arc_buf_hdr_t *hdr = buf->b_hdr;
6771 arc_write_callback_t *callback;
6772 zio_t *zio;
6773 zio_prop_t localprop = *zp;
6774
6775 ASSERT3P(ready, !=, NULL);
6776 ASSERT3P(done, !=, NULL);
6777 ASSERT(!HDR_IO_ERROR(hdr));
6778 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6779 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6780 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6781 if (l2arc)
6782 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6783
6784 if (ARC_BUF_ENCRYPTED(buf)) {
6785 ASSERT(ARC_BUF_COMPRESSED(buf));
6786 localprop.zp_encrypt = B_TRUE;
6787 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6788 /* CONSTCOND */
6789 localprop.zp_byteorder =
6790 (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
6791 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
6792 bcopy(hdr->b_crypt_hdr.b_salt, localprop.zp_salt,
6793 ZIO_DATA_SALT_LEN);
6794 bcopy(hdr->b_crypt_hdr.b_iv, localprop.zp_iv,
6795 ZIO_DATA_IV_LEN);
6796 bcopy(hdr->b_crypt_hdr.b_mac, localprop.zp_mac,
6797 ZIO_DATA_MAC_LEN);
6798 if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
6799 localprop.zp_nopwrite = B_FALSE;
6800 localprop.zp_copies =
6801 MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
6802 }
6803 zio_flags |= ZIO_FLAG_RAW;
6804 } else if (ARC_BUF_COMPRESSED(buf)) {
6805 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6806 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6807 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6808 }
6809
6810 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6811 callback->awcb_ready = ready;
6812 callback->awcb_children_ready = children_ready;
6813 callback->awcb_physdone = physdone;
6814 callback->awcb_done = done;
6815 callback->awcb_private = private;
6816 callback->awcb_buf = buf;
6817
6818 /*
6819 * The hdr's b_pabd is now stale, free it now. A new data block
6820 * will be allocated when the zio pipeline calls arc_write_ready().
6821 */
6822 if (hdr->b_l1hdr.b_pabd != NULL) {
6823 /*
6824 * If the buf is currently sharing the data block with
6825 * the hdr then we need to break that relationship here.
6826 * The hdr will remain with a NULL data pointer and the
6827 * buf will take sole ownership of the block.
6828 */
6829 if (arc_buf_is_shared(buf)) {
6830 arc_unshare_buf(hdr, buf);
6831 } else {
6832 arc_hdr_free_pabd(hdr, B_FALSE);
6833 }
6834 VERIFY3P(buf->b_data, !=, NULL);
6835 }
6836
6837 if (HDR_HAS_RABD(hdr))
6838 arc_hdr_free_pabd(hdr, B_TRUE);
6839
6840 if (!(zio_flags & ZIO_FLAG_RAW))
6841 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6842
6843 ASSERT(!arc_buf_is_shared(buf));
6844 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6845
6846 zio = zio_write(pio, spa, txg, bp,
6847 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6848 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6849 (children_ready != NULL) ? arc_write_children_ready : NULL,
6850 arc_write_physdone, arc_write_done, callback,
6851 priority, zio_flags, zb);
6852
6853 return (zio);
6854 }
6855
6856 static int
6857 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
6858 {
6859 #ifdef _KERNEL
6860 uint64_t available_memory = ptob(freemem);
6861
6862 #if defined(__i386)
6863 available_memory =
6864 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
6865 #endif
6866
6867 if (freemem > physmem * arc_lotsfree_percent / 100)
6868 return (0);
6869
6870 if (txg > spa->spa_lowmem_last_txg) {
6871 spa->spa_lowmem_last_txg = txg;
6872 spa->spa_lowmem_page_load = 0;
6873 }
6874 /*
6875 * If we are in pageout, we know that memory is already tight,
6876 * the arc is already going to be evicting, so we just want to
6877 * continue to let page writes occur as quickly as possible.
6878 */
6879 if (curproc == proc_pageout) {
6880 if (spa->spa_lowmem_page_load >
6881 MAX(ptob(minfree), available_memory) / 4)
6882 return (SET_ERROR(ERESTART));
6883 /* Note: reserve is inflated, so we deflate */
6884 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
6885 return (0);
6886 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
6887 /* memory is low, delay before restarting */
6888 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6889 return (SET_ERROR(EAGAIN));
6890 }
6891 spa->spa_lowmem_page_load = 0;
6892 #endif /* _KERNEL */
6893 return (0);
6894 }
6895
6896 void
6897 arc_tempreserve_clear(uint64_t reserve)
6898 {
6899 atomic_add_64(&arc_tempreserve, -reserve);
6900 ASSERT((int64_t)arc_tempreserve >= 0);
6901 }
6902
6903 int
6904 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
6905 {
6906 int error;
6907 uint64_t anon_size;
6908
6909 if (reserve > arc_c/4 && !arc_no_grow)
6910 arc_c = MIN(arc_c_max, reserve * 4);
6911 if (reserve > arc_c)
6912 return (SET_ERROR(ENOMEM));
6913
6914 /*
6915 * Don't count loaned bufs as in flight dirty data to prevent long
6916 * network delays from blocking transactions that are ready to be
6917 * assigned to a txg.
6918 */
6919
6920 /* assert that it has not wrapped around */
6921 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6922
6923 anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) -
6924 arc_loaned_bytes), 0);
6925
6926 /*
6927 * Writes will, almost always, require additional memory allocations
6928 * in order to compress/encrypt/etc the data. We therefore need to
6929 * make sure that there is sufficient available memory for this.
6930 */
6931 error = arc_memory_throttle(spa, reserve, txg);
6932 if (error != 0)
6933 return (error);
6934
6935 /*
6936 * Throttle writes when the amount of dirty data in the cache
6937 * gets too large. We try to keep the cache less than half full
6938 * of dirty blocks so that our sync times don't grow too large.
6939 *
6940 * In the case of one pool being built on another pool, we want
6941 * to make sure we don't end up throttling the lower (backing)
6942 * pool when the upper pool is the majority contributor to dirty
6943 * data. To insure we make forward progress during throttling, we
6944 * also check the current pool's net dirty data and only throttle
6945 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
6946 * data in the cache.
6947 *
6948 * Note: if two requests come in concurrently, we might let them
6949 * both succeed, when one of them should fail. Not a huge deal.
6950 */
6951 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
6952 uint64_t spa_dirty_anon = spa_dirty_data(spa);
6953
6954 if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 &&
6955 anon_size > arc_c * zfs_arc_anon_limit_percent / 100 &&
6956 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
6957 uint64_t meta_esize =
6958 zfs_refcount_count(
6959 &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6960 uint64_t data_esize =
6961 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6962 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6963 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6964 arc_tempreserve >> 10, meta_esize >> 10,
6965 data_esize >> 10, reserve >> 10, arc_c >> 10);
6966 return (SET_ERROR(ERESTART));
6967 }
6968 atomic_add_64(&arc_tempreserve, reserve);
6969 return (0);
6970 }
6971
6972 static void
6973 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6974 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6975 {
6976 size->value.ui64 = zfs_refcount_count(&state->arcs_size);
6977 evict_data->value.ui64 =
6978 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6979 evict_metadata->value.ui64 =
6980 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6981 }
6982
6983 static int
6984 arc_kstat_update(kstat_t *ksp, int rw)
6985 {
6986 arc_stats_t *as = ksp->ks_data;
6987
6988 if (rw == KSTAT_WRITE) {
6989 return (EACCES);
6990 } else {
6991 arc_kstat_update_state(arc_anon,
6992 &as->arcstat_anon_size,
6993 &as->arcstat_anon_evictable_data,
6994 &as->arcstat_anon_evictable_metadata);
6995 arc_kstat_update_state(arc_mru,
6996 &as->arcstat_mru_size,
6997 &as->arcstat_mru_evictable_data,
6998 &as->arcstat_mru_evictable_metadata);
6999 arc_kstat_update_state(arc_mru_ghost,
7000 &as->arcstat_mru_ghost_size,
7001 &as->arcstat_mru_ghost_evictable_data,
7002 &as->arcstat_mru_ghost_evictable_metadata);
7003 arc_kstat_update_state(arc_mfu,
7004 &as->arcstat_mfu_size,
7005 &as->arcstat_mfu_evictable_data,
7006 &as->arcstat_mfu_evictable_metadata);
7007 arc_kstat_update_state(arc_mfu_ghost,
7008 &as->arcstat_mfu_ghost_size,
7009 &as->arcstat_mfu_ghost_evictable_data,
7010 &as->arcstat_mfu_ghost_evictable_metadata);
7011
7012 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
7013 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
7014 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
7015 ARCSTAT(arcstat_metadata_size) =
7016 aggsum_value(&astat_metadata_size);
7017 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
7018 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size);
7019 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
7020 }
7021
7022 return (0);
7023 }
7024
7025 /*
7026 * This function *must* return indices evenly distributed between all
7027 * sublists of the multilist. This is needed due to how the ARC eviction
7028 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
7029 * distributed between all sublists and uses this assumption when
7030 * deciding which sublist to evict from and how much to evict from it.
7031 */
7032 unsigned int
7033 arc_state_multilist_index_func(multilist_t *ml, void *obj)
7034 {
7035 arc_buf_hdr_t *hdr = obj;
7036
7037 /*
7038 * We rely on b_dva to generate evenly distributed index
7039 * numbers using buf_hash below. So, as an added precaution,
7040 * let's make sure we never add empty buffers to the arc lists.
7041 */
7042 ASSERT(!HDR_EMPTY(hdr));
7043
7044 /*
7045 * The assumption here, is the hash value for a given
7046 * arc_buf_hdr_t will remain constant throughout its lifetime
7047 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
7048 * Thus, we don't need to store the header's sublist index
7049 * on insertion, as this index can be recalculated on removal.
7050 *
7051 * Also, the low order bits of the hash value are thought to be
7052 * distributed evenly. Otherwise, in the case that the multilist
7053 * has a power of two number of sublists, each sublists' usage
7054 * would not be evenly distributed.
7055 */
7056 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
7057 multilist_get_num_sublists(ml));
7058 }
7059
7060 static void
7061 arc_state_init(void)
7062 {
7063 arc_anon = &ARC_anon;
7064 arc_mru = &ARC_mru;
7065 arc_mru_ghost = &ARC_mru_ghost;
7066 arc_mfu = &ARC_mfu;
7067 arc_mfu_ghost = &ARC_mfu_ghost;
7068 arc_l2c_only = &ARC_l2c_only;
7069
7070 arc_mru->arcs_list[ARC_BUFC_METADATA] =
7071 multilist_create(sizeof (arc_buf_hdr_t),
7072 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7073 arc_state_multilist_index_func);
7074 arc_mru->arcs_list[ARC_BUFC_DATA] =
7075 multilist_create(sizeof (arc_buf_hdr_t),
7076 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7077 arc_state_multilist_index_func);
7078 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
7079 multilist_create(sizeof (arc_buf_hdr_t),
7080 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7081 arc_state_multilist_index_func);
7082 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
7083 multilist_create(sizeof (arc_buf_hdr_t),
7084 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7085 arc_state_multilist_index_func);
7086 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
7087 multilist_create(sizeof (arc_buf_hdr_t),
7088 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7089 arc_state_multilist_index_func);
7090 arc_mfu->arcs_list[ARC_BUFC_DATA] =
7091 multilist_create(sizeof (arc_buf_hdr_t),
7092 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7093 arc_state_multilist_index_func);
7094 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
7095 multilist_create(sizeof (arc_buf_hdr_t),
7096 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7097 arc_state_multilist_index_func);
7098 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
7099 multilist_create(sizeof (arc_buf_hdr_t),
7100 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7101 arc_state_multilist_index_func);
7102 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
7103 multilist_create(sizeof (arc_buf_hdr_t),
7104 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7105 arc_state_multilist_index_func);
7106 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
7107 multilist_create(sizeof (arc_buf_hdr_t),
7108 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
7109 arc_state_multilist_index_func);
7110
7111 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7112 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7113 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7114 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7115 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7116 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7117 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7118 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7119 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7120 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7121 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7122 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7123
7124 zfs_refcount_create(&arc_anon->arcs_size);
7125 zfs_refcount_create(&arc_mru->arcs_size);
7126 zfs_refcount_create(&arc_mru_ghost->arcs_size);
7127 zfs_refcount_create(&arc_mfu->arcs_size);
7128 zfs_refcount_create(&arc_mfu_ghost->arcs_size);
7129 zfs_refcount_create(&arc_l2c_only->arcs_size);
7130
7131 aggsum_init(&arc_meta_used, 0);
7132 aggsum_init(&arc_size, 0);
7133 aggsum_init(&astat_data_size, 0);
7134 aggsum_init(&astat_metadata_size, 0);
7135 aggsum_init(&astat_hdr_size, 0);
7136 aggsum_init(&astat_other_size, 0);
7137 aggsum_init(&astat_l2_hdr_size, 0);
7138
7139 arc_anon->arcs_state = ARC_STATE_ANON;
7140 arc_mru->arcs_state = ARC_STATE_MRU;
7141 arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST;
7142 arc_mfu->arcs_state = ARC_STATE_MFU;
7143 arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
7144 arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
7145 }
7146
7147 static void
7148 arc_state_fini(void)
7149 {
7150 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
7151 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
7152 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
7153 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
7154 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
7155 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
7156 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
7157 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
7158 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
7159 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
7160 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
7161 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
7162
7163 zfs_refcount_destroy(&arc_anon->arcs_size);
7164 zfs_refcount_destroy(&arc_mru->arcs_size);
7165 zfs_refcount_destroy(&arc_mru_ghost->arcs_size);
7166 zfs_refcount_destroy(&arc_mfu->arcs_size);
7167 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size);
7168 zfs_refcount_destroy(&arc_l2c_only->arcs_size);
7169
7170 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
7171 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
7172 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
7173 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
7174 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
7175 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
7176 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
7177 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
7178 multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
7179 multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
7180
7181 aggsum_fini(&arc_meta_used);
7182 aggsum_fini(&arc_size);
7183 aggsum_fini(&astat_data_size);
7184 aggsum_fini(&astat_metadata_size);
7185 aggsum_fini(&astat_hdr_size);
7186 aggsum_fini(&astat_other_size);
7187 aggsum_fini(&astat_l2_hdr_size);
7188
7189 }
7190
7191 uint64_t
7192 arc_max_bytes(void)
7193 {
7194 return (arc_c_max);
7195 }
7196
7197 void
7198 arc_init(void)
7199 {
7200 /*
7201 * allmem is "all memory that we could possibly use".
7202 */
7203 #ifdef _KERNEL
7204 uint64_t allmem = ptob(physmem - swapfs_minfree);
7205 #else
7206 uint64_t allmem = (physmem * PAGESIZE) / 2;
7207 #endif
7208 mutex_init(&arc_adjust_lock, NULL, MUTEX_DEFAULT, NULL);
7209 cv_init(&arc_adjust_waiters_cv, NULL, CV_DEFAULT, NULL);
7210
7211 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
7212 arc_c_min = MAX(allmem / 32, 64 << 20);
7213 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
7214 if (allmem >= 1 << 30)
7215 arc_c_max = allmem - (1 << 30);
7216 else
7217 arc_c_max = arc_c_min;
7218 arc_c_max = MAX(allmem * 3 / 4, arc_c_max);
7219
7220 /*
7221 * In userland, there's only the memory pressure that we artificially
7222 * create (see arc_available_memory()). Don't let arc_c get too
7223 * small, because it can cause transactions to be larger than
7224 * arc_c, causing arc_tempreserve_space() to fail.
