1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2019, 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 */
30
31 /*
32 * DVA-based Adjustable Replacement Cache
33 *
34 * While much of the theory of operation used here is
35 * based on the self-tuning, low overhead replacement cache
36 * presented by Megiddo and Modha at FAST 2003, there are some
37 * significant differences:
38 *
39 * 1. The Megiddo and Modha model assumes any page is evictable.
40 * Pages in its cache cannot be "locked" into memory. This makes
41 * the eviction algorithm simple: evict the last page in the list.
42 * This also make the performance characteristics easy to reason
43 * about. Our cache is not so simple. At any given moment, some
44 * subset of the blocks in the cache are un-evictable because we
45 * have handed out a reference to them. Blocks are only evictable
46 * when there are no external references active. This makes
47 * eviction far more problematic: we choose to evict the evictable
48 * blocks that are the "lowest" in the list.
49 *
50 * There are times when it is not possible to evict the requested
51 * space. In these circumstances we are unable to adjust the cache
52 * size. To prevent the cache growing unbounded at these times we
53 * implement a "cache throttle" that slows the flow of new data
54 * into the cache until we can make space available.
55 *
56 * 2. The Megiddo and Modha model assumes a fixed cache size.
57 * Pages are evicted when the cache is full and there is a cache
58 * miss. Our model has a variable sized cache. It grows with
59 * high use, but also tries to react to memory pressure from the
60 * operating system: decreasing its size when system memory is
61 * tight.
62 *
63 * 3. The Megiddo and Modha model assumes a fixed page size. All
64 * elements of the cache are therefore exactly the same size. So
65 * when adjusting the cache size following a cache miss, its simply
66 * a matter of choosing a single page to evict. In our model, we
67 * have variable sized cache blocks (rangeing from 512 bytes to
68 * 128K bytes). We therefore choose a set of blocks to evict to make
69 * space for a cache miss that approximates as closely as possible
70 * the space used by the new block.
71 *
72 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
73 * by N. Megiddo & D. Modha, FAST 2003
74 */
75
76 /*
77 * The locking model:
78 *
79 * A new reference to a cache buffer can be obtained in two
80 * ways: 1) via a hash table lookup using the DVA as a key,
81 * or 2) via one of the ARC lists. The arc_read() interface
82 * uses method 1, while the internal ARC algorithms for
83 * adjusting the cache use method 2. We therefore provide two
84 * types of locks: 1) the hash table lock array, and 2) the
85 * ARC list locks.
86 *
87 * Buffers do not have their own mutexes, rather they rely on the
88 * hash table mutexes for the bulk of their protection (i.e. most
89 * fields in the arc_buf_hdr_t are protected by these mutexes).
90 *
91 * buf_hash_find() returns the appropriate mutex (held) when it
92 * locates the requested buffer in the hash table. It returns
93 * NULL for the mutex if the buffer was not in the table.
94 *
95 * buf_hash_remove() expects the appropriate hash mutex to be
96 * already held before it is invoked.
97 *
98 * Each ARC state also has a mutex which is used to protect the
99 * buffer list associated with the state. When attempting to
100 * obtain a hash table lock while holding an ARC list lock you
101 * must use: mutex_tryenter() to avoid deadlock. Also note that
102 * the active state mutex must be held before the ghost state mutex.
103 *
104 * Note that the majority of the performance stats are manipulated
105 * with atomic operations.
106 *
107 * The L2ARC uses the l2ad_mtx on each vdev for the following:
108 *
109 * - L2ARC buflist creation
110 * - L2ARC buflist eviction
111 * - L2ARC write completion, which walks L2ARC buflists
112 * - ARC header destruction, as it removes from L2ARC buflists
113 * - ARC header release, as it removes from L2ARC buflists
114 */
115
116 /*
117 * ARC operation:
118 *
119 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure.
120 * This structure can point either to a block that is still in the cache or to
121 * one that is only accessible in an L2 ARC device, or it can provide
122 * information about a block that was recently evicted. If a block is
123 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough
124 * information to retrieve it from the L2ARC device. This information is
125 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block
126 * that is in this state cannot access the data directly.
127 *
128 * Blocks that are actively being referenced or have not been evicted
129 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within
130 * the arc_buf_hdr_t that will point to the data block in memory. A block can
131 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC
132 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and
133 * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd).
134 *
135 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the
136 * ability to store the physical data (b_pabd) associated with the DVA of the
137 * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block,
138 * it will match its on-disk compression characteristics. This behavior can be
139 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the
140 * compressed ARC functionality is disabled, the b_pabd will point to an
141 * uncompressed version of the on-disk data.
142 *
143 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each
144 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it.
145 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC
146 * consumer. The ARC will provide references to this data and will keep it
147 * cached until it is no longer in use. The ARC caches only the L1ARC's physical
148 * data block and will evict any arc_buf_t that is no longer referenced. The
149 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the
150 * "overhead_size" kstat.
151 *
152 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or
153 * compressed form. The typical case is that consumers will want uncompressed
154 * data, and when that happens a new data buffer is allocated where the data is
155 * decompressed for them to use. Currently the only consumer who wants
156 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it
157 * exists on disk. When this happens, the arc_buf_t's data buffer is shared
158 * with the arc_buf_hdr_t.
159 *
160 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The
161 * first one is owned by a compressed send consumer (and therefore references
162 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be
163 * used by any other consumer (and has its own uncompressed copy of the data
164 * buffer).
165 *
166 * arc_buf_hdr_t
167 * +-----------+
168 * | fields |
169 * | common to |
170 * | L1- and |
171 * | L2ARC |
172 * +-----------+
173 * | l2arc_buf_hdr_t
174 * | |
175 * +-----------+
176 * | l1arc_buf_hdr_t
177 * | | arc_buf_t
178 * | b_buf +------------>+-----------+ arc_buf_t
179 * | b_pabd +-+ |b_next +---->+-----------+
180 * +-----------+ | |-----------| |b_next +-->NULL
181 * | |b_comp = T | +-----------+
182 * | |b_data +-+ |b_comp = F |
183 * | +-----------+ | |b_data +-+
184 * +->+------+ | +-----------+ |
185 * compressed | | | |
186 * data | |<--------------+ | uncompressed
187 * +------+ compressed, | data
188 * shared +-->+------+
189 * data | |
190 * | |
191 * +------+
192 *
193 * When a consumer reads a block, the ARC must first look to see if the
194 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new
195 * arc_buf_t and either copies uncompressed data into a new data buffer from an
196 * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a
197 * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the
198 * hdr is compressed and the desired compression characteristics of the
199 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the
200 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be
201 * the last buffer in the hdr's b_buf list, however a shared compressed buf can
202 * be anywhere in the hdr's list.
203 *
204 * The diagram below shows an example of an uncompressed ARC hdr that is
205 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is
206 * the last element in the buf list):
207 *
208 * arc_buf_hdr_t
209 * +-----------+
210 * | |
211 * | |
212 * | |
213 * +-----------+
214 * l2arc_buf_hdr_t| |
215 * | |
216 * +-----------+
217 * l1arc_buf_hdr_t| |
218 * | | arc_buf_t (shared)
219 * | b_buf +------------>+---------+ arc_buf_t
220 * | | |b_next +---->+---------+
221 * | b_pabd +-+ |---------| |b_next +-->NULL
222 * +-----------+ | | | +---------+
223 * | |b_data +-+ | |
224 * | +---------+ | |b_data +-+
225 * +->+------+ | +---------+ |
226 * | | | |
227 * uncompressed | | | |
228 * data +------+ | |
229 * ^ +->+------+ |
230 * | uncompressed | | |
231 * | data | | |
232 * | +------+ |
233 * +---------------------------------+
234 *
235 * Writing to the ARC requires that the ARC first discard the hdr's b_pabd
236 * since the physical block is about to be rewritten. The new data contents
237 * will be contained in the arc_buf_t. As the I/O pipeline performs the write,
238 * it may compress the data before writing it to disk. The ARC will be called
239 * with the transformed data and will bcopy the transformed on-disk block into
240 * a newly allocated b_pabd. Writes are always done into buffers which have
241 * either been loaned (and hence are new and don't have other readers) or
242 * buffers which have been released (and hence have their own hdr, if there
243 * were originally other readers of the buf's original hdr). This ensures that
244 * the ARC only needs to update a single buf and its hdr after a write occurs.
245 *
246 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
247 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
248 * that when compressed ARC is enabled that the L2ARC blocks are identical
249 * to the on-disk block in the main data pool. This provides a significant
250 * advantage since the ARC can leverage the bp's checksum when reading from the
251 * L2ARC to determine if the contents are valid. However, if the compressed
252 * ARC is disabled, then the L2ARC's block must be transformed to look
253 * like the physical block in the main data pool before comparing the
254 * checksum and determining its validity.
255 *
256 * The L1ARC has a slightly different system for storing encrypted data.
257 * Raw (encrypted + possibly compressed) data has a few subtle differences from
258 * data that is just compressed. The biggest difference is that it is not
259 * possible to decrypt encrypted data (or visa versa) if the keys aren't loaded.
260 * The other difference is that encryption cannot be treated as a suggestion.
261 * If a caller would prefer compressed data, but they actually wind up with
262 * uncompressed data the worst thing that could happen is there might be a
263 * performance hit. If the caller requests encrypted data, however, we must be
264 * sure they actually get it or else secret information could be leaked. Raw
265 * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore,
266 * may have both an encrypted version and a decrypted version of its data at
267 * once. When a caller needs a raw arc_buf_t, it is allocated and the data is
268 * copied out of this header. To avoid complications with b_pabd, raw buffers
269 * cannot be shared.
270 */
271
272 #include <sys/spa.h>
273 #include <sys/zio.h>
274 #include <sys/spa_impl.h>
275 #include <sys/zio_compress.h>
276 #include <sys/zio_checksum.h>
277 #include <sys/zfs_context.h>
278 #include <sys/arc.h>
279 #include <sys/refcount.h>
280 #include <sys/vdev.h>
281 #include <sys/vdev_impl.h>
282 #include <sys/dsl_pool.h>
283 #include <sys/zio_checksum.h>
284 #include <sys/multilist.h>
285 #include <sys/abd.h>
286 #include <sys/zil.h>
287 #include <sys/fm/fs/zfs.h>
288 #ifdef _KERNEL
289 #include <sys/vmsystm.h>
290 #include <vm/anon.h>
291 #include <sys/fs/swapnode.h>
292 #include <sys/dnlc.h>
293 #endif
294 #include <sys/callb.h>
295 #include <sys/kstat.h>
296 #include <sys/zthr.h>
297 #include <zfs_fletcher.h>
298 #include <sys/arc_impl.h>
299 #include <sys/aggsum.h>
300 #include <sys/cityhash.h>
301 #include <sys/param.h>
302
303 #ifndef _KERNEL
304 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
305 boolean_t arc_watch = B_FALSE;
306 int arc_procfd;
307 #endif
308
309 /*
310 * This thread's job is to keep enough free memory in the system, by
311 * calling arc_kmem_reap_now() plus arc_shrink(), which improves
312 * arc_available_memory().
313 */
314 static zthr_t *arc_reap_zthr;
315
316 /*
317 * This thread's job is to keep arc_size under arc_c, by calling
318 * arc_adjust(), which improves arc_is_overflowing().
319 */
320 static zthr_t *arc_adjust_zthr;
321
322 static kmutex_t arc_adjust_lock;
323 static kcondvar_t arc_adjust_waiters_cv;
324 static boolean_t arc_adjust_needed = B_FALSE;
325
326 uint_t arc_reduce_dnlc_percent = 3;
327
328 /*
329 * The number of headers to evict in arc_evict_state_impl() before
330 * dropping the sublist lock and evicting from another sublist. A lower
331 * value means we're more likely to evict the "correct" header (i.e. the
332 * oldest header in the arc state), but comes with higher overhead
333 * (i.e. more invocations of arc_evict_state_impl()).
334 */
335 int zfs_arc_evict_batch_limit = 10;
336
337 /* number of seconds before growing cache again */
338 int arc_grow_retry = 60;
339
340 /*
341 * Minimum time between calls to arc_kmem_reap_soon(). Note that this will
342 * be converted to ticks, so with the default hz=100, a setting of 15 ms
343 * will actually wait 2 ticks, or 20ms.
344 */
345 int arc_kmem_cache_reap_retry_ms = 1000;
346
347 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */
348 int zfs_arc_overflow_shift = 8;
349
350 /* shift of arc_c for calculating both min and max arc_p */
351 int arc_p_min_shift = 4;
352
353 /* log2(fraction of arc to reclaim) */
354 int arc_shrink_shift = 7;
355
356 /*
357 * log2(fraction of ARC which must be free to allow growing).
358 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
359 * when reading a new block into the ARC, we will evict an equal-sized block
360 * from the ARC.
361 *
362 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
363 * we will still not allow it to grow.
364 */
365 int arc_no_grow_shift = 5;
366
367
368 /*
369 * minimum lifespan of a prefetch block in clock ticks
370 * (initialized in arc_init())
371 */
372 static int zfs_arc_min_prefetch_ms = 1;
373 static int zfs_arc_min_prescient_prefetch_ms = 6;
374
375 /*
376 * If this percent of memory is free, don't throttle.
377 */
378 int arc_lotsfree_percent = 10;
379
380 static boolean_t arc_initialized;
381
382 /*
383 * The arc has filled available memory and has now warmed up.
384 */
385 static boolean_t arc_warm;
386
387 /*
388 * log2 fraction of the zio arena to keep free.
389 */
390 int arc_zio_arena_free_shift = 2;
391
392 /*
393 * These tunables are for performance analysis.
394 */
395 uint64_t zfs_arc_max;
396 uint64_t zfs_arc_min;
397 uint64_t zfs_arc_meta_limit = 0;
398 uint64_t zfs_arc_meta_min = 0;
399 int zfs_arc_grow_retry = 0;
400 int zfs_arc_shrink_shift = 0;
401 int zfs_arc_p_min_shift = 0;
402 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
403
404 /*
405 * ARC dirty data constraints for arc_tempreserve_space() throttle
406 */
407 uint_t zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */
408 uint_t zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */
409 uint_t zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */
410
411 boolean_t zfs_compressed_arc_enabled = B_TRUE;
412
413 /* The 6 states: */
414 static arc_state_t ARC_anon;
415 static arc_state_t ARC_mru;
416 static arc_state_t ARC_mru_ghost;
417 static arc_state_t ARC_mfu;
418 static arc_state_t ARC_mfu_ghost;
419 static arc_state_t ARC_l2c_only;
420
421 arc_stats_t arc_stats = {
422 { "hits", KSTAT_DATA_UINT64 },
423 { "misses", KSTAT_DATA_UINT64 },
424 { "demand_data_hits", KSTAT_DATA_UINT64 },
425 { "demand_data_misses", KSTAT_DATA_UINT64 },
426 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
427 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
428 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
429 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
430 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
431 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
432 { "mru_hits", KSTAT_DATA_UINT64 },
433 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
434 { "mfu_hits", KSTAT_DATA_UINT64 },
435 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
436 { "deleted", KSTAT_DATA_UINT64 },
437 { "mutex_miss", KSTAT_DATA_UINT64 },
438 { "access_skip", KSTAT_DATA_UINT64 },
439 { "evict_skip", KSTAT_DATA_UINT64 },
440 { "evict_not_enough", KSTAT_DATA_UINT64 },
441 { "evict_l2_cached", KSTAT_DATA_UINT64 },
442 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
443 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
444 { "evict_l2_skip", KSTAT_DATA_UINT64 },
445 { "hash_elements", KSTAT_DATA_UINT64 },
446 { "hash_elements_max", KSTAT_DATA_UINT64 },
447 { "hash_collisions", KSTAT_DATA_UINT64 },
448 { "hash_chains", KSTAT_DATA_UINT64 },
449 { "hash_chain_max", KSTAT_DATA_UINT64 },
450 { "p", KSTAT_DATA_UINT64 },
451 { "c", KSTAT_DATA_UINT64 },
452 { "c_min", KSTAT_DATA_UINT64 },
453 { "c_max", KSTAT_DATA_UINT64 },
454 { "size", KSTAT_DATA_UINT64 },
455 { "compressed_size", KSTAT_DATA_UINT64 },
456 { "uncompressed_size", KSTAT_DATA_UINT64 },
457 { "overhead_size", KSTAT_DATA_UINT64 },
458 { "hdr_size", KSTAT_DATA_UINT64 },
459 { "data_size", KSTAT_DATA_UINT64 },
460 { "metadata_size", KSTAT_DATA_UINT64 },
461 { "other_size", KSTAT_DATA_UINT64 },
462 { "anon_size", KSTAT_DATA_UINT64 },
463 { "anon_evictable_data", KSTAT_DATA_UINT64 },
464 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
465 { "mru_size", KSTAT_DATA_UINT64 },
466 { "mru_evictable_data", KSTAT_DATA_UINT64 },
467 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
468 { "mru_ghost_size", KSTAT_DATA_UINT64 },
469 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
470 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
471 { "mfu_size", KSTAT_DATA_UINT64 },
472 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
473 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
474 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
475 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
476 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
477 { "l2_hits", KSTAT_DATA_UINT64 },
478 { "l2_misses", KSTAT_DATA_UINT64 },
479 { "l2_feeds", KSTAT_DATA_UINT64 },
480 { "l2_rw_clash", KSTAT_DATA_UINT64 },
481 { "l2_read_bytes", KSTAT_DATA_UINT64 },
482 { "l2_write_bytes", KSTAT_DATA_UINT64 },
483 { "l2_writes_sent", KSTAT_DATA_UINT64 },
484 { "l2_writes_done", KSTAT_DATA_UINT64 },
485 { "l2_writes_error", KSTAT_DATA_UINT64 },
486 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
487 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
488 { "l2_evict_reading", KSTAT_DATA_UINT64 },
489 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
490 { "l2_free_on_write", KSTAT_DATA_UINT64 },
491 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
492 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
493 { "l2_io_error", KSTAT_DATA_UINT64 },
494 { "l2_size", KSTAT_DATA_UINT64 },
495 { "l2_asize", KSTAT_DATA_UINT64 },
496 { "l2_hdr_size", KSTAT_DATA_UINT64 },
497 { "l2_log_blk_writes", KSTAT_DATA_UINT64 },
498 { "l2_log_blk_avg_asize", KSTAT_DATA_UINT64 },
499 { "l2_log_blk_asize", KSTAT_DATA_UINT64 },
500 { "l2_log_blk_count", KSTAT_DATA_UINT64 },
501 { "l2_data_to_meta_ratio", KSTAT_DATA_UINT64 },
502 { "l2_rebuild_success", KSTAT_DATA_UINT64 },
503 { "l2_rebuild_unsupported", KSTAT_DATA_UINT64 },
504 { "l2_rebuild_io_errors", KSTAT_DATA_UINT64 },
505 { "l2_rebuild_dh_errors", KSTAT_DATA_UINT64 },
506 { "l2_rebuild_cksum_lb_errors", KSTAT_DATA_UINT64 },
507 { "l2_rebuild_lowmem", KSTAT_DATA_UINT64 },
508 { "l2_rebuild_size", KSTAT_DATA_UINT64 },
509 { "l2_rebuild_asize", KSTAT_DATA_UINT64 },
510 { "l2_rebuild_bufs", KSTAT_DATA_UINT64 },
511 { "l2_rebuild_bufs_precached", KSTAT_DATA_UINT64 },
512 { "l2_rebuild_log_blks", KSTAT_DATA_UINT64 },
513 { "memory_throttle_count", KSTAT_DATA_UINT64 },
514 { "arc_meta_used", KSTAT_DATA_UINT64 },
515 { "arc_meta_limit", KSTAT_DATA_UINT64 },
516 { "arc_meta_max", KSTAT_DATA_UINT64 },
517 { "arc_meta_min", KSTAT_DATA_UINT64 },
518 { "async_upgrade_sync", KSTAT_DATA_UINT64 },
519 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
520 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 },
521 };
522
523 #define ARCSTAT_MAX(stat, val) { \
524 uint64_t m; \
525 while ((val) > (m = arc_stats.stat.value.ui64) && \
526 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
527 continue; \
528 }
529
530 #define ARCSTAT_MAXSTAT(stat) \
531 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
532
533 /*
534 * We define a macro to allow ARC hits/misses to be easily broken down by
535 * two separate conditions, giving a total of four different subtypes for
536 * each of hits and misses (so eight statistics total).
537 */
538 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
539 if (cond1) { \
540 if (cond2) { \
541 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
542 } else { \
543 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
544 } \
545 } else { \
546 if (cond2) { \
547 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
548 } else { \
549 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
550 } \
551 }
552
553 /*
554 * This macro allows us to use kstats as floating averages. Each time we
555 * update this kstat, we first factor it and the update value by
556 * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall
557 * average. This macro assumes that integer loads and stores are atomic, but
558 * is not safe for multiple writers updating the kstat in parallel (only the
559 * last writer's update will remain).
560 */
561 #define ARCSTAT_F_AVG_FACTOR 3
562 #define ARCSTAT_F_AVG(stat, value) \
563 do { \
564 uint64_t x = ARCSTAT(stat); \
565 x = x - x / ARCSTAT_F_AVG_FACTOR + \
566 (value) / ARCSTAT_F_AVG_FACTOR; \
567 ARCSTAT(stat) = x; \
568 _NOTE(CONSTCOND) \
569 } while (0)
570
571 kstat_t *arc_ksp;
572 static arc_state_t *arc_anon;
573 static arc_state_t *arc_mru;
574 static arc_state_t *arc_mru_ghost;
575 static arc_state_t *arc_mfu;
576 static arc_state_t *arc_mfu_ghost;
577 static arc_state_t *arc_l2c_only;
578
579 /*
580 * There are also some ARC variables that we want to export, but that are
581 * updated so often that having the canonical representation be the statistic
582 * variable causes a performance bottleneck. We want to use aggsum_t's for these
583 * instead, but still be able to export the kstat in the same way as before.
584 * The solution is to always use the aggsum version, except in the kstat update
585 * callback.
586 */
587 aggsum_t arc_size;
588 aggsum_t arc_meta_used;
589 aggsum_t astat_data_size;
590 aggsum_t astat_metadata_size;
591 aggsum_t astat_hdr_size;
592 aggsum_t astat_other_size;
593 aggsum_t astat_l2_hdr_size;
594
595 static int arc_no_grow; /* Don't try to grow cache size */
596 static hrtime_t arc_growtime;
597 static uint64_t arc_tempreserve;
598 static uint64_t arc_loaned_bytes;
599
600 #define GHOST_STATE(state) \
601 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
602 (state) == arc_l2c_only)
603
604 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
605 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
606 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
607 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
608 #define HDR_PRESCIENT_PREFETCH(hdr) \
609 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH)
610 #define HDR_COMPRESSION_ENABLED(hdr) \
611 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
612
613 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
614 #define HDR_L2_READING(hdr) \
615 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
616 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
617 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
618 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
619 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
620 #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED)
621 #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH)
622 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
623
624 #define HDR_ISTYPE_METADATA(hdr) \
625 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
626 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
627
628 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
629 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
630 #define HDR_HAS_RABD(hdr) \
631 (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \
632 (hdr)->b_crypt_hdr.b_rabd != NULL)
633 #define HDR_ENCRYPTED(hdr) \
634 (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
635 #define HDR_AUTHENTICATED(hdr) \
636 (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot))
637
638 /* For storing compression mode in b_flags */
639 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
640
641 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
642 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
643 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
644 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
645
646 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
647 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
648 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
649 #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED)
650
651 /*
652 * Other sizes
653 */
654
655 #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
656 #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr))
657 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
658
659 /*
660 * Hash table routines
661 */
662
663 #define HT_LOCK_PAD 64
664
665 struct ht_lock {
666 kmutex_t ht_lock;
667 #ifdef _KERNEL
668 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
669 #endif
670 };
671
672 #define BUF_LOCKS 256
673 typedef struct buf_hash_table {
674 uint64_t ht_mask;
675 arc_buf_hdr_t **ht_table;
676 struct ht_lock ht_locks[BUF_LOCKS];
677 } buf_hash_table_t;
678
679 static buf_hash_table_t buf_hash_table;
680
681 #define BUF_HASH_INDEX(spa, dva, birth) \
682 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
683 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
684 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
685 #define HDR_LOCK(hdr) \
686 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
687
688 uint64_t zfs_crc64_table[256];
689
690 /*
691 * Level 2 ARC
692 */
693
694 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
695 #define L2ARC_HEADROOM 2 /* num of writes */
696 /*
697 * If we discover during ARC scan any buffers to be compressed, we boost
698 * our headroom for the next scanning cycle by this percentage multiple.
699 */
700 #define L2ARC_HEADROOM_BOOST 200
701 #define L2ARC_FEED_SECS 1 /* caching interval secs */
702 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
703
704 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
705 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
706
707 /* L2ARC Performance Tunables */
708 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
709 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
710 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
711 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
712 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
713 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
714 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
715 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
716 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
717
718 /*
719 * L2ARC Internals
720 */
721 static list_t L2ARC_dev_list; /* device list */
722 static list_t *l2arc_dev_list; /* device list pointer */
723 static kmutex_t l2arc_dev_mtx; /* device list mutex */
724 static l2arc_dev_t *l2arc_dev_last; /* last device used */
725 static list_t L2ARC_free_on_write; /* free after write buf list */
726 static list_t *l2arc_free_on_write; /* free after write list ptr */
727 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
728 static uint64_t l2arc_ndev; /* number of devices */
729
730 typedef struct l2arc_read_callback {
731 arc_buf_hdr_t *l2rcb_hdr; /* read header */
732 blkptr_t l2rcb_bp; /* original blkptr */
733 zbookmark_phys_t l2rcb_zb; /* original bookmark */
734 int l2rcb_flags; /* original flags */
735 abd_t *l2rcb_abd; /* temporary buffer */
736 } l2arc_read_callback_t;
737
738 typedef struct l2arc_data_free {
739 /* protected by l2arc_free_on_write_mtx */
740 abd_t *l2df_abd;
741 size_t l2df_size;
742 arc_buf_contents_t l2df_type;
743 list_node_t l2df_list_node;
744 } l2arc_data_free_t;
745
746 static kmutex_t l2arc_feed_thr_lock;
747 static kcondvar_t l2arc_feed_thr_cv;
748 static uint8_t l2arc_thread_exit;
749
750 static kmutex_t l2arc_rebuild_thr_lock;
751 static kcondvar_t l2arc_rebuild_thr_cv;
752
753 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
754 typedef enum arc_fill_flags {
755 ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */
756 ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */
757 ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */
758 ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */
759 ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */
760 } arc_fill_flags_t;
761
762 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
763 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
764 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
765 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
766 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
767 static void arc_hdr_free_pabd(arc_buf_hdr_t *, boolean_t);
768 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *, boolean_t);
769 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
770 static boolean_t arc_is_overflowing();
771 static void arc_buf_watch(arc_buf_t *);
772 static l2arc_dev_t *l2arc_vdev_get(vdev_t *vd);
773
774 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
775 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
776 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
777 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
778
779 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
780 static void l2arc_read_done(zio_t *);
781
782 /*
783 * The arc_all_memory function is a ZoL enhancement that lives in their OSL
784 * code. In user-space code, which is used primarily for testing, we return
785 * half of all memory.
786 */
787 uint64_t
788 arc_all_memory(void)
789 {
790 #ifdef _KERNEL
791 return (ptob(physmem));
792 #else
793 return ((sysconf(_SC_PAGESIZE) * sysconf(_SC_PHYS_PAGES)) / 2);
794 #endif
795 }
796
797 /*
798 * We use Cityhash for this. It's fast, and has good hash properties without
799 * requiring any large static buffers.
800 */
801 static uint64_t
802 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
803 {
804 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
805 }
806
807 #define HDR_EMPTY(hdr) \
808 ((hdr)->b_dva.dva_word[0] == 0 && \
809 (hdr)->b_dva.dva_word[1] == 0)
810
811 #define HDR_EMPTY_OR_LOCKED(hdr) \
812 (HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr)))
813
814 #define HDR_EQUAL(spa, dva, birth, hdr) \
815 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
816 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
817 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
818
819 static void
820 buf_discard_identity(arc_buf_hdr_t *hdr)
821 {
822 hdr->b_dva.dva_word[0] = 0;
823 hdr->b_dva.dva_word[1] = 0;
824 hdr->b_birth = 0;
825 }
826
827 static arc_buf_hdr_t *
828 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
829 {
830 const dva_t *dva = BP_IDENTITY(bp);
831 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
832 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
833 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
834 arc_buf_hdr_t *hdr;
835
836 mutex_enter(hash_lock);
837 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
838 hdr = hdr->b_hash_next) {
839 if (HDR_EQUAL(spa, dva, birth, hdr)) {
840 *lockp = hash_lock;
841 return (hdr);
842 }
843 }
844 mutex_exit(hash_lock);
845 *lockp = NULL;
846 return (NULL);
847 }
848
849 /*
850 * Insert an entry into the hash table. If there is already an element
851 * equal to elem in the hash table, then the already existing element
852 * will be returned and the new element will not be inserted.
853 * Otherwise returns NULL.
854 * If lockp == NULL, the caller is assumed to already hold the hash lock.
855 */
856 static arc_buf_hdr_t *
857 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
858 {
859 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
860 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
861 arc_buf_hdr_t *fhdr;
862 uint32_t i;
863
864 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
865 ASSERT(hdr->b_birth != 0);
866 ASSERT(!HDR_IN_HASH_TABLE(hdr));
867
868 if (lockp != NULL) {
869 *lockp = hash_lock;
870 mutex_enter(hash_lock);
871 } else {
872 ASSERT(MUTEX_HELD(hash_lock));
873 }
874
875 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
876 fhdr = fhdr->b_hash_next, i++) {
877 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
878 return (fhdr);
879 }
880
881 hdr->b_hash_next = buf_hash_table.ht_table[idx];
882 buf_hash_table.ht_table[idx] = hdr;
883 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
884
885 /* collect some hash table performance data */
886 if (i > 0) {
887 ARCSTAT_BUMP(arcstat_hash_collisions);
888 if (i == 1)
889 ARCSTAT_BUMP(arcstat_hash_chains);
890
891 ARCSTAT_MAX(arcstat_hash_chain_max, i);
892 }
893
894 ARCSTAT_BUMP(arcstat_hash_elements);
895 ARCSTAT_MAXSTAT(arcstat_hash_elements);
896
897 return (NULL);
898 }
899
900 static void
901 buf_hash_remove(arc_buf_hdr_t *hdr)
902 {
903 arc_buf_hdr_t *fhdr, **hdrp;
904 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
905
906 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
907 ASSERT(HDR_IN_HASH_TABLE(hdr));
908
909 hdrp = &buf_hash_table.ht_table[idx];
910 while ((fhdr = *hdrp) != hdr) {
911 ASSERT3P(fhdr, !=, NULL);
912 hdrp = &fhdr->b_hash_next;
913 }
914 *hdrp = hdr->b_hash_next;
915 hdr->b_hash_next = NULL;
916 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
917
918 /* collect some hash table performance data */
919 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
920
921 if (buf_hash_table.ht_table[idx] &&
922 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
923 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
924 }
925
926 /*
927 * Global data structures and functions for the buf kmem cache.
928 */
929
930 static kmem_cache_t *hdr_full_cache;
931 static kmem_cache_t *hdr_full_crypt_cache;
932 static kmem_cache_t *hdr_l2only_cache;
933 static kmem_cache_t *buf_cache;
934
935 static void
936 buf_fini(void)
937 {
938 int i;
939
940 kmem_free(buf_hash_table.ht_table,
941 (buf_hash_table.ht_mask + 1) * sizeof (void *));
942 for (i = 0; i < BUF_LOCKS; i++)
943 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
944 kmem_cache_destroy(hdr_full_cache);
945 kmem_cache_destroy(hdr_full_crypt_cache);
946 kmem_cache_destroy(hdr_l2only_cache);
947 kmem_cache_destroy(buf_cache);
948 }
949
950 /*
951 * Constructor callback - called when the cache is empty
952 * and a new buf is requested.
953 */
954 /* ARGSUSED */
955 static int
956 hdr_full_cons(void *vbuf, void *unused, int kmflag)
957 {
958 arc_buf_hdr_t *hdr = vbuf;
959
960 bzero(hdr, HDR_FULL_SIZE);
961 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
962 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
963 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt);
964 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
965 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
966 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
967
968 return (0);
969 }
970
971 /* ARGSUSED */
972 static int
973 hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag)
974 {
975 arc_buf_hdr_t *hdr = vbuf;
976
977 (void) hdr_full_cons(vbuf, unused, kmflag);
978 bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr));
979 arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
980
981 return (0);
982 }
983
984 /* ARGSUSED */
985 static int
986 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
987 {
988 arc_buf_hdr_t *hdr = vbuf;
989
990 bzero(hdr, HDR_L2ONLY_SIZE);
991 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
992
993 return (0);
994 }
995
996 /* ARGSUSED */
997 static int
998 buf_cons(void *vbuf, void *unused, int kmflag)
999 {
1000 arc_buf_t *buf = vbuf;
1001
1002 bzero(buf, sizeof (arc_buf_t));
1003 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1004 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1005
1006 return (0);
1007 }
1008
1009 /*
1010 * Destructor callback - called when a cached buf is
1011 * no longer required.
1012 */
1013 /* ARGSUSED */
1014 static void
1015 hdr_full_dest(void *vbuf, void *unused)
1016 {
1017 arc_buf_hdr_t *hdr = vbuf;
1018
1019 ASSERT(HDR_EMPTY(hdr));
1020 cv_destroy(&hdr->b_l1hdr.b_cv);
1021 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1022 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1023 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1024 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1025 }
1026
1027 /* ARGSUSED */
1028 static void
1029 hdr_full_crypt_dest(void *vbuf, void *unused)
1030 {
1031 arc_buf_hdr_t *hdr = vbuf;
1032
1033 hdr_full_dest(hdr, unused);
1034 arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS);
1035 }
1036
1037 /* ARGSUSED */
1038 static void
1039 hdr_l2only_dest(void *vbuf, void *unused)
1040 {
1041 arc_buf_hdr_t *hdr = vbuf;
1042
1043 ASSERT(HDR_EMPTY(hdr));
1044 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1045 }
1046
1047 /* ARGSUSED */
1048 static void
1049 buf_dest(void *vbuf, void *unused)
1050 {
1051 arc_buf_t *buf = vbuf;
1052
1053 mutex_destroy(&buf->b_evict_lock);
1054 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1055 }
1056
1057 /*
1058 * Reclaim callback -- invoked when memory is low.
1059 */
1060 /* ARGSUSED */
1061 static void
1062 hdr_recl(void *unused)
1063 {
1064 dprintf("hdr_recl called\n");
1065 /*
1066 * umem calls the reclaim func when we destroy the buf cache,
1067 * which is after we do arc_fini().
1068 */
1069 if (arc_initialized)
1070 zthr_wakeup(arc_reap_zthr);
1071 }
1072
1073 static void
1074 buf_init(void)
1075 {
1076 uint64_t *ct;
1077 uint64_t hsize = 1ULL << 12;
1078 int i, j;
1079
1080 /*
1081 * The hash table is big enough to fill all of physical memory
1082 * with an average block size of zfs_arc_average_blocksize (default 8K).
1083 * By default, the table will take up
1084 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1085 */
1086 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1087 hsize <<= 1;
1088 retry:
1089 buf_hash_table.ht_mask = hsize - 1;
1090 buf_hash_table.ht_table =
1091 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1092 if (buf_hash_table.ht_table == NULL) {
1093 ASSERT(hsize > (1ULL << 8));
1094 hsize >>= 1;
1095 goto retry;
1096 }
1097
1098 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1099 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1100 hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt",
1101 HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest,
1102 hdr_recl, NULL, NULL, 0);
1103 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1104 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1105 NULL, NULL, 0);
1106 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1107 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1108
1109 for (i = 0; i < 256; i++)
1110 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1111 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1112
1113 for (i = 0; i < BUF_LOCKS; i++) {
1114 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1115 NULL, MUTEX_DEFAULT, NULL);
1116 }
1117 }
1118
1119 /*
1120 * This is the size that the buf occupies in memory. If the buf is compressed,
1121 * it will correspond to the compressed size. You should use this method of
1122 * getting the buf size unless you explicitly need the logical size.
1123 */
1124 int32_t
1125 arc_buf_size(arc_buf_t *buf)
1126 {
1127 return (ARC_BUF_COMPRESSED(buf) ?
1128 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1129 }
1130
1131 int32_t
1132 arc_buf_lsize(arc_buf_t *buf)
1133 {
1134 return (HDR_GET_LSIZE(buf->b_hdr));
1135 }
1136
1137 /*
1138 * This function will return B_TRUE if the buffer is encrypted in memory.
1139 * This buffer can be decrypted by calling arc_untransform().
1140 */
1141 boolean_t
1142 arc_is_encrypted(arc_buf_t *buf)
1143 {
1144 return (ARC_BUF_ENCRYPTED(buf) != 0);
1145 }
1146
1147 /*
1148 * Returns B_TRUE if the buffer represents data that has not had its MAC
1149 * verified yet.
1150 */
1151 boolean_t
1152 arc_is_unauthenticated(arc_buf_t *buf)
1153 {
1154 return (HDR_NOAUTH(buf->b_hdr) != 0);
1155 }
1156
1157 void
1158 arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt,
1159 uint8_t *iv, uint8_t *mac)
1160 {
1161 arc_buf_hdr_t *hdr = buf->b_hdr;
1162
1163 ASSERT(HDR_PROTECTED(hdr));
1164
1165 bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN);
1166 bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN);
1167 bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN);
1168 *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
1169 /* CONSTCOND */
1170 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
1171 }
1172
1173 /*
1174 * Indicates how this buffer is compressed in memory. If it is not compressed
1175 * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with
1176 * arc_untransform() as long as it is also unencrypted.
1177 */
1178 enum zio_compress
1179 arc_get_compression(arc_buf_t *buf)
1180 {
1181 return (ARC_BUF_COMPRESSED(buf) ?
1182 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF);
1183 }
1184
1185 #define ARC_MINTIME (hz>>4) /* 62 ms */
1186
1187 /*
1188 * Return the compression algorithm used to store this data in the ARC. If ARC
1189 * compression is enabled or this is an encrypted block, this will be the same
1190 * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF.
1191 */
1192 static inline enum zio_compress
1193 arc_hdr_get_compress(arc_buf_hdr_t *hdr)
1194 {
1195 return (HDR_COMPRESSION_ENABLED(hdr) ?
1196 HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF);
1197 }
1198
1199 static inline boolean_t
1200 arc_buf_is_shared(arc_buf_t *buf)
1201 {
1202 boolean_t shared = (buf->b_data != NULL &&
1203 buf->b_hdr->b_l1hdr.b_pabd != NULL &&
1204 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) &&
1205 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd));
1206 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr));
1207 IMPLY(shared, ARC_BUF_SHARED(buf));
1208 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1209
1210 /*
1211 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1212 * already being shared" requirement prevents us from doing that.
1213 */
1214
1215 return (shared);
1216 }
1217
1218 /*
1219 * Free the checksum associated with this header. If there is no checksum, this
1220 * is a no-op.
1221 */
1222 static inline void
1223 arc_cksum_free(arc_buf_hdr_t *hdr)
1224 {
1225 ASSERT(HDR_HAS_L1HDR(hdr));
1226
1227 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1228 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1229 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1230 hdr->b_l1hdr.b_freeze_cksum = NULL;
1231 }
1232 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1233 }
1234
1235 /*
1236 * Return true iff at least one of the bufs on hdr is not compressed.
1237 * Encrypted buffers count as compressed.
1238 */
1239 static boolean_t
1240 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1241 {
1242 ASSERT(hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY_OR_LOCKED(hdr));
1243
1244 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1245 if (!ARC_BUF_COMPRESSED(b)) {
1246 return (B_TRUE);
1247 }
1248 }
1249 return (B_FALSE);
1250 }
1251
1252 /*
1253 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1254 * matches the checksum that is stored in the hdr. If there is no checksum,
1255 * or if the buf is compressed, this is a no-op.
1256 */
1257 static void
1258 arc_cksum_verify(arc_buf_t *buf)
1259 {
1260 arc_buf_hdr_t *hdr = buf->b_hdr;
1261 zio_cksum_t zc;
1262
1263 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1264 return;
1265
1266 if (ARC_BUF_COMPRESSED(buf))
1267 return;
1268
1269 ASSERT(HDR_HAS_L1HDR(hdr));
1270
1271 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1272
1273 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1274 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1275 return;
1276 }
1277
1278 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1279 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1280 panic("buffer modified while frozen!");
1281 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1282 }
1283
1284 /*
1285 * This function makes the assumption that data stored in the L2ARC
1286 * will be transformed exactly as it is in the main pool. Because of
1287 * this we can verify the checksum against the reading process's bp.
1288 */
1289 static boolean_t
1290 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1291 {
1292 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1293 boolean_t valid_cksum;
1294
1295 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1296 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1297
1298 /*
1299 * We rely on the blkptr's checksum to determine if the block
1300 * is valid or not. When compressed arc is enabled, the l2arc
1301 * writes the block to the l2arc just as it appears in the pool.
1302 * This allows us to use the blkptr's checksum to validate the
1303 * data that we just read off of the l2arc without having to store
1304 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1305 * arc is disabled, then the data written to the l2arc is always
1306 * uncompressed and won't match the block as it exists in the main
1307 * pool. When this is the case, we must first compress it if it is
1308 * compressed on the main pool before we can validate the checksum.
1309 */
1310 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1311 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1312 uint64_t lsize = HDR_GET_LSIZE(hdr);
1313 uint64_t csize;
1314
1315 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1316 csize = zio_compress_data(compress, zio->io_abd,
1317 abd_to_buf(cdata), lsize);
1318
1319 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1320 if (csize < HDR_GET_PSIZE(hdr)) {
1321 /*
1322 * Compressed blocks are always a multiple of the
1323 * smallest ashift in the pool. Ideally, we would
1324 * like to round up the csize to the next
1325 * spa_min_ashift but that value may have changed
1326 * since the block was last written. Instead,
1327 * we rely on the fact that the hdr's psize
1328 * was set to the psize of the block when it was
1329 * last written. We set the csize to that value
1330 * and zero out any part that should not contain
1331 * data.
1332 */
1333 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
1334 csize = HDR_GET_PSIZE(hdr);
1335 }
1336 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
1337 }
1338
1339 /*
1340 * Block pointers always store the checksum for the logical data.
1341 * If the block pointer has the gang bit set, then the checksum
1342 * it represents is for the reconstituted data and not for an
1343 * individual gang member. The zio pipeline, however, must be able to
1344 * determine the checksum of each of the gang constituents so it
1345 * treats the checksum comparison differently than what we need
1346 * for l2arc blocks. This prevents us from using the
1347 * zio_checksum_error() interface directly. Instead we must call the
1348 * zio_checksum_error_impl() so that we can ensure the checksum is
1349 * generated using the correct checksum algorithm and accounts for the
1350 * logical I/O size and not just a gang fragment.
1351 */
1352 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp,
1353 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size,
1354 zio->io_offset, NULL) == 0);
1355 zio_pop_transforms(zio);
1356 return (valid_cksum);
1357 }
1358
1359 /*
1360 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1361 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1362 * isn't modified later on. If buf is compressed or there is already a checksum
1363 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1364 */
1365 static void
1366 arc_cksum_compute(arc_buf_t *buf)
1367 {
1368 arc_buf_hdr_t *hdr = buf->b_hdr;
1369
1370 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1371 return;
1372
1373 ASSERT(HDR_HAS_L1HDR(hdr));
1374
1375 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1376 if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) {
1377 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1378 return;
1379 }
1380
1381 ASSERT(!ARC_BUF_ENCRYPTED(buf));
1382 ASSERT(!ARC_BUF_COMPRESSED(buf));
1383 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1384 KM_SLEEP);
1385 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1386 hdr->b_l1hdr.b_freeze_cksum);
1387 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1388 arc_buf_watch(buf);
1389 }
1390
1391 #ifndef _KERNEL
1392 typedef struct procctl {
1393 long cmd;
1394 prwatch_t prwatch;
1395 } procctl_t;
1396 #endif
1397
1398 /* ARGSUSED */
1399 static void
1400 arc_buf_unwatch(arc_buf_t *buf)
1401 {
1402 #ifndef _KERNEL
1403 if (arc_watch) {
1404 int result;
1405 procctl_t ctl;
1406 ctl.cmd = PCWATCH;
1407 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1408 ctl.prwatch.pr_size = 0;
1409 ctl.prwatch.pr_wflags = 0;
1410 result = write(arc_procfd, &ctl, sizeof (ctl));
1411 ASSERT3U(result, ==, sizeof (ctl));
1412 }
1413 #endif
1414 }
1415
1416 /* ARGSUSED */
1417 static void
1418 arc_buf_watch(arc_buf_t *buf)
1419 {
1420 #ifndef _KERNEL
1421 if (arc_watch) {
1422 int result;
1423 procctl_t ctl;
1424 ctl.cmd = PCWATCH;
1425 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1426 ctl.prwatch.pr_size = arc_buf_size(buf);
1427 ctl.prwatch.pr_wflags = WA_WRITE;
1428 result = write(arc_procfd, &ctl, sizeof (ctl));
1429 ASSERT3U(result, ==, sizeof (ctl));
1430 }
1431 #endif
1432 }
1433
1434 static arc_buf_contents_t
1435 arc_buf_type(arc_buf_hdr_t *hdr)
1436 {
1437 arc_buf_contents_t type;
1438 if (HDR_ISTYPE_METADATA(hdr)) {
1439 type = ARC_BUFC_METADATA;
1440 } else {
1441 type = ARC_BUFC_DATA;
1442 }
1443 VERIFY3U(hdr->b_type, ==, type);
1444 return (type);
1445 }
1446
1447 boolean_t
1448 arc_is_metadata(arc_buf_t *buf)
1449 {
1450 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1451 }
1452
1453 static uint32_t
1454 arc_bufc_to_flags(arc_buf_contents_t type)
1455 {
1456 switch (type) {
1457 case ARC_BUFC_DATA:
1458 /* metadata field is 0 if buffer contains normal data */
1459 return (0);
1460 case ARC_BUFC_METADATA:
1461 return (ARC_FLAG_BUFC_METADATA);
1462 default:
1463 break;
1464 }
1465 panic("undefined ARC buffer type!");
1466 return ((uint32_t)-1);
1467 }
1468
1469 void
1470 arc_buf_thaw(arc_buf_t *buf)
1471 {
1472 arc_buf_hdr_t *hdr = buf->b_hdr;
1473
1474 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1475 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1476
1477 arc_cksum_verify(buf);
1478
1479 /*
1480 * Compressed buffers do not manipulate the b_freeze_cksum.
1481 */
1482 if (ARC_BUF_COMPRESSED(buf))
1483 return;
1484
1485 ASSERT(HDR_HAS_L1HDR(hdr));
1486 arc_cksum_free(hdr);
1487
1488 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1489 #ifdef ZFS_DEBUG
1490 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1491 if (hdr->b_l1hdr.b_thawed != NULL)
1492 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1493 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1494 }
1495 #endif
1496
1497 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1498
1499 arc_buf_unwatch(buf);
1500 }
1501
1502 void
1503 arc_buf_freeze(arc_buf_t *buf)
1504 {
1505 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1506 return;
1507
1508 if (ARC_BUF_COMPRESSED(buf))
1509 return;
1510
1511 ASSERT(HDR_HAS_L1HDR(buf->b_hdr));
1512 arc_cksum_compute(buf);
1513 }
1514
1515 /*
1516 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1517 * the following functions should be used to ensure that the flags are
1518 * updated in a thread-safe way. When manipulating the flags either
1519 * the hash_lock must be held or the hdr must be undiscoverable. This
1520 * ensures that we're not racing with any other threads when updating
1521 * the flags.
1522 */
1523 static inline void
1524 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1525 {
1526 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1527 hdr->b_flags |= flags;
1528 }
1529
1530 static inline void
1531 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1532 {
1533 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1534 hdr->b_flags &= ~flags;
1535 }
1536
1537 /*
1538 * Setting the compression bits in the arc_buf_hdr_t's b_flags is
1539 * done in a special way since we have to clear and set bits
1540 * at the same time. Consumers that wish to set the compression bits
1541 * must use this function to ensure that the flags are updated in
1542 * thread-safe manner.
1543 */
1544 static void
1545 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp)
1546 {
1547 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1548
1549 /*
1550 * Holes and embedded blocks will always have a psize = 0 so
1551 * we ignore the compression of the blkptr and set the
1552 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF.
1553 * Holes and embedded blocks remain anonymous so we don't
1554 * want to uncompress them. Mark them as uncompressed.
1555 */
1556 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) {
1557 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1558 ASSERT(!HDR_COMPRESSION_ENABLED(hdr));
1559 } else {
1560 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
1561 ASSERT(HDR_COMPRESSION_ENABLED(hdr));
1562 }
1563
1564 HDR_SET_COMPRESS(hdr, cmp);
1565 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp);
1566 }
1567
1568 /*
1569 * Looks for another buf on the same hdr which has the data decompressed, copies
1570 * from it, and returns true. If no such buf exists, returns false.
1571 */
1572 static boolean_t
1573 arc_buf_try_copy_decompressed_data(arc_buf_t *buf)
1574 {
1575 arc_buf_hdr_t *hdr = buf->b_hdr;
1576 boolean_t copied = B_FALSE;
1577
1578 ASSERT(HDR_HAS_L1HDR(hdr));
1579 ASSERT3P(buf->b_data, !=, NULL);
1580 ASSERT(!ARC_BUF_COMPRESSED(buf));
1581
1582 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1583 from = from->b_next) {
1584 /* can't use our own data buffer */
1585 if (from == buf) {
1586 continue;
1587 }
1588
1589 if (!ARC_BUF_COMPRESSED(from)) {
1590 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1591 copied = B_TRUE;
1592 break;
1593 }
1594 }
1595
1596 /*
1597 * Note: With encryption support, the following assertion is no longer
1598 * necessarily valid. If we receive two back to back raw snapshots
1599 * (send -w), the second receive can use a hdr with a cksum already
1600 * calculated. This happens via:
1601 * dmu_recv_stream() -> receive_read_record() -> arc_loan_raw_buf()
1602 * The rsend/send_mixed_raw test case exercises this code path.
1603 *
1604 * There were no decompressed bufs, so there should not be a
1605 * checksum on the hdr either.
1606 * EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1607 */
1608
1609 return (copied);
1610 }
1611
1612 /*
1613 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t.
1614 */
1615 static uint64_t
1616 arc_hdr_size(arc_buf_hdr_t *hdr)
1617 {
1618 uint64_t size;
1619
1620 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
1621 HDR_GET_PSIZE(hdr) > 0) {
1622 size = HDR_GET_PSIZE(hdr);
1623 } else {
1624 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0);
1625 size = HDR_GET_LSIZE(hdr);
1626 }
1627 return (size);
1628 }
1629
1630 static int
1631 arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj)
1632 {
1633 int ret;
1634 uint64_t csize;
1635 uint64_t lsize = HDR_GET_LSIZE(hdr);
1636 uint64_t psize = HDR_GET_PSIZE(hdr);
1637 void *tmpbuf = NULL;
1638 abd_t *abd = hdr->b_l1hdr.b_pabd;
1639
1640 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1641 ASSERT(HDR_AUTHENTICATED(hdr));
1642 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1643
1644 /*
1645 * The MAC is calculated on the compressed data that is stored on disk.
1646 * However, if compressed arc is disabled we will only have the
1647 * decompressed data available to us now. Compress it into a temporary
1648 * abd so we can verify the MAC. The performance overhead of this will
1649 * be relatively low, since most objects in an encrypted objset will
1650 * be encrypted (instead of authenticated) anyway.
1651 */
1652 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1653 !HDR_COMPRESSION_ENABLED(hdr)) {
1654 tmpbuf = zio_buf_alloc(lsize);
1655 abd = abd_get_from_buf(tmpbuf, lsize);
1656 abd_take_ownership_of_buf(abd, B_TRUE);
1657
1658 csize = zio_compress_data(HDR_GET_COMPRESS(hdr),
1659 hdr->b_l1hdr.b_pabd, tmpbuf, lsize);
1660 ASSERT3U(csize, <=, psize);
1661 abd_zero_off(abd, csize, psize - csize);
1662 }
1663
1664 /*
1665 * Authentication is best effort. We authenticate whenever the key is
1666 * available. If we succeed we clear ARC_FLAG_NOAUTH.
1667 */
1668 if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) {
1669 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1670 ASSERT3U(lsize, ==, psize);
1671 ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd,
1672 psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1673 } else {
1674 ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize,
1675 hdr->b_crypt_hdr.b_mac);
1676 }
1677
1678 if (ret == 0)
1679 arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH);
1680 else if (ret != ENOENT)
1681 goto error;
1682
1683 if (tmpbuf != NULL)
1684 abd_free(abd);
1685
1686 return (0);
1687
1688 error:
1689 if (tmpbuf != NULL)
1690 abd_free(abd);
1691
1692 return (ret);
1693 }
1694
1695 /*
1696 * This function will take a header that only has raw encrypted data in
1697 * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in
1698 * b_l1hdr.b_pabd. If designated in the header flags, this function will
1699 * also decompress the data.
1700 */
1701 static int
1702 arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb)
1703 {
1704 int ret;
1705 abd_t *cabd = NULL;
1706 void *tmp = NULL;
1707 boolean_t no_crypt = B_FALSE;
1708 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
1709
1710 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1711 ASSERT(HDR_ENCRYPTED(hdr));
1712
1713 arc_hdr_alloc_pabd(hdr, B_FALSE);
1714
1715 ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot,
1716 B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv,
1717 hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd,
1718 hdr->b_crypt_hdr.b_rabd, &no_crypt);
1719 if (ret != 0)
1720 goto error;
1721
1722 if (no_crypt) {
1723 abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd,
1724 HDR_GET_PSIZE(hdr));
1725 }
1726
1727 /*
1728 * If this header has disabled arc compression but the b_pabd is
1729 * compressed after decrypting it, we need to decompress the newly
1730 * decrypted data.
1731 */
1732 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
1733 !HDR_COMPRESSION_ENABLED(hdr)) {
1734 /*
1735 * We want to make sure that we are correctly honoring the
1736 * zfs_abd_scatter_enabled setting, so we allocate an abd here
1737 * and then loan a buffer from it, rather than allocating a
1738 * linear buffer and wrapping it in an abd later.
1739 */
1740 cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
1741 tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
1742
1743 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1744 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
1745 HDR_GET_LSIZE(hdr));
1746 if (ret != 0) {
1747 abd_return_buf(cabd, tmp, arc_hdr_size(hdr));
1748 goto error;
1749 }
1750
1751 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
1752 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
1753 arc_hdr_size(hdr), hdr);
1754 hdr->b_l1hdr.b_pabd = cabd;
1755 }
1756
1757 return (0);
1758
1759 error:
1760 arc_hdr_free_pabd(hdr, B_FALSE);
1761 if (cabd != NULL)
1762 arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr);
1763
1764 return (ret);
1765 }
1766
1767 /*
1768 * This function is called during arc_buf_fill() to prepare the header's
1769 * abd plaintext pointer for use. This involves authenticated protected
1770 * data and decrypting encrypted data into the plaintext abd.
1771 */
1772 static int
1773 arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa,
1774 const zbookmark_phys_t *zb, boolean_t noauth)
1775 {
1776 int ret;
1777
1778 ASSERT(HDR_PROTECTED(hdr));
1779
1780 if (hash_lock != NULL)
1781 mutex_enter(hash_lock);
1782
1783 if (HDR_NOAUTH(hdr) && !noauth) {
1784 /*
1785 * The caller requested authenticated data but our data has
1786 * not been authenticated yet. Verify the MAC now if we can.
1787 */
1788 ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset);
1789 if (ret != 0)
1790 goto error;
1791 } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) {
1792 /*
1793 * If we only have the encrypted version of the data, but the
1794 * unencrypted version was requested we take this opportunity
1795 * to store the decrypted version in the header for future use.
1796 */
1797 ret = arc_hdr_decrypt(hdr, spa, zb);
1798 if (ret != 0)
1799 goto error;
1800 }
1801
1802 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1803
1804 if (hash_lock != NULL)
1805 mutex_exit(hash_lock);
1806
1807 return (0);
1808
1809 error:
1810 if (hash_lock != NULL)
1811 mutex_exit(hash_lock);
1812
1813 return (ret);
1814 }
1815
1816 /*
1817 * This function is used by the dbuf code to decrypt bonus buffers in place.
1818 * The dbuf code itself doesn't have any locking for decrypting a shared dnode
1819 * block, so we use the hash lock here to protect against concurrent calls to
1820 * arc_buf_fill().
1821 */
1822 /* ARGSUSED */
1823 static void
1824 arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock)
1825 {
1826 arc_buf_hdr_t *hdr = buf->b_hdr;
1827
1828 ASSERT(HDR_ENCRYPTED(hdr));
1829 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1830 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
1831 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
1832
1833 zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data,
1834 arc_buf_size(buf));
1835 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
1836 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1837 hdr->b_crypt_hdr.b_ebufcnt -= 1;
1838 }
1839
1840 /*
1841 * Given a buf that has a data buffer attached to it, this function will
1842 * efficiently fill the buf with data of the specified compression setting from
1843 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1844 * are already sharing a data buf, no copy is performed.
1845 *
1846 * If the buf is marked as compressed but uncompressed data was requested, this
1847 * will allocate a new data buffer for the buf, remove that flag, and fill the
1848 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1849 * uncompressed data, and (since we haven't added support for it yet) if you
1850 * want compressed data your buf must already be marked as compressed and have
1851 * the correct-sized data buffer.
1852 */
1853 static int
1854 arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
1855 arc_fill_flags_t flags)
1856 {
1857 int error = 0;
1858 arc_buf_hdr_t *hdr = buf->b_hdr;
1859 boolean_t hdr_compressed =
1860 (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
1861 boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0;
1862 boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0;
1863 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap;
1864 kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr);
1865
1866 ASSERT3P(buf->b_data, !=, NULL);
1867 IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf));
1868 IMPLY(compressed, ARC_BUF_COMPRESSED(buf));
1869 IMPLY(encrypted, HDR_ENCRYPTED(hdr));
1870 IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf));
1871 IMPLY(encrypted, ARC_BUF_COMPRESSED(buf));
1872 IMPLY(encrypted, !ARC_BUF_SHARED(buf));
1873
1874 /*
1875 * If the caller wanted encrypted data we just need to copy it from
1876 * b_rabd and potentially byteswap it. We won't be able to do any
1877 * further transforms on it.
1878 */
1879 if (encrypted) {
1880 ASSERT(HDR_HAS_RABD(hdr));
1881 abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd,
1882 HDR_GET_PSIZE(hdr));
1883 goto byteswap;
1884 }
1885
1886 /*
1887 * Adjust encrypted and authenticated headers to accomodate
1888 * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are
1889 * allowed to fail decryption due to keys not being loaded
1890 * without being marked as an IO error.
1891 */
1892 if (HDR_PROTECTED(hdr)) {
1893 error = arc_fill_hdr_crypt(hdr, hash_lock, spa,
1894 zb, !!(flags & ARC_FILL_NOAUTH));
1895 if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) {
1896 return (error);
1897 } else if (error != 0) {
1898 if (hash_lock != NULL)
1899 mutex_enter(hash_lock);
1900 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
1901 if (hash_lock != NULL)
1902 mutex_exit(hash_lock);
1903 return (error);
1904 }
1905 }
1906
1907 /*
1908 * There is a special case here for dnode blocks which are
1909 * decrypting their bonus buffers. These blocks may request to
1910 * be decrypted in-place. This is necessary because there may
1911 * be many dnodes pointing into this buffer and there is
1912 * currently no method to synchronize replacing the backing
1913 * b_data buffer and updating all of the pointers. Here we use
1914 * the hash lock to ensure there are no races. If the need
1915 * arises for other types to be decrypted in-place, they must
1916 * add handling here as well.
1917 */
1918 if ((flags & ARC_FILL_IN_PLACE) != 0) {
1919 ASSERT(!hdr_compressed);
1920 ASSERT(!compressed);
1921 ASSERT(!encrypted);
1922
1923 if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) {
1924 ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE);
1925
1926 if (hash_lock != NULL)
1927 mutex_enter(hash_lock);
1928 arc_buf_untransform_in_place(buf, hash_lock);
1929 if (hash_lock != NULL)
1930 mutex_exit(hash_lock);
1931
1932 /* Compute the hdr's checksum if necessary */
1933 arc_cksum_compute(buf);
1934 }
1935
1936 return (0);
1937 }
1938
1939 if (hdr_compressed == compressed) {
1940 if (!arc_buf_is_shared(buf)) {
1941 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd,
1942 arc_buf_size(buf));
1943 }
1944 } else {
1945 ASSERT(hdr_compressed);
1946 ASSERT(!compressed);
1947 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr));
1948
1949 /*
1950 * If the buf is sharing its data with the hdr, unlink it and
1951 * allocate a new data buffer for the buf.
1952 */
1953 if (arc_buf_is_shared(buf)) {
1954 ASSERT(ARC_BUF_COMPRESSED(buf));
1955
1956 /* We need to give the buf its own b_data */
1957 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
1958 buf->b_data =
1959 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
1960 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
1961
1962 /* Previously overhead was 0; just add new overhead */
1963 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr));
1964 } else if (ARC_BUF_COMPRESSED(buf)) {
1965 /* We need to reallocate the buf's b_data */
1966 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr),
1967 buf);
1968 buf->b_data =
1969 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
1970
1971 /* We increased the size of b_data; update overhead */
1972 ARCSTAT_INCR(arcstat_overhead_size,
1973 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
1974 }
1975
1976 /*
1977 * Regardless of the buf's previous compression settings, it
1978 * should not be compressed at the end of this function.
1979 */
1980 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1981
1982 /*
1983 * Try copying the data from another buf which already has a
1984 * decompressed version. If that's not possible, it's time to
1985 * bite the bullet and decompress the data from the hdr.
1986 */
1987 if (arc_buf_try_copy_decompressed_data(buf)) {
1988 /* Skip byteswapping and checksumming (already done) */
1989 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
1990 return (0);
1991 } else {
1992 error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1993 hdr->b_l1hdr.b_pabd, buf->b_data,
1994 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
1995
1996 /*
1997 * Absent hardware errors or software bugs, this should
1998 * be impossible, but log it anyway so we can debug it.
1999 */
2000 if (error != 0) {
2001 zfs_dbgmsg(
2002 "hdr %p, compress %d, psize %d, lsize %d",
2003 hdr, arc_hdr_get_compress(hdr),
2004 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2005 if (hash_lock != NULL)
2006 mutex_enter(hash_lock);
2007 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
2008 if (hash_lock != NULL)
2009 mutex_exit(hash_lock);
2010 return (SET_ERROR(EIO));
2011 }
2012 }
2013 }
2014
2015 byteswap:
2016 /* Byteswap the buf's data if necessary */
2017 if (bswap != DMU_BSWAP_NUMFUNCS) {
2018 ASSERT(!HDR_SHARED_DATA(hdr));
2019 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS);
2020 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr));
2021 }
2022
2023 /* Compute the hdr's checksum if necessary */
2024 arc_cksum_compute(buf);
2025
2026 return (0);
2027 }
2028
2029 /*
2030 * If this function is being called to decrypt an encrypted buffer or verify an
2031 * authenticated one, the key must be loaded and a mapping must be made
2032 * available in the keystore via spa_keystore_create_mapping() or one of its
2033 * callers.
2034 */
2035 int
2036 arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb,
2037 boolean_t in_place)
2038 {
2039 int ret;
2040 arc_fill_flags_t flags = 0;
2041
2042 if (in_place)
2043 flags |= ARC_FILL_IN_PLACE;
2044
2045 ret = arc_buf_fill(buf, spa, zb, flags);
2046 if (ret == ECKSUM) {
2047 /*
2048 * Convert authentication and decryption errors to EIO
2049 * (and generate an ereport) before leaving the ARC.
2050 */
2051 ret = SET_ERROR(EIO);
2052 spa_log_error(spa, zb);
2053 (void) zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION,
2054 spa, NULL, zb, NULL, 0, 0);
2055 }
2056
2057 return (ret);
2058 }
2059
2060 /*
2061 * Increment the amount of evictable space in the arc_state_t's refcount.
2062 * We account for the space used by the hdr and the arc buf individually
2063 * so that we can add and remove them from the refcount individually.
2064 */
2065 static void
2066 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state)
2067 {
2068 arc_buf_contents_t type = arc_buf_type(hdr);
2069
2070 ASSERT(HDR_HAS_L1HDR(hdr));
2071
2072 if (GHOST_STATE(state)) {
2073 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2074 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2075 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2076 ASSERT(!HDR_HAS_RABD(hdr));
2077 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2078 HDR_GET_LSIZE(hdr), hdr);
2079 return;
2080 }
2081
2082 ASSERT(!GHOST_STATE(state));
2083 if (hdr->b_l1hdr.b_pabd != NULL) {
2084 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2085 arc_hdr_size(hdr), hdr);
2086 }
2087 if (HDR_HAS_RABD(hdr)) {
2088 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2089 HDR_GET_PSIZE(hdr), hdr);
2090 }
2091 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2092 buf = buf->b_next) {
2093 if (arc_buf_is_shared(buf))
2094 continue;
2095 (void) zfs_refcount_add_many(&state->arcs_esize[type],
2096 arc_buf_size(buf), buf);
2097 }
2098 }
2099
2100 /*
2101 * Decrement the amount of evictable space in the arc_state_t's refcount.
2102 * We account for the space used by the hdr and the arc buf individually
2103 * so that we can add and remove them from the refcount individually.
2104 */
2105 static void
2106 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state)
2107 {
2108 arc_buf_contents_t type = arc_buf_type(hdr);
2109
2110 ASSERT(HDR_HAS_L1HDR(hdr));
2111
2112 if (GHOST_STATE(state)) {
2113 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2114 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2115 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2116 ASSERT(!HDR_HAS_RABD(hdr));
2117 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2118 HDR_GET_LSIZE(hdr), hdr);
2119 return;
2120 }
2121
2122 ASSERT(!GHOST_STATE(state));
2123 if (hdr->b_l1hdr.b_pabd != NULL) {
2124 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2125 arc_hdr_size(hdr), hdr);
2126 }
2127 if (HDR_HAS_RABD(hdr)) {
2128 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2129 HDR_GET_PSIZE(hdr), hdr);
2130 }
2131 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2132 buf = buf->b_next) {
2133 if (arc_buf_is_shared(buf))
2134 continue;
2135 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2136 arc_buf_size(buf), buf);
2137 }
2138 }
2139
2140 /*
2141 * Add a reference to this hdr indicating that someone is actively
2142 * referencing that memory. When the refcount transitions from 0 to 1,
2143 * we remove it from the respective arc_state_t list to indicate that
2144 * it is not evictable.
2145 */
2146 static void
2147 add_reference(arc_buf_hdr_t *hdr, void *tag)
2148 {
2149 ASSERT(HDR_HAS_L1HDR(hdr));
2150 if (!HDR_EMPTY(hdr) && !MUTEX_HELD(HDR_LOCK(hdr))) {
2151 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
2152 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2153 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2154 }
2155
2156 arc_state_t *state = hdr->b_l1hdr.b_state;
2157
2158 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
2159 (state != arc_anon)) {
2160 /* We don't use the L2-only state list. */
2161 if (state != arc_l2c_only) {
2162 multilist_remove(state->arcs_list[arc_buf_type(hdr)],
2163 hdr);
2164 arc_evictable_space_decrement(hdr, state);
2165 }
2166 /* remove the prefetch flag if we get a reference */
2167 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH);
2168 }
2169 }
2170
2171 /*
2172 * Remove a reference from this hdr. When the reference transitions from
2173 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's
2174 * list making it eligible for eviction.
2175 */
2176 static int
2177 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
2178 {
2179 int cnt;
2180 arc_state_t *state = hdr->b_l1hdr.b_state;
2181
2182 ASSERT(HDR_HAS_L1HDR(hdr));
2183 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
2184 ASSERT(!GHOST_STATE(state));
2185
2186 /*
2187 * arc_l2c_only counts as a ghost state so we don't need to explicitly
2188 * check to prevent usage of the arc_l2c_only list.
2189 */
2190 if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
2191 (state != arc_anon)) {
2192 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr);
2193 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
2194 arc_evictable_space_increment(hdr, state);
2195 }
2196 return (cnt);
2197 }
2198
2199 /*
2200 * Move the supplied buffer to the indicated state. The hash lock
2201 * for the buffer must be held by the caller.
2202 */
2203 static void
2204 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
2205 kmutex_t *hash_lock)
2206 {
2207 arc_state_t *old_state;
2208 int64_t refcnt;
2209 uint32_t bufcnt;
2210 boolean_t update_old, update_new;
2211 arc_buf_contents_t buftype = arc_buf_type(hdr);
2212
2213 /*
2214 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
2215 * in arc_read() when bringing a buffer out of the L2ARC. However, the
2216 * L1 hdr doesn't always exist when we change state to arc_anon before
2217 * destroying a header, in which case reallocating to add the L1 hdr is
2218 * pointless.
2219 */
2220 if (HDR_HAS_L1HDR(hdr)) {
2221 old_state = hdr->b_l1hdr.b_state;
2222 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt);
2223 bufcnt = hdr->b_l1hdr.b_bufcnt;
2224
2225 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL ||
2226 HDR_HAS_RABD(hdr));
2227 } else {
2228 old_state = arc_l2c_only;
2229 refcnt = 0;
2230 bufcnt = 0;
2231 update_old = B_FALSE;
2232 }
2233 update_new = update_old;
2234
2235 ASSERT(MUTEX_HELD(hash_lock));
2236 ASSERT3P(new_state, !=, old_state);
2237 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0);
2238 ASSERT(old_state != arc_anon || bufcnt <= 1);
2239
2240 /*
2241 * If this buffer is evictable, transfer it from the
2242 * old state list to the new state list.
2243 */
2244 if (refcnt == 0) {
2245 if (old_state != arc_anon && old_state != arc_l2c_only) {
2246 ASSERT(HDR_HAS_L1HDR(hdr));
2247 multilist_remove(old_state->arcs_list[buftype], hdr);
2248
2249 if (GHOST_STATE(old_state)) {
2250 ASSERT0(bufcnt);
2251 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2252 update_old = B_TRUE;
2253 }
2254 arc_evictable_space_decrement(hdr, old_state);
2255 }
2256 if (new_state != arc_anon && new_state != arc_l2c_only) {
2257
2258 /*
2259 * An L1 header always exists here, since if we're
2260 * moving to some L1-cached state (i.e. not l2c_only or
2261 * anonymous), we realloc the header to add an L1hdr
2262 * beforehand.
2263 */
2264 ASSERT(HDR_HAS_L1HDR(hdr));
2265 multilist_insert(new_state->arcs_list[buftype], hdr);
2266
2267 if (GHOST_STATE(new_state)) {
2268 ASSERT0(bufcnt);
2269 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2270 update_new = B_TRUE;
2271 }
2272 arc_evictable_space_increment(hdr, new_state);
2273 }
2274 }
2275
2276 ASSERT(!HDR_EMPTY(hdr));
2277 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2278 buf_hash_remove(hdr);
2279
2280 /* adjust state sizes (ignore arc_l2c_only) */
2281
2282 if (update_new && new_state != arc_l2c_only) {
2283 ASSERT(HDR_HAS_L1HDR(hdr));
2284 if (GHOST_STATE(new_state)) {
2285 ASSERT0(bufcnt);
2286
2287 /*
2288 * When moving a header to a ghost state, we first
2289 * remove all arc buffers. Thus, we'll have a
2290 * bufcnt of zero, and no arc buffer to use for
2291 * the reference. As a result, we use the arc
2292 * header pointer for the reference.
2293 */
2294 (void) zfs_refcount_add_many(&new_state->arcs_size,
2295 HDR_GET_LSIZE(hdr), hdr);
2296 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2297 ASSERT(!HDR_HAS_RABD(hdr));
2298 } else {
2299 uint32_t buffers = 0;
2300
2301 /*
2302 * Each individual buffer holds a unique reference,
2303 * thus we must remove each of these references one
2304 * at a time.
2305 */
2306 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2307 buf = buf->b_next) {
2308 ASSERT3U(bufcnt, !=, 0);
2309 buffers++;
2310
2311 /*
2312 * When the arc_buf_t is sharing the data
2313 * block with the hdr, the owner of the
2314 * reference belongs to the hdr. Only
2315 * add to the refcount if the arc_buf_t is
2316 * not shared.
2317 */
2318 if (arc_buf_is_shared(buf))
2319 continue;
2320
2321 (void) zfs_refcount_add_many(
2322 &new_state->arcs_size,
2323 arc_buf_size(buf), buf);
2324 }
2325 ASSERT3U(bufcnt, ==, buffers);
2326
2327 if (hdr->b_l1hdr.b_pabd != NULL) {
2328 (void) zfs_refcount_add_many(
2329 &new_state->arcs_size,
2330 arc_hdr_size(hdr), hdr);
2331 }
2332
2333 if (HDR_HAS_RABD(hdr)) {
2334 (void) zfs_refcount_add_many(
2335 &new_state->arcs_size,
2336 HDR_GET_PSIZE(hdr), hdr);
2337 }
2338 }
2339 }
2340
2341 if (update_old && old_state != arc_l2c_only) {
2342 ASSERT(HDR_HAS_L1HDR(hdr));
2343 if (GHOST_STATE(old_state)) {
2344 ASSERT0(bufcnt);
2345 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2346 ASSERT(!HDR_HAS_RABD(hdr));
2347
2348 /*
2349 * When moving a header off of a ghost state,
2350 * the header will not contain any arc buffers.
2351 * We use the arc header pointer for the reference
2352 * which is exactly what we did when we put the
2353 * header on the ghost state.
2354 */
2355
2356 (void) zfs_refcount_remove_many(&old_state->arcs_size,
2357 HDR_GET_LSIZE(hdr), hdr);
2358 } else {
2359 uint32_t buffers = 0;
2360
2361 /*
2362 * Each individual buffer holds a unique reference,
2363 * thus we must remove each of these references one
2364 * at a time.
2365 */
2366 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
2367 buf = buf->b_next) {
2368 ASSERT3U(bufcnt, !=, 0);
2369 buffers++;
2370
2371 /*
2372 * When the arc_buf_t is sharing the data
2373 * block with the hdr, the owner of the
2374 * reference belongs to the hdr. Only
2375 * add to the refcount if the arc_buf_t is
2376 * not shared.
2377 */
2378 if (arc_buf_is_shared(buf))
2379 continue;
2380
2381 (void) zfs_refcount_remove_many(
2382 &old_state->arcs_size, arc_buf_size(buf),
2383 buf);
2384 }
2385 ASSERT3U(bufcnt, ==, buffers);
2386 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
2387 HDR_HAS_RABD(hdr));
2388
2389 if (hdr->b_l1hdr.b_pabd != NULL) {
2390 (void) zfs_refcount_remove_many(
2391 &old_state->arcs_size, arc_hdr_size(hdr),
2392 hdr);
2393 }
2394
2395 if (HDR_HAS_RABD(hdr)) {
2396 (void) zfs_refcount_remove_many(
2397 &old_state->arcs_size, HDR_GET_PSIZE(hdr),
2398 hdr);
2399 }
2400 }
2401 }
2402
2403 if (HDR_HAS_L1HDR(hdr))
2404 hdr->b_l1hdr.b_state = new_state;
2405
2406 /*
2407 * L2 headers should never be on the L2 state list since they don't
2408 * have L1 headers allocated.
2409 */
2410 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2411 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2412 }
2413
2414 void
2415 arc_space_consume(uint64_t space, arc_space_type_t type)
2416 {
2417 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2418
2419 switch (type) {
2420 case ARC_SPACE_DATA:
2421 aggsum_add(&astat_data_size, space);
2422 break;
2423 case ARC_SPACE_META:
2424 aggsum_add(&astat_metadata_size, space);
2425 break;
2426 case ARC_SPACE_OTHER:
2427 aggsum_add(&astat_other_size, space);
2428 break;
2429 case ARC_SPACE_HDRS:
2430 aggsum_add(&astat_hdr_size, space);
2431 break;
2432 case ARC_SPACE_L2HDRS:
2433 aggsum_add(&astat_l2_hdr_size, space);
2434 break;
2435 }
2436
2437 if (type != ARC_SPACE_DATA)
2438 aggsum_add(&arc_meta_used, space);
2439
2440 aggsum_add(&arc_size, space);
2441 }
2442
2443 void
2444 arc_space_return(uint64_t space, arc_space_type_t type)
2445 {
2446 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2447
2448 switch (type) {
2449 case ARC_SPACE_DATA:
2450 aggsum_add(&astat_data_size, -space);
2451 break;
2452 case ARC_SPACE_META:
2453 aggsum_add(&astat_metadata_size, -space);
2454 break;
2455 case ARC_SPACE_OTHER:
2456 aggsum_add(&astat_other_size, -space);
2457 break;
2458 case ARC_SPACE_HDRS:
2459 aggsum_add(&astat_hdr_size, -space);
2460 break;
2461 case ARC_SPACE_L2HDRS:
2462 aggsum_add(&astat_l2_hdr_size, -space);
2463 break;
2464 }
2465
2466 if (type != ARC_SPACE_DATA) {
2467 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2468 /*
2469 * We use the upper bound here rather than the precise value
2470 * because the arc_meta_max value doesn't need to be
2471 * precise. It's only consumed by humans via arcstats.
2472 */
2473 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2474 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2475 aggsum_add(&arc_meta_used, -space);
2476 }
2477
2478 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2479 aggsum_add(&arc_size, -space);
2480 }
2481
2482 /*
2483 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2484 * with the hdr's b_pabd.
2485 */
2486 static boolean_t
2487 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2488 {
2489 /*
2490 * The criteria for sharing a hdr's data are:
2491 * 1. the buffer is not encrypted
2492 * 2. the hdr's compression matches the buf's compression
2493 * 3. the hdr doesn't need to be byteswapped
2494 * 4. the hdr isn't already being shared
2495 * 5. the buf is either compressed or it is the last buf in the hdr list
2496 *
2497 * Criterion #5 maintains the invariant that shared uncompressed
2498 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2499 * might ask, "if a compressed buf is allocated first, won't that be the
2500 * last thing in the list?", but in that case it's impossible to create
2501 * a shared uncompressed buf anyway (because the hdr must be compressed
2502 * to have the compressed buf). You might also think that #3 is
2503 * sufficient to make this guarantee, however it's possible
2504 * (specifically in the rare L2ARC write race mentioned in
2505 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that
2506 * is sharable, but wasn't at the time of its allocation. Rather than
2507 * allow a new shared uncompressed buf to be created and then shuffle
2508 * the list around to make it the last element, this simply disallows
2509 * sharing if the new buf isn't the first to be added.
2510 */
2511 ASSERT3P(buf->b_hdr, ==, hdr);
2512 boolean_t hdr_compressed = arc_hdr_get_compress(hdr) !=
2513 ZIO_COMPRESS_OFF;
2514 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0;
2515 return (!ARC_BUF_ENCRYPTED(buf) &&
2516 buf_compressed == hdr_compressed &&
2517 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2518 !HDR_SHARED_DATA(hdr) &&
2519 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2520 }
2521
2522 /*
2523 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2524 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2525 * copy was made successfully, or an error code otherwise.
2526 */
2527 static int
2528 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb,
2529 void *tag, boolean_t encrypted, boolean_t compressed, boolean_t noauth,
2530 boolean_t fill, arc_buf_t **ret)
2531 {
2532 arc_buf_t *buf;
2533 arc_fill_flags_t flags = ARC_FILL_LOCKED;
2534
2535 ASSERT(HDR_HAS_L1HDR(hdr));
2536 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2537 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2538 hdr->b_type == ARC_BUFC_METADATA);
2539 ASSERT3P(ret, !=, NULL);
2540 ASSERT3P(*ret, ==, NULL);
2541 IMPLY(encrypted, compressed);
2542
2543 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2544 buf->b_hdr = hdr;
2545 buf->b_data = NULL;
2546 buf->b_next = hdr->b_l1hdr.b_buf;
2547 buf->b_flags = 0;
2548
2549 add_reference(hdr, tag);
2550
2551 /*
2552 * We're about to change the hdr's b_flags. We must either
2553 * hold the hash_lock or be undiscoverable.
2554 */
2555 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2556
2557 /*
2558 * Only honor requests for compressed bufs if the hdr is actually
2559 * compressed. This must be overriden if the buffer is encrypted since
2560 * encrypted buffers cannot be decompressed.
2561 */
2562 if (encrypted) {
2563 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2564 buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED;
2565 flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED;
2566 } else if (compressed &&
2567 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
2568 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED;
2569 flags |= ARC_FILL_COMPRESSED;
2570 }
2571
2572 if (noauth) {
2573 ASSERT0(encrypted);
2574 flags |= ARC_FILL_NOAUTH;
2575 }
2576
2577 /*
2578 * If the hdr's data can be shared then we share the data buffer and
2579 * set the appropriate bit in the hdr's b_flags to indicate the hdr is
2580 * allocate a new buffer to store the buf's data.
2581 *
2582 * There are two additional restrictions here because we're sharing
2583 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be
2584 * actively involved in an L2ARC write, because if this buf is used by
2585 * an arc_write() then the hdr's data buffer will be released when the
2586 * write completes, even though the L2ARC write might still be using it.
2587 * Second, the hdr's ABD must be linear so that the buf's user doesn't
2588 * need to be ABD-aware.
2589 */
2590 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) &&
2591 hdr->b_l1hdr.b_pabd != NULL && abd_is_linear(hdr->b_l1hdr.b_pabd);
2592
2593 /* Set up b_data and sharing */
2594 if (can_share) {
2595 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd);
2596 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2597 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2598 } else {
2599 buf->b_data =
2600 arc_get_data_buf(hdr, arc_buf_size(buf), buf);
2601 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2602 }
2603 VERIFY3P(buf->b_data, !=, NULL);
2604
2605 hdr->b_l1hdr.b_buf = buf;
2606 hdr->b_l1hdr.b_bufcnt += 1;
2607 if (encrypted)
2608 hdr->b_crypt_hdr.b_ebufcnt += 1;
2609
2610 /*
2611 * If the user wants the data from the hdr, we need to either copy or
2612 * decompress the data.
2613 */
2614 if (fill) {
2615 ASSERT3P(zb, !=, NULL);
2616 return (arc_buf_fill(buf, spa, zb, flags));
2617 }
2618
2619 return (0);
2620 }
2621
2622 static char *arc_onloan_tag = "onloan";
2623
2624 static inline void
2625 arc_loaned_bytes_update(int64_t delta)
2626 {
2627 atomic_add_64(&arc_loaned_bytes, delta);
2628
2629 /* assert that it did not wrap around */
2630 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2631 }
2632
2633 /*
2634 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2635 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2636 * buffers must be returned to the arc before they can be used by the DMU or
2637 * freed.