7225 */
7226 #ifndef _KERNEL
7227 arc_c_min = arc_c_max / 2;
7228 #endif
7229
7230 /*
7231 * Allow the tunables to override our calculations if they are
7232 * reasonable (ie. over 64MB)
7233 */
7234 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) {
7235 arc_c_max = zfs_arc_max;
7236 arc_c_min = MIN(arc_c_min, arc_c_max);
7237 }
7238 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
7239 arc_c_min = zfs_arc_min;
7240
7241 arc_c = arc_c_max;
7242 arc_p = (arc_c >> 1);
7243
7244 /* limit meta-data to 1/4 of the arc capacity */
7245 arc_meta_limit = arc_c_max / 4;
7246
7247 #ifdef _KERNEL
7248 /*
7249 * Metadata is stored in the kernel's heap. Don't let us
7250 * use more than half the heap for the ARC.
7251 */
7252 arc_meta_limit = MIN(arc_meta_limit,
7253 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
7254 #endif
7255
7256 /* Allow the tunable to override if it is reasonable */
7257 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
7258 arc_meta_limit = zfs_arc_meta_limit;
7259
7260 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
7261 arc_c_min = arc_meta_limit / 2;
7262
7263 /* On larger-memory machines, we clamp the minimum at 1GB */
7264 if (zfs_arc_min == 0)
7265 arc_c_min = MIN(arc_c_min, (1 << 30));
7266
7267 if (zfs_arc_meta_min > 0) {
7268 arc_meta_min = zfs_arc_meta_min;
7269 } else {
7270 arc_meta_min = arc_c_min / 2;
7271 }
7272
7273 if (zfs_arc_grow_retry > 0)
7274 arc_grow_retry = zfs_arc_grow_retry;
7275
7276 if (zfs_arc_shrink_shift > 0)
7277 arc_shrink_shift = zfs_arc_shrink_shift;
7278
7279 /*
7280 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
7281 */
7282 if (arc_no_grow_shift >= arc_shrink_shift)
7283 arc_no_grow_shift = arc_shrink_shift - 1;
7284
7285 if (zfs_arc_p_min_shift > 0)
7286 arc_p_min_shift = zfs_arc_p_min_shift;
7287
7288 /* if kmem_flags are set, lets try to use less memory */
7289 if (kmem_debugging())
7290 arc_c = arc_c / 2;
7291 if (arc_c < arc_c_min)
7292 arc_c = arc_c_min;
7293
7294 arc_state_init();
7295
7296 /*
7297 * The arc must be "uninitialized", so that hdr_recl() (which is
7298 * registered by buf_init()) will not access arc_reap_zthr before
7299 * it is created.
7300 */
7301 ASSERT(!arc_initialized);
7302 buf_init();
7303
7304 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7305 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7306
7307 if (arc_ksp != NULL) {
7308 arc_ksp->ks_data = &arc_stats;
7309 arc_ksp->ks_update = arc_kstat_update;
7310 kstat_install(arc_ksp);
7311 }
7312
7313 arc_adjust_zthr = zthr_create(arc_adjust_cb_check,
7314 arc_adjust_cb, NULL);
7315 arc_reap_zthr = zthr_create_timer(arc_reap_cb_check,
7316 arc_reap_cb, NULL, SEC2NSEC(1));
7317
7318 arc_initialized = B_TRUE;
7319 arc_warm = B_FALSE;
7320
7321 /*
7322 * Calculate maximum amount of dirty data per pool.
7323 *
7324 * If it has been set by /etc/system, take that.
7325 * Otherwise, use a percentage of physical memory defined by
7326 * zfs_dirty_data_max_percent (default 10%) with a cap at
7327 * zfs_dirty_data_max_max (default 4GB).
7328 */
7329 if (zfs_dirty_data_max == 0) {
7330 zfs_dirty_data_max = physmem * PAGESIZE *
7331 zfs_dirty_data_max_percent / 100;
7332 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7333 zfs_dirty_data_max_max);
7334 }
7335 }
7336
7337 void
7338 arc_fini(void)
7339 {
7340 /* Use B_TRUE to ensure *all* buffers are evicted */
7341 arc_flush(NULL, B_TRUE);
7342
7343 arc_initialized = B_FALSE;
7344
7345 if (arc_ksp != NULL) {
7346 kstat_delete(arc_ksp);
7347 arc_ksp = NULL;
7348 }
7349
7350 (void) zthr_cancel(arc_adjust_zthr);
7351 zthr_destroy(arc_adjust_zthr);
7352
7353 (void) zthr_cancel(arc_reap_zthr);
7354 zthr_destroy(arc_reap_zthr);
7355
7356 mutex_destroy(&arc_adjust_lock);
7357 cv_destroy(&arc_adjust_waiters_cv);
7358
7359 /*
7360 * buf_fini() must proceed arc_state_fini() because buf_fin() may
7361 * trigger the release of kmem magazines, which can callback to
7362 * arc_space_return() which accesses aggsums freed in act_state_fini().
7363 */
7364 buf_fini();
7365 arc_state_fini();
7366
7367 ASSERT0(arc_loaned_bytes);
7368 }
7369
7370 /*
7371 * Level 2 ARC
7372 *
7373 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7374 * It uses dedicated storage devices to hold cached data, which are populated
7375 * using large infrequent writes. The main role of this cache is to boost
7376 * the performance of random read workloads. The intended L2ARC devices
7377 * include short-stroked disks, solid state disks, and other media with
7378 * substantially faster read latency than disk.
7379 *
7380 * +-----------------------+
7381 * | ARC |
7382 * +-----------------------+
7383 * | ^ ^
7384 * | | |
7385 * l2arc_feed_thread() arc_read()
7386 * | | |
7387 * | l2arc read |
7388 * V | |
7389 * +---------------+ |
7390 * | L2ARC | |
7391 * +---------------+ |
7392 * | ^ |
7393 * l2arc_write() | |
7394 * | | |
7395 * V | |
7396 * +-------+ +-------+
7397 * | vdev | | vdev |
7398 * | cache | | cache |
7399 * +-------+ +-------+
7400 * +=========+ .-----.
7401 * : L2ARC : |-_____-|
7402 * : devices : | Disks |
7403 * +=========+ `-_____-'
7404 *
7405 * Read requests are satisfied from the following sources, in order:
7406 *
7407 * 1) ARC
7408 * 2) vdev cache of L2ARC devices
7409 * 3) L2ARC devices
7410 * 4) vdev cache of disks
7411 * 5) disks
7412 *
7413 * Some L2ARC device types exhibit extremely slow write performance.
7414 * To accommodate for this there are some significant differences between
7415 * the L2ARC and traditional cache design:
7416 *
7417 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7418 * the ARC behave as usual, freeing buffers and placing headers on ghost
7419 * lists. The ARC does not send buffers to the L2ARC during eviction as
7420 * this would add inflated write latencies for all ARC memory pressure.
7421 *
7422 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7423 * It does this by periodically scanning buffers from the eviction-end of
7424 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7425 * not already there. It scans until a headroom of buffers is satisfied,
7426 * which itself is a buffer for ARC eviction. If a compressible buffer is
7427 * found during scanning and selected for writing to an L2ARC device, we
7428 * temporarily boost scanning headroom during the next scan cycle to make
7429 * sure we adapt to compression effects (which might significantly reduce
7430 * the data volume we write to L2ARC). The thread that does this is
7431 * l2arc_feed_thread(), illustrated below; example sizes are included to
7432 * provide a better sense of ratio than this diagram:
7433 *
7434 * head --> tail
7435 * +---------------------+----------+
7436 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7437 * +---------------------+----------+ | o L2ARC eligible
7438 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7439 * +---------------------+----------+ |
7440 * 15.9 Gbytes ^ 32 Mbytes |
7441 * headroom |
7442 * l2arc_feed_thread()
7443 * |
7444 * l2arc write hand <--[oooo]--'
7445 * | 8 Mbyte
7446 * | write max
7447 * V
7448 * +==============================+
7449 * L2ARC dev |####|#|###|###| |####| ... |
7450 * +==============================+
7451 * 32 Gbytes
7452 *
7453 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7454 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7455 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7456 * safe to say that this is an uncommon case, since buffers at the end of
7457 * the ARC lists have moved there due to inactivity.
7458 *
7459 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7460 * then the L2ARC simply misses copying some buffers. This serves as a
7461 * pressure valve to prevent heavy read workloads from both stalling the ARC
7462 * with waits and clogging the L2ARC with writes. This also helps prevent
7463 * the potential for the L2ARC to churn if it attempts to cache content too
7464 * quickly, such as during backups of the entire pool.
7465 *
7466 * 5. After system boot and before the ARC has filled main memory, there are
7467 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7468 * lists can remain mostly static. Instead of searching from tail of these
7469 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7470 * for eligible buffers, greatly increasing its chance of finding them.
7471 *
7472 * The L2ARC device write speed is also boosted during this time so that
7473 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7474 * there are no L2ARC reads, and no fear of degrading read performance
7475 * through increased writes.
7476 *
7477 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7478 * the vdev queue can aggregate them into larger and fewer writes. Each
7479 * device is written to in a rotor fashion, sweeping writes through
7480 * available space then repeating.
7481 *
7482 * 7. The L2ARC does not store dirty content. It never needs to flush
7483 * write buffers back to disk based storage.
7484 *
7485 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7486 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7487 *
7488 * The performance of the L2ARC can be tweaked by a number of tunables, which
7489 * may be necessary for different workloads:
7490 *
7491 * l2arc_write_max max write bytes per interval
7492 * l2arc_write_boost extra write bytes during device warmup
7493 * l2arc_noprefetch skip caching prefetched buffers
7494 * l2arc_headroom number of max device writes to precache
7495 * l2arc_headroom_boost when we find compressed buffers during ARC
7496 * scanning, we multiply headroom by this
7497 * percentage factor for the next scan cycle,
7498 * since more compressed buffers are likely to
7499 * be present
7500 * l2arc_feed_secs seconds between L2ARC writing
7501 *
7502 * Tunables may be removed or added as future performance improvements are
7503 * integrated, and also may become zpool properties.
7504 *
7505 * There are three key functions that control how the L2ARC warms up:
7506 *
7507 * l2arc_write_eligible() check if a buffer is eligible to cache
7508 * l2arc_write_size() calculate how much to write
7509 * l2arc_write_interval() calculate sleep delay between writes
7510 *
7511 * These three functions determine what to write, how much, and how quickly
7512 * to send writes.
7513 *
7514 * L2ARC persistence:
7515 *
7516 * When writing buffers to L2ARC, we periodically add some metadata to
7517 * make sure we can pick them up after reboot, thus dramatically reducing
7518 * the impact that any downtime has on the performance of storage systems
7519 * with large caches.
7520 *
7521 * The implementation works fairly simply by integrating the following two
7522 * modifications:
7523 *
7524 * *) When writing to the L2ARC, we occasionally write a "l2arc log block",
7525 * which is an additional piece of metadata which describes what's been
7526 * written. This allows us to rebuild the arc_buf_hdr_t structures of the
7527 * main ARC buffers. There are 2 linked-lists of log blocks headed by
7528 * dh_start_lbps[2]. We alternate which chain we append to, so they are
7529 * time-wise and offset-wise interleaved, but that is an optimization rather
7530 * than for correctness. The log block also includes a pointer to the
7531 * previous block in its chain.
7532 *
7533 * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device
7534 * for our header bookkeeping purposes. This contains a device header,
7535 * which contains our top-level reference structures. We update it each
7536 * time we write a new log block, so that we're able to locate it in the
7537 * L2ARC device. If this write results in an inconsistent device header
7538 * (e.g. due to power failure), we detect this by verifying the header's
7539 * checksum and simply fail to reconstruct the L2ARC after reboot.
7540 *
7541 * Implementation diagram:
7542 *
7543 * +=== L2ARC device (not to scale) ======================================+
7544 * | ___two newest log block pointers__.__________ |
7545 * | / \dh_start_lbps[1] |
7546 * | / \ \dh_start_lbps[0]|
7547 * |.___/__. V V |
7548 * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---|
7549 * || hdr| ^ /^ /^ / / |
7550 * |+------+ ...--\-------/ \-----/--\------/ / |
7551 * | \--------------/ \--------------/ |
7552 * +======================================================================+
7553 *
7554 * As can be seen on the diagram, rather than using a simple linked list,
7555 * we use a pair of linked lists with alternating elements. This is a
7556 * performance enhancement due to the fact that we only find out the
7557 * address of the next log block access once the current block has been
7558 * completely read in. Obviously, this hurts performance, because we'd be
7559 * keeping the device's I/O queue at only a 1 operation deep, thus
7560 * incurring a large amount of I/O round-trip latency. Having two lists
7561 * allows us to fetch two log blocks ahead of where we are currently
7562 * rebuilding L2ARC buffers.
7563 *
7564 * On-device data structures:
7565 *
7566 * L2ARC device header: l2arc_dev_hdr_phys_t
7567 * L2ARC log block: l2arc_log_blk_phys_t
7568 *
7569 * L2ARC reconstruction:
7570 *
7571 * When writing data, we simply write in the standard rotary fashion,
7572 * evicting buffers as we go and simply writing new data over them (writing
7573 * a new log block every now and then). This obviously means that once we
7574 * loop around the end of the device, we will start cutting into an already
7575 * committed log block (and its referenced data buffers), like so:
7576 *
7577 * current write head__ __old tail
7578 * \ /
7579 * V V
7580 * <--|bufs |lb |bufs |lb | |bufs |lb |bufs |lb |-->
7581 * ^ ^^^^^^^^^___________________________________
7582 * | \
7583 * <<nextwrite>> may overwrite this blk and/or its bufs --'
7584 *
7585 * When importing the pool, we detect this situation and use it to stop
7586 * our scanning process (see l2arc_rebuild).
7587 *
7588 * There is one significant caveat to consider when rebuilding ARC contents
7589 * from an L2ARC device: what about invalidated buffers? Given the above
7590 * construction, we cannot update blocks which we've already written to amend
7591 * them to remove buffers which were invalidated. Thus, during reconstruction,
7592 * we might be populating the cache with buffers for data that's not on the
7593 * main pool anymore, or may have been overwritten!
7594 *
7595 * As it turns out, this isn't a problem. Every arc_read request includes
7596 * both the DVA and, crucially, the birth TXG of the BP the caller is
7597 * looking for. So even if the cache were populated by completely rotten
7598 * blocks for data that had been long deleted and/or overwritten, we'll
7599 * never actually return bad data from the cache, since the DVA with the
7600 * birth TXG uniquely identify a block in space and time - once created,
7601 * a block is immutable on disk. The worst thing we have done is wasted
7602 * some time and memory at l2arc rebuild to reconstruct outdated ARC
7603 * entries that will get dropped from the l2arc as it is being updated
7604 * with new blocks.
7605 *
7606 * L2ARC buffers that have been evicted by l2arc_evict() ahead of the write
7607 * hand are not restored. This is done by saving the offset (in bytes)
7608 * l2arc_evict() has evicted to in the L2ARC device header and taking it
7609 * into account when restoring buffers.
7610 */
7611
7612 static boolean_t
7613 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
7614 {
7615 /*
7616 * A buffer is *not* eligible for the L2ARC if it:
7617 * 1. belongs to a different spa.
7618 * 2. is already cached on the L2ARC.