2638 */
2639 arc_buf_t *
2640 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2641 {
2642 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2643 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2644
2645 arc_loaned_bytes_update(arc_buf_size(buf));
2646
2647 return (buf);
2648 }
2649
2650 arc_buf_t *
2651 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2652 enum zio_compress compression_type)
2653 {
2654 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag,
2655 psize, lsize, compression_type);
2656
2657 arc_loaned_bytes_update(arc_buf_size(buf));
2658
2659 return (buf);
2660 }
2661
2662 arc_buf_t *
2663 arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder,
2664 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
2665 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
2666 enum zio_compress compression_type)
2667 {
2668 arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj,
2669 byteorder, salt, iv, mac, ot, psize, lsize, compression_type);
2670
2671 atomic_add_64(&arc_loaned_bytes, psize);
2672 return (buf);
2673 }
2674
2675 /*
2676 * Performance tuning of L2ARC persistence:
2677 *
2678 * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding
2679 * an L2ARC device (either at pool import or later) will attempt
2680 * to rebuild L2ARC buffer contents.
2681 * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls
2682 * whether log blocks are written to the L2ARC device. If the L2ARC
2683 * device is less than 1GB, the amount of data l2arc_evict()
2684 * evicts is significant compared to the amount of restored L2ARC
2685 * data. In this case do not write log blocks in L2ARC in order
2686 * not to waste space.
2687 */
2688 int l2arc_rebuild_enabled = B_TRUE;
2689 unsigned long l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024;
2690
2691 /* L2ARC persistence rebuild control routines. */
2692 void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen);
2693 static void l2arc_dev_rebuild_start(l2arc_dev_t *dev);
2694 static int l2arc_rebuild(l2arc_dev_t *dev);
2695
2696 /* L2ARC persistence read I/O routines. */
2697 static int l2arc_dev_hdr_read(l2arc_dev_t *dev);
2698 static int l2arc_log_blk_read(l2arc_dev_t *dev,
2699 const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp,
2700 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
2701 zio_t *this_io, zio_t **next_io);
2702 static zio_t *l2arc_log_blk_fetch(vdev_t *vd,
2703 const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb);
2704 static void l2arc_log_blk_fetch_abort(zio_t *zio);
2705
2706 /* L2ARC persistence block restoration routines. */
2707 static void l2arc_log_blk_restore(l2arc_dev_t *dev,
2708 const l2arc_log_blk_phys_t *lb, uint64_t lb_asize, uint64_t lb_daddr);
2709 static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le,
2710 l2arc_dev_t *dev);
2711
2712 /* L2ARC persistence write I/O routines. */
2713 static void l2arc_dev_hdr_update(l2arc_dev_t *dev);
2714 static void l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
2715 l2arc_write_callback_t *cb);
2716
2717 /* L2ARC persistence auxilliary routines. */
2718 boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev,
2719 const l2arc_log_blkptr_t *lbp);
2720 static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev,
2721 const arc_buf_hdr_t *ab);
2722 boolean_t l2arc_range_check_overlap(uint64_t bottom,
2723 uint64_t top, uint64_t check);
2724 static void l2arc_blk_fetch_done(zio_t *zio);
2725 static inline uint64_t
2726 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev);
2727
2728 /*
2729 * Return a loaned arc buffer to the arc.
2730 */
2731 void
2732 arc_return_buf(arc_buf_t *buf, void *tag)
2733 {
2734 arc_buf_hdr_t *hdr = buf->b_hdr;
2735
2736 ASSERT3P(buf->b_data, !=, NULL);
2737 ASSERT(HDR_HAS_L1HDR(hdr));
2738 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
2739 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2740
2741 arc_loaned_bytes_update(-arc_buf_size(buf));
2742 }
2743
2744 /* Detach an arc_buf from a dbuf (tag) */
2745 void
2746 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
2747 {
2748 arc_buf_hdr_t *hdr = buf->b_hdr;
2749
2750 ASSERT3P(buf->b_data, !=, NULL);
2751 ASSERT(HDR_HAS_L1HDR(hdr));
2752 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
2753 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
2754
2755 arc_loaned_bytes_update(arc_buf_size(buf));
2756 }
2757
2758 static void
2759 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type)
2760 {
2761 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP);
2762
2763 df->l2df_abd = abd;
2764 df->l2df_size = size;
2765 df->l2df_type = type;
2766 mutex_enter(&l2arc_free_on_write_mtx);
2767 list_insert_head(l2arc_free_on_write, df);
2768 mutex_exit(&l2arc_free_on_write_mtx);
2769 }
2770
2771 static void
2772 arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata)
2773 {
2774 arc_state_t *state = hdr->b_l1hdr.b_state;
2775 arc_buf_contents_t type = arc_buf_type(hdr);
2776 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
2777
2778 /* protected by hash lock, if in the hash table */
2779 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2780 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2781 ASSERT(state != arc_anon && state != arc_l2c_only);
2782
2783 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
2784 size, hdr);
2785 }
2786 (void) zfs_refcount_remove_many(&state->arcs_size, size, hdr);
2787 if (type == ARC_BUFC_METADATA) {
2788 arc_space_return(size, ARC_SPACE_META);
2789 } else {
2790 ASSERT(type == ARC_BUFC_DATA);
2791 arc_space_return(size, ARC_SPACE_DATA);
2792 }
2793
2794 if (free_rdata) {
2795 l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type);
2796 } else {
2797 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2798 }
2799 }
2800
2801 /*
2802 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2803 * data buffer, we transfer the refcount ownership to the hdr and update
2804 * the appropriate kstats.
2805 */
2806 static void
2807 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2808 {
2809 /* LINTED */
2810 arc_state_t *state = hdr->b_l1hdr.b_state;
2811
2812 ASSERT(arc_can_share(hdr, buf));
2813 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2814 ASSERT(!ARC_BUF_ENCRYPTED(buf));
2815 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2816
2817 /*
2818 * Start sharing the data buffer. We transfer the
2819 * refcount ownership to the hdr since it always owns
2820 * the refcount whenever an arc_buf_t is shared.
2821 */
2822 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
2823 arc_hdr_size(hdr), buf, hdr);
2824 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2825 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2826 HDR_ISTYPE_METADATA(hdr));
2827 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2828 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2829
2830 /*
2831 * Since we've transferred ownership to the hdr we need
2832 * to increment its compressed and uncompressed kstats and
2833 * decrement the overhead size.
2834 */
2835 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2836 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2837 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2838 }
2839
2840 static void
2841 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2842 {
2843 /* LINTED */
2844 arc_state_t *state = hdr->b_l1hdr.b_state;
2845
2846 ASSERT(arc_buf_is_shared(buf));
2847 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2848 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2849
2850 /*
2851 * We are no longer sharing this buffer so we need
2852 * to transfer its ownership to the rightful owner.
2853 */
2854 zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size,
2855 arc_hdr_size(hdr), hdr, buf);
2856 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2857 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd);
2858 abd_put(hdr->b_l1hdr.b_pabd);
2859 hdr->b_l1hdr.b_pabd = NULL;
2860 buf->b_flags &= ~ARC_BUF_FLAG_SHARED;
2861
2862 /*
2863 * Since the buffer is no longer shared between
2864 * the arc buf and the hdr, count it as overhead.
2865 */
2866 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2867 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2868 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf));
2869 }
2870
2871 /*
2872 * Remove an arc_buf_t from the hdr's buf list and return the last
2873 * arc_buf_t on the list. If no buffers remain on the list then return
2874 * NULL.
2875 */
2876 static arc_buf_t *
2877 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2878 {
2879 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf;
2880 arc_buf_t *lastbuf = NULL;
2881
2882 ASSERT(HDR_HAS_L1HDR(hdr));
2883 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2884
2885 /*
2886 * Remove the buf from the hdr list and locate the last
2887 * remaining buffer on the list.
2888 */
2889 while (*bufp != NULL) {
2890 if (*bufp == buf)
2891 *bufp = buf->b_next;
2892
2893 /*
2894 * If we've removed a buffer in the middle of
2895 * the list then update the lastbuf and update
2896 * bufp.
2897 */
2898 if (*bufp != NULL) {
2899 lastbuf = *bufp;
2900 bufp = &(*bufp)->b_next;
2901 }
2902 }
2903 buf->b_next = NULL;
2904 ASSERT3P(lastbuf, !=, buf);
2905 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL);
2906 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL);
2907 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf));
2908
2909 return (lastbuf);
2910 }
2911
2912 /*
2913 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's
2914 * list and free it.
2915 */
2916 static void
2917 arc_buf_destroy_impl(arc_buf_t *buf)
2918 {
2919 arc_buf_hdr_t *hdr = buf->b_hdr;
2920
2921 /*
2922 * Free up the data associated with the buf but only if we're not
2923 * sharing this with the hdr. If we are sharing it with the hdr, the
2924 * hdr is responsible for doing the free.
2925 */
2926 if (buf->b_data != NULL) {
2927 /*
2928 * We're about to change the hdr's b_flags. We must either
2929 * hold the hash_lock or be undiscoverable.
2930 */
2931 ASSERT(HDR_EMPTY_OR_LOCKED(hdr));
2932
2933 arc_cksum_verify(buf);
2934 arc_buf_unwatch(buf);
2935
2936 if (arc_buf_is_shared(buf)) {
2937 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA);
2938 } else {
2939 uint64_t size = arc_buf_size(buf);
2940 arc_free_data_buf(hdr, buf->b_data, size, buf);
2941 ARCSTAT_INCR(arcstat_overhead_size, -size);
2942 }
2943 buf->b_data = NULL;
2944
2945 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
2946 hdr->b_l1hdr.b_bufcnt -= 1;
2947
2948 if (ARC_BUF_ENCRYPTED(buf)) {
2949 hdr->b_crypt_hdr.b_ebufcnt -= 1;
2950
2951 /*
2952 * If we have no more encrypted buffers and we've
2953 * already gotten a copy of the decrypted data we can
2954 * free b_rabd to save some space.
2955 */
2956 if (hdr->b_crypt_hdr.b_ebufcnt == 0 &&
2957 HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL &&
2958 !HDR_IO_IN_PROGRESS(hdr)) {
2959 arc_hdr_free_pabd(hdr, B_TRUE);
2960 }
2961 }
2962 }
2963
2964 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
2965
2966 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) {
2967 /*
2968 * If the current arc_buf_t is sharing its data buffer with the
2969 * hdr, then reassign the hdr's b_pabd to share it with the new
2970 * buffer at the end of the list. The shared buffer is always
2971 * the last one on the hdr's buffer list.
2972 *
2973 * There is an equivalent case for compressed bufs, but since
2974 * they aren't guaranteed to be the last buf in the list and
2975 * that is an exceedingly rare case, we just allow that space be
2976 * wasted temporarily. We must also be careful not to share
2977 * encrypted buffers, since they cannot be shared.
2978 */
2979 if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) {
2980 /* Only one buf can be shared at once */
2981 VERIFY(!arc_buf_is_shared(lastbuf));
2982 /* hdr is uncompressed so can't have compressed buf */
2983 VERIFY(!ARC_BUF_COMPRESSED(lastbuf));
2984
2985 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2986 arc_hdr_free_pabd(hdr, B_FALSE);
2987
2988 /*
2989 * We must setup a new shared block between the
2990 * last buffer and the hdr. The data would have
2991 * been allocated by the arc buf so we need to transfer
2992 * ownership to the hdr since it's now being shared.
2993 */
2994 arc_share_buf(hdr, lastbuf);
2995 }
2996 } else if (HDR_SHARED_DATA(hdr)) {
2997 /*
2998 * Uncompressed shared buffers are always at the end
2999 * of the list. Compressed buffers don't have the
3000 * same requirements. This makes it hard to
3001 * simply assert that the lastbuf is shared so
3002 * we rely on the hdr's compression flags to determine
3003 * if we have a compressed, shared buffer.
3004 */
3005 ASSERT3P(lastbuf, !=, NULL);
3006 ASSERT(arc_buf_is_shared(lastbuf) ||
3007 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
3008 }
3009
3010 /*
3011 * Free the checksum if we're removing the last uncompressed buf from
3012 * this hdr.
3013 */
3014 if (!arc_hdr_has_uncompressed_buf(hdr)) {
3015 arc_cksum_free(hdr);
3016 }
3017
3018 /* clean up the buf */
3019 buf->b_hdr = NULL;
3020 kmem_cache_free(buf_cache, buf);
3021 }
3022
3023 static void
3024 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr, boolean_t alloc_rdata)
3025 {
3026 uint64_t size;
3027
3028 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3029 ASSERT(HDR_HAS_L1HDR(hdr));
3030 ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata);
3031 IMPLY(alloc_rdata, HDR_PROTECTED(hdr));
3032
3033 if (alloc_rdata) {
3034 size = HDR_GET_PSIZE(hdr);
3035 ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL);
3036 hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr);
3037 ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL);
3038 } else {
3039 size = arc_hdr_size(hdr);
3040 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3041 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr);
3042 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3043 }
3044
3045 ARCSTAT_INCR(arcstat_compressed_size, size);
3046 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3047 }
3048
3049 static void
3050 arc_hdr_free_pabd(arc_buf_hdr_t *hdr, boolean_t free_rdata)
3051 {
3052 uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr);
3053
3054 ASSERT(HDR_HAS_L1HDR(hdr));
3055 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
3056 IMPLY(free_rdata, HDR_HAS_RABD(hdr));
3057
3058
3059 /*
3060 * If the hdr is currently being written to the l2arc then
3061 * we defer freeing the data by adding it to the l2arc_free_on_write
3062 * list. The l2arc will free the data once it's finished
3063 * writing it to the l2arc device.
3064 */
3065 if (HDR_L2_WRITING(hdr)) {
3066 arc_hdr_free_on_write(hdr, free_rdata);
3067 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3068 } else if (free_rdata) {
3069 arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr);
3070 } else {
3071 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3072 size, hdr);
3073 }
3074
3075 if (free_rdata) {
3076 hdr->b_crypt_hdr.b_rabd = NULL;
3077 } else {
3078 hdr->b_l1hdr.b_pabd = NULL;
3079 }
3080
3081 if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr))
3082 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3083
3084 ARCSTAT_INCR(arcstat_compressed_size, -size);
3085 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3086 }
3087
3088 static arc_buf_hdr_t *
3089 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3090 boolean_t protected, enum zio_compress compression_type,
3091 arc_buf_contents_t type, boolean_t alloc_rdata)
3092 {
3093 arc_buf_hdr_t *hdr;
3094
3095 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
3096 if (protected) {
3097 hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE);
3098 } else {
3099 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3100 }
3101 ASSERT(HDR_EMPTY(hdr));
3102 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3103 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3104 HDR_SET_PSIZE(hdr, psize);
3105 HDR_SET_LSIZE(hdr, lsize);
3106 hdr->b_spa = spa;
3107 hdr->b_type = type;
3108 hdr->b_flags = 0;
3109 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3110 arc_hdr_set_compress(hdr, compression_type);
3111 if (protected)
3112 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3113
3114 hdr->b_l1hdr.b_state = arc_anon;
3115 hdr->b_l1hdr.b_arc_access = 0;
3116 hdr->b_l1hdr.b_bufcnt = 0;
3117 hdr->b_l1hdr.b_buf = NULL;
3118
3119 /*
3120 * Allocate the hdr's buffer. This will contain either
3121 * the compressed or uncompressed data depending on the block
3122 * it references and compressed arc enablement.
3123 */
3124 arc_hdr_alloc_pabd(hdr, alloc_rdata);
3125 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3126
3127 return (hdr);
3128 }
3129
3130 /*
3131 * Transition between the two allocation states for the arc_buf_hdr struct.
3132 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
3133 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
3134 * version is used when a cache buffer is only in the L2ARC in order to reduce
3135 * memory usage.
3136 */
3137 static arc_buf_hdr_t *
3138 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
3139 {
3140 ASSERT(HDR_HAS_L2HDR(hdr));
3141
3142 arc_buf_hdr_t *nhdr;
3143 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3144
3145 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
3146 (old == hdr_l2only_cache && new == hdr_full_cache));
3147
3148 /*
3149 * if the caller wanted a new full header and the header is to be
3150 * encrypted we will actually allocate the header from the full crypt
3151 * cache instead. The same applies to freeing from the old cache.
3152 */
3153 if (HDR_PROTECTED(hdr) && new == hdr_full_cache)
3154 new = hdr_full_crypt_cache;
3155 if (HDR_PROTECTED(hdr) && old == hdr_full_cache)
3156 old = hdr_full_crypt_cache;
3157
3158 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
3159
3160 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3161 buf_hash_remove(hdr);
3162
3163 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3164
3165 if (new == hdr_full_cache || new == hdr_full_crypt_cache) {
3166 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3167 /*
3168 * arc_access and arc_change_state need to be aware that a
3169 * header has just come out of L2ARC, so we set its state to
3170 * l2c_only even though it's about to change.
3171 */
3172 nhdr->b_l1hdr.b_state = arc_l2c_only;
3173
3174 /* Verify previous threads set to NULL before freeing */
3175 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3176 ASSERT(!HDR_HAS_RABD(hdr));
3177 } else {
3178 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3179 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3180 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3181
3182 /*
3183 * If we've reached here, We must have been called from
3184 * arc_evict_hdr(), as such we should have already been
3185 * removed from any ghost list we were previously on
3186 * (which protects us from racing with arc_evict_state),
3187 * thus no locking is needed during this check.
3188 */
3189 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3190
3191 /*
3192 * A buffer must not be moved into the arc_l2c_only
3193 * state if it's not finished being written out to the
3194 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3195 * might try to be accessed, even though it was removed.
3196 */
3197 VERIFY(!HDR_L2_WRITING(hdr));
3198 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3199 ASSERT(!HDR_HAS_RABD(hdr));
3200
3201 #ifdef ZFS_DEBUG
3202 if (hdr->b_l1hdr.b_thawed != NULL) {
3203 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3204 hdr->b_l1hdr.b_thawed = NULL;
3205 }
3206 #endif
3207
3208 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3209 }
3210 /*
3211 * The header has been reallocated so we need to re-insert it into any
3212 * lists it was on.
3213 */
3214 (void) buf_hash_insert(nhdr, NULL);
3215
3216 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
3217
3218 mutex_enter(&dev->l2ad_mtx);
3219
3220 /*
3221 * We must place the realloc'ed header back into the list at
3222 * the same spot. Otherwise, if it's placed earlier in the list,
3223 * l2arc_write_buffers() could find it during the function's
3224 * write phase, and try to write it out to the l2arc.
3225 */
3226 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
3227 list_remove(&dev->l2ad_buflist, hdr);
3228
3229 mutex_exit(&dev->l2ad_mtx);
3230
3231 /*
3232 * Since we're using the pointer address as the tag when
3233 * incrementing and decrementing the l2ad_alloc refcount, we
3234 * must remove the old pointer (that we're about to destroy) and
3235 * add the new pointer to the refcount. Otherwise we'd remove
3236 * the wrong pointer address when calling arc_hdr_destroy() later.
3237 */
3238
3239 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3240 hdr);
3241 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr),
3242 nhdr);
3243
3244 buf_discard_identity(hdr);
3245 kmem_cache_free(old, hdr);
3246
3247 return (nhdr);
3248 }
3249
3250 /*
3251 * This function allows an L1 header to be reallocated as a crypt
3252 * header and vice versa. If we are going to a crypt header, the
3253 * new fields will be zeroed out.
3254 */
3255 static arc_buf_hdr_t *
3256 arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt)
3257 {
3258 arc_buf_hdr_t *nhdr;
3259 arc_buf_t *buf;
3260 kmem_cache_t *ncache, *ocache;
3261
3262 ASSERT(HDR_HAS_L1HDR(hdr));
3263 ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt);
3264 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3265 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3266 ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node));
3267 ASSERT3P(hdr->b_hash_next, ==, NULL);
3268
3269 if (need_crypt) {
3270 ncache = hdr_full_crypt_cache;
3271 ocache = hdr_full_cache;
3272 } else {
3273 ncache = hdr_full_cache;
3274 ocache = hdr_full_crypt_cache;
3275 }
3276
3277 nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE);
3278
3279 /*
3280 * Copy all members that aren't locks or condvars to the new header.
3281 * No lists are pointing to us (as we asserted above), so we don't
3282 * need to worry about the list nodes.
3283 */
3284 nhdr->b_dva = hdr->b_dva;
3285 nhdr->b_birth = hdr->b_birth;
3286 nhdr->b_type = hdr->b_type;
3287 nhdr->b_flags = hdr->b_flags;
3288 nhdr->b_psize = hdr->b_psize;
3289 nhdr->b_lsize = hdr->b_lsize;
3290 nhdr->b_spa = hdr->b_spa;
3291 nhdr->b_l2hdr.b_dev = hdr->b_l2hdr.b_dev;
3292 nhdr->b_l2hdr.b_daddr = hdr->b_l2hdr.b_daddr;
3293 nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum;
3294 nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt;
3295 nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap;
3296 nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state;
3297 nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access;
3298 nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb;
3299 nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd;
3300 #ifdef ZFS_DEBUG
3301 if (hdr->b_l1hdr.b_thawed != NULL) {
3302 nhdr->b_l1hdr.b_thawed = hdr->b_l1hdr.b_thawed;
3303 hdr->b_l1hdr.b_thawed = NULL;
3304 }
3305 #endif
3306
3307 /*
3308 * This refcount_add() exists only to ensure that the individual
3309 * arc buffers always point to a header that is referenced, avoiding
3310 * a small race condition that could trigger ASSERTs.
3311 */
3312 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG);
3313 nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf;
3314 for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) {
3315 mutex_enter(&buf->b_evict_lock);
3316 buf->b_hdr = nhdr;
3317 mutex_exit(&buf->b_evict_lock);
3318 }
3319 zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt);
3320 (void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG);
3321 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3322
3323 if (need_crypt) {
3324 arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED);
3325 } else {
3326 arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED);
3327 }
3328
3329 /* unset all members of the original hdr */
3330 bzero(&hdr->b_dva, sizeof (dva_t));
3331 hdr->b_birth = 0;
3332 hdr->b_type = ARC_BUFC_INVALID;
3333 hdr->b_flags = 0;
3334 hdr->b_psize = 0;
3335 hdr->b_lsize = 0;
3336 hdr->b_spa = 0;
3337 hdr->b_l2hdr.b_dev = NULL;
3338 hdr->b_l2hdr.b_daddr = 0;
3339 hdr->b_l1hdr.b_freeze_cksum = NULL;
3340 hdr->b_l1hdr.b_buf = NULL;
3341 hdr->b_l1hdr.b_bufcnt = 0;
3342 hdr->b_l1hdr.b_byteswap = 0;
3343 hdr->b_l1hdr.b_state = NULL;
3344 hdr->b_l1hdr.b_arc_access = 0;
3345 hdr->b_l1hdr.b_acb = NULL;
3346 hdr->b_l1hdr.b_pabd = NULL;
3347
3348 if (ocache == hdr_full_crypt_cache) {
3349 ASSERT(!HDR_HAS_RABD(hdr));
3350 hdr->b_crypt_hdr.b_ot = DMU_OT_NONE;
3351 hdr->b_crypt_hdr.b_ebufcnt = 0;
3352 hdr->b_crypt_hdr.b_dsobj = 0;
3353 bzero(hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3354 bzero(hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3355 bzero(hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3356 }
3357
3358 buf_discard_identity(hdr);
3359 kmem_cache_free(ocache, hdr);
3360
3361 return (nhdr);
3362 }
3363
3364 /*
3365 * This function is used by the send / receive code to convert a newly
3366 * allocated arc_buf_t to one that is suitable for a raw encrypted write. It
3367 * is also used to allow the root objset block to be uupdated without altering
3368 * its embedded MACs. Both block types will always be uncompressed so we do not
3369 * have to worry about compression type or psize.
3370 */
3371 void
3372 arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder,
3373 dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv,
3374 const uint8_t *mac)
3375 {
3376 arc_buf_hdr_t *hdr = buf->b_hdr;
3377
3378 ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET);
3379 ASSERT(HDR_HAS_L1HDR(hdr));
3380 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3381
3382 buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED);
3383 if (!HDR_PROTECTED(hdr))
3384 hdr = arc_hdr_realloc_crypt(hdr, B_TRUE);
3385 hdr->b_crypt_hdr.b_dsobj = dsobj;
3386 hdr->b_crypt_hdr.b_ot = ot;
3387 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3388 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3389 if (!arc_hdr_has_uncompressed_buf(hdr))
3390 arc_cksum_free(hdr);
3391
3392 if (salt != NULL)
3393 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3394 if (iv != NULL)
3395 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3396 if (mac != NULL)
3397 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3398 }
3399
3400 /*
3401 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller.
3402 * The buf is returned thawed since we expect the consumer to modify it.
3403 */
3404 arc_buf_t *
3405 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size)
3406 {
3407 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size,
3408 B_FALSE, ZIO_COMPRESS_OFF, type, B_FALSE);
3409
3410 arc_buf_t *buf = NULL;
3411 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE,
3412 B_FALSE, B_FALSE, &buf));
3413 arc_buf_thaw(buf);
3414
3415 return (buf);
3416 }
3417
3418 /*
3419 * Allocates an ARC buf header that's in an evicted & L2-cached state.
3420 * This is used during l2arc reconstruction to make empty ARC buffers
3421 * which circumvent the regular disk->arc->l2arc path and instead come
3422 * into being in the reverse order, i.e. l2arc->arc.
3423 */
3424 arc_buf_hdr_t *
3425 arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev,
3426 dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth,
3427 enum zio_compress compress, boolean_t protected, boolean_t prefetch)
3428 {
3429 arc_buf_hdr_t *hdr;
3430
3431 ASSERT(size != 0);
3432 hdr = kmem_cache_alloc(hdr_l2only_cache, KM_SLEEP);
3433 hdr->b_birth = birth;
3434 hdr->b_type = type;
3435 hdr->b_flags = 0;
3436 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR);
3437 HDR_SET_LSIZE(hdr, size);
3438 HDR_SET_PSIZE(hdr, psize);
3439 arc_hdr_set_compress(hdr, compress);
3440 if (protected)
3441 arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED);
3442 if (prefetch)
3443 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
3444 hdr->b_spa = spa_load_guid(dev->l2ad_vdev->vdev_spa);
3445
3446 hdr->b_dva = dva;
3447
3448 hdr->b_l2hdr.b_dev = dev;
3449 hdr->b_l2hdr.b_daddr = daddr;
3450
3451 return (hdr);
3452 }
3453
3454 /*
3455 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3456 * for bufs containing metadata.
3457 */
3458 arc_buf_t *
3459 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3460 enum zio_compress compression_type)
3461 {
3462 ASSERT3U(lsize, >, 0);
3463 ASSERT3U(lsize, >=, psize);
3464 ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF);
3465 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3466
3467 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3468 B_FALSE, compression_type, ARC_BUFC_DATA, B_FALSE);
3469
3470 arc_buf_t *buf = NULL;
3471 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE,
3472 B_TRUE, B_FALSE, B_FALSE, &buf));
3473 arc_buf_thaw(buf);
3474 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3475
3476 if (!arc_buf_is_shared(buf)) {
3477 /*
3478 * To ensure that the hdr has the correct data in it if we call
3479 * arc_untransform() on this buf before it's been written to
3480 * disk, it's easiest if we just set up sharing between the
3481 * buf and the hdr.
3482 */
3483 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3484 arc_hdr_free_pabd(hdr, B_FALSE);
3485 arc_share_buf(hdr, buf);
3486 }
3487
3488 return (buf);
3489 }
3490
3491 arc_buf_t *
3492 arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder,
3493 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac,
3494 dmu_object_type_t ot, uint64_t psize, uint64_t lsize,
3495 enum zio_compress compression_type)
3496 {
3497 arc_buf_hdr_t *hdr;
3498 arc_buf_t *buf;
3499 arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ?
3500 ARC_BUFC_METADATA : ARC_BUFC_DATA;
3501
3502 ASSERT3U(lsize, >, 0);
3503 ASSERT3U(lsize, >=, psize);
3504 ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF);
3505 ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS);
3506
3507 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE,
3508 compression_type, type, B_TRUE);
3509
3510 hdr->b_crypt_hdr.b_dsobj = dsobj;
3511 hdr->b_crypt_hdr.b_ot = ot;
3512 hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ?
3513 DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot);
3514 bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN);
3515 bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN);
3516 bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN);
3517
3518 /*
3519 * This buffer will be considered encrypted even if the ot is not an
3520 * encrypted type. It will become authenticated instead in
3521 * arc_write_ready().
3522 */
3523 buf = NULL;
3524 VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE,
3525 B_FALSE, B_FALSE, &buf));
3526 arc_buf_thaw(buf);
3527 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3528
3529 return (buf);
3530 }
3531
3532 static void
3533 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3534 {
3535 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3536 l2arc_dev_t *dev = l2hdr->b_dev;
3537 uint64_t psize = HDR_GET_PSIZE(hdr);
3538 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
3539
3540 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3541 ASSERT(HDR_HAS_L2HDR(hdr));
3542
3543 list_remove(&dev->l2ad_buflist, hdr);
3544
3545 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3546 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3547
3548 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
3549
3550 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr),
3551 hdr);
3552 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3553 }
3554
3555 static void
3556 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3557 {
3558 if (HDR_HAS_L1HDR(hdr)) {
3559 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3560 hdr->b_l1hdr.b_bufcnt > 0);
3561 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3562 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3563 }
3564 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3565 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3566
3567 if (HDR_HAS_L2HDR(hdr)) {
3568 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3569 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3570
3571 if (!buflist_held)
3572 mutex_enter(&dev->l2ad_mtx);
3573
3574 /*
3575 * Even though we checked this conditional above, we
3576 * need to check this again now that we have the
3577 * l2ad_mtx. This is because we could be racing with
3578 * another thread calling l2arc_evict() which might have
3579 * destroyed this header's L2 portion as we were waiting
3580 * to acquire the l2ad_mtx. If that happens, we don't
3581 * want to re-destroy the header's L2 portion.
3582 */
3583 if (HDR_HAS_L2HDR(hdr))
3584 arc_hdr_l2hdr_destroy(hdr);
3585
3586 if (!buflist_held)
3587 mutex_exit(&dev->l2ad_mtx);
3588 }
3589
3590 /*
3591 * The header's identity can only be safely discarded once it is no
3592 * longer discoverable. This requires removing it from the hash table
3593 * and the l2arc header list. After this point the hash lock can not
3594 * be used to protect the header.
3595 */
3596 if (!HDR_EMPTY(hdr))
3597 buf_discard_identity(hdr);
3598
3599 if (HDR_HAS_L1HDR(hdr)) {
3600 arc_cksum_free(hdr);
3601
3602 while (hdr->b_l1hdr.b_buf != NULL)
3603 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3604
3605 #ifdef ZFS_DEBUG
3606 if (hdr->b_l1hdr.b_thawed != NULL) {
3607 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3608 hdr->b_l1hdr.b_thawed = NULL;
3609 }
3610 #endif
3611
3612 if (hdr->b_l1hdr.b_pabd != NULL)
3613 arc_hdr_free_pabd(hdr, B_FALSE);
3614
3615 if (HDR_HAS_RABD(hdr))
3616 arc_hdr_free_pabd(hdr, B_TRUE);
3617 }
3618
3619 ASSERT3P(hdr->b_hash_next, ==, NULL);
3620 if (HDR_HAS_L1HDR(hdr)) {
3621 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3622 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
3623
3624 if (!HDR_PROTECTED(hdr)) {
3625 kmem_cache_free(hdr_full_cache, hdr);
3626 } else {
3627 kmem_cache_free(hdr_full_crypt_cache, hdr);
3628 }
3629 } else {
3630 kmem_cache_free(hdr_l2only_cache, hdr);
3631 }
3632 }
3633
3634 void
3635 arc_buf_destroy(arc_buf_t *buf, void* tag)
3636 {
3637 arc_buf_hdr_t *hdr = buf->b_hdr;
3638
3639 if (hdr->b_l1hdr.b_state == arc_anon) {
3640 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
3641 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3642 VERIFY0(remove_reference(hdr, NULL, tag));
3643 arc_hdr_destroy(hdr);
3644 return;
3645 }
3646
3647 kmutex_t *hash_lock = HDR_LOCK(hdr);
3648 mutex_enter(hash_lock);
3649
3650 ASSERT3P(hdr, ==, buf->b_hdr);
3651 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
3652 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
3653 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
3654 ASSERT3P(buf->b_data, !=, NULL);
3655
3656 (void) remove_reference(hdr, hash_lock, tag);
3657 arc_buf_destroy_impl(buf);
3658 mutex_exit(hash_lock);
3659 }
3660
3661 /*
3662 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3663 * state of the header is dependent on its state prior to entering this
3664 * function. The following transitions are possible:
3665 *
3666 * - arc_mru -> arc_mru_ghost
3667 * - arc_mfu -> arc_mfu_ghost
3668 * - arc_mru_ghost -> arc_l2c_only
3669 * - arc_mru_ghost -> deleted
3670 * - arc_mfu_ghost -> arc_l2c_only
3671 * - arc_mfu_ghost -> deleted
3672 */
3673 static int64_t
3674 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3675 {
3676 arc_state_t *evicted_state, *state;
3677 int64_t bytes_evicted = 0;
3678 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ?
3679 zfs_arc_min_prescient_prefetch_ms : zfs_arc_min_prefetch_ms;
3680
3681 ASSERT(MUTEX_HELD(hash_lock));
3682 ASSERT(HDR_HAS_L1HDR(hdr));
3683
3684 state = hdr->b_l1hdr.b_state;
3685 if (GHOST_STATE(state)) {
3686 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3687 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3688
3689 /*
3690 * l2arc_write_buffers() relies on a header's L1 portion
3691 * (i.e. its b_pabd field) during its write phase.
3692 * Thus, we cannot push a header onto the arc_l2c_only
3693 * state (removing its L1 piece) until the header is
3694 * done being written to the l2arc.
3695 */
3696 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3697 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3698 return (bytes_evicted);
3699 }
3700
3701 ARCSTAT_BUMP(arcstat_deleted);
3702 bytes_evicted += HDR_GET_LSIZE(hdr);
3703
3704 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
3705
3706 if (HDR_HAS_L2HDR(hdr)) {
3707 ASSERT(hdr->b_l1hdr.b_pabd == NULL);
3708 ASSERT(!HDR_HAS_RABD(hdr));
3709 /*
3710 * This buffer is cached on the 2nd Level ARC;
3711 * don't destroy the header.
3712 */
3713 arc_change_state(arc_l2c_only, hdr, hash_lock);
3714 /*
3715 * dropping from L1+L2 cached to L2-only,
3716 * realloc to remove the L1 header.
3717 */
3718 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
3719 hdr_l2only_cache);
3720 } else {
3721 arc_change_state(arc_anon, hdr, hash_lock);
3722 arc_hdr_destroy(hdr);
3723 }
3724 return (bytes_evicted);
3725 }
3726
3727 ASSERT(state == arc_mru || state == arc_mfu);
3728 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3729
3730 /* prefetch buffers have a minimum lifespan */
3731 if (HDR_IO_IN_PROGRESS(hdr) ||
3732 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
3733 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < min_lifetime * hz)) {
3734 ARCSTAT_BUMP(arcstat_evict_skip);
3735 return (bytes_evicted);
3736 }
3737
3738 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt));
3739 while (hdr->b_l1hdr.b_buf) {
3740 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
3741 if (!mutex_tryenter(&buf->b_evict_lock)) {
3742 ARCSTAT_BUMP(arcstat_mutex_miss);
3743 break;
3744 }
3745 if (buf->b_data != NULL)
3746 bytes_evicted += HDR_GET_LSIZE(hdr);
3747 mutex_exit(&buf->b_evict_lock);
3748 arc_buf_destroy_impl(buf);
3749 }
3750
3751 if (HDR_HAS_L2HDR(hdr)) {
3752 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr));
3753 } else {
3754 if (l2arc_write_eligible(hdr->b_spa, hdr)) {
3755 ARCSTAT_INCR(arcstat_evict_l2_eligible,
3756 HDR_GET_LSIZE(hdr));
3757 } else {
3758 ARCSTAT_INCR(arcstat_evict_l2_ineligible,
3759 HDR_GET_LSIZE(hdr));
3760 }
3761 }
3762
3763 if (hdr->b_l1hdr.b_bufcnt == 0) {
3764 arc_cksum_free(hdr);
3765
3766 bytes_evicted += arc_hdr_size(hdr);
3767
3768 /*
3769 * If this hdr is being evicted and has a compressed
3770 * buffer then we discard it here before we change states.
3771 * This ensures that the accounting is updated correctly
3772 * in arc_free_data_impl().
3773 */
3774 if (hdr->b_l1hdr.b_pabd != NULL)
3775 arc_hdr_free_pabd(hdr, B_FALSE);
3776
3777 if (HDR_HAS_RABD(hdr))
3778 arc_hdr_free_pabd(hdr, B_TRUE);
3779
3780 arc_change_state(evicted_state, hdr, hash_lock);
3781 ASSERT(HDR_IN_HASH_TABLE(hdr));
3782 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
3783 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
3784 }
3785
3786 return (bytes_evicted);
3787 }
3788
3789 static uint64_t
3790 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
3791 uint64_t spa, int64_t bytes)
3792 {
3793 multilist_sublist_t *mls;
3794 uint64_t bytes_evicted = 0;
3795 arc_buf_hdr_t *hdr;
3796 kmutex_t *hash_lock;
3797 int evict_count = 0;
3798
3799 ASSERT3P(marker, !=, NULL);
3800 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3801
3802 mls = multilist_sublist_lock(ml, idx);
3803
3804 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
3805 hdr = multilist_sublist_prev(mls, marker)) {
3806 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
3807 (evict_count >= zfs_arc_evict_batch_limit))
3808 break;
3809
3810 /*
3811 * To keep our iteration location, move the marker
3812 * forward. Since we're not holding hdr's hash lock, we
3813 * must be very careful and not remove 'hdr' from the
3814 * sublist. Otherwise, other consumers might mistake the
3815 * 'hdr' as not being on a sublist when they call the
3816 * multilist_link_active() function (they all rely on
3817 * the hash lock protecting concurrent insertions and
3818 * removals). multilist_sublist_move_forward() was
3819 * specifically implemented to ensure this is the case
3820 * (only 'marker' will be removed and re-inserted).
3821 */
3822 multilist_sublist_move_forward(mls, marker);
3823
3824 /*
3825 * The only case where the b_spa field should ever be
3826 * zero, is the marker headers inserted by
3827 * arc_evict_state(). It's possible for multiple threads
3828 * to be calling arc_evict_state() concurrently (e.g.
3829 * dsl_pool_close() and zio_inject_fault()), so we must
3830 * skip any markers we see from these other threads.
3831 */
3832 if (hdr->b_spa == 0)
3833 continue;
3834
3835 /* we're only interested in evicting buffers of a certain spa */
3836 if (spa != 0 && hdr->b_spa != spa) {
3837 ARCSTAT_BUMP(arcstat_evict_skip);
3838 continue;
3839 }
3840
3841 hash_lock = HDR_LOCK(hdr);
3842
3843 /*
3844 * We aren't calling this function from any code path
3845 * that would already be holding a hash lock, so we're
3846 * asserting on this assumption to be defensive in case
3847 * this ever changes. Without this check, it would be
3848 * possible to incorrectly increment arcstat_mutex_miss
3849 * below (e.g. if the code changed such that we called
3850 * this function with a hash lock held).
3851 */
3852 ASSERT(!MUTEX_HELD(hash_lock));
3853
3854 if (mutex_tryenter(hash_lock)) {
3855 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
3856 mutex_exit(hash_lock);
3857
3858 bytes_evicted += evicted;
3859
3860 /*
3861 * If evicted is zero, arc_evict_hdr() must have
3862 * decided to skip this header, don't increment
3863 * evict_count in this case.
3864 */
3865 if (evicted != 0)
3866 evict_count++;
3867
3868 /*
3869 * If arc_size isn't overflowing, signal any
3870 * threads that might happen to be waiting.
3871 *
3872 * For each header evicted, we wake up a single
3873 * thread. If we used cv_broadcast, we could
3874 * wake up "too many" threads causing arc_size
3875 * to significantly overflow arc_c; since
3876 * arc_get_data_impl() doesn't check for overflow
3877 * when it's woken up (it doesn't because it's
3878 * possible for the ARC to be overflowing while
3879 * full of un-evictable buffers, and the
3880 * function should proceed in this case).
3881 *
3882 * If threads are left sleeping, due to not
3883 * using cv_broadcast here, they will be woken
3884 * up via cv_broadcast in arc_adjust_cb() just
3885 * before arc_adjust_zthr sleeps.
3886 */
3887 mutex_enter(&arc_adjust_lock);
3888 if (!arc_is_overflowing())
3889 cv_signal(&arc_adjust_waiters_cv);
3890 mutex_exit(&arc_adjust_lock);
3891 } else {
3892 ARCSTAT_BUMP(arcstat_mutex_miss);
3893 }
3894 }
3895
3896 multilist_sublist_unlock(mls);
3897
3898 return (bytes_evicted);
3899 }
3900
3901 /*
3902 * Evict buffers from the given arc state, until we've removed the
3903 * specified number of bytes. Move the removed buffers to the
3904 * appropriate evict state.
3905 *
3906 * This function makes a "best effort". It skips over any buffers
3907 * it can't get a hash_lock on, and so, may not catch all candidates.
3908 * It may also return without evicting as much space as requested.
3909 *
3910 * If bytes is specified using the special value ARC_EVICT_ALL, this
3911 * will evict all available (i.e. unlocked and evictable) buffers from
3912 * the given arc state; which is used by arc_flush().
3913 */
3914 static uint64_t
3915 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
3916 arc_buf_contents_t type)
3917 {
3918 uint64_t total_evicted = 0;
3919 multilist_t *ml = state->arcs_list[type];
3920 int num_sublists;
3921 arc_buf_hdr_t **markers;
3922
3923 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
3924
3925 num_sublists = multilist_get_num_sublists(ml);
3926
3927 /*
3928 * If we've tried to evict from each sublist, made some
3929 * progress, but still have not hit the target number of bytes
3930 * to evict, we want to keep trying. The markers allow us to
3931 * pick up where we left off for each individual sublist, rather
3932 * than starting from the tail each time.
3933 */
3934 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
3935 for (int i = 0; i < num_sublists; i++) {
3936 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
3937
3938 /*
3939 * A b_spa of 0 is used to indicate that this header is
3940 * a marker. This fact is used in arc_adjust_type() and
3941 * arc_evict_state_impl().
3942 */
3943 markers[i]->b_spa = 0;
3944
3945 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
3946 multilist_sublist_insert_tail(mls, markers[i]);
3947 multilist_sublist_unlock(mls);
3948 }
3949
3950 /*
3951 * While we haven't hit our target number of bytes to evict, or
3952 * we're evicting all available buffers.
3953 */
3954 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
3955 /*
3956 * Start eviction using a randomly selected sublist,
3957 * this is to try and evenly balance eviction across all
3958 * sublists. Always starting at the same sublist
3959 * (e.g. index 0) would cause evictions to favor certain
3960 * sublists over others.
3961 */
3962 int sublist_idx = multilist_get_random_index(ml);
3963 uint64_t scan_evicted = 0;
3964
3965 for (int i = 0; i < num_sublists; i++) {
3966 uint64_t bytes_remaining;
3967 uint64_t bytes_evicted;
3968
3969 if (bytes == ARC_EVICT_ALL)
3970 bytes_remaining = ARC_EVICT_ALL;
3971 else if (total_evicted < bytes)
3972 bytes_remaining = bytes - total_evicted;
3973 else
3974 break;
3975
3976 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
3977 markers[sublist_idx], spa, bytes_remaining);
3978
3979 scan_evicted += bytes_evicted;
3980 total_evicted += bytes_evicted;
3981
3982 /* we've reached the end, wrap to the beginning */
3983 if (++sublist_idx >= num_sublists)
3984 sublist_idx = 0;
3985 }
3986
3987 /*
3988 * If we didn't evict anything during this scan, we have
3989 * no reason to believe we'll evict more during another
3990 * scan, so break the loop.
3991 */
3992 if (scan_evicted == 0) {
3993 /* This isn't possible, let's make that obvious */
3994 ASSERT3S(bytes, !=, 0);
3995
3996 /*
3997 * When bytes is ARC_EVICT_ALL, the only way to
3998 * break the loop is when scan_evicted is zero.
3999 * In that case, we actually have evicted enough,
4000 * so we don't want to increment the kstat.
4001 */
4002 if (bytes != ARC_EVICT_ALL) {
4003 ASSERT3S(total_evicted, <, bytes);
4004 ARCSTAT_BUMP(arcstat_evict_not_enough);
4005 }
4006
4007 break;
4008 }
4009 }
4010
4011 for (int i = 0; i < num_sublists; i++) {
4012 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
4013 multilist_sublist_remove(mls, markers[i]);
4014 multilist_sublist_unlock(mls);
4015
4016 kmem_cache_free(hdr_full_cache, markers[i]);
4017 }
4018 kmem_free(markers, sizeof (*markers) * num_sublists);
4019
4020 return (total_evicted);
4021 }
4022
4023 /*
4024 * Flush all "evictable" data of the given type from the arc state
4025 * specified. This will not evict any "active" buffers (i.e. referenced).
4026 *
4027 * When 'retry' is set to B_FALSE, the function will make a single pass
4028 * over the state and evict any buffers that it can. Since it doesn't
4029 * continually retry the eviction, it might end up leaving some buffers
4030 * in the ARC due to lock misses.
4031 *
4032 * When 'retry' is set to B_TRUE, the function will continually retry the
4033 * eviction until *all* evictable buffers have been removed from the
4034 * state. As a result, if concurrent insertions into the state are
4035 * allowed (e.g. if the ARC isn't shutting down), this function might
4036 * wind up in an infinite loop, continually trying to evict buffers.
4037 */
4038 static uint64_t
4039 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
4040 boolean_t retry)
4041 {
4042 uint64_t evicted = 0;
4043
4044 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) {
4045 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
4046
4047 if (!retry)
4048 break;
4049 }
4050
4051 return (evicted);
4052 }
4053
4054 /*
4055 * Evict the specified number of bytes from the state specified,
4056 * restricting eviction to the spa and type given. This function
4057 * prevents us from trying to evict more from a state's list than
4058 * is "evictable", and to skip evicting altogether when passed a
4059 * negative value for "bytes". In contrast, arc_evict_state() will
4060 * evict everything it can, when passed a negative value for "bytes".
4061 */
4062 static uint64_t
4063 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4064 arc_buf_contents_t type)
4065 {
4066 int64_t delta;
4067
4068 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) {
4069 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]),
4070 bytes);
4071 return (arc_evict_state(state, spa, delta, type));
4072 }
4073
4074 return (0);
4075 }
4076
4077 /*
4078 * Evict metadata buffers from the cache, such that arc_meta_used is
4079 * capped by the arc_meta_limit tunable.
4080 */
4081 static uint64_t
4082 arc_adjust_meta(uint64_t meta_used)
4083 {
4084 uint64_t total_evicted = 0;
4085 int64_t target;
4086
4087 /*
4088 * If we're over the meta limit, we want to evict enough
4089 * metadata to get back under the meta limit. We don't want to
4090 * evict so much that we drop the MRU below arc_p, though. If
4091 * we're over the meta limit more than we're over arc_p, we
4092 * evict some from the MRU here, and some from the MFU below.
4093 */
4094 target = MIN((int64_t)(meta_used - arc_meta_limit),
4095 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4096 zfs_refcount_count(&arc_mru->arcs_size) - arc_p));
4097
4098 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4099
4100 /*
4101 * Similar to the above, we want to evict enough bytes to get us
4102 * below the meta limit, but not so much as to drop us below the
4103 * space allotted to the MFU (which is defined as arc_c - arc_p).
4104 */
4105 target = MIN((int64_t)(meta_used - arc_meta_limit),
4106 (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) -
4107 (arc_c - arc_p)));
4108
4109 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4110
4111 return (total_evicted);
4112 }
4113
4114 /*
4115 * Return the type of the oldest buffer in the given arc state
4116 *
4117 * This function will select a random sublist of type ARC_BUFC_DATA and
4118 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
4119 * is compared, and the type which contains the "older" buffer will be
4120 * returned.
4121 */
4122 static arc_buf_contents_t
4123 arc_adjust_type(arc_state_t *state)
4124 {
4125 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4126 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4127 int data_idx = multilist_get_random_index(data_ml);
4128 int meta_idx = multilist_get_random_index(meta_ml);
4129 multilist_sublist_t *data_mls;
4130 multilist_sublist_t *meta_mls;
4131 arc_buf_contents_t type;
4132 arc_buf_hdr_t *data_hdr;
4133 arc_buf_hdr_t *meta_hdr;
4134
4135 /*
4136 * We keep the sublist lock until we're finished, to prevent
4137 * the headers from being destroyed via arc_evict_state().
4138 */
4139 data_mls = multilist_sublist_lock(data_ml, data_idx);
4140 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4141
4142 /*
4143 * These two loops are to ensure we skip any markers that
4144 * might be at the tail of the lists due to arc_evict_state().
4145 */
4146
4147 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4148 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4149 if (data_hdr->b_spa != 0)
4150 break;
4151 }
4152
4153 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4154 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4155 if (meta_hdr->b_spa != 0)
4156 break;
4157 }
4158
4159 if (data_hdr == NULL && meta_hdr == NULL) {
4160 type = ARC_BUFC_DATA;
4161 } else if (data_hdr == NULL) {
4162 ASSERT3P(meta_hdr, !=, NULL);
4163 type = ARC_BUFC_METADATA;
4164 } else if (meta_hdr == NULL) {
4165 ASSERT3P(data_hdr, !=, NULL);
4166 type = ARC_BUFC_DATA;
4167 } else {
4168 ASSERT3P(data_hdr, !=, NULL);
4169 ASSERT3P(meta_hdr, !=, NULL);
4170
4171 /* The headers can't be on the sublist without an L1 header */
4172 ASSERT(HDR_HAS_L1HDR(data_hdr));
4173 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4174
4175 if (data_hdr->b_l1hdr.b_arc_access <
4176 meta_hdr->b_l1hdr.b_arc_access) {
4177 type = ARC_BUFC_DATA;
4178 } else {
4179 type = ARC_BUFC_METADATA;
4180 }
4181 }
4182
4183 multilist_sublist_unlock(meta_mls);
4184 multilist_sublist_unlock(data_mls);
4185
4186 return (type);
4187 }
4188
4189 /*
4190 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4191 */
4192 static uint64_t
4193 arc_adjust(void)
4194 {
4195 uint64_t total_evicted = 0;
4196 uint64_t bytes;
4197 int64_t target;
4198 uint64_t asize = aggsum_value(&arc_size);
4199 uint64_t ameta = aggsum_value(&arc_meta_used);
4200
4201 /*
4202 * If we're over arc_meta_limit, we want to correct that before
4203 * potentially evicting data buffers below.
4204 */
4205 total_evicted += arc_adjust_meta(ameta);
4206
4207 /*
4208 * Adjust MRU size
4209 *
4210 * If we're over the target cache size, we want to evict enough
4211 * from the list to get back to our target size. We don't want
4212 * to evict too much from the MRU, such that it drops below
4213 * arc_p. So, if we're over our target cache size more than
4214 * the MRU is over arc_p, we'll evict enough to get back to
4215 * arc_p here, and then evict more from the MFU below.
4216 */
4217 target = MIN((int64_t)(asize - arc_c),
4218 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) +
4219 zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
4220
4221 /*
4222 * If we're below arc_meta_min, always prefer to evict data.
4223 * Otherwise, try to satisfy the requested number of bytes to
4224 * evict from the type which contains older buffers; in an
4225 * effort to keep newer buffers in the cache regardless of their
4226 * type. If we cannot satisfy the number of bytes from this
4227 * type, spill over into the next type.
4228 */
4229 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4230 ameta > arc_meta_min) {
4231 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4232 total_evicted += bytes;
4233
4234 /*
4235 * If we couldn't evict our target number of bytes from
4236 * metadata, we try to get the rest from data.
4237 */
4238 target -= bytes;
4239
4240 total_evicted +=
4241 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4242 } else {
4243 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4244 total_evicted += bytes;
4245
4246 /*
4247 * If we couldn't evict our target number of bytes from
4248 * data, we try to get the rest from metadata.
4249 */
4250 target -= bytes;
4251
4252 total_evicted +=
4253 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4254 }
4255
4256 /*
4257 * Adjust MFU size
4258 *
4259 * Now that we've tried to evict enough from the MRU to get its
4260 * size back to arc_p, if we're still above the target cache
4261 * size, we evict the rest from the MFU.
4262 */
4263 target = asize - arc_c;
4264
4265 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4266 ameta > arc_meta_min) {
4267 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4268 total_evicted += bytes;
4269
4270 /*
4271 * If we couldn't evict our target number of bytes from
4272 * metadata, we try to get the rest from data.
4273 */
4274 target -= bytes;
4275
4276 total_evicted +=
4277 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4278 } else {
4279 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4280 total_evicted += bytes;
4281
4282 /*
4283 * If we couldn't evict our target number of bytes from
4284 * data, we try to get the rest from data.
4285 */
4286 target -= bytes;
4287
4288 total_evicted +=
4289 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4290 }
4291
4292 /*
4293 * Adjust ghost lists
4294 *
4295 * In addition to the above, the ARC also defines target values
4296 * for the ghost lists. The sum of the mru list and mru ghost
4297 * list should never exceed the target size of the cache, and
4298 * the sum of the mru list, mfu list, mru ghost list, and mfu
4299 * ghost list should never exceed twice the target size of the
4300 * cache. The following logic enforces these limits on the ghost
4301 * caches, and evicts from them as needed.
4302 */
4303 target = zfs_refcount_count(&arc_mru->arcs_size) +
4304 zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4305
4306 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4307 total_evicted += bytes;
4308
4309 target -= bytes;
4310
4311 total_evicted +=
4312 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4313
4314 /*
4315 * We assume the sum of the mru list and mfu list is less than
4316 * or equal to arc_c (we enforced this above), which means we
4317 * can use the simpler of the two equations below:
4318 *
4319 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4320 * mru ghost + mfu ghost <= arc_c
4321 */
4322 target = zfs_refcount_count(&arc_mru_ghost->arcs_size) +
4323 zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4324
4325 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4326 total_evicted += bytes;
4327
4328 target -= bytes;
4329
4330 total_evicted +=
4331 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4332
4333 return (total_evicted);
4334 }
4335
4336 void
4337 arc_flush(spa_t *spa, boolean_t retry)
4338 {
4339 uint64_t guid = 0;
4340
4341 /*
4342 * If retry is B_TRUE, a spa must not be specified since we have
4343 * no good way to determine if all of a spa's buffers have been
4344 * evicted from an arc state.
4345 */
4346 ASSERT(!retry || spa == 0);
4347
4348 if (spa != NULL)
4349 guid = spa_load_guid(spa);
4350
4351 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
4352 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
4353
4354 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
4355 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
4356
4357 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
4358 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
4359
4360 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
4361 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
4362 }
4363
4364 static void
4365 arc_reduce_target_size(int64_t to_free)
4366 {
4367 uint64_t asize = aggsum_value(&arc_size);
4368 if (arc_c > arc_c_min) {
4369
4370 if (arc_c > arc_c_min + to_free)
4371 atomic_add_64(&arc_c, -to_free);
4372 else
4373 arc_c = arc_c_min;
4374
4375 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4376 if (asize < arc_c)
4377 arc_c = MAX(asize, arc_c_min);
4378 if (arc_p > arc_c)
4379 arc_p = (arc_c >> 1);
4380 ASSERT(arc_c >= arc_c_min);
4381 ASSERT((int64_t)arc_p >= 0);
4382 }
4383
4384 if (asize > arc_c) {
4385 /* See comment in arc_adjust_cb_check() on why lock+flag */
4386 mutex_enter(&arc_adjust_lock);
4387 arc_adjust_needed = B_TRUE;
4388 mutex_exit(&arc_adjust_lock);
4389 zthr_wakeup(arc_adjust_zthr);
4390 }
4391 }
4392
4393 typedef enum free_memory_reason_t {
4394 FMR_UNKNOWN,
4395 FMR_NEEDFREE,
4396 FMR_LOTSFREE,
4397 FMR_SWAPFS_MINFREE,
4398 FMR_PAGES_PP_MAXIMUM,
4399 FMR_HEAP_ARENA,
4400 FMR_ZIO_ARENA,
4401 } free_memory_reason_t;
4402
4403 int64_t last_free_memory;
4404 free_memory_reason_t last_free_reason;
4405
4406 /*
4407 * Additional reserve of pages for pp_reserve.
4408 */
4409 int64_t arc_pages_pp_reserve = 64;
4410
4411 /*
4412 * Additional reserve of pages for swapfs.
4413 */
4414 int64_t arc_swapfs_reserve = 64;
4415
4416 /*
4417 * Return the amount of memory that can be consumed before reclaim will be
4418 * needed. Positive if there is sufficient free memory, negative indicates
4419 * the amount of memory that needs to be freed up.
4420 */
4421 static int64_t
4422 arc_available_memory(void)
4423 {
4424 int64_t lowest = INT64_MAX;
4425 int64_t n;
4426 free_memory_reason_t r = FMR_UNKNOWN;
4427
4428 #ifdef _KERNEL
4429 if (needfree > 0) {
4430 n = PAGESIZE * (-needfree);
4431 if (n < lowest) {
4432 lowest = n;
4433 r = FMR_NEEDFREE;
4434 }
4435 }
4436
4437 /*
4438 * check that we're out of range of the pageout scanner. It starts to
4439 * schedule paging if freemem is less than lotsfree and needfree.
4440 * lotsfree is the high-water mark for pageout, and needfree is the
4441 * number of needed free pages. We add extra pages here to make sure
4442 * the scanner doesn't start up while we're freeing memory.
4443 */
4444 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
4445 if (n < lowest) {
4446 lowest = n;
4447 r = FMR_LOTSFREE;
4448 }
4449
4450 /*
4451 * check to make sure that swapfs has enough space so that anon
4452 * reservations can still succeed. anon_resvmem() checks that the
4453 * availrmem is greater than swapfs_minfree, and the number of reserved
4454 * swap pages. We also add a bit of extra here just to prevent
4455 * circumstances from getting really dire.
4456 */
4457 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
4458 desfree - arc_swapfs_reserve);
4459 if (n < lowest) {
4460 lowest = n;
4461 r = FMR_SWAPFS_MINFREE;
4462 }
4463
4464
4465 /*
4466 * Check that we have enough availrmem that memory locking (e.g., via
4467 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
4468 * stores the number of pages that cannot be locked; when availrmem
4469 * drops below pages_pp_maximum, page locking mechanisms such as
4470 * page_pp_lock() will fail.)
4471 */
4472 n = PAGESIZE * (availrmem - pages_pp_maximum -
4473 arc_pages_pp_reserve);
4474 if (n < lowest) {
4475 lowest = n;
4476 r = FMR_PAGES_PP_MAXIMUM;
4477 }
4478
4479 #if defined(__i386)
4480 /*
4481 * If we're on an i386 platform, it's possible that we'll exhaust the
4482 * kernel heap space before we ever run out of available physical
4483 * memory. Most checks of the size of the heap_area compare against
4484 * tune.t_minarmem, which is the minimum available real memory that we
4485 * can have in the system. However, this is generally fixed at 25 pages
4486 * which is so low that it's useless. In this comparison, we seek to
4487 * calculate the total heap-size, and reclaim if more than 3/4ths of the
4488 * heap is allocated. (Or, in the calculation, if less than 1/4th is
4489 * free)
4490 */
4491 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) -
4492 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
4493 if (n < lowest) {
4494 lowest = n;
4495 r = FMR_HEAP_ARENA;
4496 }
4497 #endif
4498
4499 /*
4500 * If zio data pages are being allocated out of a separate heap segment,
4501 * then enforce that the size of available vmem for this arena remains
4502 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free.
4503 *
4504 * Note that reducing the arc_zio_arena_free_shift keeps more virtual
4505 * memory (in the zio_arena) free, which can avoid memory
4506 * fragmentation issues.
4507 */
4508 if (zio_arena != NULL) {
4509 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) -
4510 (vmem_size(zio_arena, VMEM_ALLOC) >>
4511 arc_zio_arena_free_shift);
4512 if (n < lowest) {
4513 lowest = n;
4514 r = FMR_ZIO_ARENA;
4515 }
4516 }
4517 #else
4518 /* Every 100 calls, free a small amount */
4519 if (spa_get_random(100) == 0)
4520 lowest = -1024;
4521 #endif
4522
4523 last_free_memory = lowest;
4524 last_free_reason = r;
4525
4526 return (lowest);
4527 }
4528
4529
4530 /*
4531 * Determine if the system is under memory pressure and is asking
4532 * to reclaim memory. A return value of B_TRUE indicates that the system
4533 * is under memory pressure and that the arc should adjust accordingly.
4534 */
4535 static boolean_t
4536 arc_reclaim_needed(void)
4537 {
4538 return (arc_available_memory() < 0);
4539 }
4540
4541 static void
4542 arc_kmem_reap_soon(void)
4543 {
4544 size_t i;
4545 kmem_cache_t *prev_cache = NULL;
4546 kmem_cache_t *prev_data_cache = NULL;
4547 extern kmem_cache_t *zio_buf_cache[];
4548 extern kmem_cache_t *zio_data_buf_cache[];
4549 extern kmem_cache_t *zfs_btree_leaf_cache;
4550 extern kmem_cache_t *abd_chunk_cache;
4551
4552 #ifdef _KERNEL
4553 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4554 /*
4555 * We are exceeding our meta-data cache limit.
4556 * Purge some DNLC entries to release holds on meta-data.
4557 */
4558 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4559 }
4560 #if defined(__i386)
4561 /*
4562 * Reclaim unused memory from all kmem caches.
4563 */
4564 kmem_reap();
4565 #endif
4566 #endif
4567
4568 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
4569 if (zio_buf_cache[i] != prev_cache) {
4570 prev_cache = zio_buf_cache[i];
4571 kmem_cache_reap_soon(zio_buf_cache[i]);
4572 }
4573 if (zio_data_buf_cache[i] != prev_data_cache) {
4574 prev_data_cache = zio_data_buf_cache[i];
4575 kmem_cache_reap_soon(zio_data_buf_cache[i]);
4576 }
4577 }
4578 kmem_cache_reap_soon(abd_chunk_cache);
4579 kmem_cache_reap_soon(buf_cache);
4580 kmem_cache_reap_soon(hdr_full_cache);
4581 kmem_cache_reap_soon(hdr_l2only_cache);
4582 kmem_cache_reap_soon(zfs_btree_leaf_cache);
4583
4584 if (zio_arena != NULL) {
4585 /*
4586 * Ask the vmem arena to reclaim unused memory from its
4587 * quantum caches.
4588 */
4589 vmem_qcache_reap(zio_arena);
4590 }
4591 }
4592
4593 /* ARGSUSED */
4594 static boolean_t
4595 arc_adjust_cb_check(void *arg, zthr_t *zthr)
4596 {
4597 /*
4598 * This is necessary in order for the mdb ::arc dcmd to
4599 * show up to date information. Since the ::arc command
4600 * does not call the kstat's update function, without
4601 * this call, the command may show stale stats for the
4602 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
4603 * with this change, the data might be up to 1 second
4604 * out of date(the arc_adjust_zthr has a maximum sleep
4605 * time of 1 second); but that should suffice. The
4606 * arc_state_t structures can be queried directly if more
4607 * accurate information is needed.
4608 */
4609 if (arc_ksp != NULL)
4610 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
4611
4612 /*
4613 * We have to rely on arc_get_data_impl() to tell us when to adjust,
4614 * rather than checking if we are overflowing here, so that we are
4615 * sure to not leave arc_get_data_impl() waiting on
4616 * arc_adjust_waiters_cv. If we have become "not overflowing" since
4617 * arc_get_data_impl() checked, we need to wake it up. We could
4618 * broadcast the CV here, but arc_get_data_impl() may have not yet
4619 * gone to sleep. We would need to use a mutex to ensure that this
4620 * function doesn't broadcast until arc_get_data_impl() has gone to
4621 * sleep (e.g. the arc_adjust_lock). However, the lock ordering of
4622 * such a lock would necessarily be incorrect with respect to the
4623 * zthr_lock, which is held before this function is called, and is
4624 * held by arc_get_data_impl() when it calls zthr_wakeup().
4625 */
4626 return (arc_adjust_needed);
4627 }
4628
4629 /*
4630 * Keep arc_size under arc_c by running arc_adjust which evicts data
4631 * from the ARC.
4632 */
4633 /* ARGSUSED */
4634 static void
4635 arc_adjust_cb(void *arg, zthr_t *zthr)
4636 {
4637 uint64_t evicted = 0;
4638
4639 /* Evict from cache */
4640 evicted = arc_adjust();
4641
4642 /*
4643 * If evicted is zero, we couldn't evict anything
4644 * via arc_adjust(). This could be due to hash lock
4645 * collisions, but more likely due to the majority of
4646 * arc buffers being unevictable. Therefore, even if
4647 * arc_size is above arc_c, another pass is unlikely to
4648 * be helpful and could potentially cause us to enter an
4649 * infinite loop. Additionally, zthr_iscancelled() is
4650 * checked here so that if the arc is shutting down, the
4651 * broadcast will wake any remaining arc adjust waiters.
4652 */
4653 mutex_enter(&arc_adjust_lock);
4654 arc_adjust_needed = !zthr_iscancelled(arc_adjust_zthr) &&
4655 evicted > 0 && aggsum_compare(&arc_size, arc_c) > 0;
4656 if (!arc_adjust_needed) {
4657 /*
4658 * We're either no longer overflowing, or we
4659 * can't evict anything more, so we should wake
4660 * up any waiters.
4661 */
4662 cv_broadcast(&arc_adjust_waiters_cv);
4663 }
4664 mutex_exit(&arc_adjust_lock);
4665 }
4666
4667 /* ARGSUSED */
4668 static boolean_t
4669 arc_reap_cb_check(void *arg, zthr_t *zthr)
4670 {
4671 int64_t free_memory = arc_available_memory();
4672
4673 /*
4674 * If a kmem reap is already active, don't schedule more. We must
4675 * check for this because kmem_cache_reap_soon() won't actually
4676 * block on the cache being reaped (this is to prevent callers from
4677 * becoming implicitly blocked by a system-wide kmem reap -- which,
4678 * on a system with many, many full magazines, can take minutes).
4679 */
4680 if (!kmem_cache_reap_active() &&
4681 free_memory < 0) {
4682 arc_no_grow = B_TRUE;
4683 arc_warm = B_TRUE;
4684 /*
4685 * Wait at least zfs_grow_retry (default 60) seconds
4686 * before considering growing.
4687 */
4688 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry);
4689 return (B_TRUE);
4690 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4691 arc_no_grow = B_TRUE;
4692 } else if (gethrtime() >= arc_growtime) {
4693 arc_no_grow = B_FALSE;
4694 }
4695
4696 return (B_FALSE);
4697 }
4698
4699 /*
4700 * Keep enough free memory in the system by reaping the ARC's kmem
4701 * caches. To cause more slabs to be reapable, we may reduce the
4702 * target size of the cache (arc_c), causing the arc_adjust_cb()
4703 * to free more buffers.
4704 */
4705 /* ARGSUSED */
4706 static void
4707 arc_reap_cb(void *arg, zthr_t *zthr)
4708 {
4709 int64_t free_memory;
4710
4711 /*
4712 * Kick off asynchronous kmem_reap()'s of all our caches.
4713 */
4714 arc_kmem_reap_soon();
4715
4716 /*
4717 * Wait at least arc_kmem_cache_reap_retry_ms between
4718 * arc_kmem_reap_soon() calls. Without this check it is possible to
4719 * end up in a situation where we spend lots of time reaping
4720 * caches, while we're near arc_c_min. Waiting here also gives the
4721 * subsequent free memory check a chance of finding that the
4722 * asynchronous reap has already freed enough memory, and we don't
4723 * need to call arc_reduce_target_size().
4724 */
4725 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000);
4726
4727 /*
4728 * Reduce the target size as needed to maintain the amount of free
4729 * memory in the system at a fraction of the arc_size (1/128th by
4730 * default). If oversubscribed (free_memory < 0) then reduce the
4731 * target arc_size by the deficit amount plus the fractional
4732 * amount. If free memory is positive but less then the fractional
4733 * amount, reduce by what is needed to hit the fractional amount.
4734 */
4735 free_memory = arc_available_memory();
4736
4737 int64_t to_free =
4738 (arc_c >> arc_shrink_shift) - free_memory;
4739 if (to_free > 0) {
4740 #ifdef _KERNEL
4741 to_free = MAX(to_free, ptob(needfree));
4742 #endif
4743 arc_reduce_target_size(to_free);
4744 }
4745 }
4746
4747 /*
4748 * Adapt arc info given the number of bytes we are trying to add and
4749 * the state that we are coming from. This function is only called
4750 * when we are adding new content to the cache.
4751 */
4752 static void
4753 arc_adapt(int bytes, arc_state_t *state)
4754 {
4755 int mult;
4756 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
4757 int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size);
4758 int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size);
4759
4760 if (state == arc_l2c_only)
4761 return;
4762
4763 ASSERT(bytes > 0);
4764 /*
4765 * Adapt the target size of the MRU list:
4766 * - if we just hit in the MRU ghost list, then increase
4767 * the target size of the MRU list.
4768 * - if we just hit in the MFU ghost list, then increase
4769 * the target size of the MFU list by decreasing the
4770 * target size of the MRU list.
4771 */
4772 if (state == arc_mru_ghost) {
4773 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
4774 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
4775
4776 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
4777 } else if (state == arc_mfu_ghost) {
4778 uint64_t delta;
4779
4780 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
4781 mult = MIN(mult, 10);
4782
4783 delta = MIN(bytes * mult, arc_p);
4784 arc_p = MAX(arc_p_min, arc_p - delta);
4785 }
4786 ASSERT((int64_t)arc_p >= 0);
4787
4788 /*
4789 * Wake reap thread if we do not have any available memory
4790 */
4791 if (arc_reclaim_needed()) {
4792 zthr_wakeup(arc_reap_zthr);
4793 return;
4794 }
4795
4796
4797 if (arc_no_grow)
4798 return;
4799
4800 if (arc_c >= arc_c_max)
4801 return;
4802
4803 /*
4804 * If we're within (2 * maxblocksize) bytes of the target
4805 * cache size, increment the target cache size
4806 */
4807 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
4808 0) {
4809 atomic_add_64(&arc_c, (int64_t)bytes);
4810 if (arc_c > arc_c_max)
4811 arc_c = arc_c_max;
4812 else if (state == arc_anon)
4813 atomic_add_64(&arc_p, (int64_t)bytes);
4814 if (arc_p > arc_c)
4815 arc_p = arc_c;
4816 }
4817 ASSERT((int64_t)arc_p >= 0);
4818 }
4819
4820 /*
4821 * Check if arc_size has grown past our upper threshold, determined by
4822 * zfs_arc_overflow_shift.
4823 */
4824 static boolean_t
4825 arc_is_overflowing(void)
4826 {
4827 /* Always allow at least one block of overflow */
4828 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4829 arc_c >> zfs_arc_overflow_shift);
4830
4831 /*
4832 * We just compare the lower bound here for performance reasons. Our
4833 * primary goals are to make sure that the arc never grows without
4834 * bound, and that it can reach its maximum size. This check
4835 * accomplishes both goals. The maximum amount we could run over by is
4836 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4837 * in the ARC. In practice, that's in the tens of MB, which is low
4838 * enough to be safe.
4839 */
4840 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
4841 }
4842
4843 static abd_t *
4844 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4845 {
4846 arc_buf_contents_t type = arc_buf_type(hdr);
4847
4848 arc_get_data_impl(hdr, size, tag);
4849 if (type == ARC_BUFC_METADATA) {
4850 return (abd_alloc(size, B_TRUE));
4851 } else {
4852 ASSERT(type == ARC_BUFC_DATA);
4853 return (abd_alloc(size, B_FALSE));
4854 }
4855 }
4856
4857 static void *
4858 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4859 {
4860 arc_buf_contents_t type = arc_buf_type(hdr);
4861
4862 arc_get_data_impl(hdr, size, tag);
4863 if (type == ARC_BUFC_METADATA) {
4864 return (zio_buf_alloc(size));
4865 } else {
4866 ASSERT(type == ARC_BUFC_DATA);
4867 return (zio_data_buf_alloc(size));
4868 }
4869 }
4870
4871 /*
4872 * Allocate a block and return it to the caller. If we are hitting the
4873 * hard limit for the cache size, we must sleep, waiting for the eviction
4874 * thread to catch up. If we're past the target size but below the hard
4875 * limit, we'll only signal the reclaim thread and continue on.
4876 */
4877 static void
4878 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4879 {
4880 arc_state_t *state = hdr->b_l1hdr.b_state;
4881 arc_buf_contents_t type = arc_buf_type(hdr);
4882
4883 arc_adapt(size, state);
4884
4885 /*
4886 * If arc_size is currently overflowing, and has grown past our
4887 * upper limit, we must be adding data faster than the evict
4888 * thread can evict. Thus, to ensure we don't compound the
4889 * problem by adding more data and forcing arc_size to grow even
4890 * further past its target size, we halt and wait for the
4891 * eviction thread to catch up.
4892 *
4893 * It's also possible that the reclaim thread is unable to evict
4894 * enough buffers to get arc_size below the overflow limit (e.g.
4895 * due to buffers being un-evictable, or hash lock collisions).
4896 * In this case, we want to proceed regardless if we're
4897 * overflowing; thus we don't use a while loop here.
4898 */
4899 if (arc_is_overflowing()) {
4900 mutex_enter(&arc_adjust_lock);
4901
4902 /*
4903 * Now that we've acquired the lock, we may no longer be
4904 * over the overflow limit, lets check.
4905 *
4906 * We're ignoring the case of spurious wake ups. If that
4907 * were to happen, it'd let this thread consume an ARC
4908 * buffer before it should have (i.e. before we're under
4909 * the overflow limit and were signalled by the reclaim
4910 * thread). As long as that is a rare occurrence, it
4911 * shouldn't cause any harm.
4912 */
4913 if (arc_is_overflowing()) {
4914 arc_adjust_needed = B_TRUE;
4915 zthr_wakeup(arc_adjust_zthr);
4916 (void) cv_wait(&arc_adjust_waiters_cv,
4917 &arc_adjust_lock);
4918 }
4919 mutex_exit(&arc_adjust_lock);
4920 }
4921
4922 VERIFY3U(hdr->b_type, ==, type);
4923 if (type == ARC_BUFC_METADATA) {
4924 arc_space_consume(size, ARC_SPACE_META);
4925 } else {
4926 arc_space_consume(size, ARC_SPACE_DATA);
4927 }
4928
4929 /*
4930 * Update the state size. Note that ghost states have a
4931 * "ghost size" and so don't need to be updated.
4932 */
4933 if (!GHOST_STATE(state)) {
4934
4935 (void) zfs_refcount_add_many(&state->arcs_size, size, tag);
4936
4937 /*
4938 * If this is reached via arc_read, the link is
4939 * protected by the hash lock. If reached via
4940 * arc_buf_alloc, the header should not be accessed by
4941 * any other thread. And, if reached via arc_read_done,
4942 * the hash lock will protect it if it's found in the
4943 * hash table; otherwise no other thread should be
4944 * trying to [add|remove]_reference it.
4945 */
4946 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4947 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4948 (void) zfs_refcount_add_many(&state->arcs_esize[type],
4949 size, tag);
4950 }
4951
4952 /*
4953 * If we are growing the cache, and we are adding anonymous
4954 * data, and we have outgrown arc_p, update arc_p
4955 */
4956 if (aggsum_compare(&arc_size, arc_c) < 0 &&
4957 hdr->b_l1hdr.b_state == arc_anon &&
4958 (zfs_refcount_count(&arc_anon->arcs_size) +
4959 zfs_refcount_count(&arc_mru->arcs_size) > arc_p))
4960 arc_p = MIN(arc_c, arc_p + size);
4961 }
4962 }
4963
4964 static void
4965 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
4966 {
4967 arc_free_data_impl(hdr, size, tag);
4968 abd_free(abd);
4969 }
4970
4971 static void
4972 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
4973 {
4974 arc_buf_contents_t type = arc_buf_type(hdr);
4975
4976 arc_free_data_impl(hdr, size, tag);
4977 if (type == ARC_BUFC_METADATA) {
4978 zio_buf_free(buf, size);
4979 } else {
4980 ASSERT(type == ARC_BUFC_DATA);
4981 zio_data_buf_free(buf, size);
4982 }
4983 }
4984
4985 /*
4986 * Free the arc data buffer.
4987 */
4988 static void
4989 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4990 {
4991 arc_state_t *state = hdr->b_l1hdr.b_state;
4992 arc_buf_contents_t type = arc_buf_type(hdr);
4993
4994 /* protected by hash lock, if in the hash table */
4995 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4996 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4997 ASSERT(state != arc_anon && state != arc_l2c_only);
4998
4999 (void) zfs_refcount_remove_many(&state->arcs_esize[type],
5000 size, tag);
5001 }
5002 (void) zfs_refcount_remove_many(&state->arcs_size, size, tag);
5003
5004 VERIFY3U(hdr->b_type, ==, type);
5005 if (type == ARC_BUFC_METADATA) {
5006 arc_space_return(size, ARC_SPACE_META);
5007 } else {
5008 ASSERT(type == ARC_BUFC_DATA);
5009 arc_space_return(size, ARC_SPACE_DATA);
5010 }
5011 }
5012
5013 /*
5014 * This routine is called whenever a buffer is accessed.
5015 * NOTE: the hash lock is dropped in this function.
5016 */
5017 static void
5018 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5019 {
5020 clock_t now;
5021
5022 ASSERT(MUTEX_HELD(hash_lock));
5023 ASSERT(HDR_HAS_L1HDR(hdr));
5024
5025 if (hdr->b_l1hdr.b_state == arc_anon) {
5026 /*
5027 * This buffer is not in the cache, and does not
5028 * appear in our "ghost" list. Add the new buffer
5029 * to the MRU state.
5030 */
5031
5032 ASSERT0(hdr->b_l1hdr.b_arc_access);
5033 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5034 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5035 arc_change_state(arc_mru, hdr, hash_lock);
5036
5037 } else if (hdr->b_l1hdr.b_state == arc_mru) {
5038 now = ddi_get_lbolt();
5039
5040 /*
5041 * If this buffer is here because of a prefetch, then either:
5042 * - clear the flag if this is a "referencing" read
5043 * (any subsequent access will bump this into the MFU state).
5044 * or
5045 * - move the buffer to the head of the list if this is
5046 * another prefetch (to make it less likely to be evicted).
5047 */
5048 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5049 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5050 /* link protected by hash lock */
5051 ASSERT(multilist_link_active(
5052 &hdr->b_l1hdr.b_arc_node));
5053 } else {
5054 arc_hdr_clear_flags(hdr,
5055 ARC_FLAG_PREFETCH |
5056 ARC_FLAG_PRESCIENT_PREFETCH);
5057 ARCSTAT_BUMP(arcstat_mru_hits);
5058 }
5059 hdr->b_l1hdr.b_arc_access = now;
5060 return;
5061 }
5062
5063 /*
5064 * This buffer has been "accessed" only once so far,
5065 * but it is still in the cache. Move it to the MFU
5066 * state.
5067 */
5068 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
5069 /*
5070 * More than 125ms have passed since we
5071 * instantiated this buffer. Move it to the
5072 * most frequently used state.
5073 */
5074 hdr->b_l1hdr.b_arc_access = now;
5075 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5076 arc_change_state(arc_mfu, hdr, hash_lock);
5077 }
5078 ARCSTAT_BUMP(arcstat_mru_hits);
5079 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
5080 arc_state_t *new_state;
5081 /*
5082 * This buffer has been "accessed" recently, but
5083 * was evicted from the cache. Move it to the
5084 * MFU state.
5085 */
5086
5087 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5088 new_state = arc_mru;
5089 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) {
5090 arc_hdr_clear_flags(hdr,
5091 ARC_FLAG_PREFETCH |
5092 ARC_FLAG_PRESCIENT_PREFETCH);
5093 }
5094 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
5095 } else {
5096 new_state = arc_mfu;
5097 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5098 }
5099
5100 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5101 arc_change_state(new_state, hdr, hash_lock);
5102
5103 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
5104 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
5105 /*
5106 * This buffer has been accessed more than once and is
5107 * still in the cache. Keep it in the MFU state.
5108 *
5109 * NOTE: an add_reference() that occurred when we did
5110 * the arc_read() will have kicked this off the list.
5111 * If it was a prefetch, we will explicitly move it to
5112 * the head of the list now.
5113 */
5114 ARCSTAT_BUMP(arcstat_mfu_hits);
5115 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5116 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
5117 arc_state_t *new_state = arc_mfu;
5118 /*
5119 * This buffer has been accessed more than once but has
5120 * been evicted from the cache. Move it back to the
5121 * MFU state.
5122 */
5123
5124 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) {
5125 /*
5126 * This is a prefetch access...
5127 * move this block back to the MRU state.
5128 */
5129 new_state = arc_mru;
5130 }
5131
5132 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5133 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5134 arc_change_state(new_state, hdr, hash_lock);
5135
5136 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5137 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5138 /*
5139 * This buffer is on the 2nd Level ARC.
5140 */
5141
5142 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5143 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5144 arc_change_state(arc_mfu, hdr, hash_lock);
5145 } else {
5146 ASSERT(!"invalid arc state");
5147 }
5148 }
5149
5150 /*
5151 * This routine is called by dbuf_hold() to update the arc_access() state
5152 * which otherwise would be skipped for entries in the dbuf cache.
5153 */
5154 void
5155 arc_buf_access(arc_buf_t *buf)
5156 {
5157 mutex_enter(&buf->b_evict_lock);
5158 arc_buf_hdr_t *hdr = buf->b_hdr;
5159
5160 /*
5161 * Avoid taking the hash_lock when possible as an optimization.
5162 * The header must be checked again under the hash_lock in order
5163 * to handle the case where it is concurrently being released.
5164 */
5165 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5166 mutex_exit(&buf->b_evict_lock);
5167 return;
5168 }
5169
5170 kmutex_t *hash_lock = HDR_LOCK(hdr);
5171 mutex_enter(hash_lock);
5172
5173 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5174 mutex_exit(hash_lock);
5175 mutex_exit(&buf->b_evict_lock);
5176 ARCSTAT_BUMP(arcstat_access_skip);
5177 return;
5178 }
5179
5180 mutex_exit(&buf->b_evict_lock);
5181
5182 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5183 hdr->b_l1hdr.b_state == arc_mfu);
5184
5185 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5186 arc_access(hdr, hash_lock);
5187 mutex_exit(hash_lock);
5188
5189 ARCSTAT_BUMP(arcstat_hits);
5190 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5191 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits);
5192 }
5193
5194 /* a generic arc_read_done_func_t which you can use */
5195 /* ARGSUSED */
5196 void
5197 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5198 arc_buf_t *buf, void *arg)
5199 {
5200 if (buf == NULL)
5201 return;
5202
5203 bcopy(buf->b_data, arg, arc_buf_size(buf));
5204 arc_buf_destroy(buf, arg);
5205 }
5206
5207 /* a generic arc_read_done_func_t */
5208 void
5209 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
5210 arc_buf_t *buf, void *arg)
5211 {
5212 arc_buf_t **bufp = arg;
5213
5214 if (buf == NULL) {
5215 ASSERT(zio == NULL || zio->io_error != 0);
5216 *bufp = NULL;
5217 } else {
5218 ASSERT(zio == NULL || zio->io_error == 0);
5219 *bufp = buf;
5220 ASSERT(buf->b_data != NULL);
5221 }
5222 }
5223
5224 static void
5225 arc_hdr_verify(arc_buf_hdr_t *hdr, const blkptr_t *bp)
5226 {
5227 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
5228 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0);
5229 ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF);
5230 } else {
5231 if (HDR_COMPRESSION_ENABLED(hdr)) {
5232 ASSERT3U(arc_hdr_get_compress(hdr), ==,
5233 BP_GET_COMPRESS(bp));
5234 }
5235 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
5236 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp));
5237 ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp));
5238 }
5239 }
5240
5241 /*
5242 * XXX this should be changed to return an error, and callers
5243 * re-read from disk on failure (on nondebug bits).