7619 * 3. has an I/O in progress (it may be an incomplete read).
7620 * 4. is flagged not eligible (zfs property).
7621 * 5. is a prefetch and l2arc_noprefetch is set.
7622 */
7623 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
7624 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr) ||
7625 (l2arc_noprefetch && HDR_PREFETCH(hdr)))
7626 return (B_FALSE);
7627
7628 return (B_TRUE);
7629 }
7630
7631 static uint64_t
7632 l2arc_write_size(l2arc_dev_t *dev)
7633 {
7634 uint64_t size, dev_size;
7635
7636 /*
7637 * Make sure our globals have meaningful values in case the user
7638 * altered them.
7639 */
7640 size = l2arc_write_max;
7641 if (size == 0) {
7642 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
7643 "be greater than zero, resetting it to the default (%d)",
7644 L2ARC_WRITE_SIZE);
7645 size = l2arc_write_max = L2ARC_WRITE_SIZE;
7646 }
7647
7648 if (arc_warm == B_FALSE)
7649 size += l2arc_write_boost;
7650
7651 /*
7652 * Make sure the write size does not exceed the size of the cache
7653 * device. This is important in l2arc_evict(), otherwise infinite
7654 * iteration can occur.
7655 */
7656 dev_size = dev->l2ad_end - dev->l2ad_start;
7657 if ((size + l2arc_log_blk_overhead(size, dev)) >= dev_size) {
7658 cmn_err(CE_NOTE, "l2arc_write_max or l2arc_write_boost "
7659 "plus the overhead of log blocks (persistent L2ARC, "
7660 "%" PRIu64 " bytes) exceeds the size of the cache device "
7661 "(guid %" PRIu64 "), resetting them to the default (%d)",
7662 l2arc_log_blk_overhead(size, dev),
7663 dev->l2ad_vdev->vdev_guid, L2ARC_WRITE_SIZE);
7664 size = l2arc_write_max = l2arc_write_boost = L2ARC_WRITE_SIZE;
7665
7666 if (arc_warm == B_FALSE)
7667 size += l2arc_write_boost;
7668 }
7669
7670 return (size);
7671
7672 }
7673
7674 static clock_t
7675 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
7676 {
7677 clock_t interval, next, now;
7678
7679 /*
7680 * If the ARC lists are busy, increase our write rate; if the
7681 * lists are stale, idle back. This is achieved by checking
7682 * how much we previously wrote - if it was more than half of
7683 * what we wanted, schedule the next write much sooner.
7684 */
7685 if (l2arc_feed_again && wrote > (wanted / 2))
7686 interval = (hz * l2arc_feed_min_ms) / 1000;
7687 else
7688 interval = hz * l2arc_feed_secs;
7689
7690 now = ddi_get_lbolt();
7691 next = MAX(now, MIN(now + interval, began + interval));
7692
7693 return (next);
7694 }
7695
7696 /*
7697 * Cycle through L2ARC devices. This is how L2ARC load balances.
7698 * If a device is returned, this also returns holding the spa config lock.
7699 */
7700 static l2arc_dev_t *
7701 l2arc_dev_get_next(void)
7702 {
7703 l2arc_dev_t *first, *next = NULL;
7704
7705 /*
7706 * Lock out the removal of spas (spa_namespace_lock), then removal
7707 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7708 * both locks will be dropped and a spa config lock held instead.
7709 */
7710 mutex_enter(&spa_namespace_lock);
7711 mutex_enter(&l2arc_dev_mtx);
7712
7713 /* if there are no vdevs, there is nothing to do */
7714 if (l2arc_ndev == 0)
7715 goto out;
7716
7717 first = NULL;
7718 next = l2arc_dev_last;
7719 do {
7720 /* loop around the list looking for a non-faulted vdev */
7721 if (next == NULL) {
7722 next = list_head(l2arc_dev_list);
7723 } else {
7724 next = list_next(l2arc_dev_list, next);
7725 if (next == NULL)
7726 next = list_head(l2arc_dev_list);
7727 }
7728
7729 /* if we have come back to the start, bail out */
7730 if (first == NULL)
7731 first = next;
7732 else if (next == first)
7733 break;
7734
7735 } while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild);
7736
7737 /* if we were unable to find any usable vdevs, return NULL */
7738 if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild)
7739 next = NULL;
7740
7741 l2arc_dev_last = next;
7742
7743 out:
7744 mutex_exit(&l2arc_dev_mtx);
7745
7746 /*
7747 * Grab the config lock to prevent the 'next' device from being
7748 * removed while we are writing to it.
7749 */
7750 if (next != NULL)
7751 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
7752 mutex_exit(&spa_namespace_lock);
7753
7754 return (next);
7755 }
7756
7757 /*
7758 * Free buffers that were tagged for destruction.
7759 */
7760 static void
7761 l2arc_do_free_on_write()
7762 {
7763 list_t *buflist;
7764 l2arc_data_free_t *df, *df_prev;
7765
7766 mutex_enter(&l2arc_free_on_write_mtx);
7767 buflist = l2arc_free_on_write;
7768
7769 for (df = list_tail(buflist); df; df = df_prev) {
7770 df_prev = list_prev(buflist, df);
7771 ASSERT3P(df->l2df_abd, !=, NULL);
7772 abd_free(df->l2df_abd);
7773 list_remove(buflist, df);
7774 kmem_free(df, sizeof (l2arc_data_free_t));
7775 }
7776
7777 mutex_exit(&l2arc_free_on_write_mtx);
7778 }
7779
7780 /*
7781 * A write to a cache device has completed. Update all headers to allow
7782 * reads from these buffers to begin.
7783 */
7784 static void
7785 l2arc_write_done(zio_t *zio)
7786 {
7787 l2arc_write_callback_t *cb;
7788 l2arc_lb_abd_buf_t *abd_buf;
7789 l2arc_lb_ptr_buf_t *lb_ptr_buf;
7790 l2arc_dev_t *dev;
7791 l2arc_dev_hdr_phys_t *l2dhdr;
7792 list_t *buflist;
7793 arc_buf_hdr_t *head, *hdr, *hdr_prev;
7794 kmutex_t *hash_lock;
7795 int64_t bytes_dropped = 0;
7796
7797 cb = zio->io_private;
7798 ASSERT3P(cb, !=, NULL);
7799 dev = cb->l2wcb_dev;
7800 l2dhdr = dev->l2ad_dev_hdr;
7801 ASSERT3P(dev, !=, NULL);
7802 head = cb->l2wcb_head;
7803 ASSERT3P(head, !=, NULL);
7804 buflist = &dev->l2ad_buflist;
7805 ASSERT3P(buflist, !=, NULL);
7806 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7807 l2arc_write_callback_t *, cb);
7808
7809 /*
7810 * All writes completed, or an error was hit.
7811 */
7812 top:
7813 mutex_enter(&dev->l2ad_mtx);
7814 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7815 hdr_prev = list_prev(buflist, hdr);
7816
7817 hash_lock = HDR_LOCK(hdr);
7818
7819 /*
7820 * We cannot use mutex_enter or else we can deadlock
7821 * with l2arc_write_buffers (due to swapping the order
7822 * the hash lock and l2ad_mtx are taken).
7823 */
7824 if (!mutex_tryenter(hash_lock)) {
7825 /*
7826 * Missed the hash lock. We must retry so we
7827 * don't leave the ARC_FLAG_L2_WRITING bit set.
7828 */
7829 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7830
7831 /*
7832 * We don't want to rescan the headers we've
7833 * already marked as having been written out, so
7834 * we reinsert the head node so we can pick up
7835 * where we left off.
7836 */
7837 list_remove(buflist, head);
7838 list_insert_after(buflist, hdr, head);
7839
7840 mutex_exit(&dev->l2ad_mtx);
7841
7842 /*
7843 * We wait for the hash lock to become available
7844 * to try and prevent busy waiting, and increase
7845 * the chance we'll be able to acquire the lock
7846 * the next time around.
7847 */
7848 mutex_enter(hash_lock);
7849 mutex_exit(hash_lock);
7850 goto top;
7851 }
7852
7853 /*
7854 * We could not have been moved into the arc_l2c_only
7855 * state while in-flight due to our ARC_FLAG_L2_WRITING
7856 * bit being set. Let's just ensure that's being enforced.
7857 */
7858 ASSERT(HDR_HAS_L1HDR(hdr));
7859
7860 if (zio->io_error != 0) {
7861 /*
7862 * Error - drop L2ARC entry.
7863 */
7864 list_remove(buflist, hdr);
7865 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7866
7867 uint64_t psize = HDR_GET_PSIZE(hdr);
7868 l2arc_hdr_arcstats_decrement(hdr);
7869
7870 bytes_dropped +=
7871 vdev_psize_to_asize(dev->l2ad_vdev, psize);
7872 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
7873 arc_hdr_size(hdr), hdr);
7874 }
7875
7876 /*
7877 * Allow ARC to begin reads and ghost list evictions to
7878 * this L2ARC entry.
7879 */
7880 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7881
7882 mutex_exit(hash_lock);
7883 }
7884
7885 /*
7886 * Free the allocated abd buffers for writing the log blocks.
7887 * If the zio failed reclaim the allocated space and remove the
7888 * pointers to these log blocks from the log block pointer list
7889 * of the L2ARC device.
7890 */
7891 while ((abd_buf = list_remove_tail(&cb->l2wcb_abd_list)) != NULL) {
7892 abd_free(abd_buf->abd);
7893 zio_buf_free(abd_buf, sizeof (*abd_buf));
7894 if (zio->io_error != 0) {
7895 lb_ptr_buf = list_remove_head(&dev->l2ad_lbptr_list);
7896 /*
7897 * L2BLK_GET_PSIZE returns aligned size for log
7898 * blocks.
7899 */
7900 uint64_t asize =
7901 L2BLK_GET_PSIZE((lb_ptr_buf->lb_ptr)->lbp_prop);
7902 bytes_dropped += asize;
7903 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
7904 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
7905 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
7906 lb_ptr_buf);
7907 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
7908 kmem_free(lb_ptr_buf->lb_ptr,
7909 sizeof (l2arc_log_blkptr_t));
7910 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
7911 }
7912 }
7913 list_destroy(&cb->l2wcb_abd_list);
7914
7915 if (zio->io_error != 0) {
7916 ARCSTAT_BUMP(arcstat_l2_writes_error);
7917
7918 /*
7919 * Restore the lbps array in the header to its previous state.
7920 * If the list of log block pointers is empty, zero out the
7921 * log block pointers in the device header.
7922 */
7923 lb_ptr_buf = list_head(&dev->l2ad_lbptr_list);
7924 for (int i = 0; i < 2; i++) {
7925 if (lb_ptr_buf == NULL) {
7926 /*
7927 * If the list is empty zero out the device
7928 * header. Otherwise zero out the second log
7929 * block pointer in the header.
7930 */
7931 if (i == 0) {
7932 bzero(l2dhdr, dev->l2ad_dev_hdr_asize);
7933 } else {
7934 bzero(&l2dhdr->dh_start_lbps[i],
7935 sizeof (l2arc_log_blkptr_t));
7936 }
7937 break;
7938 }
7939 bcopy(lb_ptr_buf->lb_ptr, &l2dhdr->dh_start_lbps[i],
7940 sizeof (l2arc_log_blkptr_t));
7941 lb_ptr_buf = list_next(&dev->l2ad_lbptr_list,
7942 lb_ptr_buf);
7943 }
7944 }
7945
7946 atomic_inc_64(&l2arc_writes_done);
7947 list_remove(buflist, head);
7948 ASSERT(!HDR_HAS_L1HDR(head));
7949 kmem_cache_free(hdr_l2only_cache, head);
7950 mutex_exit(&dev->l2ad_mtx);
7951
7952 ASSERT(dev->l2ad_vdev != NULL);
7953 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7954
7955 l2arc_do_free_on_write();
7956
7957 kmem_free(cb, sizeof (l2arc_write_callback_t));
7958 }
7959
7960 static int
7961 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
7962 {
7963 int ret;
7964 spa_t *spa = zio->io_spa;
7965 arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
7966 blkptr_t *bp = zio->io_bp;
7967 uint8_t salt[ZIO_DATA_SALT_LEN];
7968 uint8_t iv[ZIO_DATA_IV_LEN];
7969 uint8_t mac[ZIO_DATA_MAC_LEN];
7970 boolean_t no_crypt = B_FALSE;
7971
7972 /*
7973 * ZIL data is never be written to the L2ARC, so we don't need
7974 * special handling for its unique MAC storage.
7975 */
7976 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
7977 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
7978 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
7979
7980 /*
7981 * If the data was encrypted, decrypt it now. Note that
7982 * we must check the bp here and not the hdr, since the
7983 * hdr does not have its encryption parameters updated
7984 * until arc_read_done().
7985 */
7986 if (BP_IS_ENCRYPTED(bp)) {
7987 abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
7988 B_TRUE);
7989
7990 zio_crypt_decode_params_bp(bp, salt, iv);
7991 zio_crypt_decode_mac_bp(bp, mac);
7992
7993 ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb,
7994 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
7995 salt, iv, mac, HDR_GET_PSIZE(hdr), eabd,
7996 hdr->b_l1hdr.b_pabd, &no_crypt);
7997 if (ret != 0) {
7998 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
7999 goto error;
8000 }
8001
8002 /*
8003 * If we actually performed decryption, replace b_pabd
8004 * with the decrypted data. Otherwise we can just throw
8005 * our decryption buffer away.
8006 */
8007 if (!no_crypt) {
8008 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8009 arc_hdr_size(hdr), hdr);
8010 hdr->b_l1hdr.b_pabd = eabd;
8011 zio->io_abd = eabd;
8012 } else {
8013 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
8014 }
8015 }
8016
8017 /*
8018 * If the L2ARC block was compressed, but ARC compression
8019 * is disabled we decompress the data into a new buffer and
8020 * replace the existing data.
8021 */
8022 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8023 !HDR_COMPRESSION_ENABLED(hdr)) {
8024 abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr,
8025 B_TRUE);
8026 void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
8027
8028 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
8029 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
8030 HDR_GET_LSIZE(hdr));
8031 if (ret != 0) {
8032 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8033 arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
8034 goto error;
8035 }
8036
8037 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
8038 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
8039 arc_hdr_size(hdr), hdr);
8040 hdr->b_l1hdr.b_pabd = cabd;
8041 zio->io_abd = cabd;
8042 zio->io_size = HDR_GET_LSIZE(hdr);
8043 }
8044
8045 return (0);
8046
8047 error:
8048 return (ret);
8049 }
8050
8051
8052 /*
8053 * A read to a cache device completed. Validate buffer contents before
8054 * handing over to the regular ARC routines.
8055 */
8056 static void
8057 l2arc_read_done(zio_t *zio)
8058 {
8059 int tfm_error = 0;
8060 l2arc_read_callback_t *cb = zio->io_private;
8061 arc_buf_hdr_t *hdr;
8062 kmutex_t *hash_lock;
8063 boolean_t valid_cksum;
8064 boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
8065 (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT));
8066
8067 ASSERT3P(zio->io_vd, !=, NULL);
8068 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
8069
8070 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
8071
8072 ASSERT3P(cb, !=, NULL);
8073 hdr = cb->l2rcb_hdr;
8074 ASSERT3P(hdr, !=, NULL);
8075
8076 hash_lock = HDR_LOCK(hdr);
8077 mutex_enter(hash_lock);
8078 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
8079
8080 /*
8081 * If the data was read into a temporary buffer,
8082 * move it and free the buffer.