5244 */
5245 static void
5246 arc_hdr_verify_checksum(spa_t *spa, arc_buf_hdr_t *hdr, const blkptr_t *bp)
5247 {
5248 arc_hdr_verify(hdr, bp);
5249 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
5250 return;
5251 int err = 0;
5252 abd_t *abd = NULL;
5253 if (BP_IS_ENCRYPTED(bp)) {
5254 if (HDR_HAS_RABD(hdr)) {
5255 abd = hdr->b_crypt_hdr.b_rabd;
5256 }
5257 } else if (HDR_COMPRESSION_ENABLED(hdr)) {
5258 abd = hdr->b_l1hdr.b_pabd;
5259 }
5260 if (abd != NULL) {
5261 /*
5262 * The offset is only used for labels, which are not
5263 * cached in the ARC, so it doesn't matter what we
5264 * pass for the offset parameter.
5265 */
5266 int psize = HDR_GET_PSIZE(hdr);
5267 err = zio_checksum_error_impl(spa, bp,
5268 BP_GET_CHECKSUM(bp), abd, psize, 0, NULL);
5269 if (err != 0) {
5270 /*
5271 * Use abd_copy_to_buf() rather than
5272 * abd_borrow_buf_copy() so that we are sure to
5273 * include the buf in crash dumps.
5274 */
5275 void *buf = kmem_alloc(psize, KM_SLEEP);
5276 abd_copy_to_buf(buf, abd, psize);
5277 panic("checksum of cached data doesn't match BP "
5278 "err=%u hdr=%p bp=%p abd=%p buf=%p",
5279 err, (void *)hdr, (void *)bp, (void *)abd, buf);
5280 }
5281 }
5282 }
5283
5284 static void
5285 arc_read_done(zio_t *zio)
5286 {
5287 blkptr_t *bp = zio->io_bp;
5288 arc_buf_hdr_t *hdr = zio->io_private;
5289 kmutex_t *hash_lock = NULL;
5290 arc_callback_t *callback_list;
5291 arc_callback_t *acb;
5292 boolean_t freeable = B_FALSE;
5293
5294 /*
5295 * The hdr was inserted into hash-table and removed from lists
5296 * prior to starting I/O. We should find this header, since
5297 * it's in the hash table, and it should be legit since it's
5298 * not possible to evict it during the I/O. The only possible
5299 * reason for it not to be found is if we were freed during the
5300 * read.
5301 */
5302 if (HDR_IN_HASH_TABLE(hdr)) {
5303 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
5304 ASSERT3U(hdr->b_dva.dva_word[0], ==,
5305 BP_IDENTITY(zio->io_bp)->dva_word[0]);
5306 ASSERT3U(hdr->b_dva.dva_word[1], ==,
5307 BP_IDENTITY(zio->io_bp)->dva_word[1]);
5308
5309 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
5310 &hash_lock);
5311
5312 ASSERT((found == hdr &&
5313 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
5314 (found == hdr && HDR_L2_READING(hdr)));
5315 ASSERT3P(hash_lock, !=, NULL);
5316 }
5317
5318 if (BP_IS_PROTECTED(bp)) {
5319 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
5320 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
5321 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
5322 hdr->b_crypt_hdr.b_iv);
5323
5324 if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) {
5325 void *tmpbuf;
5326
5327 tmpbuf = abd_borrow_buf_copy(zio->io_abd,
5328 sizeof (zil_chain_t));
5329 zio_crypt_decode_mac_zil(tmpbuf,
5330 hdr->b_crypt_hdr.b_mac);
5331 abd_return_buf(zio->io_abd, tmpbuf,
5332 sizeof (zil_chain_t));
5333 } else {
5334 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
5335 }
5336 }
5337
5338 if (zio->io_error == 0) {
5339 /* byteswap if necessary */
5340 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
5341 if (BP_GET_LEVEL(zio->io_bp) > 0) {
5342 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
5343 } else {
5344 hdr->b_l1hdr.b_byteswap =
5345 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
5346 }
5347 } else {
5348 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
5349 }
5350 }
5351
5352 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED);
5353 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
5354 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE);
5355
5356 callback_list = hdr->b_l1hdr.b_acb;
5357 ASSERT3P(callback_list, !=, NULL);
5358
5359 if (hash_lock && zio->io_error == 0 &&
5360 hdr->b_l1hdr.b_state == arc_anon) {
5361 /*
5362 * Only call arc_access on anonymous buffers. This is because
5363 * if we've issued an I/O for an evicted buffer, we've already
5364 * called arc_access (to prevent any simultaneous readers from
5365 * getting confused).
5366 */
5367 arc_access(hdr, hash_lock);
5368 }
5369
5370 /*
5371 * If a read request has a callback (i.e. acb_done is not NULL), then we
5372 * make a buf containing the data according to the parameters which were
5373 * passed in. The implementation of arc_buf_alloc_impl() ensures that we
5374 * aren't needlessly decompressing the data multiple times.
5375 */
5376 int callback_cnt = 0;
5377 for (acb = callback_list; acb != NULL; acb = acb->acb_next) {
5378 if (!acb->acb_done)
5379 continue;
5380
5381 callback_cnt++;
5382
5383 if (zio->io_error != 0)
5384 continue;
5385
5386 int error = arc_buf_alloc_impl(hdr, zio->io_spa,
5387 &acb->acb_zb, acb->acb_private, acb->acb_encrypted,
5388 acb->acb_compressed, acb->acb_noauth, B_TRUE,
5389 &acb->acb_buf);
5390
5391 /*
5392 * Assert non-speculative zios didn't fail because an
5393 * encryption key wasn't loaded
5394 */
5395 ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) ||
5396 error != EACCES);
5397
5398 /*
5399 * If we failed to decrypt, report an error now (as the zio
5400 * layer would have done if it had done the transforms).
5401 */
5402 if (error == ECKSUM) {
5403 ASSERT(BP_IS_PROTECTED(bp));
5404 error = SET_ERROR(EIO);
5405 if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5406 spa_log_error(zio->io_spa, &acb->acb_zb);
5407 (void) zfs_ereport_post(
5408 FM_EREPORT_ZFS_AUTHENTICATION,
5409 zio->io_spa, NULL, &acb->acb_zb, zio, 0, 0);
5410 }
5411 }
5412
5413 if (error != 0) {
5414 /*
5415 * Decompression failed. Set io_error
5416 * so that when we call acb_done (below),
5417 * we will indicate that the read failed.
5418 * Note that in the unusual case where one
5419 * callback is compressed and another
5420 * uncompressed, we will mark all of them
5421 * as failed, even though the uncompressed
5422 * one can't actually fail. In this case,
5423 * the hdr will not be anonymous, because
5424 * if there are multiple callbacks, it's
5425 * because multiple threads found the same
5426 * arc buf in the hash table.
5427 */
5428 zio->io_error = error;
5429 }
5430 }
5431
5432 /*
5433 * If there are multiple callbacks, we must have the hash lock,
5434 * because the only way for multiple threads to find this hdr is
5435 * in the hash table. This ensures that if there are multiple
5436 * callbacks, the hdr is not anonymous. If it were anonymous,
5437 * we couldn't use arc_buf_destroy() in the error case below.
5438 */
5439 ASSERT(callback_cnt < 2 || hash_lock != NULL);
5440
5441 hdr->b_l1hdr.b_acb = NULL;
5442 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5443 if (callback_cnt == 0)
5444 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
5445
5446 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5447 callback_list != NULL);
5448
5449 if (zio->io_error == 0) {
5450 arc_hdr_verify(hdr, zio->io_bp);
5451 } else {
5452 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5453 if (hdr->b_l1hdr.b_state != arc_anon)
5454 arc_change_state(arc_anon, hdr, hash_lock);
5455 if (HDR_IN_HASH_TABLE(hdr))
5456 buf_hash_remove(hdr);
5457 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5458 }
5459
5460 /*
5461 * Broadcast before we drop the hash_lock to avoid the possibility
5462 * that the hdr (and hence the cv) might be freed before we get to
5463 * the cv_broadcast().
5464 */
5465 cv_broadcast(&hdr->b_l1hdr.b_cv);
5466
5467 if (hash_lock != NULL) {
5468 mutex_exit(hash_lock);
5469 } else {
5470 /*
5471 * This block was freed while we waited for the read to
5472 * complete. It has been removed from the hash table and
5473 * moved to the anonymous state (so that it won't show up
5474 * in the cache).
5475 */
5476 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5477 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5478 }
5479
5480 /* execute each callback and free its structure */
5481 while ((acb = callback_list) != NULL) {
5482
5483 if (acb->acb_done != NULL) {
5484 if (zio->io_error != 0 && acb->acb_buf != NULL) {
5485 /*
5486 * If arc_buf_alloc_impl() fails during
5487 * decompression, the buf will still be
5488 * allocated, and needs to be freed here.
5489 */
5490 arc_buf_destroy(acb->acb_buf, acb->acb_private);
5491 acb->acb_buf = NULL;
5492 }
5493 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp,
5494 acb->acb_buf, acb->acb_private);
5495 }
5496
5497 if (acb->acb_zio_dummy != NULL) {
5498 acb->acb_zio_dummy->io_error = zio->io_error;
5499 zio_nowait(acb->acb_zio_dummy);
5500 }
5501
5502 callback_list = acb->acb_next;
5503 kmem_free(acb, sizeof (arc_callback_t));
5504 }
5505
5506 if (freeable)
5507 arc_hdr_destroy(hdr);
5508 }
5509
5510 /*
5511 * "Read" the block at the specified DVA (in bp) via the
5512 * cache. If the block is found in the cache, invoke the provided
5513 * callback immediately and return. Note that the `zio' parameter
5514 * in the callback will be NULL in this case, since no IO was
5515 * required. If the block is not in the cache pass the read request
5516 * on to the spa with a substitute callback function, so that the
5517 * requested block will be added to the cache.
5518 *
5519 * If a read request arrives for a block that has a read in-progress,
5520 * either wait for the in-progress read to complete (and return the
5521 * results); or, if this is a read with a "done" func, add a record
5522 * to the read to invoke the "done" func when the read completes,
5523 * and return; or just return.
5524 *
5525 * arc_read_done() will invoke all the requested "done" functions
5526 * for readers of this block.
5527 */
5528 int
5529 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_read_done_func_t *done,
5530 void *private, zio_priority_t priority, int zio_flags,
5531 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
5532 {
5533 arc_buf_hdr_t *hdr = NULL;
5534 kmutex_t *hash_lock = NULL;
5535 zio_t *rzio;
5536 uint64_t guid = spa_load_guid(spa);
5537 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0;
5538 boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) &&
5539 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5540 boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) &&
5541 (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0;
5542 int rc = 0;
5543
5544 ASSERT(!BP_IS_EMBEDDED(bp) ||
5545 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
5546
5547 top:
5548 if (!BP_IS_EMBEDDED(bp)) {
5549 /*
5550 * Embedded BP's have no DVA and require no I/O to "read".
5551 * Create an anonymous arc buf to back it.
5552 */
5553 hdr = buf_hash_find(guid, bp, &hash_lock);
5554 }
5555
5556 /*
5557 * Determine if we have an L1 cache hit or a cache miss. For simplicity
5558 * we maintain encrypted data seperately from compressed / uncompressed
5559 * data. If the user is requesting raw encrypted data and we don't have
5560 * that in the header we will read from disk to guarantee that we can
5561 * get it even if the encryption keys aren't loaded.
5562 */
5563 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) ||
5564 (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) {
5565 arc_buf_t *buf = NULL;
5566 *arc_flags |= ARC_FLAG_CACHED;
5567
5568 if (HDR_IO_IN_PROGRESS(hdr)) {
5569 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head;
5570
5571 ASSERT3P(head_zio, !=, NULL);
5572 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) &&
5573 priority == ZIO_PRIORITY_SYNC_READ) {
5574 /*
5575 * This is a sync read that needs to wait for
5576 * an in-flight async read. Request that the
5577 * zio have its priority upgraded.
5578 */
5579 zio_change_priority(head_zio, priority);
5580 DTRACE_PROBE1(arc__async__upgrade__sync,
5581 arc_buf_hdr_t *, hdr);
5582 ARCSTAT_BUMP(arcstat_async_upgrade_sync);
5583 }
5584 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5585 arc_hdr_clear_flags(hdr,
5586 ARC_FLAG_PREDICTIVE_PREFETCH);
5587 }
5588
5589 if (*arc_flags & ARC_FLAG_WAIT) {
5590 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5591 mutex_exit(hash_lock);
5592 goto top;
5593 }
5594 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5595
5596 if (done) {
5597 arc_callback_t *acb = NULL;
5598
5599 acb = kmem_zalloc(sizeof (arc_callback_t),
5600 KM_SLEEP);
5601 acb->acb_done = done;
5602 acb->acb_private = private;
5603 acb->acb_compressed = compressed_read;
5604 acb->acb_encrypted = encrypted_read;
5605 acb->acb_noauth = noauth_read;
5606 acb->acb_zb = *zb;
5607 if (pio != NULL)
5608 acb->acb_zio_dummy = zio_null(pio,
5609 spa, NULL, NULL, NULL, zio_flags);
5610
5611 ASSERT3P(acb->acb_done, !=, NULL);
5612 acb->acb_zio_head = head_zio;
5613 acb->acb_next = hdr->b_l1hdr.b_acb;
5614 hdr->b_l1hdr.b_acb = acb;
5615 mutex_exit(hash_lock);
5616 return (0);
5617 }
5618 mutex_exit(hash_lock);
5619 return (0);
5620 }
5621
5622 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5623 hdr->b_l1hdr.b_state == arc_mfu);
5624
5625 if (done) {
5626 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) {
5627 /*
5628 * This is a demand read which does not have to
5629 * wait for i/o because we did a predictive
5630 * prefetch i/o for it, which has completed.
5631 */
5632 DTRACE_PROBE1(
5633 arc__demand__hit__predictive__prefetch,
5634 arc_buf_hdr_t *, hdr);
5635 ARCSTAT_BUMP(
5636 arcstat_demand_hit_predictive_prefetch);
5637 arc_hdr_clear_flags(hdr,
5638 ARC_FLAG_PREDICTIVE_PREFETCH);
5639 }
5640
5641 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) {
5642 ARCSTAT_BUMP(
5643 arcstat_demand_hit_prescient_prefetch);
5644 arc_hdr_clear_flags(hdr,
5645 ARC_FLAG_PRESCIENT_PREFETCH);
5646 }
5647
5648 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5649
5650 arc_hdr_verify_checksum(spa, hdr, bp);
5651
5652 /* Get a buf with the desired data in it. */
5653 rc = arc_buf_alloc_impl(hdr, spa, zb, private,
5654 encrypted_read, compressed_read, noauth_read,
5655 B_TRUE, &buf);
5656 if (rc == ECKSUM) {
5657 /*
5658 * Convert authentication and decryption errors
5659 * to EIO (and generate an ereport if needed)
5660 * before leaving the ARC.
5661 */
5662 rc = SET_ERROR(EIO);
5663 if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) {
5664 spa_log_error(spa, zb);
5665 (void) zfs_ereport_post(
5666 FM_EREPORT_ZFS_AUTHENTICATION,
5667 spa, NULL, zb, NULL, 0, 0);
5668 }
5669 }
5670 if (rc != 0) {
5671 (void) remove_reference(hdr, hash_lock,
5672 private);
5673 arc_buf_destroy_impl(buf);
5674 buf = NULL;
5675 }
5676 /* assert any errors weren't due to unloaded keys */
5677 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) ||
5678 rc != EACCES);
5679 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5680 zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5681 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5682 }
5683 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5684 arc_access(hdr, hash_lock);
5685 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5686 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5687 if (*arc_flags & ARC_FLAG_L2CACHE)
5688 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5689 mutex_exit(hash_lock);
5690 ARCSTAT_BUMP(arcstat_hits);
5691 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5692 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5693 data, metadata, hits);
5694
5695 if (done)
5696 done(NULL, zb, bp, buf, private);
5697 } else {
5698 uint64_t lsize = BP_GET_LSIZE(bp);
5699 uint64_t psize = BP_GET_PSIZE(bp);
5700 arc_callback_t *acb;
5701 vdev_t *vd = NULL;
5702 uint64_t addr = 0;
5703 boolean_t devw = B_FALSE;
5704 uint64_t size;
5705 abd_t *hdr_abd;
5706
5707 if (hdr == NULL) {
5708 /* this block is not in the cache */
5709 arc_buf_hdr_t *exists = NULL;
5710 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5711 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5712 BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), type,
5713 encrypted_read);
5714
5715 if (!BP_IS_EMBEDDED(bp)) {
5716 hdr->b_dva = *BP_IDENTITY(bp);
5717 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5718 exists = buf_hash_insert(hdr, &hash_lock);
5719 }
5720 if (exists != NULL) {
5721 /* somebody beat us to the hash insert */
5722 mutex_exit(hash_lock);
5723 buf_discard_identity(hdr);
5724 arc_hdr_destroy(hdr);
5725 goto top; /* restart the IO request */
5726 }
5727 } else {
5728 /*
5729 * This block is in the ghost cache or encrypted data
5730 * was requested and we didn't have it. If it was
5731 * L2-only (and thus didn't have an L1 hdr),
5732 * we realloc the header to add an L1 hdr.
5733 */
5734 if (!HDR_HAS_L1HDR(hdr)) {
5735 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5736 hdr_full_cache);
5737 }
5738
5739 if (GHOST_STATE(hdr->b_l1hdr.b_state)) {
5740 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5741 ASSERT(!HDR_HAS_RABD(hdr));
5742 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5743 ASSERT0(zfs_refcount_count(
5744 &hdr->b_l1hdr.b_refcnt));
5745 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5746 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5747 } else if (HDR_IO_IN_PROGRESS(hdr)) {
5748 /*
5749 * If this header already had an IO in progress
5750 * and we are performing another IO to fetch
5751 * encrypted data we must wait until the first
5752 * IO completes so as not to confuse
5753 * arc_read_done(). This should be very rare
5754 * and so the performance impact shouldn't
5755 * matter.
5756 */
5757 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5758 mutex_exit(hash_lock);
5759 goto top;
5760 }
5761
5762 /*
5763 * This is a delicate dance that we play here.
5764 * This hdr might be in the ghost list so we access
5765 * it to move it out of the ghost list before we
5766 * initiate the read. If it's a prefetch then
5767 * it won't have a callback so we'll remove the
5768 * reference that arc_buf_alloc_impl() created. We
5769 * do this after we've called arc_access() to
5770 * avoid hitting an assert in remove_reference().
5771 */
5772 arc_access(hdr, hash_lock);
5773 arc_hdr_alloc_pabd(hdr, encrypted_read);
5774 }
5775
5776 if (encrypted_read) {
5777 ASSERT(HDR_HAS_RABD(hdr));
5778 size = HDR_GET_PSIZE(hdr);
5779 hdr_abd = hdr->b_crypt_hdr.b_rabd;
5780 zio_flags |= ZIO_FLAG_RAW;
5781 } else {
5782 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5783 size = arc_hdr_size(hdr);
5784 hdr_abd = hdr->b_l1hdr.b_pabd;
5785
5786 if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) {
5787 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
5788 }
5789
5790 /*
5791 * For authenticated bp's, we do not ask the ZIO layer
5792 * to authenticate them since this will cause the entire
5793 * IO to fail if the key isn't loaded. Instead, we
5794 * defer authentication until arc_buf_fill(), which will
5795 * verify the data when the key is available.
5796 */
5797 if (BP_IS_AUTHENTICATED(bp))
5798 zio_flags |= ZIO_FLAG_RAW_ENCRYPT;
5799 }
5800
5801 if (*arc_flags & ARC_FLAG_PREFETCH &&
5802 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))
5803 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5804 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH)
5805 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH);
5806
5807 if (*arc_flags & ARC_FLAG_L2CACHE)
5808 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5809 if (BP_IS_AUTHENTICATED(bp))
5810 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
5811 if (BP_GET_LEVEL(bp) > 0)
5812 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5813 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5814 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5815 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5816
5817 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5818 acb->acb_done = done;
5819 acb->acb_private = private;
5820 acb->acb_compressed = compressed_read;
5821 acb->acb_encrypted = encrypted_read;
5822 acb->acb_noauth = noauth_read;
5823 acb->acb_zb = *zb;
5824
5825 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5826 hdr->b_l1hdr.b_acb = acb;
5827 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5828
5829 if (HDR_HAS_L2HDR(hdr) &&
5830 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5831 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5832 addr = hdr->b_l2hdr.b_daddr;
5833 /*
5834 * Lock out L2ARC device removal.
5835 */
5836 if (vdev_is_dead(vd) ||
5837 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5838 vd = NULL;
5839 }
5840
5841 /*
5842 * We count both async reads and scrub IOs as asynchronous so
5843 * that both can be upgraded in the event of a cache hit while
5844 * the read IO is still in-flight.
5845 */
5846 if (priority == ZIO_PRIORITY_ASYNC_READ ||
5847 priority == ZIO_PRIORITY_SCRUB)
5848 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5849 else
5850 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5851
5852 /*
5853 * At this point, we have a level 1 cache miss. Try again in
5854 * L2ARC if possible.
5855 */
5856 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5857
5858 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5859 uint64_t, lsize, zbookmark_phys_t *, zb);
5860 ARCSTAT_BUMP(arcstat_misses);
5861 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5862 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5863 data, metadata, misses);
5864
5865 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5866 /*
5867 * Read from the L2ARC if the following are true:
5868 * 1. The L2ARC vdev was previously cached.
5869 * 2. This buffer still has L2ARC metadata.
5870 * 3. This buffer isn't currently writing to the L2ARC.
5871 * 4. The L2ARC entry wasn't evicted, which may
5872 * also have invalidated the vdev.
5873 * 5. This isn't prefetch and l2arc_noprefetch is set.
5874 */
5875 if (HDR_HAS_L2HDR(hdr) &&
5876 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5877 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5878 l2arc_read_callback_t *cb;
5879 abd_t *abd;
5880 uint64_t asize;
5881
5882 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5883 ARCSTAT_BUMP(arcstat_l2_hits);
5884
5885 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5886 KM_SLEEP);
5887 cb->l2rcb_hdr = hdr;
5888 cb->l2rcb_bp = *bp;
5889 cb->l2rcb_zb = *zb;
5890 cb->l2rcb_flags = zio_flags;
5891
5892 asize = vdev_psize_to_asize(vd, size);
5893 if (asize != size) {
5894 abd = abd_alloc_for_io(asize,
5895 HDR_ISTYPE_METADATA(hdr));
5896 cb->l2rcb_abd = abd;
5897 } else {
5898 abd = hdr_abd;
5899 }
5900
5901 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5902 addr + asize <= vd->vdev_psize -
5903 VDEV_LABEL_END_SIZE);
5904
5905 /*
5906 * l2arc read. The SCL_L2ARC lock will be
5907 * released by l2arc_read_done().
5908 * Issue a null zio if the underlying buffer
5909 * was squashed to zero size by compression.
5910 */
5911 ASSERT3U(arc_hdr_get_compress(hdr), !=,
5912 ZIO_COMPRESS_EMPTY);
5913 rzio = zio_read_phys(pio, vd, addr,
5914 asize, abd,
5915 ZIO_CHECKSUM_OFF,
5916 l2arc_read_done, cb, priority,
5917 zio_flags | ZIO_FLAG_DONT_CACHE |
5918 ZIO_FLAG_CANFAIL |
5919 ZIO_FLAG_DONT_PROPAGATE |
5920 ZIO_FLAG_DONT_RETRY, B_FALSE);
5921 acb->acb_zio_head = rzio;
5922
5923 if (hash_lock != NULL)
5924 mutex_exit(hash_lock);
5925
5926 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5927 zio_t *, rzio);
5928 ARCSTAT_INCR(arcstat_l2_read_bytes,
5929 HDR_GET_PSIZE(hdr));
5930
5931 if (*arc_flags & ARC_FLAG_NOWAIT) {
5932 zio_nowait(rzio);
5933 return (0);
5934 }
5935
5936 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5937 if (zio_wait(rzio) == 0)
5938 return (0);
5939
5940 /* l2arc read error; goto zio_read() */
5941 if (hash_lock != NULL)
5942 mutex_enter(hash_lock);
5943 } else {
5944 DTRACE_PROBE1(l2arc__miss,
5945 arc_buf_hdr_t *, hdr);
5946 ARCSTAT_BUMP(arcstat_l2_misses);
5947 if (HDR_L2_WRITING(hdr))
5948 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5949 spa_config_exit(spa, SCL_L2ARC, vd);
5950 }
5951 } else {
5952 if (vd != NULL)
5953 spa_config_exit(spa, SCL_L2ARC, vd);
5954 if (l2arc_ndev != 0) {
5955 DTRACE_PROBE1(l2arc__miss,
5956 arc_buf_hdr_t *, hdr);
5957 ARCSTAT_BUMP(arcstat_l2_misses);
5958 }
5959 }
5960
5961 rzio = zio_read(pio, spa, bp, hdr_abd, size,
5962 arc_read_done, hdr, priority, zio_flags, zb);
5963 acb->acb_zio_head = rzio;
5964
5965 if (hash_lock != NULL)
5966 mutex_exit(hash_lock);
5967
5968 if (*arc_flags & ARC_FLAG_WAIT)
5969 return (zio_wait(rzio));
5970
5971 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
5972 zio_nowait(rzio);
5973 }
5974 return (rc);
5975 }
5976
5977 /*
5978 * Notify the arc that a block was freed, and thus will never be used again.
5979 */
5980 void
5981 arc_freed(spa_t *spa, const blkptr_t *bp)
5982 {
5983 arc_buf_hdr_t *hdr;
5984 kmutex_t *hash_lock;
5985 uint64_t guid = spa_load_guid(spa);
5986
5987 ASSERT(!BP_IS_EMBEDDED(bp));
5988
5989 hdr = buf_hash_find(guid, bp, &hash_lock);
5990 if (hdr == NULL)
5991 return;
5992
5993 /*
5994 * We might be trying to free a block that is still doing I/O
5995 * (i.e. prefetch) or has a reference (i.e. a dedup-ed,
5996 * dmu_sync-ed block). If this block is being prefetched, then it
5997 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr
5998 * until the I/O completes. A block may also have a reference if it is
5999 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would
6000 * have written the new block to its final resting place on disk but
6001 * without the dedup flag set. This would have left the hdr in the MRU
6002 * state and discoverable. When the txg finally syncs it detects that
6003 * the block was overridden in open context and issues an override I/O.
6004 * Since this is a dedup block, the override I/O will determine if the
6005 * block is already in the DDT. If so, then it will replace the io_bp
6006 * with the bp from the DDT and allow the I/O to finish. When the I/O
6007 * reaches the done callback, dbuf_write_override_done, it will
6008 * check to see if the io_bp and io_bp_override are identical.
6009 * If they are not, then it indicates that the bp was replaced with
6010 * the bp in the DDT and the override bp is freed. This allows
6011 * us to arrive here with a reference on a block that is being
6012 * freed. So if we have an I/O in progress, or a reference to
6013 * this hdr, then we don't destroy the hdr.
6014 */
6015 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) &&
6016 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) {
6017 arc_change_state(arc_anon, hdr, hash_lock);
6018 arc_hdr_destroy(hdr);
6019 mutex_exit(hash_lock);
6020 } else {
6021 mutex_exit(hash_lock);
6022 }
6023
6024 }
6025
6026 /*
6027 * Release this buffer from the cache, making it an anonymous buffer. This
6028 * must be done after a read and prior to modifying the buffer contents.
6029 * If the buffer has more than one reference, we must make
6030 * a new hdr for the buffer.
6031 */
6032 void
6033 arc_release(arc_buf_t *buf, void *tag)
6034 {
6035 arc_buf_hdr_t *hdr = buf->b_hdr;
6036
6037 /*
6038 * It would be nice to assert that if its DMU metadata (level >
6039 * 0 || it's the dnode file), then it must be syncing context.
6040 * But we don't know that information at this level.
6041 */
6042
6043 mutex_enter(&buf->b_evict_lock);
6044
6045 ASSERT(HDR_HAS_L1HDR(hdr));
6046
6047 /*
6048 * We don't grab the hash lock prior to this check, because if
6049 * the buffer's header is in the arc_anon state, it won't be
6050 * linked into the hash table.
6051 */
6052 if (hdr->b_l1hdr.b_state == arc_anon) {
6053 mutex_exit(&buf->b_evict_lock);
6054 /*
6055 * If we are called from dmu_convert_mdn_block_to_raw(),
6056 * a write might be in progress. This is OK because
6057 * the caller won't change the content of this buffer,
6058 * only the flags (via arc_convert_to_raw()).
6059 */
6060 /* ASSERT(!HDR_IO_IN_PROGRESS(hdr)); */
6061 ASSERT(!HDR_IN_HASH_TABLE(hdr));
6062 ASSERT(!HDR_HAS_L2HDR(hdr));
6063 ASSERT(HDR_EMPTY(hdr));
6064
6065 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6066 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
6067 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
6068
6069 hdr->b_l1hdr.b_arc_access = 0;
6070
6071 /*
6072 * If the buf is being overridden then it may already
6073 * have a hdr that is not empty.
6074 */
6075 buf_discard_identity(hdr);
6076 arc_buf_thaw(buf);
6077
6078 return;
6079 }
6080
6081 kmutex_t *hash_lock = HDR_LOCK(hdr);
6082 mutex_enter(hash_lock);
6083
6084 /*
6085 * This assignment is only valid as long as the hash_lock is
6086 * held, we must be careful not to reference state or the
6087 * b_state field after dropping the lock.
6088 */
6089 arc_state_t *state = hdr->b_l1hdr.b_state;
6090 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
6091 ASSERT3P(state, !=, arc_anon);
6092
6093 /* this buffer is not on any list */
6094 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0);
6095
6096 if (HDR_HAS_L2HDR(hdr)) {
6097 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6098
6099 /*
6100 * We have to recheck this conditional again now that
6101 * we're holding the l2ad_mtx to prevent a race with
6102 * another thread which might be concurrently calling
6103 * l2arc_evict(). In that case, l2arc_evict() might have
6104 * destroyed the header's L2 portion as we were waiting
6105 * to acquire the l2ad_mtx.
6106 */
6107 if (HDR_HAS_L2HDR(hdr))
6108 arc_hdr_l2hdr_destroy(hdr);
6109
6110 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
6111 }
6112
6113 /*
6114 * Do we have more than one buf?
6115 */
6116 if (hdr->b_l1hdr.b_bufcnt > 1) {
6117 arc_buf_hdr_t *nhdr;
6118 uint64_t spa = hdr->b_spa;
6119 uint64_t psize = HDR_GET_PSIZE(hdr);
6120 uint64_t lsize = HDR_GET_LSIZE(hdr);
6121 boolean_t protected = HDR_PROTECTED(hdr);
6122 enum zio_compress compress = arc_hdr_get_compress(hdr);
6123 arc_buf_contents_t type = arc_buf_type(hdr);
6124 VERIFY3U(hdr->b_type, ==, type);
6125
6126 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
6127 (void) remove_reference(hdr, hash_lock, tag);
6128
6129 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
6130 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6131 ASSERT(ARC_BUF_LAST(buf));
6132 }
6133
6134 /*
6135 * Pull the data off of this hdr and attach it to
6136 * a new anonymous hdr. Also find the last buffer
6137 * in the hdr's buffer list.
6138 */
6139 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf);
6140 ASSERT3P(lastbuf, !=, NULL);
6141
6142 /*
6143 * If the current arc_buf_t and the hdr are sharing their data
6144 * buffer, then we must stop sharing that block.
6145 */
6146 if (arc_buf_is_shared(buf)) {
6147 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
6148 VERIFY(!arc_buf_is_shared(lastbuf));
6149
6150 /*
6151 * First, sever the block sharing relationship between
6152 * buf and the arc_buf_hdr_t.
6153 */
6154 arc_unshare_buf(hdr, buf);
6155
6156 /*
6157 * Now we need to recreate the hdr's b_pabd. Since we
6158 * have lastbuf handy, we try to share with it, but if
6159 * we can't then we allocate a new b_pabd and copy the
6160 * data from buf into it.
6161 */
6162 if (arc_can_share(hdr, lastbuf)) {
6163 arc_share_buf(hdr, lastbuf);
6164 } else {
6165 arc_hdr_alloc_pabd(hdr, B_FALSE);
6166 abd_copy_from_buf(hdr->b_l1hdr.b_pabd,
6167 buf->b_data, psize);
6168 }
6169 VERIFY3P(lastbuf->b_data, !=, NULL);
6170 } else if (HDR_SHARED_DATA(hdr)) {
6171 /*
6172 * Uncompressed shared buffers are always at the end
6173 * of the list. Compressed buffers don't have the
6174 * same requirements. This makes it hard to
6175 * simply assert that the lastbuf is shared so
6176 * we rely on the hdr's compression flags to determine
6177 * if we have a compressed, shared buffer.
6178 */
6179 ASSERT(arc_buf_is_shared(lastbuf) ||
6180 arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF);
6181 ASSERT(!ARC_BUF_SHARED(buf));
6182 }
6183 ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr));
6184 ASSERT3P(state, !=, arc_l2c_only);
6185
6186 (void) zfs_refcount_remove_many(&state->arcs_size,
6187 arc_buf_size(buf), buf);
6188
6189 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
6190 ASSERT3P(state, !=, arc_l2c_only);
6191 (void) zfs_refcount_remove_many(
6192 &state->arcs_esize[type],
6193 arc_buf_size(buf), buf);
6194 }
6195
6196 hdr->b_l1hdr.b_bufcnt -= 1;
6197 if (ARC_BUF_ENCRYPTED(buf))
6198 hdr->b_crypt_hdr.b_ebufcnt -= 1;
6199
6200 arc_cksum_verify(buf);
6201 arc_buf_unwatch(buf);
6202
6203 /* if this is the last uncompressed buf free the checksum */
6204 if (!arc_hdr_has_uncompressed_buf(hdr))
6205 arc_cksum_free(hdr);
6206
6207 mutex_exit(hash_lock);
6208
6209 /*
6210 * Allocate a new hdr. The new hdr will contain a b_pabd
6211 * buffer which will be freed in arc_write().
6212 */
6213 nhdr = arc_hdr_alloc(spa, psize, lsize, protected,
6214 compress, type, HDR_HAS_RABD(hdr));
6215 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6216 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6217 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt));
6218 VERIFY3U(nhdr->b_type, ==, type);
6219 ASSERT(!HDR_SHARED_DATA(nhdr));
6220
6221 nhdr->b_l1hdr.b_buf = buf;
6222 nhdr->b_l1hdr.b_bufcnt = 1;
6223 if (ARC_BUF_ENCRYPTED(buf))
6224 nhdr->b_crypt_hdr.b_ebufcnt = 1;
6225 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6226 buf->b_hdr = nhdr;
6227
6228 mutex_exit(&buf->b_evict_lock);
6229 (void) zfs_refcount_add_many(&arc_anon->arcs_size,
6230 arc_buf_size(buf), buf);
6231 } else {
6232 mutex_exit(&buf->b_evict_lock);
6233 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6234 /* protected by hash lock, or hdr is on arc_anon */
6235 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6236 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6237 arc_change_state(arc_anon, hdr, hash_lock);
6238 hdr->b_l1hdr.b_arc_access = 0;
6239
6240 mutex_exit(hash_lock);
6241 buf_discard_identity(hdr);
6242 arc_buf_thaw(buf);
6243 }
6244 }
6245
6246 int
6247 arc_released(arc_buf_t *buf)
6248 {
6249 int released;
6250
6251 mutex_enter(&buf->b_evict_lock);
6252 released = (buf->b_data != NULL &&
6253 buf->b_hdr->b_l1hdr.b_state == arc_anon);
6254 mutex_exit(&buf->b_evict_lock);
6255 return (released);
6256 }
6257
6258 #ifdef ZFS_DEBUG
6259 int
6260 arc_referenced(arc_buf_t *buf)
6261 {
6262 int referenced;
6263
6264 mutex_enter(&buf->b_evict_lock);
6265 referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
6266 mutex_exit(&buf->b_evict_lock);
6267 return (referenced);
6268 }
6269 #endif
6270
6271 static void
6272 arc_write_ready(zio_t *zio)
6273 {
6274 arc_write_callback_t *callback = zio->io_private;
6275 arc_buf_t *buf = callback->awcb_buf;
6276 arc_buf_hdr_t *hdr = buf->b_hdr;
6277 blkptr_t *bp = zio->io_bp;
6278 uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp);
6279
6280 ASSERT(HDR_HAS_L1HDR(hdr));
6281 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
6282 ASSERT(hdr->b_l1hdr.b_bufcnt > 0);
6283
6284 /*
6285 * If we're reexecuting this zio because the pool suspended, then
6286 * cleanup any state that was previously set the first time the
6287 * callback was invoked.
6288 */
6289 if (zio->io_flags & ZIO_FLAG_REEXECUTED) {
6290 arc_cksum_free(hdr);
6291 arc_buf_unwatch(buf);
6292 if (hdr->b_l1hdr.b_pabd != NULL) {
6293 if (arc_buf_is_shared(buf)) {
6294 arc_unshare_buf(hdr, buf);
6295 } else {
6296 arc_hdr_free_pabd(hdr, B_FALSE);
6297 }
6298 }
6299
6300 if (HDR_HAS_RABD(hdr))
6301 arc_hdr_free_pabd(hdr, B_TRUE);
6302 }
6303 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6304 ASSERT(!HDR_HAS_RABD(hdr));
6305 ASSERT(!HDR_SHARED_DATA(hdr));
6306 ASSERT(!arc_buf_is_shared(buf));
6307
6308 callback->awcb_ready(zio, buf, callback->awcb_private);
6309
6310 if (HDR_IO_IN_PROGRESS(hdr))
6311 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED);
6312
6313 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6314
6315 if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr))
6316 hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp));
6317
6318 if (BP_IS_PROTECTED(bp)) {
6319 /* ZIL blocks are written through zio_rewrite */
6320 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
6321 ASSERT(HDR_PROTECTED(hdr));
6322
6323 if (BP_SHOULD_BYTESWAP(bp)) {
6324 if (BP_GET_LEVEL(bp) > 0) {
6325 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64;
6326 } else {
6327 hdr->b_l1hdr.b_byteswap =
6328 DMU_OT_BYTESWAP(BP_GET_TYPE(bp));
6329 }
6330 } else {
6331 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
6332 }
6333
6334 hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp);
6335 hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset;
6336 zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt,
6337 hdr->b_crypt_hdr.b_iv);
6338 zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac);
6339 }
6340
6341 /*
6342 * If this block was written for raw encryption but the zio layer
6343 * ended up only authenticating it, adjust the buffer flags now.
6344 */
6345 if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) {
6346 arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH);
6347 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6348 if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF)
6349 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6350 } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) {
6351 buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED;
6352 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
6353 }
6354
6355 /* this must be done after the buffer flags are adjusted */
6356 arc_cksum_compute(buf);
6357
6358 enum zio_compress compress;
6359 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) {
6360 compress = ZIO_COMPRESS_OFF;
6361 } else {
6362 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp));
6363 compress = BP_GET_COMPRESS(bp);
6364 }
6365 HDR_SET_PSIZE(hdr, psize);
6366 arc_hdr_set_compress(hdr, compress);
6367
6368 if (zio->io_error != 0 || psize == 0)
6369 goto out;
6370
6371 /*
6372 * Fill the hdr with data. If the buffer is encrypted we have no choice
6373 * but to copy the data into b_rabd. If the hdr is compressed, the data
6374 * we want is available from the zio, otherwise we can take it from
6375 * the buf.
6376 *
6377 * We might be able to share the buf's data with the hdr here. However,
6378 * doing so would cause the ARC to be full of linear ABDs if we write a
6379 * lot of shareable data. As a compromise, we check whether scattered
6380 * ABDs are allowed, and assume that if they are then the user wants
6381 * the ARC to be primarily filled with them regardless of the data being
6382 * written. Therefore, if they're allowed then we allocate one and copy
6383 * the data into it; otherwise, we share the data directly if we can.
6384 */
6385 if (ARC_BUF_ENCRYPTED(buf)) {
6386 ASSERT3U(psize, >, 0);
6387 ASSERT(ARC_BUF_COMPRESSED(buf));
6388 arc_hdr_alloc_pabd(hdr, B_TRUE);
6389 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6390 } else if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) {
6391 /*
6392 * Ideally, we would always copy the io_abd into b_pabd, but the
6393 * user may have disabled compressed ARC, thus we must check the
6394 * hdr's compression setting rather than the io_bp's.