8083 */
8084 if (cb->l2rcb_abd != NULL) {
8085 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
8086 if (zio->io_error == 0) {
8087 if (using_rdata) {
8088 abd_copy(hdr->b_crypt_hdr.b_rabd,
8089 cb->l2rcb_abd, arc_hdr_size(hdr));
8090 } else {
8091 abd_copy(hdr->b_l1hdr.b_pabd,
8092 cb->l2rcb_abd, arc_hdr_size(hdr));
8093 }
8094 }
8095
8096 /*
8097 * The following must be done regardless of whether
8098 * there was an error:
8099 * - free the temporary buffer
8100 * - point zio to the real ARC buffer
8101 * - set zio size accordingly
8102 * These are required because zio is either re-used for
8103 * an I/O of the block in the case of the error
8104 * or the zio is passed to arc_read_done() and it
8105 * needs real data.
8106 */
8107 abd_free(cb->l2rcb_abd);
8108 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
8109
8110 if (using_rdata) {
8111 ASSERT(HDR_HAS_RABD(hdr));
8112 zio->io_abd = zio->io_orig_abd =
8113 hdr->b_crypt_hdr.b_rabd;
8114 } else {
8115 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
8116 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
8117 }
8118 }
8119
8120 ASSERT3P(zio->io_abd, !=, NULL);
8121
8122 /*
8123 * Check this survived the L2ARC journey.
8124 */
8125 ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
8126 (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
8127 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
8128 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
8129
8130 valid_cksum = arc_cksum_is_equal(hdr, zio);
8131
8132 /*
8133 * b_rabd will always match the data as it exists on disk if it is
8134 * being used. Therefore if we are reading into b_rabd we do not
8135 * attempt to untransform the data.
8136 */
8137 if (valid_cksum && !using_rdata)
8138 tfm_error = l2arc_untransform(zio, cb);
8139
8140 if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
8141 !HDR_L2_EVICTED(hdr)) {
8142 mutex_exit(hash_lock);
8143 zio->io_private = hdr;
8144 arc_read_done(zio);
8145 } else {
8146 /*
8147 * Buffer didn't survive caching. Increment stats and
8148 * reissue to the original storage device.
8149 */
8150 if (zio->io_error != 0) {
8151 ARCSTAT_BUMP(arcstat_l2_io_error);
8152 } else {
8153 zio->io_error = SET_ERROR(EIO);
8154 }
8155 if (!valid_cksum || tfm_error != 0)
8156 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
8157
8158 /*
8159 * If there's no waiter, issue an async i/o to the primary
8160 * storage now. If there *is* a waiter, the caller must
8161 * issue the i/o in a context where it's OK to block.
8162 */
8163 if (zio->io_waiter == NULL) {
8164 zio_t *pio = zio_unique_parent(zio);
8165 void *abd = (using_rdata) ?
8166 hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
8167
8168 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
8169
8170 zio = zio_read(pio, zio->io_spa, zio->io_bp,
8171 abd, zio->io_size, arc_read_done,
8172 hdr, zio->io_priority, cb->l2rcb_flags,
8173 &cb->l2rcb_zb);
8174
8175 /*
8176 * Original ZIO will be freed, so we need to update
8177 * ARC header with the new ZIO pointer to be used
8178 * by zio_change_priority() in arc_read().
8179 */
8180 for (struct arc_callback *acb = hdr->b_l1hdr.b_acb;
8181 acb != NULL; acb = acb->acb_next)
8182 acb->acb_zio_head = zio;
8183
8184 mutex_exit(hash_lock);
8185 zio_nowait(zio);
8186 } else {
8187 mutex_exit(hash_lock);
8188 }
8189 }
8190
8191 kmem_free(cb, sizeof (l2arc_read_callback_t));
8192 }
8193
8194 /*
8195 * This is the list priority from which the L2ARC will search for pages to
8196 * cache. This is used within loops (0..3) to cycle through lists in the
8197 * desired order. This order can have a significant effect on cache
8198 * performance.
8199 *
8200 * Currently the metadata lists are hit first, MFU then MRU, followed by
8201 * the data lists. This function returns a locked list, and also returns
8202 * the lock pointer.
8203 */
8204 static multilist_sublist_t *
8205 l2arc_sublist_lock(int list_num)
8206 {
8207 multilist_t *ml = NULL;
8208 unsigned int idx;
8209
8210 ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES);
8211
8212 switch (list_num) {
8213 case 0:
8214 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
8215 break;
8216 case 1:
8217 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
8218 break;
8219 case 2:
8220 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
8221 break;
8222 case 3:
8223 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
8224 break;
8225 default:
8226 return (NULL);
8227 }
8228
8229 /*
8230 * Return a randomly-selected sublist. This is acceptable
8231 * because the caller feeds only a little bit of data for each
8232 * call (8MB). Subsequent calls will result in different
8233 * sublists being selected.
8234 */
8235 idx = multilist_get_random_index(ml);
8236 return (multilist_sublist_lock(ml, idx));
8237 }
8238
8239 /*
8240 * Calculates the maximum overhead of L2ARC metadata log blocks for a given
8241 * L2ARC write size. l2arc_evict and l2arc_write_size need to include this
8242 * overhead in processing to make sure there is enough headroom available
8243 * when writing buffers.
8244 */
8245 static inline uint64_t
8246 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev)
8247 {
8248 if (dev->l2ad_log_entries == 0) {
8249 return (0);
8250 } else {
8251 uint64_t log_entries = write_sz >> SPA_MINBLOCKSHIFT;
8252
8253 uint64_t log_blocks = (log_entries +
8254 dev->l2ad_log_entries - 1) /
8255 dev->l2ad_log_entries;
8256
8257 return (vdev_psize_to_asize(dev->l2ad_vdev,
8258 sizeof (l2arc_log_blk_phys_t)) * log_blocks);
8259 }
8260 }
8261
8262 /*
8263 * Evict buffers from the device write hand to the distance specified in
8264 * bytes. This distance may span populated buffers, it may span nothing.
8265 * This is clearing a region on the L2ARC device ready for writing.
8266 * If the 'all' boolean is set, every buffer is evicted.
8267 */
8268 static void
8269 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
8270 {
8271 list_t *buflist;
8272 arc_buf_hdr_t *hdr, *hdr_prev;
8273 kmutex_t *hash_lock;
8274 uint64_t taddr;
8275 l2arc_lb_ptr_buf_t *lb_ptr_buf, *lb_ptr_buf_prev;
8276 boolean_t rerun;
8277
8278 buflist = &dev->l2ad_buflist;
8279
8280 /*
8281 * We need to add in the worst case scenario of log block overhead.
8282 */
8283 distance += l2arc_log_blk_overhead(distance, dev);
8284
8285 top:
8286 rerun = B_FALSE;
8287 if (dev->l2ad_hand >= (dev->l2ad_end - distance)) {
8288 /*
8289 * When there is no space to accommodate upcoming writes,
8290 * evict to the end. Then bump the write and evict hands
8291 * to the start and iterate. This iteration does not
8292 * happen indefinitely as we make sure in
8293 * l2arc_write_size() that when the write hand is reset,
8294 * the write size does not exceed the end of the device.
8295 */
8296 rerun = B_TRUE;
8297 taddr = dev->l2ad_end;
8298 } else {
8299 taddr = dev->l2ad_hand + distance;
8300 }
8301 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
8302 uint64_t, taddr, boolean_t, all);
8303
8304 /*
8305 * This check has to be placed after deciding whether to iterate
8306 * (rerun).
8307 */
8308 if (!all && dev->l2ad_first) {
8309 /*
8310 * This is the first sweep through the device. There is
8311 * nothing to evict.
8312 */
8313 goto out;
8314 }
8315
8316 /*
8317 * When rebuilding L2ARC we retrieve the evict hand from the header of
8318 * the device. Of note, l2arc_evict() does not actually delete buffers
8319 * from the cache device, but keeping track of the evict hand will be
8320 * useful when TRIM is implemented.
8321 */
8322 dev->l2ad_evict = MAX(dev->l2ad_evict, taddr);
8323
8324 retry:
8325 mutex_enter(&dev->l2ad_mtx);
8326 /*
8327 * We have to account for evicted log blocks. Run vdev_space_update()
8328 * on log blocks whose offset (in bytes) is before the evicted offset
8329 * (in bytes) by searching in the list of pointers to log blocks
8330 * present in the L2ARC device.
8331 */
8332 for (lb_ptr_buf = list_tail(&dev->l2ad_lbptr_list); lb_ptr_buf;
8333 lb_ptr_buf = lb_ptr_buf_prev) {
8334
8335 lb_ptr_buf_prev = list_prev(&dev->l2ad_lbptr_list, lb_ptr_buf);
8336
8337 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
8338 uint64_t asize = L2BLK_GET_PSIZE(
8339 (lb_ptr_buf->lb_ptr)->lbp_prop);
8340
8341 /*
8342 * We don't worry about log blocks left behind (ie
8343 * lbp_payload_start < l2ad_hand) because l2arc_write_buffers()
8344 * will never write more than l2arc_evict() evicts.
8345 */
8346 if (!all && l2arc_log_blkptr_valid(dev, lb_ptr_buf->lb_ptr)) {
8347 break;
8348 } else {
8349 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
8350 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8351 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8352 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8353 lb_ptr_buf);
8354 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
8355 list_remove(&dev->l2ad_lbptr_list, lb_ptr_buf);
8356 kmem_free(lb_ptr_buf->lb_ptr,
8357 sizeof (l2arc_log_blkptr_t));
8358 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8359 }
8360 }
8361
8362 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
8363 hdr_prev = list_prev(buflist, hdr);
8364
8365 ASSERT(!HDR_EMPTY(hdr));
8366 hash_lock = HDR_LOCK(hdr);
8367
8368 /*
8369 * We cannot use mutex_enter or else we can deadlock
8370 * with l2arc_write_buffers (due to swapping the order
8371 * the hash lock and l2ad_mtx are taken).
8372 */
8373 if (!mutex_tryenter(hash_lock)) {
8374 /*
8375 * Missed the hash lock. Retry.
8376 */
8377 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
8378 mutex_exit(&dev->l2ad_mtx);
8379 mutex_enter(hash_lock);
8380 mutex_exit(hash_lock);
8381 goto retry;
8382 }
8383
8384 /*
8385 * A header can't be on this list if it doesn't have L2 header.
8386 */
8387 ASSERT(HDR_HAS_L2HDR(hdr));
8388
8389 /* Ensure this header has finished being written. */
8390 ASSERT(!HDR_L2_WRITING(hdr));
8391 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
8392
8393 if (!all && (hdr->b_l2hdr.b_daddr >= dev->l2ad_evict ||
8394 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
8395 /*
8396 * We've evicted to the target address,
8397 * or the end of the device.
8398 */
8399 mutex_exit(hash_lock);
8400 break;
8401 }
8402
8403 if (!HDR_HAS_L1HDR(hdr)) {
8404 ASSERT(!HDR_L2_READING(hdr));
8405 /*
8406 * This doesn't exist in the ARC. Destroy.
8407 * arc_hdr_destroy() will call list_remove()
8408 * and decrement arcstat_l2_lsize.
8409 */
8410 arc_change_state(arc_anon, hdr, hash_lock);
8411 arc_hdr_destroy(hdr);
8412 } else {
8413 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
8414 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
8415 /*
8416 * Invalidate issued or about to be issued
8417 * reads, since we may be about to write
8418 * over this location.
8419 */
8420 if (HDR_L2_READING(hdr)) {
8421 ARCSTAT_BUMP(arcstat_l2_evict_reading);
8422 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
8423 }
8424
8425 arc_hdr_l2hdr_destroy(hdr);
8426 }
8427 mutex_exit(hash_lock);
8428 }
8429 mutex_exit(&dev->l2ad_mtx);
8430
8431 out:
8432 /*
8433 * We need to check if we evict all buffers, otherwise we may iterate
8434 * unnecessarily.
8435 */
8436 if (!all && rerun) {
8437 /*
8438 * Bump device hand to the device start if it is approaching the
8439 * end. l2arc_evict() has already evicted ahead for this case.
8440 */
8441 dev->l2ad_hand = dev->l2ad_start;
8442 dev->l2ad_evict = dev->l2ad_start;
8443 dev->l2ad_first = B_FALSE;
8444 goto top;
8445 }
8446
8447 ASSERT3U(dev->l2ad_hand + distance, <, dev->l2ad_end);
8448 if (!dev->l2ad_first)
8449 ASSERT3U(dev->l2ad_hand, <, dev->l2ad_evict);
8450 }
8451
8452 /*
8453 * Handle any abd transforms that might be required for writing to the L2ARC.
8454 * If successful, this function will always return an abd with the data
8455 * transformed as it is on disk in a new abd of asize bytes.
8456 */
8457 static int
8458 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
8459 abd_t **abd_out)
8460 {
8461 int ret;
8462 void *tmp = NULL;
8463 abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
8464 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
8465 uint64_t psize = HDR_GET_PSIZE(hdr);
8466 uint64_t size = arc_hdr_size(hdr);
8467 boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
8468 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
8469 dsl_crypto_key_t *dck = NULL;
8470 uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
8471 boolean_t no_crypt = B_FALSE;
8472
8473 ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8474 !HDR_COMPRESSION_ENABLED(hdr)) ||
8475 HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
8476 ASSERT3U(psize, <=, asize);
8477
8478 /*
8479 * If this data simply needs its own buffer, we simply allocate it
8480 * and copy the data. This may be done to eliminate a dependency on a
8481 * shared buffer or to reallocate the buffer to match asize.
8482 */
8483 if (HDR_HAS_RABD(hdr) && asize != psize) {
8484 ASSERT3U(asize, >=, psize);
8485 to_write = abd_alloc_for_io(asize, ismd);
8486 abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize);
8487 if (psize != asize)
8488 abd_zero_off(to_write, psize, asize - psize);
8489 goto out;
8490 }
8491
8492 if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
8493 !HDR_ENCRYPTED(hdr)) {
8494 ASSERT3U(size, ==, psize);
8495 to_write = abd_alloc_for_io(asize, ismd);
8496 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
8497 if (size != asize)
8498 abd_zero_off(to_write, size, asize - size);
8499 goto out;
8500 }
8501
8502 if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
8503 cabd = abd_alloc_for_io(asize, ismd);
8504 tmp = abd_borrow_buf(cabd, asize);
8505
8506 psize = zio_compress_data(compress, to_write, tmp, size);
8507 ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr));
8508 if (psize < asize)
8509 bzero((char *)tmp + psize, asize - psize);
8510 psize = HDR_GET_PSIZE(hdr);
8511 abd_return_buf_copy(cabd, tmp, asize);
8512 to_write = cabd;
8513 }
8514
8515 if (HDR_ENCRYPTED(hdr)) {
8516 eabd = abd_alloc_for_io(asize, ismd);
8517
8518 /*
8519 * If the dataset was disowned before the buffer
8520 * made it to this point, the key to re-encrypt
8521 * it won't be available. In this case we simply
8522 * won't write the buffer to the L2ARC.