6395 */
6396 if (BP_IS_ENCRYPTED(bp)) {
6397 ASSERT3U(psize, >, 0);
6398 arc_hdr_alloc_pabd(hdr, B_TRUE);
6399 abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
6400 } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF &&
6401 !ARC_BUF_COMPRESSED(buf)) {
6402 ASSERT3U(psize, >, 0);
6403 arc_hdr_alloc_pabd(hdr, B_FALSE);
6404 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize);
6405 } else {
6406 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr));
6407 arc_hdr_alloc_pabd(hdr, B_FALSE);
6408 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data,
6409 arc_buf_size(buf));
6410 }
6411 } else {
6412 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd));
6413 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf));
6414 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
6415 arc_share_buf(hdr, buf);
6416 }
6417
6418 out:
6419 arc_hdr_verify(hdr, bp);
6420 }
6421
6422 static void
6423 arc_write_children_ready(zio_t *zio)
6424 {
6425 arc_write_callback_t *callback = zio->io_private;
6426 arc_buf_t *buf = callback->awcb_buf;
6427
6428 callback->awcb_children_ready(zio, buf, callback->awcb_private);
6429 }
6430
6431 /*
6432 * The SPA calls this callback for each physical write that happens on behalf
6433 * of a logical write. See the comment in dbuf_write_physdone() for details.
6434 */
6435 static void
6436 arc_write_physdone(zio_t *zio)
6437 {
6438 arc_write_callback_t *cb = zio->io_private;
6439 if (cb->awcb_physdone != NULL)
6440 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
6441 }
6442
6443 static void
6444 arc_write_done(zio_t *zio)
6445 {
6446 arc_write_callback_t *callback = zio->io_private;
6447 arc_buf_t *buf = callback->awcb_buf;
6448 arc_buf_hdr_t *hdr = buf->b_hdr;
6449
6450 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6451
6452 if (zio->io_error == 0) {
6453 arc_hdr_verify(hdr, zio->io_bp);
6454
6455 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
6456 buf_discard_identity(hdr);
6457 } else {
6458 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
6459 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
6460 }
6461 } else {
6462 ASSERT(HDR_EMPTY(hdr));
6463 }
6464
6465 /*
6466 * If the block to be written was all-zero or compressed enough to be
6467 * embedded in the BP, no write was performed so there will be no
6468 * dva/birth/checksum. The buffer must therefore remain anonymous
6469 * (and uncached).
6470 */
6471 if (!HDR_EMPTY(hdr)) {
6472 arc_buf_hdr_t *exists;
6473 kmutex_t *hash_lock;
6474
6475 ASSERT3U(zio->io_error, ==, 0);
6476
6477 arc_cksum_verify(buf);
6478
6479 exists = buf_hash_insert(hdr, &hash_lock);
6480 if (exists != NULL) {
6481 /*
6482 * This can only happen if we overwrite for
6483 * sync-to-convergence, because we remove
6484 * buffers from the hash table when we arc_free().
6485 */
6486 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6487 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6488 panic("bad overwrite, hdr=%p exists=%p",
6489 (void *)hdr, (void *)exists);
6490 ASSERT(zfs_refcount_is_zero(
6491 &exists->b_l1hdr.b_refcnt));
6492 arc_change_state(arc_anon, exists, hash_lock);
6493 arc_hdr_destroy(exists);
6494 mutex_exit(hash_lock);
6495 exists = buf_hash_insert(hdr, &hash_lock);
6496 ASSERT3P(exists, ==, NULL);
6497 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6498 /* nopwrite */
6499 ASSERT(zio->io_prop.zp_nopwrite);
6500 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6501 panic("bad nopwrite, hdr=%p exists=%p",
6502 (void *)hdr, (void *)exists);
6503 } else {
6504 /* Dedup */
6505 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6506 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6507 ASSERT(BP_GET_DEDUP(zio->io_bp));
6508 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6509 }
6510 }
6511 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6512 /* if it's not anon, we are doing a scrub */
6513 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6514 arc_access(hdr, hash_lock);
6515 mutex_exit(hash_lock);
6516 } else {
6517 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6518 }
6519
6520 ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6521 callback->awcb_done(zio, buf, callback->awcb_private);
6522
6523 abd_put(zio->io_abd);
6524 kmem_free(callback, sizeof (arc_write_callback_t));
6525 }
6526
6527 zio_t *
6528 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6529 boolean_t l2arc, const zio_prop_t *zp, arc_write_done_func_t *ready,
6530 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
6531 arc_write_done_func_t *done, void *private, zio_priority_t priority,
6532 int zio_flags, const zbookmark_phys_t *zb)
6533 {
6534 arc_buf_hdr_t *hdr = buf->b_hdr;
6535 arc_write_callback_t *callback;
6536 zio_t *zio;
6537 zio_prop_t localprop = *zp;
6538
6539 ASSERT3P(ready, !=, NULL);
6540 ASSERT3P(done, !=, NULL);
6541 ASSERT(!HDR_IO_ERROR(hdr));
6542 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6543 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6544 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6545 if (l2arc)
6546 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6547
6548 if (ARC_BUF_ENCRYPTED(buf)) {
6549 ASSERT(ARC_BUF_COMPRESSED(buf));
6550 localprop.zp_encrypt = B_TRUE;
6551 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6552 /* CONSTCOND */
6553 localprop.zp_byteorder =
6554 (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ?
6555 ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER;
6556 bcopy(hdr->b_crypt_hdr.b_salt, localprop.zp_salt,
6557 ZIO_DATA_SALT_LEN);
6558 bcopy(hdr->b_crypt_hdr.b_iv, localprop.zp_iv,
6559 ZIO_DATA_IV_LEN);
6560 bcopy(hdr->b_crypt_hdr.b_mac, localprop.zp_mac,
6561 ZIO_DATA_MAC_LEN);
6562 if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) {
6563 localprop.zp_nopwrite = B_FALSE;
6564 localprop.zp_copies =
6565 MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1);
6566 }
6567 zio_flags |= ZIO_FLAG_RAW;
6568 } else if (ARC_BUF_COMPRESSED(buf)) {
6569 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf));
6570 localprop.zp_compress = HDR_GET_COMPRESS(hdr);
6571 zio_flags |= ZIO_FLAG_RAW_COMPRESS;
6572 }
6573
6574 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
6575 callback->awcb_ready = ready;
6576 callback->awcb_children_ready = children_ready;
6577 callback->awcb_physdone = physdone;
6578 callback->awcb_done = done;
6579 callback->awcb_private = private;
6580 callback->awcb_buf = buf;
6581
6582 /*
6583 * The hdr's b_pabd is now stale, free it now. A new data block
6584 * will be allocated when the zio pipeline calls arc_write_ready().
6585 */
6586 if (hdr->b_l1hdr.b_pabd != NULL) {
6587 /*
6588 * If the buf is currently sharing the data block with
6589 * the hdr then we need to break that relationship here.
6590 * The hdr will remain with a NULL data pointer and the
6591 * buf will take sole ownership of the block.
6592 */
6593 if (arc_buf_is_shared(buf)) {
6594 arc_unshare_buf(hdr, buf);
6595 } else {
6596 arc_hdr_free_pabd(hdr, B_FALSE);
6597 }
6598 VERIFY3P(buf->b_data, !=, NULL);
6599 }
6600
6601 if (HDR_HAS_RABD(hdr))
6602 arc_hdr_free_pabd(hdr, B_TRUE);
6603
6604 if (!(zio_flags & ZIO_FLAG_RAW))
6605 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6606
6607 ASSERT(!arc_buf_is_shared(buf));
6608 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6609
6610 zio = zio_write(pio, spa, txg, bp,
6611 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6612 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6613 (children_ready != NULL) ? arc_write_children_ready : NULL,
6614 arc_write_physdone, arc_write_done, callback,
6615 priority, zio_flags, zb);
6616
6617 return (zio);
6618 }
6619
6620 static int
6621 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
6622 {
6623 #ifdef _KERNEL
6624 uint64_t available_memory = ptob(freemem);
6625
6626 #if defined(__i386)
6627 available_memory =
6628 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
6629 #endif
6630
6631 if (freemem > physmem * arc_lotsfree_percent / 100)
6632 return (0);
6633
6634 if (txg > spa->spa_lowmem_last_txg) {
6635 spa->spa_lowmem_last_txg = txg;
6636 spa->spa_lowmem_page_load = 0;
6637 }
6638 /*
6639 * If we are in pageout, we know that memory is already tight,
6640 * the arc is already going to be evicting, so we just want to
6641 * continue to let page writes occur as quickly as possible.
6642 */
6643 if (curproc == proc_pageout) {
6644 if (spa->spa_lowmem_page_load >
6645 MAX(ptob(minfree), available_memory) / 4)
6646 return (SET_ERROR(ERESTART));
6647 /* Note: reserve is inflated, so we deflate */
6648 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
6649 return (0);
6650 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
6651 /* memory is low, delay before restarting */
6652 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
6653 return (SET_ERROR(EAGAIN));
6654 }
6655 spa->spa_lowmem_page_load = 0;
6656 #endif /* _KERNEL */
6657 return (0);
6658 }
6659
6660 void
6661 arc_tempreserve_clear(uint64_t reserve)
6662 {
6663 atomic_add_64(&arc_tempreserve, -reserve);
6664 ASSERT((int64_t)arc_tempreserve >= 0);
6665 }
6666
6667 int
6668 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg)
6669 {
6670 int error;
6671 uint64_t anon_size;
6672
6673 if (reserve > arc_c/4 && !arc_no_grow)
6674 arc_c = MIN(arc_c_max, reserve * 4);
6675 if (reserve > arc_c)
6676 return (SET_ERROR(ENOMEM));
6677
6678 /*
6679 * Don't count loaned bufs as in flight dirty data to prevent long
6680 * network delays from blocking transactions that are ready to be
6681 * assigned to a txg.
6682 */
6683
6684 /* assert that it has not wrapped around */
6685 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
6686
6687 anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) -
6688 arc_loaned_bytes), 0);
6689
6690 /*
6691 * Writes will, almost always, require additional memory allocations
6692 * in order to compress/encrypt/etc the data. We therefore need to
6693 * make sure that there is sufficient available memory for this.
6694 */
6695 error = arc_memory_throttle(spa, reserve, txg);
6696 if (error != 0)
6697 return (error);
6698
6699 /*
6700 * Throttle writes when the amount of dirty data in the cache
6701 * gets too large. We try to keep the cache less than half full
6702 * of dirty blocks so that our sync times don't grow too large.
6703 *
6704 * In the case of one pool being built on another pool, we want
6705 * to make sure we don't end up throttling the lower (backing)
6706 * pool when the upper pool is the majority contributor to dirty
6707 * data. To insure we make forward progress during throttling, we
6708 * also check the current pool's net dirty data and only throttle
6709 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty
6710 * data in the cache.
6711 *
6712 * Note: if two requests come in concurrently, we might let them
6713 * both succeed, when one of them should fail. Not a huge deal.
6714 */
6715 uint64_t total_dirty = reserve + arc_tempreserve + anon_size;
6716 uint64_t spa_dirty_anon = spa_dirty_data(spa);
6717
6718 if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 &&
6719 anon_size > arc_c * zfs_arc_anon_limit_percent / 100 &&
6720 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) {
6721 uint64_t meta_esize =
6722 zfs_refcount_count(
6723 &arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6724 uint64_t data_esize =
6725 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6726 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6727 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6728 arc_tempreserve >> 10, meta_esize >> 10,
6729 data_esize >> 10, reserve >> 10, arc_c >> 10);
6730 return (SET_ERROR(ERESTART));
6731 }
6732 atomic_add_64(&arc_tempreserve, reserve);
6733 return (0);
6734 }
6735
6736 static void
6737 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6738 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
6739 {
6740 size->value.ui64 = zfs_refcount_count(&state->arcs_size);
6741 evict_data->value.ui64 =
6742 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6743 evict_metadata->value.ui64 =
6744 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6745 }
6746
6747 static int
6748 arc_kstat_update(kstat_t *ksp, int rw)
6749 {
6750 arc_stats_t *as = ksp->ks_data;
6751
6752 if (rw == KSTAT_WRITE) {
6753 return (EACCES);
6754 } else {
6755 arc_kstat_update_state(arc_anon,
6756 &as->arcstat_anon_size,
6757 &as->arcstat_anon_evictable_data,
6758 &as->arcstat_anon_evictable_metadata);
6759 arc_kstat_update_state(arc_mru,
6760 &as->arcstat_mru_size,
6761 &as->arcstat_mru_evictable_data,
6762 &as->arcstat_mru_evictable_metadata);
6763 arc_kstat_update_state(arc_mru_ghost,
6764 &as->arcstat_mru_ghost_size,
6765 &as->arcstat_mru_ghost_evictable_data,
6766 &as->arcstat_mru_ghost_evictable_metadata);
6767 arc_kstat_update_state(arc_mfu,
6768 &as->arcstat_mfu_size,
6769 &as->arcstat_mfu_evictable_data,
6770 &as->arcstat_mfu_evictable_metadata);
6771 arc_kstat_update_state(arc_mfu_ghost,
6772 &as->arcstat_mfu_ghost_size,
6773 &as->arcstat_mfu_ghost_evictable_data,
6774 &as->arcstat_mfu_ghost_evictable_metadata);
6775
6776 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
6777 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
6778 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
6779 ARCSTAT(arcstat_metadata_size) =
6780 aggsum_value(&astat_metadata_size);
6781 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
6782 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size);
6783 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
6784 }
6785
6786 return (0);
6787 }
6788
6789 /*
6790 * This function *must* return indices evenly distributed between all
6791 * sublists of the multilist. This is needed due to how the ARC eviction
6792 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6793 * distributed between all sublists and uses this assumption when
6794 * deciding which sublist to evict from and how much to evict from it.
6795 */
6796 unsigned int
6797 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6798 {
6799 arc_buf_hdr_t *hdr = obj;
6800
6801 /*
6802 * We rely on b_dva to generate evenly distributed index
6803 * numbers using buf_hash below. So, as an added precaution,
6804 * let's make sure we never add empty buffers to the arc lists.
6805 */
6806 ASSERT(!HDR_EMPTY(hdr));
6807
6808 /*
6809 * The assumption here, is the hash value for a given
6810 * arc_buf_hdr_t will remain constant throughout its lifetime
6811 * (i.e. its b_spa, b_dva, and b_birth fields don't change).
6812 * Thus, we don't need to store the header's sublist index
6813 * on insertion, as this index can be recalculated on removal.
6814 *
6815 * Also, the low order bits of the hash value are thought to be
6816 * distributed evenly. Otherwise, in the case that the multilist
6817 * has a power of two number of sublists, each sublists' usage
6818 * would not be evenly distributed.
6819 */
6820 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6821 multilist_get_num_sublists(ml));
6822 }
6823
6824 static void
6825 arc_state_init(void)
6826 {
6827 arc_anon = &ARC_anon;
6828 arc_mru = &ARC_mru;
6829 arc_mru_ghost = &ARC_mru_ghost;
6830 arc_mfu = &ARC_mfu;
6831 arc_mfu_ghost = &ARC_mfu_ghost;
6832 arc_l2c_only = &ARC_l2c_only;
6833
6834 arc_mru->arcs_list[ARC_BUFC_METADATA] =
6835 multilist_create(sizeof (arc_buf_hdr_t),
6836 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6837 arc_state_multilist_index_func);
6838 arc_mru->arcs_list[ARC_BUFC_DATA] =
6839 multilist_create(sizeof (arc_buf_hdr_t),
6840 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6841 arc_state_multilist_index_func);
6842 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
6843 multilist_create(sizeof (arc_buf_hdr_t),
6844 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6845 arc_state_multilist_index_func);
6846 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
6847 multilist_create(sizeof (arc_buf_hdr_t),
6848 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6849 arc_state_multilist_index_func);
6850 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
6851 multilist_create(sizeof (arc_buf_hdr_t),
6852 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6853 arc_state_multilist_index_func);
6854 arc_mfu->arcs_list[ARC_BUFC_DATA] =
6855 multilist_create(sizeof (arc_buf_hdr_t),
6856 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6857 arc_state_multilist_index_func);
6858 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
6859 multilist_create(sizeof (arc_buf_hdr_t),
6860 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6861 arc_state_multilist_index_func);
6862 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
6863 multilist_create(sizeof (arc_buf_hdr_t),
6864 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6865 arc_state_multilist_index_func);
6866 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
6867 multilist_create(sizeof (arc_buf_hdr_t),
6868 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6869 arc_state_multilist_index_func);
6870 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6871 multilist_create(sizeof (arc_buf_hdr_t),
6872 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6873 arc_state_multilist_index_func);
6874
6875 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6876 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6877 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6878 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6879 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6880 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6881 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6882 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6883 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6884 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6885 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6886 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6887
6888 zfs_refcount_create(&arc_anon->arcs_size);
6889 zfs_refcount_create(&arc_mru->arcs_size);
6890 zfs_refcount_create(&arc_mru_ghost->arcs_size);
6891 zfs_refcount_create(&arc_mfu->arcs_size);
6892 zfs_refcount_create(&arc_mfu_ghost->arcs_size);
6893 zfs_refcount_create(&arc_l2c_only->arcs_size);
6894
6895 aggsum_init(&arc_meta_used, 0);
6896 aggsum_init(&arc_size, 0);
6897 aggsum_init(&astat_data_size, 0);
6898 aggsum_init(&astat_metadata_size, 0);
6899 aggsum_init(&astat_hdr_size, 0);
6900 aggsum_init(&astat_other_size, 0);
6901 aggsum_init(&astat_l2_hdr_size, 0);
6902 }
6903
6904 static void
6905 arc_state_fini(void)
6906 {
6907 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6908 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6909 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6910 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6911 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6912 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6913 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6914 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6915 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6916 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6917 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6918 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6919
6920 zfs_refcount_destroy(&arc_anon->arcs_size);
6921 zfs_refcount_destroy(&arc_mru->arcs_size);
6922 zfs_refcount_destroy(&arc_mru_ghost->arcs_size);
6923 zfs_refcount_destroy(&arc_mfu->arcs_size);
6924 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size);
6925 zfs_refcount_destroy(&arc_l2c_only->arcs_size);
6926
6927 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6928 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6929 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6930 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6931 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
6932 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6933 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
6934 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6935 multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
6936 multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
6937
6938 aggsum_fini(&arc_meta_used);
6939 aggsum_fini(&arc_size);
6940 aggsum_fini(&astat_data_size);
6941 aggsum_fini(&astat_metadata_size);
6942 aggsum_fini(&astat_hdr_size);
6943 aggsum_fini(&astat_other_size);
6944 aggsum_fini(&astat_l2_hdr_size);
6945
6946 }
6947
6948 uint64_t
6949 arc_max_bytes(void)
6950 {
6951 return (arc_c_max);
6952 }
6953
6954 void
6955 arc_init(void)
6956 {
6957 /*
6958 * allmem is "all memory that we could possibly use".
6959 */
6960 #ifdef _KERNEL
6961 uint64_t allmem = ptob(physmem - swapfs_minfree);
6962 #else
6963 uint64_t allmem = (physmem * PAGESIZE) / 2;
6964 #endif
6965 mutex_init(&arc_adjust_lock, NULL, MUTEX_DEFAULT, NULL);
6966 cv_init(&arc_adjust_waiters_cv, NULL, CV_DEFAULT, NULL);
6967
6968 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
6969 arc_c_min = MAX(allmem / 32, 64 << 20);
6970 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
6971 if (allmem >= 1 << 30)
6972 arc_c_max = allmem - (1 << 30);
6973 else
6974 arc_c_max = arc_c_min;
6975 arc_c_max = MAX(allmem * 3 / 4, arc_c_max);
6976
6977 /*
6978 * In userland, there's only the memory pressure that we artificially
6979 * create (see arc_available_memory()). Don't let arc_c get too
6980 * small, because it can cause transactions to be larger than
6981 * arc_c, causing arc_tempreserve_space() to fail.
6982 */
6983 #ifndef _KERNEL
6984 arc_c_min = arc_c_max / 2;
6985 #endif
6986
6987 /*
6988 * Allow the tunables to override our calculations if they are
6989 * reasonable (ie. over 64MB)
6990 */
6991 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) {
6992 arc_c_max = zfs_arc_max;
6993 arc_c_min = MIN(arc_c_min, arc_c_max);
6994 }
6995 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
6996 arc_c_min = zfs_arc_min;
6997
6998 arc_c = arc_c_max;
6999 arc_p = (arc_c >> 1);
7000
7001 /* limit meta-data to 1/4 of the arc capacity */
7002 arc_meta_limit = arc_c_max / 4;
7003
7004 #ifdef _KERNEL
7005 /*
7006 * Metadata is stored in the kernel's heap. Don't let us
7007 * use more than half the heap for the ARC.
7008 */
7009 arc_meta_limit = MIN(arc_meta_limit,
7010 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
7011 #endif
7012
7013 /* Allow the tunable to override if it is reasonable */
7014 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
7015 arc_meta_limit = zfs_arc_meta_limit;
7016
7017 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
7018 arc_c_min = arc_meta_limit / 2;
7019
7020 if (zfs_arc_meta_min > 0) {
7021 arc_meta_min = zfs_arc_meta_min;
7022 } else {
7023 arc_meta_min = arc_c_min / 2;
7024 }
7025
7026 if (zfs_arc_grow_retry > 0)
7027 arc_grow_retry = zfs_arc_grow_retry;
7028
7029 if (zfs_arc_shrink_shift > 0)
7030 arc_shrink_shift = zfs_arc_shrink_shift;
7031
7032 /*
7033 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
7034 */
7035 if (arc_no_grow_shift >= arc_shrink_shift)
7036 arc_no_grow_shift = arc_shrink_shift - 1;
7037
7038 if (zfs_arc_p_min_shift > 0)
7039 arc_p_min_shift = zfs_arc_p_min_shift;
7040
7041 /* if kmem_flags are set, lets try to use less memory */
7042 if (kmem_debugging())
7043 arc_c = arc_c / 2;
7044 if (arc_c < arc_c_min)
7045 arc_c = arc_c_min;
7046
7047 arc_state_init();
7048
7049 /*
7050 * The arc must be "uninitialized", so that hdr_recl() (which is
7051 * registered by buf_init()) will not access arc_reap_zthr before
7052 * it is created.
7053 */
7054 ASSERT(!arc_initialized);
7055 buf_init();
7056
7057 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
7058 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
7059
7060 if (arc_ksp != NULL) {
7061 arc_ksp->ks_data = &arc_stats;
7062 arc_ksp->ks_update = arc_kstat_update;
7063 kstat_install(arc_ksp);
7064 }
7065
7066 arc_adjust_zthr = zthr_create(arc_adjust_cb_check,
7067 arc_adjust_cb, NULL);
7068 arc_reap_zthr = zthr_create_timer(arc_reap_cb_check,
7069 arc_reap_cb, NULL, SEC2NSEC(1));
7070
7071 arc_initialized = B_TRUE;
7072 arc_warm = B_FALSE;
7073
7074 /*
7075 * Calculate maximum amount of dirty data per pool.
7076 *
7077 * If it has been set by /etc/system, take that.
7078 * Otherwise, use a percentage of physical memory defined by
7079 * zfs_dirty_data_max_percent (default 10%) with a cap at
7080 * zfs_dirty_data_max_max (default 4GB).
7081 */
7082 if (zfs_dirty_data_max == 0) {
7083 zfs_dirty_data_max = physmem * PAGESIZE *
7084 zfs_dirty_data_max_percent / 100;
7085 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
7086 zfs_dirty_data_max_max);
7087 }
7088 }
7089
7090 void
7091 arc_fini(void)
7092 {
7093 /* Use B_TRUE to ensure *all* buffers are evicted */
7094 arc_flush(NULL, B_TRUE);
7095
7096 arc_initialized = B_FALSE;
7097
7098 if (arc_ksp != NULL) {
7099 kstat_delete(arc_ksp);
7100 arc_ksp = NULL;
7101 }
7102
7103 (void) zthr_cancel(arc_adjust_zthr);
7104 zthr_destroy(arc_adjust_zthr);
7105
7106 (void) zthr_cancel(arc_reap_zthr);
7107 zthr_destroy(arc_reap_zthr);
7108
7109 mutex_destroy(&arc_adjust_lock);
7110 cv_destroy(&arc_adjust_waiters_cv);
7111
7112 /*
7113 * buf_fini() must proceed arc_state_fini() because buf_fin() may
7114 * trigger the release of kmem magazines, which can callback to
7115 * arc_space_return() which accesses aggsums freed in act_state_fini().
7116 */
7117 buf_fini();
7118 arc_state_fini();
7119
7120 ASSERT0(arc_loaned_bytes);
7121 }
7122
7123 /*
7124 * Level 2 ARC
7125 *
7126 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7127 * It uses dedicated storage devices to hold cached data, which are populated
7128 * using large infrequent writes. The main role of this cache is to boost
7129 * the performance of random read workloads. The intended L2ARC devices
7130 * include short-stroked disks, solid state disks, and other media with
7131 * substantially faster read latency than disk.
7132 *
7133 * +-----------------------+
7134 * | ARC |
7135 * +-----------------------+
7136 * | ^ ^
7137 * | | |
7138 * l2arc_feed_thread() arc_read()
7139 * | | |
7140 * | l2arc read |
7141 * V | |
7142 * +---------------+ |
7143 * | L2ARC | |
7144 * +---------------+ |
7145 * | ^ |
7146 * l2arc_write() | |
7147 * | | |
7148 * V | |
7149 * +-------+ +-------+
7150 * | vdev | | vdev |
7151 * | cache | | cache |
7152 * +-------+ +-------+
7153 * +=========+ .-----.
7154 * : L2ARC : |-_____-|
7155 * : devices : | Disks |
7156 * +=========+ `-_____-'
7157 *
7158 * Read requests are satisfied from the following sources, in order:
7159 *
7160 * 1) ARC
7161 * 2) vdev cache of L2ARC devices
7162 * 3) L2ARC devices
7163 * 4) vdev cache of disks
7164 * 5) disks
7165 *
7166 * Some L2ARC device types exhibit extremely slow write performance.
7167 * To accommodate for this there are some significant differences between
7168 * the L2ARC and traditional cache design:
7169 *
7170 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
7171 * the ARC behave as usual, freeing buffers and placing headers on ghost
7172 * lists. The ARC does not send buffers to the L2ARC during eviction as
7173 * this would add inflated write latencies for all ARC memory pressure.
7174 *
7175 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
7176 * It does this by periodically scanning buffers from the eviction-end of
7177 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
7178 * not already there. It scans until a headroom of buffers is satisfied,
7179 * which itself is a buffer for ARC eviction. If a compressible buffer is
7180 * found during scanning and selected for writing to an L2ARC device, we
7181 * temporarily boost scanning headroom during the next scan cycle to make
7182 * sure we adapt to compression effects (which might significantly reduce
7183 * the data volume we write to L2ARC). The thread that does this is
7184 * l2arc_feed_thread(), illustrated below; example sizes are included to
7185 * provide a better sense of ratio than this diagram:
7186 *
7187 * head --> tail
7188 * +---------------------+----------+
7189 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
7190 * +---------------------+----------+ | o L2ARC eligible
7191 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
7192 * +---------------------+----------+ |
7193 * 15.9 Gbytes ^ 32 Mbytes |
7194 * headroom |
7195 * l2arc_feed_thread()
7196 * |
7197 * l2arc write hand <--[oooo]--'
7198 * | 8 Mbyte
7199 * | write max
7200 * V
7201 * +==============================+
7202 * L2ARC dev |####|#|###|###| |####| ... |
7203 * +==============================+
7204 * 32 Gbytes
7205 *
7206 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
7207 * evicted, then the L2ARC has cached a buffer much sooner than it probably
7208 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
7209 * safe to say that this is an uncommon case, since buffers at the end of
7210 * the ARC lists have moved there due to inactivity.
7211 *
7212 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
7213 * then the L2ARC simply misses copying some buffers. This serves as a
7214 * pressure valve to prevent heavy read workloads from both stalling the ARC
7215 * with waits and clogging the L2ARC with writes. This also helps prevent
7216 * the potential for the L2ARC to churn if it attempts to cache content too
7217 * quickly, such as during backups of the entire pool.
7218 *
7219 * 5. After system boot and before the ARC has filled main memory, there are
7220 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
7221 * lists can remain mostly static. Instead of searching from tail of these
7222 * lists as pictured, the l2arc_feed_thread() will search from the list heads
7223 * for eligible buffers, greatly increasing its chance of finding them.
7224 *
7225 * The L2ARC device write speed is also boosted during this time so that
7226 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
7227 * there are no L2ARC reads, and no fear of degrading read performance
7228 * through increased writes.
7229 *
7230 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
7231 * the vdev queue can aggregate them into larger and fewer writes. Each
7232 * device is written to in a rotor fashion, sweeping writes through
7233 * available space then repeating.
7234 *
7235 * 7. The L2ARC does not store dirty content. It never needs to flush
7236 * write buffers back to disk based storage.
7237 *
7238 * 8. If an ARC buffer is written (and dirtied) which also exists in the
7239 * L2ARC, the now stale L2ARC buffer is immediately dropped.
7240 *
7241 * The performance of the L2ARC can be tweaked by a number of tunables, which
7242 * may be necessary for different workloads:
7243 *
7244 * l2arc_write_max max write bytes per interval
7245 * l2arc_write_boost extra write bytes during device warmup
7246 * l2arc_noprefetch skip caching prefetched buffers
7247 * l2arc_headroom number of max device writes to precache
7248 * l2arc_headroom_boost when we find compressed buffers during ARC
7249 * scanning, we multiply headroom by this
7250 * percentage factor for the next scan cycle,
7251 * since more compressed buffers are likely to
7252 * be present
7253 * l2arc_feed_secs seconds between L2ARC writing
7254 *
7255 * Tunables may be removed or added as future performance improvements are
7256 * integrated, and also may become zpool properties.
7257 *
7258 * There are three key functions that control how the L2ARC warms up:
7259 *
7260 * l2arc_write_eligible() check if a buffer is eligible to cache
7261 * l2arc_write_size() calculate how much to write
7262 * l2arc_write_interval() calculate sleep delay between writes
7263 *
7264 * These three functions determine what to write, how much, and how quickly
7265 * to send writes.
7266 *
7267 * L2ARC persistence:
7268 *
7269 * When writing buffers to L2ARC, we periodically add some metadata to
7270 * make sure we can pick them up after reboot, thus dramatically reducing
7271 * the impact that any downtime has on the performance of storage systems
7272 * with large caches.
7273 *
7274 * The implementation works fairly simply by integrating the following two
7275 * modifications:
7276 *
7277 * *) When writing to the L2ARC, we occasionally write a "l2arc log block",
7278 * which is an additional piece of metadata which describes what's been
7279 * written. This allows us to rebuild the arc_buf_hdr_t structures of the
7280 * main ARC buffers. There are 2 linked-lists of log blocks headed by
7281 * dh_start_lbps[2]. We alternate which chain we append to, so they are
7282 * time-wise and offset-wise interleaved, but that is an optimization rather
7283 * than for correctness. The log block also includes a pointer to the
7284 * previous block in its chain.
7285 *
7286 * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device
7287 * for our header bookkeeping purposes. This contains a device header,
7288 * which contains our top-level reference structures. We update it each
7289 * time we write a new log block, so that we're able to locate it in the
7290 * L2ARC device. If this write results in an inconsistent device header
7291 * (e.g. due to power failure), we detect this by verifying the header's
7292 * checksum and simply fail to reconstruct the L2ARC after reboot.
7293 *
7294 * Implementation diagram:
7295 *
7296 * +=== L2ARC device (not to scale) ======================================+
7297 * | ___two newest log block pointers__.__________ |
7298 * | / \dh_start_lbps[1] |
7299 * | / \ \dh_start_lbps[0]|
7300 * |.___/__. V V |
7301 * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---|
7302 * || hdr| ^ /^ /^ / / |
7303 * |+------+ ...--\-------/ \-----/--\------/ / |
7304 * | \--------------/ \--------------/ |
7305 * +======================================================================+
7306 *
7307 * As can be seen on the diagram, rather than using a simple linked list,
7308 * we use a pair of linked lists with alternating elements. This is a
7309 * performance enhancement due to the fact that we only find out the
7310 * address of the next log block access once the current block has been
7311 * completely read in. Obviously, this hurts performance, because we'd be
7312 * keeping the device's I/O queue at only a 1 operation deep, thus
7313 * incurring a large amount of I/O round-trip latency. Having two lists
7314 * allows us to fetch two log blocks ahead of where we are currently
7315 * rebuilding L2ARC buffers.
7316 *
7317 * On-device data structures:
7318 *
7319 * L2ARC device header: l2arc_dev_hdr_phys_t
7320 * L2ARC log block: l2arc_log_blk_phys_t
7321 *
7322 * L2ARC reconstruction:
7323 *
7324 * When writing data, we simply write in the standard rotary fashion,
7325 * evicting buffers as we go and simply writing new data over them (writing
7326 * a new log block every now and then). This obviously means that once we
7327 * loop around the end of the device, we will start cutting into an already
7328 * committed log block (and its referenced data buffers), like so:
7329 *
7330 * current write head__ __old tail
7331 * \ /
7332 * V V
7333 * <--|bufs |lb |bufs |lb | |bufs |lb |bufs |lb |-->
7334 * ^ ^^^^^^^^^___________________________________
7335 * | \
7336 * <<nextwrite>> may overwrite this blk and/or its bufs --'
7337 *
7338 * When importing the pool, we detect this situation and use it to stop
7339 * our scanning process (see l2arc_rebuild).
7340 *
7341 * There is one significant caveat to consider when rebuilding ARC contents
7342 * from an L2ARC device: what about invalidated buffers? Given the above
7343 * construction, we cannot update blocks which we've already written to amend
7344 * them to remove buffers which were invalidated. Thus, during reconstruction,
7345 * we might be populating the cache with buffers for data that's not on the
7346 * main pool anymore, or may have been overwritten!
7347 *
7348 * As it turns out, this isn't a problem. Every arc_read request includes
7349 * both the DVA and, crucially, the birth TXG of the BP the caller is
7350 * looking for. So even if the cache were populated by completely rotten
7351 * blocks for data that had been long deleted and/or overwritten, we'll
7352 * never actually return bad data from the cache, since the DVA with the
7353 * birth TXG uniquely identify a block in space and time - once created,
7354 * a block is immutable on disk. The worst thing we have done is wasted
7355 * some time and memory at l2arc rebuild to reconstruct outdated ARC
7356 * entries that will get dropped from the l2arc as it is being updated
7357 * with new blocks.
7358 *
7359 * L2ARC buffers that have been evicted by l2arc_evict() ahead of the write
7360 * hand are not restored. This is done by saving the offset (in bytes)
7361 * l2arc_evict() has evicted to in the L2ARC device header and taking it
7362 * into account when restoring buffers.
7363 */
7364
7365 static boolean_t
7366 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
7367 {
7368 /*
7369 * A buffer is *not* eligible for the L2ARC if it:
7370 * 1. belongs to a different spa.
7371 * 2. is already cached on the L2ARC.
7372 * 3. has an I/O in progress (it may be an incomplete read).
7373 * 4. is flagged not eligible (zfs property).
7374 */
7375 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
7376 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
7377 return (B_FALSE);
7378
7379 return (B_TRUE);
7380 }
7381
7382 static uint64_t
7383 l2arc_write_size(l2arc_dev_t *dev)
7384 {
7385 uint64_t size, dev_size;
7386
7387 /*
7388 * Make sure our globals have meaningful values in case the user
7389 * altered them.
7390 */
7391 size = l2arc_write_max;
7392 if (size == 0) {
7393 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
7394 "be greater than zero, resetting it to the default (%d)",
7395 L2ARC_WRITE_SIZE);
7396 size = l2arc_write_max = L2ARC_WRITE_SIZE;
7397 }
7398
7399 if (arc_warm == B_FALSE)
7400 size += l2arc_write_boost;
7401
7402 /*
7403 * Make sure the write size does not exceed the size of the cache
7404 * device. This is important in l2arc_evict(), otherwise infinite
7405 * iteration can occur.
7406 */
7407 dev_size = dev->l2ad_end - dev->l2ad_start;
7408 if ((size + l2arc_log_blk_overhead(size, dev)) >= dev_size) {
7409 cmn_err(CE_NOTE, "l2arc_write_max or l2arc_write_boost "
7410 "plus the overhead of log blocks (persistent L2ARC, "
7411 "%" PRIu64 " bytes) exceeds the size of the cache device "
7412 "(guid %" PRIu64 "), resetting them to the default (%d)",
7413 l2arc_log_blk_overhead(size, dev),
7414 dev->l2ad_vdev->vdev_guid, L2ARC_WRITE_SIZE);
7415 size = l2arc_write_max = l2arc_write_boost = L2ARC_WRITE_SIZE;
7416
7417 if (arc_warm == B_FALSE)
7418 size += l2arc_write_boost;
7419 }
7420
7421 return (size);
7422
7423 }
7424
7425 static clock_t
7426 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
7427 {
7428 clock_t interval, next, now;
7429
7430 /*
7431 * If the ARC lists are busy, increase our write rate; if the
7432 * lists are stale, idle back. This is achieved by checking
7433 * how much we previously wrote - if it was more than half of
7434 * what we wanted, schedule the next write much sooner.
7435 */
7436 if (l2arc_feed_again && wrote > (wanted / 2))
7437 interval = (hz * l2arc_feed_min_ms) / 1000;
7438 else
7439 interval = hz * l2arc_feed_secs;
7440
7441 now = ddi_get_lbolt();
7442 next = MAX(now, MIN(now + interval, began + interval));
7443
7444 return (next);
7445 }
7446
7447 /*
7448 * Cycle through L2ARC devices. This is how L2ARC load balances.
7449 * If a device is returned, this also returns holding the spa config lock.
7450 */
7451 static l2arc_dev_t *
7452 l2arc_dev_get_next(void)
7453 {
7454 l2arc_dev_t *first, *next = NULL;
7455
7456 /*
7457 * Lock out the removal of spas (spa_namespace_lock), then removal
7458 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7459 * both locks will be dropped and a spa config lock held instead.
7460 */
7461 mutex_enter(&spa_namespace_lock);
7462 mutex_enter(&l2arc_dev_mtx);
7463
7464 /* if there are no vdevs, there is nothing to do */
7465 if (l2arc_ndev == 0)
7466 goto out;
7467
7468 first = NULL;
7469 next = l2arc_dev_last;
7470 do {
7471 /* loop around the list looking for a non-faulted vdev */
7472 if (next == NULL) {
7473 next = list_head(l2arc_dev_list);
7474 } else {
7475 next = list_next(l2arc_dev_list, next);
7476 if (next == NULL)
7477 next = list_head(l2arc_dev_list);
7478 }
7479
7480 /* if we have come back to the start, bail out */
7481 if (first == NULL)
7482 first = next;
7483 else if (next == first)
7484 break;
7485
7486 } while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild);
7487
7488 /* if we were unable to find any usable vdevs, return NULL */
7489 if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild)
7490 next = NULL;
7491
7492 l2arc_dev_last = next;
7493
7494 out:
7495 mutex_exit(&l2arc_dev_mtx);
7496
7497 /*
7498 * Grab the config lock to prevent the 'next' device from being
7499 * removed while we are writing to it.
7500 */
7501 if (next != NULL)
7502 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
7503 mutex_exit(&spa_namespace_lock);
7504
7505 return (next);
7506 }
7507
7508 /*
7509 * Free buffers that were tagged for destruction.
7510 */
7511 static void
7512 l2arc_do_free_on_write()
7513 {
7514 list_t *buflist;
7515 l2arc_data_free_t *df, *df_prev;
7516
7517 mutex_enter(&l2arc_free_on_write_mtx);
7518 buflist = l2arc_free_on_write;
7519
7520 for (df = list_tail(buflist); df; df = df_prev) {
7521 df_prev = list_prev(buflist, df);
7522 ASSERT3P(df->l2df_abd, !=, NULL);
7523 abd_free(df->l2df_abd);
7524 list_remove(buflist, df);
7525 kmem_free(df, sizeof (l2arc_data_free_t));
7526 }
7527
7528 mutex_exit(&l2arc_free_on_write_mtx);
7529 }
7530
7531 /*
7532 * A write to a cache device has completed. Update all headers to allow
7533 * reads from these buffers to begin.
7534 */
7535 static void
7536 l2arc_write_done(zio_t *zio)
7537 {
7538 l2arc_write_callback_t *cb;
7539 l2arc_lb_abd_buf_t *abd_buf;
7540 l2arc_lb_ptr_buf_t *lb_ptr_buf;
7541 l2arc_dev_t *dev;
7542 l2arc_dev_hdr_phys_t *l2dhdr;
7543 list_t *buflist;
7544 arc_buf_hdr_t *head, *hdr, *hdr_prev;
7545 kmutex_t *hash_lock;
7546 int64_t bytes_dropped = 0;
7547
7548 cb = zio->io_private;
7549 ASSERT3P(cb, !=, NULL);
7550 dev = cb->l2wcb_dev;
7551 l2dhdr = dev->l2ad_dev_hdr;
7552 ASSERT3P(dev, !=, NULL);
7553 head = cb->l2wcb_head;
7554 ASSERT3P(head, !=, NULL);
7555 buflist = &dev->l2ad_buflist;
7556 ASSERT3P(buflist, !=, NULL);
7557 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7558 l2arc_write_callback_t *, cb);
7559
7560 if (zio->io_error != 0)
7561 ARCSTAT_BUMP(arcstat_l2_writes_error);
7562
7563 /*
7564 * All writes completed, or an error was hit.
7565 */
7566 top:
7567 mutex_enter(&dev->l2ad_mtx);
7568 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
7569 hdr_prev = list_prev(buflist, hdr);
7570
7571 hash_lock = HDR_LOCK(hdr);
7572
7573 /*
7574 * We cannot use mutex_enter or else we can deadlock
7575 * with l2arc_write_buffers (due to swapping the order
7576 * the hash lock and l2ad_mtx are taken).
7577 */
7578 if (!mutex_tryenter(hash_lock)) {
7579 /*
7580 * Missed the hash lock. We must retry so we
7581 * don't leave the ARC_FLAG_L2_WRITING bit set.