8523 */
8524 ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
8525 FTAG, &dck);
8526 if (ret != 0)
8527 goto error;
8528
8529 ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
8530 hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
8531 hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
8532 &no_crypt);
8533 if (ret != 0)
8534 goto error;
8535
8536 if (no_crypt)
8537 abd_copy(eabd, to_write, psize);
8538
8539 if (psize != asize)
8540 abd_zero_off(eabd, psize, asize - psize);
8541
8542 /* assert that the MAC we got here matches the one we saved */
8543 ASSERT0(bcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
8544 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8545
8546 if (to_write == cabd)
8547 abd_free(cabd);
8548
8549 to_write = eabd;
8550 }
8551
8552 out:
8553 ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
8554 *abd_out = to_write;
8555 return (0);
8556
8557 error:
8558 if (dck != NULL)
8559 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8560 if (cabd != NULL)
8561 abd_free(cabd);
8562 if (eabd != NULL)
8563 abd_free(eabd);
8564
8565 *abd_out = NULL;
8566 return (ret);
8567 }
8568
8569 static void
8570 l2arc_blk_fetch_done(zio_t *zio)
8571 {
8572 l2arc_read_callback_t *cb;
8573
8574 cb = zio->io_private;
8575 if (cb->l2rcb_abd != NULL)
8576 abd_put(cb->l2rcb_abd);
8577 kmem_free(cb, sizeof (l2arc_read_callback_t));
8578 }
8579
8580 /*
8581 * Find and write ARC buffers to the L2ARC device.
8582 *
8583 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
8584 * for reading until they have completed writing.
8585 * The headroom_boost is an in-out parameter used to maintain headroom boost
8586 * state between calls to this function.
8587 *
8588 * Returns the number of bytes actually written (which may be smaller than
8589 * the delta by which the device hand has changed due to alignment and the
8590 * writing of log blocks).
8591 */
8592 static uint64_t
8593 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
8594 {
8595 arc_buf_hdr_t *hdr, *hdr_prev, *head;
8596 uint64_t write_asize, write_psize, write_lsize, headroom;
8597 boolean_t full;
8598 l2arc_write_callback_t *cb = NULL;
8599 zio_t *pio, *wzio;
8600 uint64_t guid = spa_load_guid(spa);
8601 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
8602
8603 ASSERT3P(dev->l2ad_vdev, !=, NULL);
8604
8605 pio = NULL;
8606 write_lsize = write_asize = write_psize = 0;
8607 full = B_FALSE;
8608 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
8609 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
8610
8611 /*
8612 * Copy buffers for L2ARC writing.
8613 */
8614 for (int try = 0; try < L2ARC_FEED_TYPES; try++) {
8615 /*
8616 * If try == 1 or 3, we cache MRU metadata and data
8617 * respectively.
8618 */
8619 if (l2arc_mfuonly) {
8620 if (try == 1 || try == 3)
8621 continue;
8622 }
8623
8624 multilist_sublist_t *mls = l2arc_sublist_lock(try);
8625 uint64_t passed_sz = 0;
8626
8627 VERIFY3P(mls, !=, NULL);
8628
8629 /*
8630 * L2ARC fast warmup.
8631 *
8632 * Until the ARC is warm and starts to evict, read from the
8633 * head of the ARC lists rather than the tail.
8634 */
8635 if (arc_warm == B_FALSE)
8636 hdr = multilist_sublist_head(mls);
8637 else
8638 hdr = multilist_sublist_tail(mls);
8639
8640 headroom = target_sz * l2arc_headroom;
8641 if (zfs_compressed_arc_enabled)
8642 headroom = (headroom * l2arc_headroom_boost) / 100;
8643
8644 for (; hdr; hdr = hdr_prev) {
8645 kmutex_t *hash_lock;
8646 abd_t *to_write = NULL;
8647
8648 if (arc_warm == B_FALSE)
8649 hdr_prev = multilist_sublist_next(mls, hdr);
8650 else
8651 hdr_prev = multilist_sublist_prev(mls, hdr);
8652
8653 hash_lock = HDR_LOCK(hdr);
8654 if (!mutex_tryenter(hash_lock)) {
8655 /*
8656 * Skip this buffer rather than waiting.
8657 */
8658 continue;
8659 }
8660
8661 passed_sz += HDR_GET_LSIZE(hdr);
8662 if (l2arc_headroom != 0 && passed_sz > headroom) {
8663 /*
8664 * Searched too far.
8665 */
8666 mutex_exit(hash_lock);
8667 break;
8668 }
8669
8670 if (!l2arc_write_eligible(guid, hdr)) {
8671 mutex_exit(hash_lock);
8672 continue;
8673 }
8674
8675 /*
8676 * We rely on the L1 portion of the header below, so
8677 * it's invalid for this header to have been evicted out
8678 * of the ghost cache, prior to being written out. The
8679 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8680 */
8681 ASSERT(HDR_HAS_L1HDR(hdr));
8682
8683 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8684 ASSERT3U(arc_hdr_size(hdr), >, 0);
8685 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
8686 HDR_HAS_RABD(hdr));
8687 uint64_t psize = HDR_GET_PSIZE(hdr);
8688 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
8689 psize);
8690
8691 if ((write_asize + asize) > target_sz) {
8692 full = B_TRUE;
8693 mutex_exit(hash_lock);
8694 break;
8695 }
8696
8697 /*
8698 * We rely on the L1 portion of the header below, so
8699 * it's invalid for this header to have been evicted out
8700 * of the ghost cache, prior to being written out. The
8701 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8702 */
8703 arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING);
8704 ASSERT(HDR_HAS_L1HDR(hdr));
8705
8706 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8707 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
8708 HDR_HAS_RABD(hdr));
8709 ASSERT3U(arc_hdr_size(hdr), >, 0);
8710
8711 /*
8712 * If this header has b_rabd, we can use this since it
8713 * must always match the data exactly as it exists on
8714 * disk. Otherwise, the L2ARC can normally use the
8715 * hdr's data, but if we're sharing data between the
8716 * hdr and one of its bufs, L2ARC needs its own copy of
8717 * the data so that the ZIO below can't race with the
8718 * buf consumer. To ensure that this copy will be
8719 * available for the lifetime of the ZIO and be cleaned
8720 * up afterwards, we add it to the l2arc_free_on_write
8721 * queue. If we need to apply any transforms to the
8722 * data (compression, encryption) we will also need the
8723 * extra buffer.
8724 */
8725 if (HDR_HAS_RABD(hdr) && psize == asize) {
8726 to_write = hdr->b_crypt_hdr.b_rabd;
8727 } else if ((HDR_COMPRESSION_ENABLED(hdr) ||
8728 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
8729 !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
8730 psize == asize) {
8731 to_write = hdr->b_l1hdr.b_pabd;
8732 } else {
8733 int ret;
8734 arc_buf_contents_t type = arc_buf_type(hdr);
8735
8736 ret = l2arc_apply_transforms(spa, hdr, asize,
8737 &to_write);
8738 if (ret != 0) {
8739 arc_hdr_clear_flags(hdr,
8740 ARC_FLAG_L2_WRITING);
8741 mutex_exit(hash_lock);
8742 continue;
8743 }
8744
8745 l2arc_free_abd_on_write(to_write, asize, type);
8746 }
8747
8748 if (pio == NULL) {
8749 /*
8750 * Insert a dummy header on the buflist so
8751 * l2arc_write_done() can find where the
8752 * write buffers begin without searching.
8753 */
8754 mutex_enter(&dev->l2ad_mtx);
8755 list_insert_head(&dev->l2ad_buflist, head);
8756 mutex_exit(&dev->l2ad_mtx);
8757
8758 cb = kmem_alloc(
8759 sizeof (l2arc_write_callback_t), KM_SLEEP);
8760 cb->l2wcb_dev = dev;
8761 cb->l2wcb_head = head;
8762 /*
8763 * Create a list to save allocated abd buffers
8764 * for l2arc_log_blk_commit().
8765 */
8766 list_create(&cb->l2wcb_abd_list,
8767 sizeof (l2arc_lb_abd_buf_t),
8768 offsetof(l2arc_lb_abd_buf_t, node));
8769 pio = zio_root(spa, l2arc_write_done, cb,
8770 ZIO_FLAG_CANFAIL);
8771 }
8772
8773 hdr->b_l2hdr.b_dev = dev;
8774 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
8775 hdr->b_l2hdr.b_arcs_state =
8776 hdr->b_l1hdr.b_state->arcs_state;
8777 arc_hdr_set_flags(hdr,
8778 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
8779
8780 mutex_enter(&dev->l2ad_mtx);
8781 list_insert_head(&dev->l2ad_buflist, hdr);
8782 mutex_exit(&dev->l2ad_mtx);
8783
8784 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
8785 arc_hdr_size(hdr), hdr);
8786
8787 wzio = zio_write_phys(pio, dev->l2ad_vdev,
8788 hdr->b_l2hdr.b_daddr, asize, to_write,
8789 ZIO_CHECKSUM_OFF, NULL, hdr,
8790 ZIO_PRIORITY_ASYNC_WRITE,
8791 ZIO_FLAG_CANFAIL, B_FALSE);
8792
8793 write_lsize += HDR_GET_LSIZE(hdr);
8794 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
8795 zio_t *, wzio);
8796
8797 write_psize += psize;
8798 write_asize += asize;
8799 dev->l2ad_hand += asize;
8800 l2arc_hdr_arcstats_increment(hdr);
8801 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
8802
8803 mutex_exit(hash_lock);
8804
8805 /*
8806 * Append buf info to current log and commit if full.
8807 * arcstat_l2_{size,asize} kstats are updated
8808 * internally.
8809 */
8810 if (l2arc_log_blk_insert(dev, hdr))
8811 l2arc_log_blk_commit(dev, pio, cb);
8812
8813 (void) zio_nowait(wzio);
8814 }
8815
8816 multilist_sublist_unlock(mls);
8817
8818 if (full == B_TRUE)
8819 break;
8820 }
8821
8822 /* No buffers selected for writing? */
8823 if (pio == NULL) {
8824 ASSERT0(write_lsize);
8825 ASSERT(!HDR_HAS_L1HDR(head));
8826 kmem_cache_free(hdr_l2only_cache, head);
8827
8828 /*
8829 * Although we did not write any buffers l2ad_evict may
8830 * have advanced.
8831 */
8832 if (dev->l2ad_evict != l2dhdr->dh_evict)
8833 l2arc_dev_hdr_update(dev);
8834
8835 return (0);
8836 }
8837
8838 if (!dev->l2ad_first)
8839 ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
8840
8841 ASSERT3U(write_asize, <=, target_sz);
8842 ARCSTAT_BUMP(arcstat_l2_writes_sent);
8843 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
8844
8845 dev->l2ad_writing = B_TRUE;
8846 (void) zio_wait(pio);
8847 dev->l2ad_writing = B_FALSE;
8848
8849 /*
8850 * Update the device header after the zio completes as
8851 * l2arc_write_done() may have updated the memory holding the log block
8852 * pointers in the device header.
8853 */
8854 l2arc_dev_hdr_update(dev);
8855
8856 return (write_asize);
8857 }
8858
8859 static boolean_t
8860 l2arc_hdr_limit_reached(void)
8861 {
8862 int64_t s = aggsum_upper_bound(&astat_l2_hdr_size);
8863
8864 return (arc_reclaim_needed() || (s > arc_meta_limit * 3 / 4) ||
8865 (s > (arc_warm ? arc_c : arc_c_max) * l2arc_meta_percent / 100));
8866 }
8867
8868 /*
8869 * This thread feeds the L2ARC at regular intervals. This is the beating
8870 * heart of the L2ARC.
8871 */
8872 /* ARGSUSED */
8873 static void
8874 l2arc_feed_thread(void *unused)
8875 {
8876 callb_cpr_t cpr;
8877 l2arc_dev_t *dev;
8878 spa_t *spa;
8879 uint64_t size, wrote;
8880 clock_t begin, next = ddi_get_lbolt();
8881
8882 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
8883
8884 mutex_enter(&l2arc_feed_thr_lock);
8885
8886 while (l2arc_thread_exit == 0) {
8887 CALLB_CPR_SAFE_BEGIN(&cpr);
8888 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
8889 next);
8890 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
8891 next = ddi_get_lbolt() + hz;
8892
8893 /*
8894 * Quick check for L2ARC devices.
8895 */
8896 mutex_enter(&l2arc_dev_mtx);
8897 if (l2arc_ndev == 0) {
8898 mutex_exit(&l2arc_dev_mtx);
8899 continue;
8900 }
8901 mutex_exit(&l2arc_dev_mtx);
8902 begin = ddi_get_lbolt();
8903
8904 /*
8905 * This selects the next l2arc device to write to, and in
8906 * doing so the next spa to feed from: dev->l2ad_spa. This
8907 * will return NULL if there are now no l2arc devices or if
8908 * they are all faulted.
8909 *
8910 * If a device is returned, its spa's config lock is also
8911 * held to prevent device removal. l2arc_dev_get_next()
8912 * will grab and release l2arc_dev_mtx.
8913 */
8914 if ((dev = l2arc_dev_get_next()) == NULL)
8915 continue;
8916
8917 spa = dev->l2ad_spa;
8918 ASSERT3P(spa, !=, NULL);
8919
8920 /*
8921 * If the pool is read-only then force the feed thread to
8922 * sleep a little longer.
8923 */
8924 if (!spa_writeable(spa)) {
8925 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
8926 spa_config_exit(spa, SCL_L2ARC, dev);
8927 continue;
8928 }
8929
8930 /*
8931 * Avoid contributing to memory pressure.
8932 */
8933 if (l2arc_hdr_limit_reached()) {
8934 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
8935 spa_config_exit(spa, SCL_L2ARC, dev);
8936 continue;
8937 }
8938
8939 ARCSTAT_BUMP(arcstat_l2_feeds);
8940
8941 size = l2arc_write_size(dev);
8942
8943 /*
8944 * Evict L2ARC buffers that will be overwritten.
8945 */
8946 l2arc_evict(dev, size, B_FALSE);
8947
8948 /*
8949 * Write ARC buffers.
8950 */
8951 wrote = l2arc_write_buffers(spa, dev, size);
8952
8953 /*
8954 * Calculate interval between writes.
8955 */
8956 next = l2arc_write_interval(begin, size, wrote);
8957 spa_config_exit(spa, SCL_L2ARC, dev);
8958 }
8959
8960 l2arc_thread_exit = 0;
8961 cv_broadcast(&l2arc_feed_thr_cv);
8962 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
8963 thread_exit();
8964 }
8965
8966 boolean_t
8967 l2arc_vdev_present(vdev_t *vd)
8968 {
8969 return (l2arc_vdev_get(vd) != NULL);
8970 }
8971
8972 /*
8973 * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if
8974 * the vdev_t isn't an L2ARC device.
8975 */
8976 static l2arc_dev_t *
8977 l2arc_vdev_get(vdev_t *vd)
8978 {
8979 l2arc_dev_t *dev;
8980
8981 mutex_enter(&l2arc_dev_mtx);
8982 for (dev = list_head(l2arc_dev_list); dev != NULL;
8983 dev = list_next(l2arc_dev_list, dev)) {
8984 if (dev->l2ad_vdev == vd)
8985 break;
8986 }
8987 mutex_exit(&l2arc_dev_mtx);
8988
8989 return (dev);
8990 }
8991
8992 /*
8993 * Add a vdev for use by the L2ARC. By this point the spa has already
8994 * validated the vdev and opened it.
8995 */
8996 void
8997 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
8998 {
8999 l2arc_dev_t *adddev;
9000 uint64_t l2dhdr_asize;
9001
9002 ASSERT(!l2arc_vdev_present(vd));
9003
9004 /*
9005 * Create a new l2arc device entry.
9006 */
9007 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
9008 adddev->l2ad_spa = spa;
9009 adddev->l2ad_vdev = vd;
9010 /* leave extra size for an l2arc device header */
9011 l2dhdr_asize = adddev->l2ad_dev_hdr_asize =
9012 MAX(sizeof (*adddev->l2ad_dev_hdr), 1 << vd->vdev_ashift);
9013 adddev->l2ad_start = VDEV_LABEL_START_SIZE + l2dhdr_asize;
9014 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
9015 ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end);
9016 adddev->l2ad_hand = adddev->l2ad_start;
9017 adddev->l2ad_evict = adddev->l2ad_start;
9018 adddev->l2ad_first = B_TRUE;
9019 adddev->l2ad_writing = B_FALSE;
9020 adddev->l2ad_dev_hdr = kmem_zalloc(l2dhdr_asize, KM_SLEEP);
9021
9022 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
9023 /*
9024 * This is a list of all ARC buffers that are still valid on the
9025 * device.
9026 */
9027 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
9028 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
9029
9030 /*
9031 * This is a list of pointers to log blocks that are still present
9032 * on the device.
9033 */
9034 list_create(&adddev->l2ad_lbptr_list, sizeof (l2arc_lb_ptr_buf_t),
9035 offsetof(l2arc_lb_ptr_buf_t, node));
9036
9037 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
9038 zfs_refcount_create(&adddev->l2ad_alloc);
9039 zfs_refcount_create(&adddev->l2ad_lb_asize);
9040 zfs_refcount_create(&adddev->l2ad_lb_count);
9041
9042 /*
9043 * Add device to global list
9044 */
9045 mutex_enter(&l2arc_dev_mtx);
9046 list_insert_head(l2arc_dev_list, adddev);
9047 atomic_inc_64(&l2arc_ndev);
9048 mutex_exit(&l2arc_dev_mtx);
9049
9050 /*
9051 * Decide if vdev is eligible for L2ARC rebuild
9052 */
9053 l2arc_rebuild_vdev(adddev->l2ad_vdev, B_FALSE);
9054 }
9055
9056 void
9057 l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen)
9058 {
9059 l2arc_dev_t *dev = NULL;
9060 l2arc_dev_hdr_phys_t *l2dhdr;
9061 uint64_t l2dhdr_asize;
9062 spa_t *spa;
9063
9064 dev = l2arc_vdev_get(vd);
9065 ASSERT3P(dev, !=, NULL);
9066 spa = dev->l2ad_spa;
9067 l2dhdr = dev->l2ad_dev_hdr;
9068 l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9069
9070 /*
9071 * The L2ARC has to hold at least the payload of one log block for
9072 * them to be restored (persistent L2ARC). The payload of a log block
9073 * depends on the amount of its log entries. We always write log blocks
9074 * with 1022 entries. How many of them are committed or restored depends
9075 * on the size of the L2ARC device. Thus the maximum payload of
9076 * one log block is 1022 * SPA_MAXBLOCKSIZE = 16GB. If the L2ARC device
9077 * is less than that, we reduce the amount of committed and restored
9078 * log entries per block so as to enable persistence.
9079 */
9080 if (dev->l2ad_end < l2arc_rebuild_blocks_min_l2size) {
9081 dev->l2ad_log_entries = 0;
9082 } else {
9083 dev->l2ad_log_entries = MIN((dev->l2ad_end -
9084 dev->l2ad_start) >> SPA_MAXBLOCKSHIFT,
9085 L2ARC_LOG_BLK_MAX_ENTRIES);
9086 }
9087
9088 /*
9089 * Read the device header, if an error is returned do not rebuild L2ARC.
9090 */
9091 if (l2arc_dev_hdr_read(dev) == 0 && dev->l2ad_log_entries > 0) {
9092 /*
9093 * If we are onlining a cache device (vdev_reopen) that was
9094 * still present (l2arc_vdev_present()) and rebuild is enabled,
9095 * we should evict all ARC buffers and pointers to log blocks
9096 * and reclaim their space before restoring its contents to
9097 * L2ARC.
9098 */
9099 if (reopen) {
9100 if (!l2arc_rebuild_enabled) {
9101 return;
9102 } else {
9103 l2arc_evict(dev, 0, B_TRUE);
9104 /* start a new log block */
9105 dev->l2ad_log_ent_idx = 0;
9106 dev->l2ad_log_blk_payload_asize = 0;
9107 dev->l2ad_log_blk_payload_start = 0;
9108 }
9109 }
9110 /*
9111 * Just mark the device as pending for a rebuild. We won't
9112 * be starting a rebuild in line here as it would block pool
9113 * import. Instead spa_load_impl will hand that off to an
9114 * async task which will call l2arc_spa_rebuild_start.
9115 */
9116 dev->l2ad_rebuild = B_TRUE;
9117 } else if (spa_writeable(spa)) {
9118 /*
9119 * In this case create a new header. We zero out the memory
9120 * holding the header to reset dh_start_lbps.
9121 */
9122 bzero(l2dhdr, l2dhdr_asize);
9123 l2arc_dev_hdr_update(dev);
9124 }
9125 }
9126
9127 /*
9128 * Remove a vdev from the L2ARC.
9129 */
9130 void
9131 l2arc_remove_vdev(vdev_t *vd)
9132 {
9133 l2arc_dev_t *remdev = NULL;
9134
9135 /*
9136 * Find the device by vdev
9137 */
9138 remdev = l2arc_vdev_get(vd);
9139 ASSERT3P(remdev, !=, NULL);
9140
9141 /*
9142 * Cancel any ongoing or scheduled rebuild.
9143 */
9144 mutex_enter(&l2arc_rebuild_thr_lock);
9145 if (remdev->l2ad_rebuild_began == B_TRUE) {
9146 remdev->l2ad_rebuild_cancel = B_TRUE;
9147 while (remdev->l2ad_rebuild == B_TRUE)
9148 cv_wait(&l2arc_rebuild_thr_cv, &l2arc_rebuild_thr_lock);
9149 }
9150 mutex_exit(&l2arc_rebuild_thr_lock);
9151
9152 /*
9153 * Remove device from global list
9154 */
9155 mutex_enter(&l2arc_dev_mtx);
9156 list_remove(l2arc_dev_list, remdev);
9157 l2arc_dev_last = NULL; /* may have been invalidated */
9158 atomic_dec_64(&l2arc_ndev);
9159 mutex_exit(&l2arc_dev_mtx);
9160
9161 /*
9162 * Clear all buflists and ARC references. L2ARC device flush.
9163 */
9164 l2arc_evict(remdev, 0, B_TRUE);
9165 list_destroy(&remdev->l2ad_buflist);
9166 ASSERT(list_is_empty(&remdev->l2ad_lbptr_list));
9167 list_destroy(&remdev->l2ad_lbptr_list);
9168 mutex_destroy(&remdev->l2ad_mtx);
9169 zfs_refcount_destroy(&remdev->l2ad_alloc);
9170 zfs_refcount_destroy(&remdev->l2ad_lb_asize);
9171 zfs_refcount_destroy(&remdev->l2ad_lb_count);
9172 kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize);
9173 kmem_free(remdev, sizeof (l2arc_dev_t));
9174 }
9175
9176 void
9177 l2arc_init(void)
9178 {
9179 l2arc_thread_exit = 0;
9180 l2arc_ndev = 0;
9181 l2arc_writes_sent = 0;
9182 l2arc_writes_done = 0;
9183
9184 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9185 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
9186 mutex_init(&l2arc_rebuild_thr_lock, NULL, MUTEX_DEFAULT, NULL);
9187 cv_init(&l2arc_rebuild_thr_cv, NULL, CV_DEFAULT, NULL);
9188 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
9189 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
9190
9191 l2arc_dev_list = &L2ARC_dev_list;
9192 l2arc_free_on_write = &L2ARC_free_on_write;
9193 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
9194 offsetof(l2arc_dev_t, l2ad_node));
9195 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
9196 offsetof(l2arc_data_free_t, l2df_list_node));
9197 }
9198
9199 void
9200 l2arc_fini(void)
9201 {
9202 /*
9203 * This is called from dmu_fini(), which is called from spa_fini();
9204 * Because of this, we can assume that all l2arc devices have
9205 * already been removed when the pools themselves were removed.
9206 */
9207
9208 l2arc_do_free_on_write();
9209
9210 mutex_destroy(&l2arc_feed_thr_lock);
9211 cv_destroy(&l2arc_feed_thr_cv);
9212 mutex_destroy(&l2arc_rebuild_thr_lock);
9213 cv_destroy(&l2arc_rebuild_thr_cv);
9214 mutex_destroy(&l2arc_dev_mtx);
9215 mutex_destroy(&l2arc_free_on_write_mtx);
9216
9217 list_destroy(l2arc_dev_list);
9218 list_destroy(l2arc_free_on_write);
9219 }
9220
9221 void
9222 l2arc_start(void)
9223 {
9224 if (!(spa_mode_global & FWRITE))
9225 return;
9226
9227 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
9228 TS_RUN, minclsyspri);
9229 }
9230
9231 void
9232 l2arc_stop(void)
9233 {
9234 if (!(spa_mode_global & FWRITE))
9235 return;
9236
9237 mutex_enter(&l2arc_feed_thr_lock);
9238 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
9239 l2arc_thread_exit = 1;
9240 while (l2arc_thread_exit != 0)
9241 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
9242 mutex_exit(&l2arc_feed_thr_lock);
9243 }
9244
9245 /*
9246 * Punches out rebuild threads for the L2ARC devices in a spa. This should
9247 * be called after pool import from the spa async thread, since starting
9248 * these threads directly from spa_import() will make them part of the
9249 * "zpool import" context and delay process exit (and thus pool import).
9250 */
9251 void
9252 l2arc_spa_rebuild_start(spa_t *spa)
9253 {
9254 ASSERT(MUTEX_HELD(&spa_namespace_lock));
9255
9256 /*
9257 * Locate the spa's l2arc devices and kick off rebuild threads.
9258 */
9259 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
9260 l2arc_dev_t *dev =
9261 l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]);
9262 if (dev == NULL) {
9263 /* Don't attempt a rebuild if the vdev is UNAVAIL */
9264 continue;
9265 }
9266 mutex_enter(&l2arc_rebuild_thr_lock);
9267 if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) {
9268 dev->l2ad_rebuild_began = B_TRUE;
9269 (void) thread_create(NULL, 0,
9270 (void (*)(void *))l2arc_dev_rebuild_start,
9271 dev, 0, &p0, TS_RUN, minclsyspri);
9272 }
9273 mutex_exit(&l2arc_rebuild_thr_lock);
9274 }
9275 }
9276
9277 /*
9278 * Main entry point for L2ARC rebuilding.
9279 */
9280 static void
9281 l2arc_dev_rebuild_start(l2arc_dev_t *dev)
9282 {
9283 VERIFY(!dev->l2ad_rebuild_cancel);
9284 VERIFY(dev->l2ad_rebuild);
9285 (void) l2arc_rebuild(dev);
9286 mutex_enter(&l2arc_rebuild_thr_lock);
9287 dev->l2ad_rebuild_began = B_FALSE;
9288 dev->l2ad_rebuild = B_FALSE;
9289 mutex_exit(&l2arc_rebuild_thr_lock);
9290
9291 thread_exit();
9292 }
9293
9294 /*
9295 * This function implements the actual L2ARC metadata rebuild. It:
9296 * starts reading the log block chain and restores each block's contents
9297 * to memory (reconstructing arc_buf_hdr_t's).
9298 *
9299 * Operation stops under any of the following conditions:
9300 *
9301 * 1) We reach the end of the log block chain.
9302 * 2) We encounter *any* error condition (cksum errors, io errors)
9303 */
9304 static int
9305 l2arc_rebuild(l2arc_dev_t *dev)
9306 {
9307 vdev_t *vd = dev->l2ad_vdev;
9308 spa_t *spa = vd->vdev_spa;
9309 int err = 0;
9310 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9311 l2arc_log_blk_phys_t *this_lb, *next_lb;
9312 zio_t *this_io = NULL, *next_io = NULL;
9313 l2arc_log_blkptr_t lbps[2];
9314 l2arc_lb_ptr_buf_t *lb_ptr_buf;
9315 boolean_t lock_held;
9316
9317 this_lb = kmem_zalloc(sizeof (*this_lb), KM_SLEEP);
9318 next_lb = kmem_zalloc(sizeof (*next_lb), KM_SLEEP);
9319
9320 /*
9321 * We prevent device removal while issuing reads to the device,
9322 * then during the rebuilding phases we drop this lock again so
9323 * that a spa_unload or device remove can be initiated - this is
9324 * safe, because the spa will signal us to stop before removing
9325 * our device and wait for us to stop.
9326 */
9327 spa_config_enter(spa, SCL_L2ARC, vd, RW_READER);
9328 lock_held = B_TRUE;
9329
9330 /*
9331 * Retrieve the persistent L2ARC device state.
9332 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9333 */
9334 dev->l2ad_evict = MAX(l2dhdr->dh_evict, dev->l2ad_start);
9335 dev->l2ad_hand = MAX(l2dhdr->dh_start_lbps[0].lbp_daddr +
9336 L2BLK_GET_PSIZE((&l2dhdr->dh_start_lbps[0])->lbp_prop),
9337 dev->l2ad_start);
9338 dev->l2ad_first = !!(l2dhdr->dh_flags & L2ARC_DEV_HDR_EVICT_FIRST);
9339
9340 /*
9341 * In case the zfs module parameter l2arc_rebuild_enabled is false
9342 * we do not start the rebuild process.
9343 */
9344 if (!l2arc_rebuild_enabled)
9345 goto out;
9346
9347 /* Prepare the rebuild process */
9348 bcopy(l2dhdr->dh_start_lbps, lbps, sizeof (lbps));
9349
9350 /* Start the rebuild process */
9351 for (;;) {
9352 if (!l2arc_log_blkptr_valid(dev, &lbps[0]))
9353 break;
9354
9355 if ((err = l2arc_log_blk_read(dev, &lbps[0], &lbps[1],
9356 this_lb, next_lb, this_io, &next_io)) != 0)
9357 goto out;
9358
9359 /*
9360 * Our memory pressure valve. If the system is running low
9361 * on memory, rather than swamping memory with new ARC buf
9362 * hdrs, we opt not to rebuild the L2ARC. At this point,
9363 * however, we have already set up our L2ARC dev to chain in
9364 * new metadata log blocks, so the user may choose to offline/
9365 * online the L2ARC dev at a later time (or re-import the pool)
9366 * to reconstruct it (when there's less memory pressure).
9367 */
9368 if (l2arc_hdr_limit_reached()) {
9369 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem);
9370 cmn_err(CE_NOTE, "System running low on memory, "
9371 "aborting L2ARC rebuild.");
9372 err = SET_ERROR(ENOMEM);
9373 goto out;
9374 }
9375
9376 spa_config_exit(spa, SCL_L2ARC, vd);
9377 lock_held = B_FALSE;
9378
9379 /*
9380 * Now that we know that the next_lb checks out alright, we
9381 * can start reconstruction from this log block.
9382 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9383 */
9384 uint64_t asize = L2BLK_GET_PSIZE((&lbps[0])->lbp_prop);
9385 l2arc_log_blk_restore(dev, this_lb, asize);
9386
9387 /*
9388 * log block restored, include its pointer in the list of
9389 * pointers to log blocks present in the L2ARC device.