7582 */
7583 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
7584
7585 /*
7586 * We don't want to rescan the headers we've
7587 * already marked as having been written out, so
7588 * we reinsert the head node so we can pick up
7589 * where we left off.
7590 */
7591 list_remove(buflist, head);
7592 list_insert_after(buflist, hdr, head);
7593
7594 mutex_exit(&dev->l2ad_mtx);
7595
7596 /*
7597 * We wait for the hash lock to become available
7598 * to try and prevent busy waiting, and increase
7599 * the chance we'll be able to acquire the lock
7600 * the next time around.
7601 */
7602 mutex_enter(hash_lock);
7603 mutex_exit(hash_lock);
7604 goto top;
7605 }
7606
7607 /*
7608 * We could not have been moved into the arc_l2c_only
7609 * state while in-flight due to our ARC_FLAG_L2_WRITING
7610 * bit being set. Let's just ensure that's being enforced.
7611 */
7612 ASSERT(HDR_HAS_L1HDR(hdr));
7613
7614 if (zio->io_error != 0) {
7615 /*
7616 * Error - drop L2ARC entry.
7617 */
7618 list_remove(buflist, hdr);
7619 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
7620
7621 uint64_t psize = HDR_GET_PSIZE(hdr);
7622 ARCSTAT_INCR(arcstat_l2_psize, -psize);
7623 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
7624
7625 bytes_dropped +=
7626 vdev_psize_to_asize(dev->l2ad_vdev, psize);
7627 (void) zfs_refcount_remove_many(&dev->l2ad_alloc,
7628 arc_hdr_size(hdr), hdr);
7629 }
7630
7631 /*
7632 * Allow ARC to begin reads and ghost list evictions to
7633 * this L2ARC entry.
7634 */
7635 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7636
7637 mutex_exit(hash_lock);
7638 }
7639
7640 /*
7641 * Free the allocated abd buffers for writing the log blocks.
7642 * If the zio failed reclaim the allocated space and remove the
7643 * pointers to these log blocks from the log block pointer list
7644 * of the L2ARC device.
7645 */
7646 while ((abd_buf = list_remove_tail(&cb->l2wcb_abd_list)) != NULL) {
7647 abd_free(abd_buf->abd);
7648 zio_buf_free(abd_buf, sizeof (*abd_buf));
7649 if (zio->io_error != 0) {
7650 lb_ptr_buf = list_remove_head(&dev->l2ad_lbptr_list);
7651 /*
7652 * L2BLK_GET_PSIZE returns aligned size for log
7653 * blocks.
7654 */
7655 uint64_t asize =
7656 L2BLK_GET_PSIZE((lb_ptr_buf->lb_ptr)->lbp_prop);
7657 bytes_dropped += asize;
7658 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
7659 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
7660 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
7661 lb_ptr_buf);
7662 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
7663 kmem_free(lb_ptr_buf->lb_ptr,
7664 sizeof (l2arc_log_blkptr_t));
7665 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
7666 }
7667 }
7668 list_destroy(&cb->l2wcb_abd_list);
7669
7670 if (zio->io_error != 0) {
7671 /*
7672 * Restore the lbps array in the header to its previous state.
7673 * If the list of log block pointers is empty, zero out the
7674 * log block pointers in the device header.
7675 */
7676 lb_ptr_buf = list_head(&dev->l2ad_lbptr_list);
7677 for (int i = 0; i < 2; i++) {
7678 if (lb_ptr_buf == NULL) {
7679 /*
7680 * If the list is empty zero out the device
7681 * header. Otherwise zero out the second log
7682 * block pointer in the header.
7683 */
7684 if (i == 0) {
7685 bzero(l2dhdr, dev->l2ad_dev_hdr_asize);
7686 } else {
7687 bzero(&l2dhdr->dh_start_lbps[i],
7688 sizeof (l2arc_log_blkptr_t));
7689 }
7690 break;
7691 }
7692 bcopy(lb_ptr_buf->lb_ptr, &l2dhdr->dh_start_lbps[i],
7693 sizeof (l2arc_log_blkptr_t));
7694 lb_ptr_buf = list_next(&dev->l2ad_lbptr_list,
7695 lb_ptr_buf);
7696 }
7697 }
7698
7699 atomic_inc_64(&l2arc_writes_done);
7700 list_remove(buflist, head);
7701 ASSERT(!HDR_HAS_L1HDR(head));
7702 kmem_cache_free(hdr_l2only_cache, head);
7703 mutex_exit(&dev->l2ad_mtx);
7704
7705 ASSERT(dev->l2ad_vdev != NULL);
7706 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7707
7708 l2arc_do_free_on_write();
7709
7710 kmem_free(cb, sizeof (l2arc_write_callback_t));
7711 }
7712
7713 static int
7714 l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb)
7715 {
7716 int ret;
7717 spa_t *spa = zio->io_spa;
7718 arc_buf_hdr_t *hdr = cb->l2rcb_hdr;
7719 blkptr_t *bp = zio->io_bp;
7720 uint8_t salt[ZIO_DATA_SALT_LEN];
7721 uint8_t iv[ZIO_DATA_IV_LEN];
7722 uint8_t mac[ZIO_DATA_MAC_LEN];
7723 boolean_t no_crypt = B_FALSE;
7724
7725 /*
7726 * ZIL data is never be written to the L2ARC, so we don't need
7727 * special handling for its unique MAC storage.
7728 */
7729 ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG);
7730 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
7731 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
7732
7733 /*
7734 * If the data was encrypted, decrypt it now. Note that
7735 * we must check the bp here and not the hdr, since the
7736 * hdr does not have its encryption parameters updated
7737 * until arc_read_done().
7738 */
7739 if (BP_IS_ENCRYPTED(bp)) {
7740 abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
7741
7742 zio_crypt_decode_params_bp(bp, salt, iv);
7743 zio_crypt_decode_mac_bp(bp, mac);
7744
7745 ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb,
7746 BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp),
7747 salt, iv, mac, HDR_GET_PSIZE(hdr), eabd,
7748 hdr->b_l1hdr.b_pabd, &no_crypt);
7749 if (ret != 0) {
7750 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
7751 goto error;
7752 }
7753
7754 /*
7755 * If we actually performed decryption, replace b_pabd
7756 * with the decrypted data. Otherwise we can just throw
7757 * our decryption buffer away.
7758 */
7759 if (!no_crypt) {
7760 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
7761 arc_hdr_size(hdr), hdr);
7762 hdr->b_l1hdr.b_pabd = eabd;
7763 zio->io_abd = eabd;
7764 } else {
7765 arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr);
7766 }
7767 }
7768
7769 /*
7770 * If the L2ARC block was compressed, but ARC compression
7771 * is disabled we decompress the data into a new buffer and
7772 * replace the existing data.
7773 */
7774 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
7775 !HDR_COMPRESSION_ENABLED(hdr)) {
7776 abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
7777 void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr));
7778
7779 ret = zio_decompress_data(HDR_GET_COMPRESS(hdr),
7780 hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr),
7781 HDR_GET_LSIZE(hdr));
7782 if (ret != 0) {
7783 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
7784 arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr);
7785 goto error;
7786 }
7787
7788 abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr));
7789 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
7790 arc_hdr_size(hdr), hdr);
7791 hdr->b_l1hdr.b_pabd = cabd;
7792 zio->io_abd = cabd;
7793 zio->io_size = HDR_GET_LSIZE(hdr);
7794 }
7795
7796 return (0);
7797
7798 error:
7799 return (ret);
7800 }
7801
7802
7803 /*
7804 * A read to a cache device completed. Validate buffer contents before
7805 * handing over to the regular ARC routines.
7806 */
7807 static void
7808 l2arc_read_done(zio_t *zio)
7809 {
7810 int tfm_error = 0;
7811 l2arc_read_callback_t *cb = zio->io_private;
7812 arc_buf_hdr_t *hdr;
7813 kmutex_t *hash_lock;
7814 boolean_t valid_cksum;
7815 boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) &&
7816 (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT));
7817
7818 ASSERT3P(zio->io_vd, !=, NULL);
7819 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7820
7821 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7822
7823 ASSERT3P(cb, !=, NULL);
7824 hdr = cb->l2rcb_hdr;
7825 ASSERT3P(hdr, !=, NULL);
7826
7827 hash_lock = HDR_LOCK(hdr);
7828 mutex_enter(hash_lock);
7829 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
7830
7831 /*
7832 * If the data was read into a temporary buffer,
7833 * move it and free the buffer.
7834 */
7835 if (cb->l2rcb_abd != NULL) {
7836 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size);
7837 if (zio->io_error == 0) {
7838 if (using_rdata) {
7839 abd_copy(hdr->b_crypt_hdr.b_rabd,
7840 cb->l2rcb_abd, arc_hdr_size(hdr));
7841 } else {
7842 abd_copy(hdr->b_l1hdr.b_pabd,
7843 cb->l2rcb_abd, arc_hdr_size(hdr));
7844 }
7845 }
7846
7847 /*
7848 * The following must be done regardless of whether
7849 * there was an error:
7850 * - free the temporary buffer
7851 * - point zio to the real ARC buffer
7852 * - set zio size accordingly
7853 * These are required because zio is either re-used for
7854 * an I/O of the block in the case of the error
7855 * or the zio is passed to arc_read_done() and it
7856 * needs real data.
7857 */
7858 abd_free(cb->l2rcb_abd);
7859 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr);
7860
7861 if (using_rdata) {
7862 ASSERT(HDR_HAS_RABD(hdr));
7863 zio->io_abd = zio->io_orig_abd =
7864 hdr->b_crypt_hdr.b_rabd;
7865 } else {
7866 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
7867 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd;
7868 }
7869 }
7870
7871 ASSERT3P(zio->io_abd, !=, NULL);
7872
7873 /*
7874 * Check this survived the L2ARC journey.
7875 */
7876 ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd ||
7877 (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd));
7878 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
7879 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
7880
7881 valid_cksum = arc_cksum_is_equal(hdr, zio);
7882
7883 /*
7884 * b_rabd will always match the data as it exists on disk if it is
7885 * being used. Therefore if we are reading into b_rabd we do not
7886 * attempt to untransform the data.
7887 */
7888 if (valid_cksum && !using_rdata)
7889 tfm_error = l2arc_untransform(zio, cb);
7890
7891 if (valid_cksum && tfm_error == 0 && zio->io_error == 0 &&
7892 !HDR_L2_EVICTED(hdr)) {
7893 mutex_exit(hash_lock);
7894 zio->io_private = hdr;
7895 arc_read_done(zio);
7896 } else {
7897 /*
7898 * Buffer didn't survive caching. Increment stats and
7899 * reissue to the original storage device.
7900 */
7901 if (zio->io_error != 0) {
7902 ARCSTAT_BUMP(arcstat_l2_io_error);
7903 } else {
7904 zio->io_error = SET_ERROR(EIO);
7905 }
7906 if (!valid_cksum || tfm_error != 0)
7907 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
7908
7909 /*
7910 * If there's no waiter, issue an async i/o to the primary
7911 * storage now. If there *is* a waiter, the caller must
7912 * issue the i/o in a context where it's OK to block.
7913 */
7914 if (zio->io_waiter == NULL) {
7915 zio_t *pio = zio_unique_parent(zio);
7916 void *abd = (using_rdata) ?
7917 hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd;
7918
7919 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
7920
7921 zio = zio_read(pio, zio->io_spa, zio->io_bp,
7922 abd, zio->io_size, arc_read_done,
7923 hdr, zio->io_priority, cb->l2rcb_flags,
7924 &cb->l2rcb_zb);
7925
7926 /*
7927 * Original ZIO will be freed, so we need to update
7928 * ARC header with the new ZIO pointer to be used
7929 * by zio_change_priority() in arc_read().
7930 */
7931 for (struct arc_callback *acb = hdr->b_l1hdr.b_acb;
7932 acb != NULL; acb = acb->acb_next)
7933 acb->acb_zio_head = zio;
7934
7935 mutex_exit(hash_lock);
7936 zio_nowait(zio);
7937 } else {
7938 mutex_exit(hash_lock);
7939 }
7940 }
7941
7942 kmem_free(cb, sizeof (l2arc_read_callback_t));
7943 }
7944
7945 /*
7946 * This is the list priority from which the L2ARC will search for pages to
7947 * cache. This is used within loops (0..3) to cycle through lists in the
7948 * desired order. This order can have a significant effect on cache
7949 * performance.
7950 *
7951 * Currently the metadata lists are hit first, MFU then MRU, followed by
7952 * the data lists. This function returns a locked list, and also returns
7953 * the lock pointer.
7954 */
7955 static multilist_sublist_t *
7956 l2arc_sublist_lock(int list_num)
7957 {
7958 multilist_t *ml = NULL;
7959 unsigned int idx;
7960
7961 ASSERT(list_num >= 0 && list_num <= 3);
7962
7963 switch (list_num) {
7964 case 0:
7965 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
7966 break;
7967 case 1:
7968 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
7969 break;
7970 case 2:
7971 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
7972 break;
7973 case 3:
7974 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
7975 break;
7976 }
7977
7978 /*
7979 * Return a randomly-selected sublist. This is acceptable
7980 * because the caller feeds only a little bit of data for each
7981 * call (8MB). Subsequent calls will result in different
7982 * sublists being selected.
7983 */
7984 idx = multilist_get_random_index(ml);
7985 return (multilist_sublist_lock(ml, idx));
7986 }
7987
7988 /*
7989 * Calculates the maximum overhead of L2ARC metadata log blocks for a given
7990 * L2ARC write size. l2arc_evict and l2arc_write_size need to include this
7991 * overhead in processing to make sure there is enough headroom available
7992 * when writing buffers.
7993 */
7994 static inline uint64_t
7995 l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev)
7996 {
7997 if (dev->l2ad_log_entries == 0) {
7998 return (0);
7999 } else {
8000 uint64_t log_entries = write_sz >> SPA_MINBLOCKSHIFT;
8001
8002 uint64_t log_blocks = (log_entries +
8003 dev->l2ad_log_entries - 1) /
8004 dev->l2ad_log_entries;
8005
8006 return (vdev_psize_to_asize(dev->l2ad_vdev,
8007 sizeof (l2arc_log_blk_phys_t)) * log_blocks);
8008 }
8009 }
8010
8011 /*
8012 * Evict buffers from the device write hand to the distance specified in
8013 * bytes. This distance may span populated buffers, it may span nothing.
8014 * This is clearing a region on the L2ARC device ready for writing.
8015 * If the 'all' boolean is set, every buffer is evicted.
8016 */
8017 static void
8018 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
8019 {
8020 list_t *buflist;
8021 arc_buf_hdr_t *hdr, *hdr_prev;
8022 kmutex_t *hash_lock;
8023 uint64_t taddr;
8024 l2arc_lb_ptr_buf_t *lb_ptr_buf, *lb_ptr_buf_prev;
8025 boolean_t rerun;
8026
8027 buflist = &dev->l2ad_buflist;
8028
8029 /*
8030 * We need to add in the worst case scenario of log block overhead.
8031 */
8032 distance += l2arc_log_blk_overhead(distance, dev);
8033
8034 top:
8035 rerun = B_FALSE;
8036 if (dev->l2ad_hand >= (dev->l2ad_end - distance)) {
8037 /*
8038 * When there is no space to accommodate upcoming writes,
8039 * evict to the end. Then bump the write and evict hands
8040 * to the start and iterate. This iteration does not
8041 * happen indefinitely as we make sure in
8042 * l2arc_write_size() that when the write hand is reset,
8043 * the write size does not exceed the end of the device.
8044 */
8045 rerun = B_TRUE;
8046 taddr = dev->l2ad_end;
8047 } else {
8048 taddr = dev->l2ad_hand + distance;
8049 }
8050 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
8051 uint64_t, taddr, boolean_t, all);
8052
8053 /*
8054 * This check has to be placed after deciding whether to iterate
8055 * (rerun).
8056 */
8057 if (!all && dev->l2ad_first) {
8058 /*
8059 * This is the first sweep through the device. There is
8060 * nothing to evict.
8061 */
8062 goto out;
8063 }
8064
8065 /*
8066 * When rebuilding L2ARC we retrieve the evict hand from the header of
8067 * the device. Of note, l2arc_evict() does not actually delete buffers
8068 * from the cache device, but keeping track of the evict hand will be
8069 * useful when TRIM is implemented.
8070 */
8071 dev->l2ad_evict = MAX(dev->l2ad_evict, taddr);
8072
8073 retry:
8074 mutex_enter(&dev->l2ad_mtx);
8075 /*
8076 * We have to account for evicted log blocks. Run vdev_space_update()
8077 * on log blocks whose offset (in bytes) is before the evicted offset
8078 * (in bytes) by searching in the list of pointers to log blocks
8079 * present in the L2ARC device.
8080 */
8081 for (lb_ptr_buf = list_tail(&dev->l2ad_lbptr_list); lb_ptr_buf;
8082 lb_ptr_buf = lb_ptr_buf_prev) {
8083
8084 lb_ptr_buf_prev = list_prev(&dev->l2ad_lbptr_list, lb_ptr_buf);
8085
8086 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
8087 uint64_t asize = L2BLK_GET_PSIZE(
8088 (lb_ptr_buf->lb_ptr)->lbp_prop);
8089
8090 /*
8091 * We don't worry about log blocks left behind (ie
8092 * lbp_payload_start < l2ad_hand) because l2arc_write_buffers()
8093 * will never write more than l2arc_evict() evicts.
8094 */
8095 if (!all && l2arc_log_blkptr_valid(dev, lb_ptr_buf->lb_ptr)) {
8096 break;
8097 } else {
8098 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0);
8099 ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize);
8100 ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count);
8101 zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize,
8102 lb_ptr_buf);
8103 zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf);
8104 list_remove(&dev->l2ad_lbptr_list, lb_ptr_buf);
8105 kmem_free(lb_ptr_buf->lb_ptr,
8106 sizeof (l2arc_log_blkptr_t));
8107 kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t));
8108 }
8109 }
8110
8111 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
8112 hdr_prev = list_prev(buflist, hdr);
8113
8114 ASSERT(!HDR_EMPTY(hdr));
8115 hash_lock = HDR_LOCK(hdr);
8116
8117 /*
8118 * We cannot use mutex_enter or else we can deadlock
8119 * with l2arc_write_buffers (due to swapping the order
8120 * the hash lock and l2ad_mtx are taken).
8121 */
8122 if (!mutex_tryenter(hash_lock)) {
8123 /*
8124 * Missed the hash lock. Retry.
8125 */
8126 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
8127 mutex_exit(&dev->l2ad_mtx);
8128 mutex_enter(hash_lock);
8129 mutex_exit(hash_lock);
8130 goto retry;
8131 }
8132
8133 /*
8134 * A header can't be on this list if it doesn't have L2 header.
8135 */
8136 ASSERT(HDR_HAS_L2HDR(hdr));
8137
8138 /* Ensure this header has finished being written. */
8139 ASSERT(!HDR_L2_WRITING(hdr));
8140 ASSERT(!HDR_L2_WRITE_HEAD(hdr));
8141
8142 if (!all && (hdr->b_l2hdr.b_daddr >= dev->l2ad_evict ||
8143 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
8144 /*
8145 * We've evicted to the target address,
8146 * or the end of the device.
8147 */
8148 mutex_exit(hash_lock);
8149 break;
8150 }
8151
8152 if (!HDR_HAS_L1HDR(hdr)) {
8153 ASSERT(!HDR_L2_READING(hdr));
8154 /*
8155 * This doesn't exist in the ARC. Destroy.
8156 * arc_hdr_destroy() will call list_remove()
8157 * and decrement arcstat_l2_lsize.
8158 */
8159 arc_change_state(arc_anon, hdr, hash_lock);
8160 arc_hdr_destroy(hdr);
8161 } else {
8162 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
8163 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
8164 /*
8165 * Invalidate issued or about to be issued
8166 * reads, since we may be about to write
8167 * over this location.
8168 */
8169 if (HDR_L2_READING(hdr)) {
8170 ARCSTAT_BUMP(arcstat_l2_evict_reading);
8171 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
8172 }
8173
8174 arc_hdr_l2hdr_destroy(hdr);
8175 }
8176 mutex_exit(hash_lock);
8177 }
8178 mutex_exit(&dev->l2ad_mtx);
8179
8180 out:
8181 /*
8182 * We need to check if we evict all buffers, otherwise we may iterate
8183 * unnecessarily.
8184 */
8185 if (!all && rerun) {
8186 /*
8187 * Bump device hand to the device start if it is approaching the
8188 * end. l2arc_evict() has already evicted ahead for this case.
8189 */
8190 dev->l2ad_hand = dev->l2ad_start;
8191 dev->l2ad_evict = dev->l2ad_start;
8192 dev->l2ad_first = B_FALSE;
8193 goto top;
8194 }
8195
8196 ASSERT3U(dev->l2ad_hand + distance, <, dev->l2ad_end);
8197 if (!dev->l2ad_first)
8198 ASSERT3U(dev->l2ad_hand, <, dev->l2ad_evict);
8199 }
8200
8201 /*
8202 * Handle any abd transforms that might be required for writing to the L2ARC.
8203 * If successful, this function will always return an abd with the data
8204 * transformed as it is on disk in a new abd of asize bytes.
8205 */
8206 static int
8207 l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
8208 abd_t **abd_out)
8209 {
8210 int ret;
8211 void *tmp = NULL;
8212 abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd;
8213 enum zio_compress compress = HDR_GET_COMPRESS(hdr);
8214 uint64_t psize = HDR_GET_PSIZE(hdr);
8215 uint64_t size = arc_hdr_size(hdr);
8216 boolean_t ismd = HDR_ISTYPE_METADATA(hdr);
8217 boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS);
8218 dsl_crypto_key_t *dck = NULL;
8219 uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 };
8220 boolean_t no_crypt = B_FALSE;
8221
8222 ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF &&
8223 !HDR_COMPRESSION_ENABLED(hdr)) ||
8224 HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize);
8225 ASSERT3U(psize, <=, asize);
8226
8227 /*
8228 * If this data simply needs its own buffer, we simply allocate it
8229 * and copy the data. This may be done to eliminate a dependency on a
8230 * shared buffer or to reallocate the buffer to match asize.
8231 */
8232 if (HDR_HAS_RABD(hdr) && asize != psize) {
8233 ASSERT3U(asize, >=, psize);
8234 to_write = abd_alloc_for_io(asize, ismd);
8235 abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize);
8236 if (psize != asize)
8237 abd_zero_off(to_write, psize, asize - psize);
8238 goto out;
8239 }
8240
8241 if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) &&
8242 !HDR_ENCRYPTED(hdr)) {
8243 ASSERT3U(size, ==, psize);
8244 to_write = abd_alloc_for_io(asize, ismd);
8245 abd_copy(to_write, hdr->b_l1hdr.b_pabd, size);
8246 if (size != asize)
8247 abd_zero_off(to_write, size, asize - size);
8248 goto out;
8249 }
8250
8251 if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) {
8252 cabd = abd_alloc_for_io(asize, ismd);
8253 tmp = abd_borrow_buf(cabd, asize);
8254
8255 psize = zio_compress_data(compress, to_write, tmp, size);
8256 ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr));
8257 if (psize < asize)
8258 bzero((char *)tmp + psize, asize - psize);
8259 psize = HDR_GET_PSIZE(hdr);
8260 abd_return_buf_copy(cabd, tmp, asize);
8261 to_write = cabd;
8262 }
8263
8264 if (HDR_ENCRYPTED(hdr)) {
8265 eabd = abd_alloc_for_io(asize, ismd);
8266
8267 /*
8268 * If the dataset was disowned before the buffer
8269 * made it to this point, the key to re-encrypt
8270 * it won't be available. In this case we simply
8271 * won't write the buffer to the L2ARC.
8272 */
8273 ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj,
8274 FTAG, &dck);
8275 if (ret != 0)
8276 goto error;
8277
8278 ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
8279 hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
8280 hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
8281 &no_crypt);
8282 if (ret != 0)
8283 goto error;
8284
8285 if (no_crypt)
8286 abd_copy(eabd, to_write, psize);
8287
8288 if (psize != asize)
8289 abd_zero_off(eabd, psize, asize - psize);
8290
8291 /* assert that the MAC we got here matches the one we saved */
8292 ASSERT0(bcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN));
8293 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8294
8295 if (to_write == cabd)
8296 abd_free(cabd);
8297
8298 to_write = eabd;
8299 }
8300
8301 out:
8302 ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd);
8303 *abd_out = to_write;
8304 return (0);
8305
8306 error:
8307 if (dck != NULL)
8308 spa_keystore_dsl_key_rele(spa, dck, FTAG);
8309 if (cabd != NULL)
8310 abd_free(cabd);
8311 if (eabd != NULL)
8312 abd_free(eabd);
8313
8314 *abd_out = NULL;
8315 return (ret);
8316 }
8317
8318 static void
8319 l2arc_blk_fetch_done(zio_t *zio)
8320 {
8321 l2arc_read_callback_t *cb;
8322
8323 cb = zio->io_private;
8324 if (cb->l2rcb_abd != NULL)
8325 abd_put(cb->l2rcb_abd);
8326 kmem_free(cb, sizeof (l2arc_read_callback_t));
8327 }
8328
8329 /*
8330 * Find and write ARC buffers to the L2ARC device.
8331 *
8332 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
8333 * for reading until they have completed writing.
8334 * The headroom_boost is an in-out parameter used to maintain headroom boost
8335 * state between calls to this function.
8336 *
8337 * Returns the number of bytes actually written (which may be smaller than
8338 * the delta by which the device hand has changed due to alignment and the
8339 * writing of log blocks).
8340 */
8341 static uint64_t
8342 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
8343 {
8344 arc_buf_hdr_t *hdr, *hdr_prev, *head;
8345 uint64_t write_asize, write_psize, write_lsize, headroom;
8346 boolean_t full;
8347 l2arc_write_callback_t *cb = NULL;
8348 zio_t *pio, *wzio;
8349 uint64_t guid = spa_load_guid(spa);
8350
8351 ASSERT3P(dev->l2ad_vdev, !=, NULL);
8352
8353 pio = NULL;
8354 write_lsize = write_asize = write_psize = 0;
8355 full = B_FALSE;
8356 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
8357 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
8358
8359 /*
8360 * Copy buffers for L2ARC writing.
8361 */
8362 for (int try = 0; try <= 3; try++) {
8363 multilist_sublist_t *mls = l2arc_sublist_lock(try);
8364 uint64_t passed_sz = 0;
8365
8366 VERIFY3P(mls, !=, NULL);
8367
8368 /*
8369 * L2ARC fast warmup.
8370 *
8371 * Until the ARC is warm and starts to evict, read from the
8372 * head of the ARC lists rather than the tail.
8373 */
8374 if (arc_warm == B_FALSE)
8375 hdr = multilist_sublist_head(mls);
8376 else
8377 hdr = multilist_sublist_tail(mls);
8378
8379 headroom = target_sz * l2arc_headroom;
8380 if (zfs_compressed_arc_enabled)
8381 headroom = (headroom * l2arc_headroom_boost) / 100;
8382
8383 for (; hdr; hdr = hdr_prev) {
8384 kmutex_t *hash_lock;
8385 abd_t *to_write = NULL;
8386
8387 if (arc_warm == B_FALSE)
8388 hdr_prev = multilist_sublist_next(mls, hdr);
8389 else
8390 hdr_prev = multilist_sublist_prev(mls, hdr);
8391
8392 hash_lock = HDR_LOCK(hdr);
8393 if (!mutex_tryenter(hash_lock)) {
8394 /*
8395 * Skip this buffer rather than waiting.
8396 */
8397 continue;
8398 }
8399
8400 passed_sz += HDR_GET_LSIZE(hdr);
8401 if (l2arc_headroom != 0 && passed_sz > headroom) {
8402 /*
8403 * Searched too far.
8404 */
8405 mutex_exit(hash_lock);
8406 break;
8407 }
8408
8409 if (!l2arc_write_eligible(guid, hdr)) {
8410 mutex_exit(hash_lock);
8411 continue;
8412 }
8413
8414 /*
8415 * We rely on the L1 portion of the header below, so
8416 * it's invalid for this header to have been evicted out
8417 * of the ghost cache, prior to being written out. The
8418 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8419 */
8420 ASSERT(HDR_HAS_L1HDR(hdr));
8421
8422 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8423 ASSERT3U(arc_hdr_size(hdr), >, 0);
8424 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
8425 HDR_HAS_RABD(hdr));
8426 uint64_t psize = HDR_GET_PSIZE(hdr);
8427 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
8428 psize);
8429
8430 if ((write_asize + asize) > target_sz) {
8431 full = B_TRUE;
8432 mutex_exit(hash_lock);
8433 break;
8434 }
8435
8436 /*
8437 * We rely on the L1 portion of the header below, so
8438 * it's invalid for this header to have been evicted out
8439 * of the ghost cache, prior to being written out. The
8440 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
8441 */
8442 arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING);
8443 ASSERT(HDR_HAS_L1HDR(hdr));
8444
8445 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
8446 ASSERT(hdr->b_l1hdr.b_pabd != NULL ||
8447 HDR_HAS_RABD(hdr));
8448 ASSERT3U(arc_hdr_size(hdr), >, 0);
8449
8450 /*
8451 * If this header has b_rabd, we can use this since it
8452 * must always match the data exactly as it exists on
8453 * disk. Otherwise, the L2ARC can normally use the
8454 * hdr's data, but if we're sharing data between the
8455 * hdr and one of its bufs, L2ARC needs its own copy of
8456 * the data so that the ZIO below can't race with the
8457 * buf consumer. To ensure that this copy will be
8458 * available for the lifetime of the ZIO and be cleaned
8459 * up afterwards, we add it to the l2arc_free_on_write
8460 * queue. If we need to apply any transforms to the
8461 * data (compression, encryption) we will also need the
8462 * extra buffer.
8463 */
8464 if (HDR_HAS_RABD(hdr) && psize == asize) {
8465 to_write = hdr->b_crypt_hdr.b_rabd;
8466 } else if ((HDR_COMPRESSION_ENABLED(hdr) ||
8467 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) &&
8468 !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) &&
8469 psize == asize) {
8470 to_write = hdr->b_l1hdr.b_pabd;
8471 } else {
8472 int ret;
8473 arc_buf_contents_t type = arc_buf_type(hdr);
8474
8475 ret = l2arc_apply_transforms(spa, hdr, asize,
8476 &to_write);
8477 if (ret != 0) {
8478 arc_hdr_clear_flags(hdr,
8479 ARC_FLAG_L2_WRITING);
8480 mutex_exit(hash_lock);
8481 continue;
8482 }
8483
8484 l2arc_free_abd_on_write(to_write, asize, type);
8485 }
8486
8487 if (pio == NULL) {
8488 /*
8489 * Insert a dummy header on the buflist so
8490 * l2arc_write_done() can find where the
8491 * write buffers begin without searching.
8492 */
8493 mutex_enter(&dev->l2ad_mtx);
8494 list_insert_head(&dev->l2ad_buflist, head);
8495 mutex_exit(&dev->l2ad_mtx);
8496
8497 cb = kmem_alloc(
8498 sizeof (l2arc_write_callback_t), KM_SLEEP);
8499 cb->l2wcb_dev = dev;
8500 cb->l2wcb_head = head;
8501 /*
8502 * Create a list to save allocated abd buffers
8503 * for l2arc_log_blk_commit().
8504 */
8505 list_create(&cb->l2wcb_abd_list,
8506 sizeof (l2arc_lb_abd_buf_t),
8507 offsetof(l2arc_lb_abd_buf_t, node));
8508 pio = zio_root(spa, l2arc_write_done, cb,
8509 ZIO_FLAG_CANFAIL);
8510 }
8511
8512 hdr->b_l2hdr.b_dev = dev;
8513 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
8514 arc_hdr_set_flags(hdr,
8515 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
8516
8517 mutex_enter(&dev->l2ad_mtx);
8518 list_insert_head(&dev->l2ad_buflist, hdr);
8519 mutex_exit(&dev->l2ad_mtx);
8520
8521 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
8522 arc_hdr_size(hdr), hdr);
8523
8524 wzio = zio_write_phys(pio, dev->l2ad_vdev,
8525 hdr->b_l2hdr.b_daddr, asize, to_write,
8526 ZIO_CHECKSUM_OFF, NULL, hdr,
8527 ZIO_PRIORITY_ASYNC_WRITE,
8528 ZIO_FLAG_CANFAIL, B_FALSE);
8529
8530 write_lsize += HDR_GET_LSIZE(hdr);
8531 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
8532 zio_t *, wzio);
8533
8534 write_psize += psize;
8535 write_asize += asize;
8536 dev->l2ad_hand += asize;
8537 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
8538
8539 mutex_exit(hash_lock);
8540
8541 /*
8542 * Append buf info to current log and commit if full.
8543 * arcstat_l2_{size,asize} kstats are updated
8544 * internally.
8545 */
8546 if (l2arc_log_blk_insert(dev, hdr))
8547 l2arc_log_blk_commit(dev, pio, cb);
8548
8549 (void) zio_nowait(wzio);
8550 }
8551
8552 multilist_sublist_unlock(mls);
8553
8554 if (full == B_TRUE)
8555 break;
8556 }
8557
8558 /* No buffers selected for writing? */
8559 if (pio == NULL) {
8560 ASSERT0(write_lsize);
8561 ASSERT(!HDR_HAS_L1HDR(head));
8562 kmem_cache_free(hdr_l2only_cache, head);
8563
8564 /*
8565 * Although we did not write any buffers l2ad_evict may
8566 * have advanced.
8567 */
8568 l2arc_dev_hdr_update(dev);
8569
8570 return (0);
8571 }
8572
8573 if (!dev->l2ad_first)
8574 ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict);
8575
8576 ASSERT3U(write_asize, <=, target_sz);
8577 ARCSTAT_BUMP(arcstat_l2_writes_sent);
8578 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
8579 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
8580 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
8581
8582 dev->l2ad_writing = B_TRUE;
8583 (void) zio_wait(pio);
8584 dev->l2ad_writing = B_FALSE;
8585
8586 /*
8587 * Update the device header after the zio completes as
8588 * l2arc_write_done() may have updated the memory holding the log block
8589 * pointers in the device header.
8590 */
8591 l2arc_dev_hdr_update(dev);
8592
8593 return (write_asize);
8594 }
8595
8596 /*
8597 * This thread feeds the L2ARC at regular intervals. This is the beating
8598 * heart of the L2ARC.
8599 */
8600 /* ARGSUSED */
8601 static void
8602 l2arc_feed_thread(void *unused)
8603 {
8604 callb_cpr_t cpr;
8605 l2arc_dev_t *dev;
8606 spa_t *spa;
8607 uint64_t size, wrote;
8608 clock_t begin, next = ddi_get_lbolt();
8609
8610 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
8611
8612 mutex_enter(&l2arc_feed_thr_lock);
8613
8614 while (l2arc_thread_exit == 0) {
8615 CALLB_CPR_SAFE_BEGIN(&cpr);
8616 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
8617 next);
8618 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
8619 next = ddi_get_lbolt() + hz;
8620
8621 /*
8622 * Quick check for L2ARC devices.
8623 */
8624 mutex_enter(&l2arc_dev_mtx);
8625 if (l2arc_ndev == 0) {
8626 mutex_exit(&l2arc_dev_mtx);
8627 continue;
8628 }
8629 mutex_exit(&l2arc_dev_mtx);
8630 begin = ddi_get_lbolt();
8631
8632 /*
8633 * This selects the next l2arc device to write to, and in
8634 * doing so the next spa to feed from: dev->l2ad_spa. This
8635 * will return NULL if there are now no l2arc devices or if
8636 * they are all faulted.
8637 *
8638 * If a device is returned, its spa's config lock is also
8639 * held to prevent device removal. l2arc_dev_get_next()
8640 * will grab and release l2arc_dev_mtx.
8641 */
8642 if ((dev = l2arc_dev_get_next()) == NULL)
8643 continue;
8644
8645 spa = dev->l2ad_spa;
8646 ASSERT3P(spa, !=, NULL);
8647
8648 /*
8649 * If the pool is read-only then force the feed thread to
8650 * sleep a little longer.
8651 */
8652 if (!spa_writeable(spa)) {
8653 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
8654 spa_config_exit(spa, SCL_L2ARC, dev);
8655 continue;
8656 }
8657
8658 /*
8659 * Avoid contributing to memory pressure.
8660 */
8661 if (arc_reclaim_needed()) {
8662 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
8663 spa_config_exit(spa, SCL_L2ARC, dev);
8664 continue;
8665 }
8666
8667 ARCSTAT_BUMP(arcstat_l2_feeds);
8668
8669 size = l2arc_write_size(dev);
8670
8671 /*
8672 * Evict L2ARC buffers that will be overwritten.
8673 */
8674 l2arc_evict(dev, size, B_FALSE);
8675
8676 /*
8677 * Write ARC buffers.
8678 */
8679 wrote = l2arc_write_buffers(spa, dev, size);
8680
8681 /*
8682 * Calculate interval between writes.
8683 */
8684 next = l2arc_write_interval(begin, size, wrote);
8685 spa_config_exit(spa, SCL_L2ARC, dev);
8686 }
8687
8688 l2arc_thread_exit = 0;
8689 cv_broadcast(&l2arc_feed_thr_cv);
8690 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
8691 thread_exit();
8692 }
8693
8694 boolean_t
8695 l2arc_vdev_present(vdev_t *vd)
8696 {
8697 return (l2arc_vdev_get(vd) != NULL);
8698 }
8699
8700 /*
8701 * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if
8702 * the vdev_t isn't an L2ARC device.
8703 */
8704 static l2arc_dev_t *
8705 l2arc_vdev_get(vdev_t *vd)
8706 {
8707 l2arc_dev_t *dev;
8708
8709 mutex_enter(&l2arc_dev_mtx);
8710 for (dev = list_head(l2arc_dev_list); dev != NULL;
8711 dev = list_next(l2arc_dev_list, dev)) {
8712 if (dev->l2ad_vdev == vd)
8713 break;
8714 }
8715 mutex_exit(&l2arc_dev_mtx);
8716
8717 return (dev);
8718 }
8719
8720 /*
8721 * Add a vdev for use by the L2ARC. By this point the spa has already
8722 * validated the vdev and opened it.
8723 */
8724 void
8725 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
8726 {
8727 l2arc_dev_t *adddev;
8728 uint64_t l2dhdr_asize;
8729
8730 ASSERT(!l2arc_vdev_present(vd));
8731
8732 /*
8733 * Create a new l2arc device entry.
8734 */
8735 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
8736 adddev->l2ad_spa = spa;
8737 adddev->l2ad_vdev = vd;
8738 /* leave extra size for an l2arc device header */
8739 l2dhdr_asize = adddev->l2ad_dev_hdr_asize =
8740 MAX(sizeof (*adddev->l2ad_dev_hdr), 1 << vd->vdev_ashift);
8741 adddev->l2ad_start = VDEV_LABEL_START_SIZE + l2dhdr_asize;
8742 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
8743 ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end);
8744 adddev->l2ad_hand = adddev->l2ad_start;
8745 adddev->l2ad_evict = adddev->l2ad_start;
8746 adddev->l2ad_first = B_TRUE;
8747 adddev->l2ad_writing = B_FALSE;
8748 adddev->l2ad_dev_hdr = kmem_zalloc(l2dhdr_asize, KM_SLEEP);
8749
8750 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
8751 /*
8752 * This is a list of all ARC buffers that are still valid on the
8753 * device.
8754 */
8755 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
8756 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
8757
8758 /*
8759 * This is a list of pointers to log blocks that are still present
8760 * on the device.
8761 */
8762 list_create(&adddev->l2ad_lbptr_list, sizeof (l2arc_lb_ptr_buf_t),
8763 offsetof(l2arc_lb_ptr_buf_t, node));
8764
8765 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
8766 zfs_refcount_create(&adddev->l2ad_alloc);
8767 zfs_refcount_create(&adddev->l2ad_lb_asize);
8768 zfs_refcount_create(&adddev->l2ad_lb_count);
8769
8770 /*
8771 * Add device to global list
8772 */
8773 mutex_enter(&l2arc_dev_mtx);
8774 list_insert_head(l2arc_dev_list, adddev);
8775 atomic_inc_64(&l2arc_ndev);
8776 mutex_exit(&l2arc_dev_mtx);
8777
8778 /*
8779 * Decide if vdev is eligible for L2ARC rebuild
8780 */
8781 l2arc_rebuild_vdev(adddev->l2ad_vdev, B_FALSE);
8782 }
8783
8784 void
8785 l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen)
8786 {
8787 l2arc_dev_t *dev = NULL;
8788 l2arc_dev_hdr_phys_t *l2dhdr;
8789 uint64_t l2dhdr_asize;
8790 spa_t *spa;
8791 int err;
8792 boolean_t l2dhdr_valid = B_TRUE;
8793
8794 dev = l2arc_vdev_get(vd);
8795 ASSERT3P(dev, !=, NULL);
8796 spa = dev->l2ad_spa;
8797 l2dhdr = dev->l2ad_dev_hdr;
8798 l2dhdr_asize = dev->l2ad_dev_hdr_asize;
8799
8800 /*
8801 * The L2ARC has to hold at least the payload of one log block for
8802 * them to be restored (persistent L2ARC). The payload of a log block
8803 * depends on the amount of its log entries. We always write log blocks
8804 * with 1022 entries. How many of them are committed or restored depends
8805 * on the size of the L2ARC device. Thus the maximum payload of
8806 * one log block is 1022 * SPA_MAXBLOCKSIZE = 16GB. If the L2ARC device
8807 * is less than that, we reduce the amount of committed and restored
8808 * log entries per block so as to enable persistence.