9390 */
9391 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
9392 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t),
9393 KM_SLEEP);
9394 bcopy(&lbps[0], lb_ptr_buf->lb_ptr,
9395 sizeof (l2arc_log_blkptr_t));
9396 mutex_enter(&dev->l2ad_mtx);
9397 list_insert_tail(&dev->l2ad_lbptr_list, lb_ptr_buf);
9398 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
9399 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
9400 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
9401 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
9402 mutex_exit(&dev->l2ad_mtx);
9403 vdev_space_update(vd, asize, 0, 0);
9404
9405 /* BEGIN CSTYLED */
9406 /*
9407 * Protection against loops of log blocks:
9408 *
9409 * l2ad_hand l2ad_evict
9410 * V V
9411 * l2ad_start |=======================================| l2ad_end
9412 * -----|||----|||---|||----|||
9413 * (3) (2) (1) (0)
9414 * ---|||---|||----|||---|||
9415 * (7) (6) (5) (4)
9416 *
9417 * In this situation the pointer of log block (4) passes
9418 * l2arc_log_blkptr_valid() but the log block should not be
9419 * restored as it is overwritten by the payload of log block
9420 * (0). Only log blocks (0)-(3) should be restored. We check
9421 * whether l2ad_evict lies in between the payload starting
9422 * offset of the next log block (lbps[1].lbp_payload_start)
9423 * and the payload starting offset of the present log block
9424 * (lbps[0].lbp_payload_start). If true and this isn't the
9425 * first pass, we are looping from the beginning and we should
9426 * stop.
9427 */
9428 /* END CSTYLED */
9429 if (l2arc_range_check_overlap(lbps[1].lbp_payload_start,
9430 lbps[0].lbp_payload_start, dev->l2ad_evict) &&
9431 !dev->l2ad_first)
9432 goto out;
9433
9434 for (;;) {
9435 mutex_enter(&l2arc_rebuild_thr_lock);
9436 if (dev->l2ad_rebuild_cancel) {
9437 dev->l2ad_rebuild = B_FALSE;
9438 cv_signal(&l2arc_rebuild_thr_cv);
9439 mutex_exit(&l2arc_rebuild_thr_lock);
9440 err = SET_ERROR(ECANCELED);
9441 goto out;
9442 }
9443 mutex_exit(&l2arc_rebuild_thr_lock);
9444 if (spa_config_tryenter(spa, SCL_L2ARC, vd,
9445 RW_READER)) {
9446 lock_held = B_TRUE;
9447 break;
9448 }
9449 /*
9450 * L2ARC config lock held by somebody in writer,
9451 * possibly due to them trying to remove us. They'll
9452 * likely to want us to shut down, so after a little
9453 * delay, we check l2ad_rebuild_cancel and retry
9454 * the lock again.
9455 */
9456 delay(1);
9457 }
9458
9459 /*
9460 * Continue with the next log block.
9461 */
9462 lbps[0] = lbps[1];
9463 lbps[1] = this_lb->lb_prev_lbp;
9464 PTR_SWAP(this_lb, next_lb);
9465 this_io = next_io;
9466 next_io = NULL;
9467 }
9468
9469 if (this_io != NULL)
9470 l2arc_log_blk_fetch_abort(this_io);
9471 out:
9472 if (next_io != NULL)
9473 l2arc_log_blk_fetch_abort(next_io);
9474 kmem_free(this_lb, sizeof (*this_lb));
9475 kmem_free(next_lb, sizeof (*next_lb));
9476
9477 if (!l2arc_rebuild_enabled) {
9478 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9479 "disabled");
9480 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) > 0) {
9481 ARCSTAT_BUMP(arcstat_l2_rebuild_success);
9482 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9483 "successful, restored %llu blocks",
9484 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
9485 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) == 0) {
9486 /*
9487 * No error but also nothing restored, meaning the lbps array
9488 * in the device header points to invalid/non-present log
9489 * blocks. Reset the header.
9490 */
9491 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9492 "no valid log blocks");
9493 bzero(l2dhdr, dev->l2ad_dev_hdr_asize);
9494 l2arc_dev_hdr_update(dev);
9495 } else if (err == ECANCELED) {
9496 /*
9497 * In case the rebuild was canceled do not log to spa history
9498 * log as the pool may be in the process of being removed.
9499 */
9500 zfs_dbgmsg("L2ARC rebuild aborted, restored %llu blocks",
9501 zfs_refcount_count(&dev->l2ad_lb_count));
9502 } else if (err != 0) {
9503 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9504 "aborted, restored %llu blocks",
9505 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
9506 }
9507
9508 if (lock_held)
9509 spa_config_exit(spa, SCL_L2ARC, vd);
9510
9511 return (err);
9512 }
9513
9514 /*
9515 * Attempts to read the device header on the provided L2ARC device and writes
9516 * it to `hdr'. On success, this function returns 0, otherwise the appropriate
9517 * error code is returned.
9518 */
9519 static int
9520 l2arc_dev_hdr_read(l2arc_dev_t *dev)
9521 {
9522 int err;
9523 uint64_t guid;
9524 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9525 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9526 abd_t *abd;
9527
9528 guid = spa_guid(dev->l2ad_vdev->vdev_spa);
9529
9530 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
9531
9532 err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev,
9533 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd,
9534 ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_SYNC_READ,
9535 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
9536 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY |
9537 ZIO_FLAG_SPECULATIVE, B_FALSE));
9538
9539 abd_put(abd);
9540
9541 if (err != 0) {
9542 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_dh_errors);
9543 zfs_dbgmsg("L2ARC IO error (%d) while reading device header, "
9544 "vdev guid: %llu", err, dev->l2ad_vdev->vdev_guid);
9545 return (err);
9546 }
9547
9548 if (l2dhdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC))
9549 byteswap_uint64_array(l2dhdr, sizeof (*l2dhdr));
9550
9551 if (l2dhdr->dh_magic != L2ARC_DEV_HDR_MAGIC ||
9552 l2dhdr->dh_spa_guid != guid ||
9553 l2dhdr->dh_vdev_guid != dev->l2ad_vdev->vdev_guid ||
9554 l2dhdr->dh_version != L2ARC_PERSISTENT_VERSION ||
9555 l2dhdr->dh_log_entries != dev->l2ad_log_entries ||
9556 l2dhdr->dh_end != dev->l2ad_end ||
9557 !l2arc_range_check_overlap(dev->l2ad_start, dev->l2ad_end,
9558 l2dhdr->dh_evict)) {
9559 /*
9560 * Attempt to rebuild a device containing no actual dev hdr
9561 * or containing a header from some other pool or from another
9562 * version of persistent L2ARC.
9563 */
9564 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported);
9565 return (SET_ERROR(ENOTSUP));
9566 }
9567
9568 return (0);
9569 }
9570
9571 /*
9572 * Reads L2ARC log blocks from storage and validates their contents.
9573 *
9574 * This function implements a simple fetcher to make sure that while
9575 * we're processing one buffer the L2ARC is already fetching the next
9576 * one in the chain.
9577 *
9578 * The arguments this_lp and next_lp point to the current and next log block
9579 * address in the block chain. Similarly, this_lb and next_lb hold the
9580 * l2arc_log_blk_phys_t's of the current and next L2ARC blk.
9581 *
9582 * The `this_io' and `next_io' arguments are used for block fetching.
9583 * When issuing the first blk IO during rebuild, you should pass NULL for
9584 * `this_io'. This function will then issue a sync IO to read the block and
9585 * also issue an async IO to fetch the next block in the block chain. The
9586 * fetched IO is returned in `next_io'. On subsequent calls to this
9587 * function, pass the value returned in `next_io' from the previous call
9588 * as `this_io' and a fresh `next_io' pointer to hold the next fetch IO.
9589 * Prior to the call, you should initialize your `next_io' pointer to be
9590 * NULL. If no fetch IO was issued, the pointer is left set at NULL.
9591 *
9592 * On success, this function returns 0, otherwise it returns an appropriate
9593 * error code. On error the fetching IO is aborted and cleared before
9594 * returning from this function. Therefore, if we return `success', the
9595 * caller can assume that we have taken care of cleanup of fetch IOs.
9596 */
9597 static int
9598 l2arc_log_blk_read(l2arc_dev_t *dev,
9599 const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp,
9600 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
9601 zio_t *this_io, zio_t **next_io)
9602 {
9603 int err = 0;
9604 zio_cksum_t cksum;
9605 abd_t *abd = NULL;
9606 uint64_t asize;
9607
9608 ASSERT(this_lbp != NULL && next_lbp != NULL);
9609 ASSERT(this_lb != NULL && next_lb != NULL);
9610 ASSERT(next_io != NULL && *next_io == NULL);
9611 ASSERT(l2arc_log_blkptr_valid(dev, this_lbp));
9612
9613 /*
9614 * Check to see if we have issued the IO for this log block in a
9615 * previous run. If not, this is the first call, so issue it now.
9616 */
9617 if (this_io == NULL) {
9618 this_io = l2arc_log_blk_fetch(dev->l2ad_vdev, this_lbp,
9619 this_lb);
9620 }
9621
9622 /*
9623 * Peek to see if we can start issuing the next IO immediately.
9624 */
9625 if (l2arc_log_blkptr_valid(dev, next_lbp)) {
9626 /*
9627 * Start issuing IO for the next log block early - this
9628 * should help keep the L2ARC device busy while we
9629 * decompress and restore this log block.
9630 */
9631 *next_io = l2arc_log_blk_fetch(dev->l2ad_vdev, next_lbp,
9632 next_lb);
9633 }
9634
9635 /* Wait for the IO to read this log block to complete */
9636 if ((err = zio_wait(this_io)) != 0) {
9637 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors);
9638 zfs_dbgmsg("L2ARC IO error (%d) while reading log block, "
9639 "offset: %llu, vdev guid: %llu", err, this_lbp->lbp_daddr,
9640 dev->l2ad_vdev->vdev_guid);
9641 goto cleanup;
9642 }
9643
9644 /*
9645 * Make sure the buffer checks out.
9646 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9647 */
9648 asize = L2BLK_GET_PSIZE((this_lbp)->lbp_prop);
9649 fletcher_4_native(this_lb, asize, NULL, &cksum);
9650 if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) {
9651 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_lb_errors);
9652 zfs_dbgmsg("L2ARC log block cksum failed, offset: %llu, "
9653 "vdev guid: %llu, l2ad_hand: %llu, l2ad_evict: %llu",
9654 this_lbp->lbp_daddr, dev->l2ad_vdev->vdev_guid,
9655 dev->l2ad_hand, dev->l2ad_evict);
9656 err = SET_ERROR(ECKSUM);
9657 goto cleanup;
9658 }
9659
9660 /* Now we can take our time decoding this buffer */
9661 switch (L2BLK_GET_COMPRESS((this_lbp)->lbp_prop)) {
9662 case ZIO_COMPRESS_OFF:
9663 break;
9664 case ZIO_COMPRESS_LZ4:
9665 abd = abd_alloc_for_io(asize, B_TRUE);
9666 abd_copy_from_buf_off(abd, this_lb, 0, asize);
9667 if ((err = zio_decompress_data(
9668 L2BLK_GET_COMPRESS((this_lbp)->lbp_prop),
9669 abd, this_lb, asize, sizeof (*this_lb))) != 0) {
9670 err = SET_ERROR(EINVAL);
9671 goto cleanup;
9672 }
9673 break;
9674 default:
9675 err = SET_ERROR(EINVAL);
9676 goto cleanup;
9677 }
9678 if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC))
9679 byteswap_uint64_array(this_lb, sizeof (*this_lb));
9680 if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) {
9681 err = SET_ERROR(EINVAL);
9682 goto cleanup;
9683 }
9684 cleanup:
9685 /* Abort an in-flight fetch I/O in case of error */
9686 if (err != 0 && *next_io != NULL) {
9687 l2arc_log_blk_fetch_abort(*next_io);
9688 *next_io = NULL;
9689 }
9690 if (abd != NULL)
9691 abd_free(abd);
9692 return (err);
9693 }
9694
9695 /*
9696 * Restores the payload of a log block to ARC. This creates empty ARC hdr
9697 * entries which only contain an l2arc hdr, essentially restoring the
9698 * buffers to their L2ARC evicted state. This function also updates space
9699 * usage on the L2ARC vdev to make sure it tracks restored buffers.
9700 */
9701 static void
9702 l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb,
9703 uint64_t lb_asize)
9704 {
9705 uint64_t size = 0, asize = 0;
9706 uint64_t log_entries = dev->l2ad_log_entries;
9707
9708 /*
9709 * Usually arc_adapt() is called only for data, not headers, but
9710 * since we may allocate significant amount of memory here, let ARC
9711 * grow its arc_c.
9712 */
9713 arc_adapt(log_entries * HDR_L2ONLY_SIZE, arc_l2c_only);
9714
9715 for (int i = log_entries - 1; i >= 0; i--) {
9716 /*
9717 * Restore goes in the reverse temporal direction to preserve
9718 * correct temporal ordering of buffers in the l2ad_buflist.
9719 * l2arc_hdr_restore also does a list_insert_tail instead of
9720 * list_insert_head on the l2ad_buflist:
9721 *
9722 * LIST l2ad_buflist LIST
9723 * HEAD <------ (time) ------ TAIL
9724 * direction +-----+-----+-----+-----+-----+ direction
9725 * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild
9726 * fill +-----+-----+-----+-----+-----+
9727 * ^ ^
9728 * | |
9729 * | |
9730 * l2arc_feed_thread l2arc_rebuild
9731 * will place new bufs here restores bufs here
9732 *
9733 * During l2arc_rebuild() the device is not used by
9734 * l2arc_feed_thread() as dev->l2ad_rebuild is set to true.
9735 */
9736 size += L2BLK_GET_LSIZE((&lb->lb_entries[i])->le_prop);
9737 asize += vdev_psize_to_asize(dev->l2ad_vdev,
9738 L2BLK_GET_PSIZE((&lb->lb_entries[i])->le_prop));
9739 l2arc_hdr_restore(&lb->lb_entries[i], dev);
9740 }
9741
9742 /*
9743 * Record rebuild stats:
9744 * size Logical size of restored buffers in the L2ARC
9745 * asize Aligned size of restored buffers in the L2ARC
9746 */
9747 ARCSTAT_INCR(arcstat_l2_rebuild_size, size);
9748 ARCSTAT_INCR(arcstat_l2_rebuild_asize, asize);
9749 ARCSTAT_INCR(arcstat_l2_rebuild_bufs, log_entries);
9750 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, lb_asize);
9751 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, asize / lb_asize);
9752 ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks);
9753 }
9754
9755 /*
9756 * Restores a single ARC buf hdr from a log entry. The ARC buffer is put
9757 * into a state indicating that it has been evicted to L2ARC.
9758 */
9759 static void
9760 l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev)
9761 {
9762 arc_buf_hdr_t *hdr, *exists;
9763 kmutex_t *hash_lock;
9764 arc_buf_contents_t type = L2BLK_GET_TYPE((le)->le_prop);
9765 uint64_t asize;
9766
9767 /*
9768 * Do all the allocation before grabbing any locks, this lets us
9769 * sleep if memory is full and we don't have to deal with failed
9770 * allocations.
9771 */
9772 hdr = arc_buf_alloc_l2only(L2BLK_GET_LSIZE((le)->le_prop), type,
9773 dev, le->le_dva, le->le_daddr,
9774 L2BLK_GET_PSIZE((le)->le_prop), le->le_birth,
9775 L2BLK_GET_COMPRESS((le)->le_prop),
9776 L2BLK_GET_PROTECTED((le)->le_prop),
9777 L2BLK_GET_PREFETCH((le)->le_prop),
9778 L2BLK_GET_STATE((le)->le_prop));
9779 asize = vdev_psize_to_asize(dev->l2ad_vdev,
9780 L2BLK_GET_PSIZE((le)->le_prop));
9781
9782 /*
9783 * vdev_space_update() has to be called before arc_hdr_destroy() to
9784 * avoid underflow since the latter also calls vdev_space_update().
9785 */
9786 l2arc_hdr_arcstats_increment(hdr);
9787 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9788
9789 mutex_enter(&dev->l2ad_mtx);
9790 list_insert_tail(&dev->l2ad_buflist, hdr);
9791 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
9792 mutex_exit(&dev->l2ad_mtx);
9793
9794 exists = buf_hash_insert(hdr, &hash_lock);
9795 if (exists) {
9796 /* Buffer was already cached, no need to restore it. */
9797 arc_hdr_destroy(hdr);
9798 /*
9799 * If the buffer is already cached, check whether it has
9800 * L2ARC metadata. If not, enter them and update the flag.
9801 * This is important is case of onlining a cache device, since
9802 * we previously evicted all L2ARC metadata from ARC.
9803 */
9804 if (!HDR_HAS_L2HDR(exists)) {
9805 arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR);
9806 exists->b_l2hdr.b_dev = dev;
9807 exists->b_l2hdr.b_daddr = le->le_daddr;
9808 exists->b_l2hdr.b_arcs_state =
9809 L2BLK_GET_STATE((le)->le_prop);
9810 mutex_enter(&dev->l2ad_mtx);
9811 list_insert_tail(&dev->l2ad_buflist, exists);
9812 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
9813 arc_hdr_size(exists), exists);
9814 mutex_exit(&dev->l2ad_mtx);
9815 l2arc_hdr_arcstats_increment(exists);
9816 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9817 }
9818 ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached);
9819 }
9820
9821 mutex_exit(hash_lock);
9822 }
9823
9824 /*
9825 * Starts an asynchronous read IO to read a log block. This is used in log
9826 * block reconstruction to start reading the next block before we are done
9827 * decoding and reconstructing the current block, to keep the l2arc device
9828 * nice and hot with read IO to process.
9829 * The returned zio will contain newly allocated memory buffers for the IO
9830 * data which should then be freed by the caller once the zio is no longer
9831 * needed (i.e. due to it having completed). If you wish to abort this
9832 * zio, you should do so using l2arc_log_blk_fetch_abort, which takes
9833 * care of disposing of the allocated buffers correctly.
9834 */
9835 static zio_t *
9836 l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp,
9837 l2arc_log_blk_phys_t *lb)
9838 {
9839 uint32_t asize;
9840 zio_t *pio;
9841 l2arc_read_callback_t *cb;
9842
9843 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
9844 asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
9845 ASSERT(asize <= sizeof (l2arc_log_blk_phys_t));
9846
9847 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP);
9848 cb->l2rcb_abd = abd_get_from_buf(lb, asize);
9849 pio = zio_root(vd->vdev_spa, l2arc_blk_fetch_done, cb,
9850 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
9851 ZIO_FLAG_DONT_RETRY);
9852 (void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, asize,
9853 cb->l2rcb_abd, ZIO_CHECKSUM_OFF, NULL, NULL,
9854 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
9855 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE));
9856
9857 return (pio);
9858 }
9859
9860 /*
9861 * Aborts a zio returned from l2arc_log_blk_fetch and frees the data
9862 * buffers allocated for it.
9863 */
9864 static void
9865 l2arc_log_blk_fetch_abort(zio_t *zio)
9866 {
9867 (void) zio_wait(zio);
9868 }
9869
9870 /*
9871 * Creates a zio to update the device header on an l2arc device.
9872 */
9873 static void
9874 l2arc_dev_hdr_update(l2arc_dev_t *dev)
9875 {
9876 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9877 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9878 abd_t *abd;
9879 int err;
9880
9881 VERIFY(spa_config_held(dev->l2ad_spa, SCL_STATE_ALL, RW_READER));
9882
9883 l2dhdr->dh_magic = L2ARC_DEV_HDR_MAGIC;
9884 l2dhdr->dh_version = L2ARC_PERSISTENT_VERSION;
9885 l2dhdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa);
9886 l2dhdr->dh_vdev_guid = dev->l2ad_vdev->vdev_guid;
9887 l2dhdr->dh_log_entries = dev->l2ad_log_entries;
9888 l2dhdr->dh_evict = dev->l2ad_evict;
9889 l2dhdr->dh_start = dev->l2ad_start;
9890 l2dhdr->dh_end = dev->l2ad_end;
9891 l2dhdr->dh_lb_asize = zfs_refcount_count(&dev->l2ad_lb_asize);
9892 l2dhdr->dh_lb_count = zfs_refcount_count(&dev->l2ad_lb_count);
9893 l2dhdr->dh_flags = 0;
9894 if (dev->l2ad_first)
9895 l2dhdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST;
9896
9897 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
9898
9899 err = zio_wait(zio_write_phys(NULL, dev->l2ad_vdev,
9900 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL,
9901 NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE));
9902
9903 abd_put(abd);
9904
9905 if (err != 0) {
9906 zfs_dbgmsg("L2ARC IO error (%d) while writing device header, "
9907 "vdev guid: %llu", err, dev->l2ad_vdev->vdev_guid);
9908 }
9909 }
9910
9911 /*
9912 * Commits a log block to the L2ARC device. This routine is invoked from
9913 * l2arc_write_buffers when the log block fills up.
9914 * This function allocates some memory to temporarily hold the serialized
9915 * buffer to be written. This is then released in l2arc_write_done.
9916 */
9917 static void
9918 l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb)
9919 {
9920 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
9921 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9922 uint64_t psize, asize;
9923 zio_t *wzio;
9924 l2arc_lb_abd_buf_t *abd_buf;
9925 uint8_t *tmpbuf;
9926 l2arc_lb_ptr_buf_t *lb_ptr_buf;
9927
9928 VERIFY3S(dev->l2ad_log_ent_idx, ==, dev->l2ad_log_entries);
9929
9930 tmpbuf = zio_buf_alloc(sizeof (*lb));
9931 abd_buf = zio_buf_alloc(sizeof (*abd_buf));
9932 abd_buf->abd = abd_get_from_buf(lb, sizeof (*lb));
9933 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
9934 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP);
9935
9936 /* link the buffer into the block chain */
9937 lb->lb_prev_lbp = l2dhdr->dh_start_lbps[1];
9938 lb->lb_magic = L2ARC_LOG_BLK_MAGIC;
9939
9940 /*
9941 * l2arc_log_blk_commit() may be called multiple times during a single
9942 * l2arc_write_buffers() call. Save the allocated abd buffers in a list
9943 * so we can free them in l2arc_write_done() later on.
9944 */
9945 list_insert_tail(&cb->l2wcb_abd_list, abd_buf);
9946
9947 /* try to compress the buffer */
9948 psize = zio_compress_data(ZIO_COMPRESS_LZ4,
9949 abd_buf->abd, tmpbuf, sizeof (*lb));
9950
9951 /* a log block is never entirely zero */
9952 ASSERT(psize != 0);
9953 asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
9954 ASSERT(asize <= sizeof (*lb));
9955
9956 /*
9957 * Update the start log block pointer in the device header to point
9958 * to the log block we're about to write.
9959 */
9960 l2dhdr->dh_start_lbps[1] = l2dhdr->dh_start_lbps[0];
9961 l2dhdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand;
9962 l2dhdr->dh_start_lbps[0].lbp_payload_asize =
9963 dev->l2ad_log_blk_payload_asize;
9964 l2dhdr->dh_start_lbps[0].lbp_payload_start =
9965 dev->l2ad_log_blk_payload_start;
9966 _NOTE(CONSTCOND)
9967 L2BLK_SET_LSIZE(
9968 (&l2dhdr->dh_start_lbps[0])->lbp_prop, sizeof (*lb));
9969 L2BLK_SET_PSIZE(
9970 (&l2dhdr->dh_start_lbps[0])->lbp_prop, asize);
9971 L2BLK_SET_CHECKSUM(
9972 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
9973 ZIO_CHECKSUM_FLETCHER_4);
9974 if (asize < sizeof (*lb)) {
9975 /* compression succeeded */
9976 bzero(tmpbuf + psize, asize - psize);
9977 L2BLK_SET_COMPRESS(
9978 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
9979 ZIO_COMPRESS_LZ4);
9980 } else {
9981 /* compression failed */
9982 bcopy(lb, tmpbuf, sizeof (*lb));
9983 L2BLK_SET_COMPRESS(
9984 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
9985 ZIO_COMPRESS_OFF);
9986 }
9987
9988 /* checksum what we're about to write */
9989 fletcher_4_native(tmpbuf, asize, NULL,
9990 &l2dhdr->dh_start_lbps[0].lbp_cksum);
9991
9992 abd_put(abd_buf->abd);
9993
9994 /* perform the write itself */
9995 abd_buf->abd = abd_get_from_buf(tmpbuf, sizeof (*lb));
9996 abd_take_ownership_of_buf(abd_buf->abd, B_TRUE);
9997 wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand,
9998 asize, abd_buf->abd, ZIO_CHECKSUM_OFF, NULL, NULL,
9999 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE);
10000 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio);
10001 (void) zio_nowait(wzio);
10002
10003 dev->l2ad_hand += asize;
10004 /*
10005 * Include the committed log block's pointer in the list of pointers
10006 * to log blocks present in the L2ARC device.
10007 */
10008 bcopy(&l2dhdr->dh_start_lbps[0], lb_ptr_buf->lb_ptr,
10009 sizeof (l2arc_log_blkptr_t));
10010 mutex_enter(&dev->l2ad_mtx);
10011 list_insert_head(&dev->l2ad_lbptr_list, lb_ptr_buf);
10012 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
10013 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
10014 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
10015 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
10016 mutex_exit(&dev->l2ad_mtx);
10017 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
10018
10019 /* bump the kstats */
10020 ARCSTAT_INCR(arcstat_l2_write_bytes, asize);
10021 ARCSTAT_BUMP(arcstat_l2_log_blk_writes);
10022 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, asize);
10023 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio,
10024 dev->l2ad_log_blk_payload_asize / asize);
10025
10026 /* start a new log block */
10027 dev->l2ad_log_ent_idx = 0;
10028 dev->l2ad_log_blk_payload_asize = 0;
10029 dev->l2ad_log_blk_payload_start = 0;
10030 }
10031
10032 /*
10033 * Validates an L2ARC log block address to make sure that it can be read
10034 * from the provided L2ARC device.
10035 */
10036 boolean_t
10037 l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp)
10038 {
10039 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
10040 uint64_t asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
10041 uint64_t end = lbp->lbp_daddr + asize - 1;
10042 uint64_t start = lbp->lbp_payload_start;
10043 boolean_t evicted = B_FALSE;
10044
10045 /* BEGIN CSTYLED */
10046 /*
10047 * A log block is valid if all of the following conditions are true:
10048 * - it fits entirely (including its payload) between l2ad_start and
10049 * l2ad_end
10050 * - it has a valid size
10051 * - neither the log block itself nor part of its payload was evicted
10052 * by l2arc_evict():
10053 *
10054 * l2ad_hand l2ad_evict
10055 * | | lbp_daddr
10056 * | start | | end
10057 * | | | | |
10058 * V V V V V
10059 * l2ad_start ============================================ l2ad_end
10060 * --------------------------||||
10061 * ^ ^
10062 * | log block
10063 * payload
10064 */
10065 /* END CSTYLED */
10066 evicted =
10067 l2arc_range_check_overlap(start, end, dev->l2ad_hand) ||
10068 l2arc_range_check_overlap(start, end, dev->l2ad_evict) ||
10069 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, start) ||
10070 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, end);
10071
10072 return (start >= dev->l2ad_start && end <= dev->l2ad_end &&
10073 asize > 0 && asize <= sizeof (l2arc_log_blk_phys_t) &&
10074 (!evicted || dev->l2ad_first));
10075 }
10076
10077 /*
10078 * Inserts ARC buffer header `hdr' into the current L2ARC log block on
10079 * the device. The buffer being inserted must be present in L2ARC.
10080 * Returns B_TRUE if the L2ARC log block is full and needs to be committed
10081 * to L2ARC, or B_FALSE if it still has room for more ARC buffers.
10082 */
10083 static boolean_t
10084 l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr)
10085 {
10086 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
10087 l2arc_log_ent_phys_t *le;
10088
10089 if (dev->l2ad_log_entries == 0)
10090 return (B_FALSE);
10091
10092 int index = dev->l2ad_log_ent_idx++;
10093
10094 ASSERT3S(index, <, dev->l2ad_log_entries);
10095 ASSERT(HDR_HAS_L2HDR(hdr));
10096
10097 le = &lb->lb_entries[index];
10098 bzero(le, sizeof (*le));
10099 le->le_dva = hdr->b_dva;
10100 le->le_birth = hdr->b_birth;
10101 le->le_daddr = hdr->b_l2hdr.b_daddr;
10102 if (index == 0)
10103 dev->l2ad_log_blk_payload_start = le->le_daddr;
10104 L2BLK_SET_LSIZE((le)->le_prop, HDR_GET_LSIZE(hdr));
10105 L2BLK_SET_PSIZE((le)->le_prop, HDR_GET_PSIZE(hdr));
10106 L2BLK_SET_COMPRESS((le)->le_prop, HDR_GET_COMPRESS(hdr));
10107 L2BLK_SET_TYPE((le)->le_prop, hdr->b_type);
10108 L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr)));
10109 L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr)));
10110 L2BLK_SET_STATE((le)->le_prop, hdr->b_l1hdr.b_state->arcs_state);
10111
10112 dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev,
10113 HDR_GET_PSIZE(hdr));
10114
10115 return (dev->l2ad_log_ent_idx == dev->l2ad_log_entries);
10116 }
10117
10118 /*
10119 * Checks whether a given L2ARC device address sits in a time-sequential
10120 * range. The trick here is that the L2ARC is a rotary buffer, so we can't
10121 * just do a range comparison, we need to handle the situation in which the
10122 * range wraps around the end of the L2ARC device. Arguments:
10123 * bottom -- Lower end of the range to check (written to earlier).
10124 * top -- Upper end of the range to check (written to later).
10125 * check -- The address for which we want to determine if it sits in
10126 * between the top and bottom.
10127 *
10128 * The 3-way conditional below represents the following cases:
10129 *
10130 * bottom < top : Sequentially ordered case:
10131 * <check>--------+-------------------+
10132 * | (overlap here?) |
10133 * L2ARC dev V V
10134 * |---------------<bottom>============<top>--------------|
10135 *
10136 * bottom > top: Looped-around case:
10137 * <check>--------+------------------+
10138 * | (overlap here?) |
10139 * L2ARC dev V V
10140 * |===============<top>---------------<bottom>===========|
10141 * ^ ^
10142 * | (or here?) |
10143 * +---------------+---------<check>
10144 *
10145 * top == bottom : Just a single address comparison.
10146 */
10147 boolean_t
10148 l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check)
10149 {
10150 if (bottom < top)
10151 return (bottom <= check && check <= top);
10152 else if (bottom > top)
10153 return (check <= top || bottom <= check);
10154 else
10155 return (check == top);
10156 }