8809 */
8810 if (dev->l2ad_end < l2arc_rebuild_blocks_min_l2size) {
8811 dev->l2ad_log_entries = 0;
8812 } else {
8813 dev->l2ad_log_entries = MIN((dev->l2ad_end -
8814 dev->l2ad_start) >> SPA_MAXBLOCKSHIFT,
8815 L2ARC_LOG_BLK_MAX_ENTRIES);
8816 }
8817
8818 /*
8819 * Read the device header, if an error is returned do not rebuild L2ARC.
8820 */
8821 if ((err = l2arc_dev_hdr_read(dev)) != 0)
8822 l2dhdr_valid = B_FALSE;
8823
8824 if (l2dhdr_valid && dev->l2ad_log_entries > 0) {
8825 /*
8826 * If we are onlining a cache device (vdev_reopen) that was
8827 * still present (l2arc_vdev_present()) and rebuild is enabled,
8828 * we should evict all ARC buffers and pointers to log blocks
8829 * and reclaim their space before restoring its contents to
8830 * L2ARC.
8831 */
8832 if (reopen) {
8833 if (!l2arc_rebuild_enabled) {
8834 return;
8835 } else {
8836 l2arc_evict(dev, 0, B_TRUE);
8837 /* start a new log block */
8838 dev->l2ad_log_ent_idx = 0;
8839 dev->l2ad_log_blk_payload_asize = 0;
8840 dev->l2ad_log_blk_payload_start = 0;
8841 }
8842 }
8843 /*
8844 * Just mark the device as pending for a rebuild. We won't
8845 * be starting a rebuild in line here as it would block pool
8846 * import. Instead spa_load_impl will hand that off to an
8847 * async task which will call l2arc_spa_rebuild_start.
8848 */
8849 dev->l2ad_rebuild = B_TRUE;
8850 } else if (spa_writeable(spa)) {
8851 /*
8852 * In this case create a new header. We zero out the memory
8853 * holding the header to reset dh_start_lbps.
8854 */
8855 bzero(l2dhdr, l2dhdr_asize);
8856 l2arc_dev_hdr_update(dev);
8857 }
8858 }
8859
8860 /*
8861 * Remove a vdev from the L2ARC.
8862 */
8863 void
8864 l2arc_remove_vdev(vdev_t *vd)
8865 {
8866 l2arc_dev_t *remdev = NULL;
8867
8868 /*
8869 * Find the device by vdev
8870 */
8871 remdev = l2arc_vdev_get(vd);
8872 ASSERT3P(remdev, !=, NULL);
8873
8874 /*
8875 * Cancel any ongoing or scheduled rebuild.
8876 */
8877 mutex_enter(&l2arc_rebuild_thr_lock);
8878 if (remdev->l2ad_rebuild_began == B_TRUE) {
8879 remdev->l2ad_rebuild_cancel = B_TRUE;
8880 while (remdev->l2ad_rebuild == B_TRUE)
8881 cv_wait(&l2arc_rebuild_thr_cv, &l2arc_rebuild_thr_lock);
8882 }
8883 mutex_exit(&l2arc_rebuild_thr_lock);
8884
8885 /*
8886 * Remove device from global list
8887 */
8888 mutex_enter(&l2arc_dev_mtx);
8889 list_remove(l2arc_dev_list, remdev);
8890 l2arc_dev_last = NULL; /* may have been invalidated */
8891 atomic_dec_64(&l2arc_ndev);
8892 mutex_exit(&l2arc_dev_mtx);
8893
8894 /*
8895 * Clear all buflists and ARC references. L2ARC device flush.
8896 */
8897 l2arc_evict(remdev, 0, B_TRUE);
8898 list_destroy(&remdev->l2ad_buflist);
8899 ASSERT(list_is_empty(&remdev->l2ad_lbptr_list));
8900 list_destroy(&remdev->l2ad_lbptr_list);
8901 mutex_destroy(&remdev->l2ad_mtx);
8902 zfs_refcount_destroy(&remdev->l2ad_alloc);
8903 zfs_refcount_destroy(&remdev->l2ad_lb_asize);
8904 zfs_refcount_destroy(&remdev->l2ad_lb_count);
8905 kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize);
8906 kmem_free(remdev, sizeof (l2arc_dev_t));
8907 }
8908
8909 void
8910 l2arc_init(void)
8911 {
8912 l2arc_thread_exit = 0;
8913 l2arc_ndev = 0;
8914 l2arc_writes_sent = 0;
8915 l2arc_writes_done = 0;
8916
8917 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
8918 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
8919 mutex_init(&l2arc_rebuild_thr_lock, NULL, MUTEX_DEFAULT, NULL);
8920 cv_init(&l2arc_rebuild_thr_cv, NULL, CV_DEFAULT, NULL);
8921 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
8922 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
8923
8924 l2arc_dev_list = &L2ARC_dev_list;
8925 l2arc_free_on_write = &L2ARC_free_on_write;
8926 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
8927 offsetof(l2arc_dev_t, l2ad_node));
8928 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
8929 offsetof(l2arc_data_free_t, l2df_list_node));
8930 }
8931
8932 void
8933 l2arc_fini(void)
8934 {
8935 /*
8936 * This is called from dmu_fini(), which is called from spa_fini();
8937 * Because of this, we can assume that all l2arc devices have
8938 * already been removed when the pools themselves were removed.
8939 */
8940
8941 l2arc_do_free_on_write();
8942
8943 mutex_destroy(&l2arc_feed_thr_lock);
8944 cv_destroy(&l2arc_feed_thr_cv);
8945 mutex_destroy(&l2arc_rebuild_thr_lock);
8946 cv_destroy(&l2arc_rebuild_thr_cv);
8947 mutex_destroy(&l2arc_dev_mtx);
8948 mutex_destroy(&l2arc_free_on_write_mtx);
8949
8950 list_destroy(l2arc_dev_list);
8951 list_destroy(l2arc_free_on_write);
8952 }
8953
8954 void
8955 l2arc_start(void)
8956 {
8957 if (!(spa_mode_global & FWRITE))
8958 return;
8959
8960 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
8961 TS_RUN, minclsyspri);
8962 }
8963
8964 void
8965 l2arc_stop(void)
8966 {
8967 if (!(spa_mode_global & FWRITE))
8968 return;
8969
8970 mutex_enter(&l2arc_feed_thr_lock);
8971 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
8972 l2arc_thread_exit = 1;
8973 while (l2arc_thread_exit != 0)
8974 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
8975 mutex_exit(&l2arc_feed_thr_lock);
8976 }
8977
8978 /*
8979 * Punches out rebuild threads for the L2ARC devices in a spa. This should
8980 * be called after pool import from the spa async thread, since starting
8981 * these threads directly from spa_import() will make them part of the
8982 * "zpool import" context and delay process exit (and thus pool import).
8983 */
8984 void
8985 l2arc_spa_rebuild_start(spa_t *spa)
8986 {
8987 ASSERT(MUTEX_HELD(&spa_namespace_lock));
8988
8989 /*
8990 * Locate the spa's l2arc devices and kick off rebuild threads.
8991 */
8992 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
8993 l2arc_dev_t *dev =
8994 l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]);
8995 if (dev == NULL) {
8996 /* Don't attempt a rebuild if the vdev is UNAVAIL */
8997 continue;
8998 }
8999 mutex_enter(&l2arc_rebuild_thr_lock);
9000 if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) {
9001 dev->l2ad_rebuild_began = B_TRUE;
9002 (void) thread_create(NULL, 0,
9003 (void (*)(void *))l2arc_dev_rebuild_start,
9004 dev, 0, &p0, TS_RUN, minclsyspri);
9005 }
9006 mutex_exit(&l2arc_rebuild_thr_lock);
9007 }
9008 }
9009
9010 /*
9011 * Main entry point for L2ARC rebuilding.
9012 */
9013 static void
9014 l2arc_dev_rebuild_start(l2arc_dev_t *dev)
9015 {
9016 VERIFY(!dev->l2ad_rebuild_cancel);
9017 VERIFY(dev->l2ad_rebuild);
9018 (void) l2arc_rebuild(dev);
9019 mutex_enter(&l2arc_rebuild_thr_lock);
9020 dev->l2ad_rebuild_began = B_FALSE;
9021 dev->l2ad_rebuild = B_FALSE;
9022 mutex_exit(&l2arc_rebuild_thr_lock);
9023
9024 thread_exit();
9025 }
9026
9027 /*
9028 * This function implements the actual L2ARC metadata rebuild. It:
9029 * starts reading the log block chain and restores each block's contents
9030 * to memory (reconstructing arc_buf_hdr_t's).
9031 *
9032 * Operation stops under any of the following conditions:
9033 *
9034 * 1) We reach the end of the log block chain.
9035 * 2) We encounter *any* error condition (cksum errors, io errors)
9036 */
9037 static int
9038 l2arc_rebuild(l2arc_dev_t *dev)
9039 {
9040 vdev_t *vd = dev->l2ad_vdev;
9041 spa_t *spa = vd->vdev_spa;
9042 int err = 0;
9043 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9044 l2arc_log_blk_phys_t *this_lb, *next_lb;
9045 zio_t *this_io = NULL, *next_io = NULL;
9046 l2arc_log_blkptr_t lbps[2];
9047 l2arc_lb_ptr_buf_t *lb_ptr_buf;
9048 boolean_t lock_held;
9049
9050 this_lb = kmem_zalloc(sizeof (*this_lb), KM_SLEEP);
9051 next_lb = kmem_zalloc(sizeof (*next_lb), KM_SLEEP);
9052
9053 /*
9054 * We prevent device removal while issuing reads to the device,
9055 * then during the rebuilding phases we drop this lock again so
9056 * that a spa_unload or device remove can be initiated - this is
9057 * safe, because the spa will signal us to stop before removing
9058 * our device and wait for us to stop.
9059 */
9060 spa_config_enter(spa, SCL_L2ARC, vd, RW_READER);
9061 lock_held = B_TRUE;
9062
9063 /*
9064 * Retrieve the persistent L2ARC device state.
9065 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9066 */
9067 dev->l2ad_evict = MAX(l2dhdr->dh_evict, dev->l2ad_start);
9068 dev->l2ad_hand = MAX(l2dhdr->dh_start_lbps[0].lbp_daddr +
9069 L2BLK_GET_PSIZE((&l2dhdr->dh_start_lbps[0])->lbp_prop),
9070 dev->l2ad_start);
9071 dev->l2ad_first = !!(l2dhdr->dh_flags & L2ARC_DEV_HDR_EVICT_FIRST);
9072
9073 /*
9074 * In case the zfs module parameter l2arc_rebuild_enabled is false
9075 * we do not start the rebuild process.
9076 */
9077 if (!l2arc_rebuild_enabled)
9078 goto out;
9079
9080 /* Prepare the rebuild process */
9081 bcopy(l2dhdr->dh_start_lbps, lbps, sizeof (lbps));
9082
9083 /* Start the rebuild process */
9084 for (;;) {
9085 if (!l2arc_log_blkptr_valid(dev, &lbps[0]))
9086 break;
9087
9088 if ((err = l2arc_log_blk_read(dev, &lbps[0], &lbps[1],
9089 this_lb, next_lb, this_io, &next_io)) != 0)
9090 goto out;
9091
9092 /*
9093 * Our memory pressure valve. If the system is running low
9094 * on memory, rather than swamping memory with new ARC buf
9095 * hdrs, we opt not to rebuild the L2ARC. At this point,
9096 * however, we have already set up our L2ARC dev to chain in
9097 * new metadata log blocks, so the user may choose to offline/
9098 * online the L2ARC dev at a later time (or re-import the pool)
9099 * to reconstruct it (when there's less memory pressure).
9100 */
9101 if (arc_reclaim_needed()) {
9102 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem);
9103 cmn_err(CE_NOTE, "System running low on memory, "
9104 "aborting L2ARC rebuild.");
9105 err = SET_ERROR(ENOMEM);
9106 goto out;
9107 }
9108
9109 spa_config_exit(spa, SCL_L2ARC, vd);
9110 lock_held = B_FALSE;
9111
9112 /*
9113 * Now that we know that the next_lb checks out alright, we
9114 * can start reconstruction from this log block.
9115 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9116 */
9117 uint64_t asize = L2BLK_GET_PSIZE((&lbps[0])->lbp_prop);
9118 l2arc_log_blk_restore(dev, this_lb, asize, lbps[0].lbp_daddr);
9119
9120 /*
9121 * log block restored, include its pointer in the list of
9122 * pointers to log blocks present in the L2ARC device.
9123 */
9124 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
9125 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t),
9126 KM_SLEEP);
9127 bcopy(&lbps[0], lb_ptr_buf->lb_ptr,
9128 sizeof (l2arc_log_blkptr_t));
9129 mutex_enter(&dev->l2ad_mtx);
9130 list_insert_tail(&dev->l2ad_lbptr_list, lb_ptr_buf);
9131 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
9132 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
9133 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
9134 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
9135 mutex_exit(&dev->l2ad_mtx);
9136 vdev_space_update(vd, asize, 0, 0);
9137
9138 /* BEGIN CSTYLED */
9139 /*
9140 * Protection against loops of log blocks:
9141 *
9142 * l2ad_hand l2ad_evict
9143 * V V
9144 * l2ad_start |=======================================| l2ad_end
9145 * -----|||----|||---|||----|||
9146 * (3) (2) (1) (0)
9147 * ---|||---|||----|||---|||
9148 * (7) (6) (5) (4)
9149 *
9150 * In this situation the pointer of log block (4) passes
9151 * l2arc_log_blkptr_valid() but the log block should not be
9152 * restored as it is overwritten by the payload of log block
9153 * (0). Only log blocks (0)-(3) should be restored. We check
9154 * whether l2ad_evict lies in between the payload starting
9155 * offset of the next log block (lbps[1].lbp_payload_start)
9156 * and the payload starting offset of the present log block
9157 * (lbps[0].lbp_payload_start). If true and this isn't the
9158 * first pass, we are looping from the beginning and we should
9159 * stop.
9160 */
9161 /* END CSTYLED */
9162 if (l2arc_range_check_overlap(lbps[1].lbp_payload_start,
9163 lbps[0].lbp_payload_start, dev->l2ad_evict) &&
9164 !dev->l2ad_first)
9165 goto out;
9166
9167 for (;;) {
9168 mutex_enter(&l2arc_rebuild_thr_lock);
9169 if (dev->l2ad_rebuild_cancel) {
9170 dev->l2ad_rebuild = B_FALSE;
9171 cv_signal(&l2arc_rebuild_thr_cv);
9172 mutex_exit(&l2arc_rebuild_thr_lock);
9173 err = SET_ERROR(ECANCELED);
9174 goto out;
9175 }
9176 mutex_exit(&l2arc_rebuild_thr_lock);
9177 if (spa_config_tryenter(spa, SCL_L2ARC, vd,
9178 RW_READER)) {
9179 lock_held = B_TRUE;
9180 break;
9181 }
9182 /*
9183 * L2ARC config lock held by somebody in writer,
9184 * possibly due to them trying to remove us. They'll
9185 * likely to want us to shut down, so after a little
9186 * delay, we check l2ad_rebuild_cancel and retry
9187 * the lock again.
9188 */
9189 delay(1);
9190 }
9191
9192 /*
9193 * Continue with the next log block.
9194 */
9195 lbps[0] = lbps[1];
9196 lbps[1] = this_lb->lb_prev_lbp;
9197 PTR_SWAP(this_lb, next_lb);
9198 this_io = next_io;
9199 next_io = NULL;
9200 }
9201
9202 if (this_io != NULL)
9203 l2arc_log_blk_fetch_abort(this_io);
9204 out:
9205 if (next_io != NULL)
9206 l2arc_log_blk_fetch_abort(next_io);
9207 kmem_free(this_lb, sizeof (*this_lb));
9208 kmem_free(next_lb, sizeof (*next_lb));
9209
9210 if (!l2arc_rebuild_enabled) {
9211 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9212 "disabled");
9213 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) > 0) {
9214 ARCSTAT_BUMP(arcstat_l2_rebuild_success);
9215 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9216 "successful, restored %llu blocks",
9217 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
9218 } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) == 0) {
9219 /*
9220 * No error but also nothing restored, meaning the lbps array
9221 * in the device header points to invalid/non-present log
9222 * blocks. Reset the header.
9223 */
9224 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9225 "no valid log blocks");
9226 bzero(l2dhdr, dev->l2ad_dev_hdr_asize);
9227 l2arc_dev_hdr_update(dev);
9228 } else if (err != 0) {
9229 spa_history_log_internal(spa, "L2ARC rebuild", NULL,
9230 "aborted, restored %llu blocks",
9231 (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count));
9232 }
9233
9234 if (lock_held)
9235 spa_config_exit(spa, SCL_L2ARC, vd);
9236
9237 return (err);
9238 }
9239
9240 /*
9241 * Attempts to read the device header on the provided L2ARC device and writes
9242 * it to `hdr'. On success, this function returns 0, otherwise the appropriate
9243 * error code is returned.
9244 */
9245 static int
9246 l2arc_dev_hdr_read(l2arc_dev_t *dev)
9247 {
9248 int err;
9249 uint64_t guid;
9250 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9251 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9252 abd_t *abd;
9253
9254 guid = spa_guid(dev->l2ad_vdev->vdev_spa);
9255
9256 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
9257
9258 err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev,
9259 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd,
9260 ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
9261 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
9262 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY |
9263 ZIO_FLAG_SPECULATIVE, B_FALSE));
9264
9265 abd_put(abd);
9266
9267 if (err != 0) {
9268 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_dh_errors);
9269 zfs_dbgmsg("L2ARC IO error (%d) while reading device header, "
9270 "vdev guid: %llu", err, dev->l2ad_vdev->vdev_guid);
9271 return (err);
9272 }
9273
9274 if (l2dhdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC))
9275 byteswap_uint64_array(l2dhdr, sizeof (*l2dhdr));
9276
9277 if (l2dhdr->dh_magic != L2ARC_DEV_HDR_MAGIC ||
9278 l2dhdr->dh_spa_guid != guid ||
9279 l2dhdr->dh_vdev_guid != dev->l2ad_vdev->vdev_guid ||
9280 l2dhdr->dh_version != L2ARC_PERSISTENT_VERSION ||
9281 l2dhdr->dh_log_entries != dev->l2ad_log_entries ||
9282 l2dhdr->dh_end != dev->l2ad_end ||
9283 !l2arc_range_check_overlap(dev->l2ad_start, dev->l2ad_end,
9284 l2dhdr->dh_evict)) {
9285 /*
9286 * Attempt to rebuild a device containing no actual dev hdr
9287 * or containing a header from some other pool or from another
9288 * version of persistent L2ARC.
9289 */
9290 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported);
9291 return (SET_ERROR(ENOTSUP));
9292 }
9293
9294 return (0);
9295 }
9296
9297 /*
9298 * Reads L2ARC log blocks from storage and validates their contents.
9299 *
9300 * This function implements a simple fetcher to make sure that while
9301 * we're processing one buffer the L2ARC is already fetching the next
9302 * one in the chain.
9303 *
9304 * The arguments this_lp and next_lp point to the current and next log block
9305 * address in the block chain. Similarly, this_lb and next_lb hold the
9306 * l2arc_log_blk_phys_t's of the current and next L2ARC blk.
9307 *
9308 * The `this_io' and `next_io' arguments are used for block fetching.
9309 * When issuing the first blk IO during rebuild, you should pass NULL for
9310 * `this_io'. This function will then issue a sync IO to read the block and
9311 * also issue an async IO to fetch the next block in the block chain. The
9312 * fetched IO is returned in `next_io'. On subsequent calls to this
9313 * function, pass the value returned in `next_io' from the previous call
9314 * as `this_io' and a fresh `next_io' pointer to hold the next fetch IO.
9315 * Prior to the call, you should initialize your `next_io' pointer to be
9316 * NULL. If no fetch IO was issued, the pointer is left set at NULL.
9317 *
9318 * On success, this function returns 0, otherwise it returns an appropriate
9319 * error code. On error the fetching IO is aborted and cleared before
9320 * returning from this function. Therefore, if we return `success', the
9321 * caller can assume that we have taken care of cleanup of fetch IOs.
9322 */
9323 static int
9324 l2arc_log_blk_read(l2arc_dev_t *dev,
9325 const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp,
9326 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
9327 zio_t *this_io, zio_t **next_io)
9328 {
9329 int err = 0;
9330 zio_cksum_t cksum;
9331 abd_t *abd = NULL;
9332 uint64_t asize;
9333
9334 ASSERT(this_lbp != NULL && next_lbp != NULL);
9335 ASSERT(this_lb != NULL && next_lb != NULL);
9336 ASSERT(next_io != NULL && *next_io == NULL);
9337 ASSERT(l2arc_log_blkptr_valid(dev, this_lbp));
9338
9339 /*
9340 * Check to see if we have issued the IO for this log block in a
9341 * previous run. If not, this is the first call, so issue it now.
9342 */
9343 if (this_io == NULL) {
9344 this_io = l2arc_log_blk_fetch(dev->l2ad_vdev, this_lbp,
9345 this_lb);
9346 }
9347
9348 /*
9349 * Peek to see if we can start issuing the next IO immediately.
9350 */
9351 if (l2arc_log_blkptr_valid(dev, next_lbp)) {
9352 /*
9353 * Start issuing IO for the next log block early - this
9354 * should help keep the L2ARC device busy while we
9355 * decompress and restore this log block.
9356 */
9357 *next_io = l2arc_log_blk_fetch(dev->l2ad_vdev, next_lbp,
9358 next_lb);
9359 }
9360
9361 /* Wait for the IO to read this log block to complete */
9362 if ((err = zio_wait(this_io)) != 0) {
9363 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors);
9364 zfs_dbgmsg("L2ARC IO error (%d) while reading log block, "
9365 "offset: %llu, vdev guid: %llu", err, this_lbp->lbp_daddr,
9366 dev->l2ad_vdev->vdev_guid);
9367 goto cleanup;
9368 }
9369
9370 /*
9371 * Make sure the buffer checks out.
9372 * L2BLK_GET_PSIZE returns aligned size for log blocks.
9373 */
9374 asize = L2BLK_GET_PSIZE((this_lbp)->lbp_prop);
9375 fletcher_4_native(this_lb, asize, NULL, &cksum);
9376 if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) {
9377 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_lb_errors);
9378 zfs_dbgmsg("L2ARC log block cksum failed, offset: %llu, "
9379 "vdev guid: %llu, l2ad_hand: %llu, l2ad_evict: %llu",
9380 this_lbp->lbp_daddr, dev->l2ad_vdev->vdev_guid,
9381 dev->l2ad_hand, dev->l2ad_evict);
9382 err = SET_ERROR(ECKSUM);
9383 goto cleanup;
9384 }
9385
9386 /* Now we can take our time decoding this buffer */
9387 switch (L2BLK_GET_COMPRESS((this_lbp)->lbp_prop)) {
9388 case ZIO_COMPRESS_OFF:
9389 break;
9390 case ZIO_COMPRESS_LZ4:
9391 abd = abd_alloc_for_io(asize, B_TRUE);
9392 abd_copy_from_buf_off(abd, this_lb, 0, asize);
9393 if ((err = zio_decompress_data(
9394 L2BLK_GET_COMPRESS((this_lbp)->lbp_prop),
9395 abd, this_lb, asize, sizeof (*this_lb))) != 0) {
9396 err = SET_ERROR(EINVAL);
9397 goto cleanup;
9398 }
9399 break;
9400 default:
9401 err = SET_ERROR(EINVAL);
9402 goto cleanup;
9403 }
9404 if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC))
9405 byteswap_uint64_array(this_lb, sizeof (*this_lb));
9406 if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) {
9407 err = SET_ERROR(EINVAL);
9408 goto cleanup;
9409 }
9410 cleanup:
9411 /* Abort an in-flight fetch I/O in case of error */
9412 if (err != 0 && *next_io != NULL) {
9413 l2arc_log_blk_fetch_abort(*next_io);
9414 *next_io = NULL;
9415 }
9416 if (abd != NULL)
9417 abd_free(abd);
9418 return (err);
9419 }
9420
9421 /*
9422 * Restores the payload of a log block to ARC. This creates empty ARC hdr
9423 * entries which only contain an l2arc hdr, essentially restoring the
9424 * buffers to their L2ARC evicted state. This function also updates space
9425 * usage on the L2ARC vdev to make sure it tracks restored buffers.
9426 */
9427 static void
9428 l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb,
9429 uint64_t lb_asize, uint64_t lb_daddr)
9430 {
9431 uint64_t size = 0, asize = 0;
9432 uint64_t log_entries = dev->l2ad_log_entries;
9433
9434 for (int i = log_entries - 1; i >= 0; i--) {
9435 /*
9436 * Restore goes in the reverse temporal direction to preserve
9437 * correct temporal ordering of buffers in the l2ad_buflist.
9438 * l2arc_hdr_restore also does a list_insert_tail instead of
9439 * list_insert_head on the l2ad_buflist:
9440 *
9441 * LIST l2ad_buflist LIST
9442 * HEAD <------ (time) ------ TAIL
9443 * direction +-----+-----+-----+-----+-----+ direction
9444 * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild
9445 * fill +-----+-----+-----+-----+-----+
9446 * ^ ^
9447 * | |
9448 * | |
9449 * l2arc_feed_thread l2arc_rebuild
9450 * will place new bufs here restores bufs here
9451 *
9452 * During l2arc_rebuild() the device is not used by
9453 * l2arc_feed_thread() as dev->l2ad_rebuild is set to true.
9454 */
9455 size += L2BLK_GET_LSIZE((&lb->lb_entries[i])->le_prop);
9456 asize += vdev_psize_to_asize(dev->l2ad_vdev,
9457 L2BLK_GET_PSIZE((&lb->lb_entries[i])->le_prop));
9458 l2arc_hdr_restore(&lb->lb_entries[i], dev);
9459 }
9460
9461 /*
9462 * Record rebuild stats:
9463 * size Logical size of restored buffers in the L2ARC
9464 * asize Aligned size of restored buffers in the L2ARC
9465 */
9466 ARCSTAT_INCR(arcstat_l2_rebuild_size, size);
9467 ARCSTAT_INCR(arcstat_l2_rebuild_asize, asize);
9468 ARCSTAT_INCR(arcstat_l2_rebuild_bufs, log_entries);
9469 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, lb_asize);
9470 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, asize / lb_asize);
9471 ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks);
9472 }
9473
9474 /*
9475 * Restores a single ARC buf hdr from a log entry. The ARC buffer is put
9476 * into a state indicating that it has been evicted to L2ARC.
9477 */
9478 static void
9479 l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev)
9480 {
9481 arc_buf_hdr_t *hdr, *exists;
9482 kmutex_t *hash_lock;
9483 arc_buf_contents_t type = L2BLK_GET_TYPE((le)->le_prop);
9484 uint64_t asize;
9485
9486 /*
9487 * Do all the allocation before grabbing any locks, this lets us
9488 * sleep if memory is full and we don't have to deal with failed
9489 * allocations.
9490 */
9491 hdr = arc_buf_alloc_l2only(L2BLK_GET_LSIZE((le)->le_prop), type,
9492 dev, le->le_dva, le->le_daddr,
9493 L2BLK_GET_PSIZE((le)->le_prop), le->le_birth,
9494 L2BLK_GET_COMPRESS((le)->le_prop),
9495 L2BLK_GET_PROTECTED((le)->le_prop),
9496 L2BLK_GET_PREFETCH((le)->le_prop));
9497 asize = vdev_psize_to_asize(dev->l2ad_vdev,
9498 L2BLK_GET_PSIZE((le)->le_prop));
9499
9500 /*
9501 * vdev_space_update() has to be called before arc_hdr_destroy() to
9502 * avoid underflow since the latter also calls the former.
9503 */
9504 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9505
9506 ARCSTAT_INCR(arcstat_l2_lsize, HDR_GET_LSIZE(hdr));
9507 ARCSTAT_INCR(arcstat_l2_psize, HDR_GET_PSIZE(hdr));
9508
9509 mutex_enter(&dev->l2ad_mtx);
9510 list_insert_tail(&dev->l2ad_buflist, hdr);
9511 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
9512 mutex_exit(&dev->l2ad_mtx);
9513
9514 exists = buf_hash_insert(hdr, &hash_lock);
9515 if (exists) {
9516 /* Buffer was already cached, no need to restore it. */
9517 arc_hdr_destroy(hdr);
9518 /*
9519 * If the buffer is already cached, check whether it has
9520 * L2ARC metadata. If not, enter them and update the flag.
9521 * This is important is case of onlining a cache device, since
9522 * we previously evicted all L2ARC metadata from ARC.
9523 */
9524 if (!HDR_HAS_L2HDR(exists)) {
9525 arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR);
9526 exists->b_l2hdr.b_dev = dev;
9527 exists->b_l2hdr.b_daddr = le->le_daddr;
9528 mutex_enter(&dev->l2ad_mtx);
9529 list_insert_tail(&dev->l2ad_buflist, exists);
9530 (void) zfs_refcount_add_many(&dev->l2ad_alloc,
9531 arc_hdr_size(exists), exists);
9532 mutex_exit(&dev->l2ad_mtx);
9533 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9534 ARCSTAT_INCR(arcstat_l2_lsize, HDR_GET_LSIZE(exists));
9535 ARCSTAT_INCR(arcstat_l2_psize, HDR_GET_PSIZE(exists));
9536 }
9537 ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached);
9538 }
9539
9540 mutex_exit(hash_lock);
9541 }
9542
9543 /*
9544 * Starts an asynchronous read IO to read a log block. This is used in log
9545 * block reconstruction to start reading the next block before we are done
9546 * decoding and reconstructing the current block, to keep the l2arc device
9547 * nice and hot with read IO to process.
9548 * The returned zio will contain newly allocated memory buffers for the IO
9549 * data which should then be freed by the caller once the zio is no longer
9550 * needed (i.e. due to it having completed). If you wish to abort this
9551 * zio, you should do so using l2arc_log_blk_fetch_abort, which takes
9552 * care of disposing of the allocated buffers correctly.
9553 */
9554 static zio_t *
9555 l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp,
9556 l2arc_log_blk_phys_t *lb)
9557 {
9558 uint32_t asize;
9559 zio_t *pio;
9560 l2arc_read_callback_t *cb;
9561
9562 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
9563 asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
9564 ASSERT(asize <= sizeof (l2arc_log_blk_phys_t));
9565
9566 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP);
9567 cb->l2rcb_abd = abd_get_from_buf(lb, asize);
9568 pio = zio_root(vd->vdev_spa, l2arc_blk_fetch_done, cb,
9569 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
9570 ZIO_FLAG_DONT_RETRY);
9571 (void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, asize,
9572 cb->l2rcb_abd, ZIO_CHECKSUM_OFF, NULL, NULL,
9573 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
9574 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE));
9575
9576 return (pio);
9577 }
9578
9579 /*
9580 * Aborts a zio returned from l2arc_log_blk_fetch and frees the data
9581 * buffers allocated for it.
9582 */
9583 static void
9584 l2arc_log_blk_fetch_abort(zio_t *zio)
9585 {
9586 (void) zio_wait(zio);
9587 }
9588
9589 /*
9590 * Creates a zio to update the device header on an l2arc device.
9591 */
9592 static void
9593 l2arc_dev_hdr_update(l2arc_dev_t *dev)
9594 {
9595 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9596 const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize;
9597 abd_t *abd;
9598 int err;
9599
9600 VERIFY(spa_config_held(dev->l2ad_spa, SCL_STATE_ALL, RW_READER));
9601
9602 l2dhdr->dh_magic = L2ARC_DEV_HDR_MAGIC;
9603 l2dhdr->dh_version = L2ARC_PERSISTENT_VERSION;
9604 l2dhdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa);
9605 l2dhdr->dh_vdev_guid = dev->l2ad_vdev->vdev_guid;
9606 l2dhdr->dh_log_entries = dev->l2ad_log_entries;
9607 l2dhdr->dh_evict = dev->l2ad_evict;
9608 l2dhdr->dh_start = dev->l2ad_start;
9609 l2dhdr->dh_end = dev->l2ad_end;
9610 l2dhdr->dh_lb_asize = zfs_refcount_count(&dev->l2ad_lb_asize);
9611 l2dhdr->dh_lb_count = zfs_refcount_count(&dev->l2ad_lb_count);
9612 l2dhdr->dh_flags = 0;
9613 if (dev->l2ad_first)
9614 l2dhdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST;
9615
9616 abd = abd_get_from_buf(l2dhdr, l2dhdr_asize);
9617
9618 err = zio_wait(zio_write_phys(NULL, dev->l2ad_vdev,
9619 VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL,
9620 NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE));
9621
9622 abd_put(abd);
9623
9624 if (err != 0) {
9625 zfs_dbgmsg("L2ARC IO error (%d) while writing device header, "
9626 "vdev guid: %llu", err, dev->l2ad_vdev->vdev_guid);
9627 }
9628 }
9629
9630 /*
9631 * Commits a log block to the L2ARC device. This routine is invoked from
9632 * l2arc_write_buffers when the log block fills up.
9633 * This function allocates some memory to temporarily hold the serialized
9634 * buffer to be written. This is then released in l2arc_write_done.
9635 */
9636 static void
9637 l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb)
9638 {
9639 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
9640 l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr;
9641 uint64_t psize, asize;
9642 zio_t *wzio;
9643 l2arc_lb_abd_buf_t *abd_buf;
9644 uint8_t *tmpbuf;
9645 l2arc_lb_ptr_buf_t *lb_ptr_buf;
9646
9647 VERIFY3S(dev->l2ad_log_ent_idx, ==, dev->l2ad_log_entries);
9648
9649 tmpbuf = zio_buf_alloc(sizeof (*lb));
9650 abd_buf = zio_buf_alloc(sizeof (*abd_buf));
9651 abd_buf->abd = abd_get_from_buf(lb, sizeof (*lb));
9652 lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP);
9653 lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP);
9654
9655 /* link the buffer into the block chain */
9656 lb->lb_prev_lbp = l2dhdr->dh_start_lbps[1];
9657 lb->lb_magic = L2ARC_LOG_BLK_MAGIC;
9658
9659 /*
9660 * l2arc_log_blk_commit() may be called multiple times during a single
9661 * l2arc_write_buffers() call. Save the allocated abd buffers in a list
9662 * so we can free them in l2arc_write_done() later on.
9663 */
9664 list_insert_tail(&cb->l2wcb_abd_list, abd_buf);
9665
9666 /* try to compress the buffer */
9667 psize = zio_compress_data(ZIO_COMPRESS_LZ4,
9668 abd_buf->abd, tmpbuf, sizeof (*lb));
9669
9670 /* a log block is never entirely zero */
9671 ASSERT(psize != 0);
9672 asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
9673 ASSERT(asize <= sizeof (*lb));
9674
9675 /*
9676 * Update the start log block pointer in the device header to point
9677 * to the log block we're about to write.
9678 */
9679 l2dhdr->dh_start_lbps[1] = l2dhdr->dh_start_lbps[0];
9680 l2dhdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand;
9681 l2dhdr->dh_start_lbps[0].lbp_payload_asize =
9682 dev->l2ad_log_blk_payload_asize;
9683 l2dhdr->dh_start_lbps[0].lbp_payload_start =
9684 dev->l2ad_log_blk_payload_start;
9685 _NOTE(CONSTCOND)
9686 L2BLK_SET_LSIZE(
9687 (&l2dhdr->dh_start_lbps[0])->lbp_prop, sizeof (*lb));
9688 L2BLK_SET_PSIZE(
9689 (&l2dhdr->dh_start_lbps[0])->lbp_prop, asize);
9690 L2BLK_SET_CHECKSUM(
9691 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
9692 ZIO_CHECKSUM_FLETCHER_4);
9693 if (asize < sizeof (*lb)) {
9694 /* compression succeeded */
9695 bzero(tmpbuf + psize, asize - psize);
9696 L2BLK_SET_COMPRESS(
9697 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
9698 ZIO_COMPRESS_LZ4);
9699 } else {
9700 /* compression failed */
9701 bcopy(lb, tmpbuf, sizeof (*lb));
9702 L2BLK_SET_COMPRESS(
9703 (&l2dhdr->dh_start_lbps[0])->lbp_prop,
9704 ZIO_COMPRESS_OFF);
9705 }
9706
9707 /* checksum what we're about to write */
9708 fletcher_4_native(tmpbuf, asize, NULL,
9709 &l2dhdr->dh_start_lbps[0].lbp_cksum);
9710
9711 abd_put(abd_buf->abd);
9712
9713 /* perform the write itself */
9714 abd_buf->abd = abd_get_from_buf(tmpbuf, sizeof (*lb));
9715 abd_take_ownership_of_buf(abd_buf->abd, B_TRUE);
9716 wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand,
9717 asize, abd_buf->abd, ZIO_CHECKSUM_OFF, NULL, NULL,
9718 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE);
9719 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio);
9720 (void) zio_nowait(wzio);
9721
9722 dev->l2ad_hand += asize;
9723 /*
9724 * Include the committed log block's pointer in the list of pointers
9725 * to log blocks present in the L2ARC device.
9726 */
9727 bcopy(&l2dhdr->dh_start_lbps[0], lb_ptr_buf->lb_ptr,
9728 sizeof (l2arc_log_blkptr_t));
9729 mutex_enter(&dev->l2ad_mtx);
9730 list_insert_head(&dev->l2ad_lbptr_list, lb_ptr_buf);
9731 ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize);
9732 ARCSTAT_BUMP(arcstat_l2_log_blk_count);
9733 zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf);
9734 zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf);
9735 mutex_exit(&dev->l2ad_mtx);
9736 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9737
9738 /* bump the kstats */
9739 ARCSTAT_INCR(arcstat_l2_write_bytes, asize);
9740 ARCSTAT_BUMP(arcstat_l2_log_blk_writes);
9741 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, asize);
9742 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio,
9743 dev->l2ad_log_blk_payload_asize / asize);
9744
9745 /* start a new log block */
9746 dev->l2ad_log_ent_idx = 0;
9747 dev->l2ad_log_blk_payload_asize = 0;
9748 dev->l2ad_log_blk_payload_start = 0;
9749 }
9750
9751 /*
9752 * Validates an L2ARC log block address to make sure that it can be read
9753 * from the provided L2ARC device.
9754 */
9755 boolean_t
9756 l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp)
9757 {
9758 /* L2BLK_GET_PSIZE returns aligned size for log blocks */
9759 uint64_t asize = L2BLK_GET_PSIZE((lbp)->lbp_prop);
9760 uint64_t end = lbp->lbp_daddr + asize - 1;
9761 uint64_t start = lbp->lbp_payload_start;
9762 boolean_t evicted = B_FALSE;
9763
9764 /* BEGIN CSTYLED */
9765 /*
9766 * A log block is valid if all of the following conditions are true:
9767 * - it fits entirely (including its payload) between l2ad_start and
9768 * l2ad_end
9769 * - it has a valid size
9770 * - neither the log block itself nor part of its payload was evicted
9771 * by l2arc_evict():
9772 *
9773 * l2ad_hand l2ad_evict
9774 * | | lbp_daddr
9775 * | start | | end
9776 * | | | | |
9777 * V V V V V
9778 * l2ad_start ============================================ l2ad_end
9779 * --------------------------||||
9780 * ^ ^
9781 * | log block
9782 * payload
9783 */
9784 /* END CSTYLED */
9785 evicted =
9786 l2arc_range_check_overlap(start, end, dev->l2ad_hand) ||
9787 l2arc_range_check_overlap(start, end, dev->l2ad_evict) ||
9788 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, start) ||
9789 l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, end);
9790
9791 return (start >= dev->l2ad_start && end <= dev->l2ad_end &&
9792 asize > 0 && asize <= sizeof (l2arc_log_blk_phys_t) &&
9793 (!evicted || dev->l2ad_first));
9794 }
9795
9796 /*
9797 * Inserts ARC buffer header `hdr' into the current L2ARC log block on
9798 * the device. The buffer being inserted must be present in L2ARC.
9799 * Returns B_TRUE if the L2ARC log block is full and needs to be committed
9800 * to L2ARC, or B_FALSE if it still has room for more ARC buffers.
9801 */
9802 static boolean_t
9803 l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr)
9804 {
9805 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
9806 l2arc_log_ent_phys_t *le;
9807
9808 if (dev->l2ad_log_entries == 0)
9809 return (B_FALSE);
9810
9811 int index = dev->l2ad_log_ent_idx++;
9812
9813 ASSERT3S(index, <, dev->l2ad_log_entries);
9814 ASSERT(HDR_HAS_L2HDR(hdr));
9815
9816 le = &lb->lb_entries[index];
9817 bzero(le, sizeof (*le));
9818 le->le_dva = hdr->b_dva;
9819 le->le_birth = hdr->b_birth;
9820 le->le_daddr = hdr->b_l2hdr.b_daddr;
9821 if (index == 0)
9822 dev->l2ad_log_blk_payload_start = le->le_daddr;
9823 L2BLK_SET_LSIZE((le)->le_prop, HDR_GET_LSIZE(hdr));
9824 L2BLK_SET_PSIZE((le)->le_prop, HDR_GET_PSIZE(hdr));
9825 L2BLK_SET_COMPRESS((le)->le_prop, HDR_GET_COMPRESS(hdr));
9826 L2BLK_SET_TYPE((le)->le_prop, hdr->b_type);
9827 L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr)));
9828 L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr)));
9829
9830 dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev,
9831 HDR_GET_PSIZE(hdr));
9832
9833 return (dev->l2ad_log_ent_idx == dev->l2ad_log_entries);
9834 }
9835
9836 /*
9837 * Checks whether a given L2ARC device address sits in a time-sequential
9838 * range. The trick here is that the L2ARC is a rotary buffer, so we can't
9839 * just do a range comparison, we need to handle the situation in which the
9840 * range wraps around the end of the L2ARC device. Arguments:
9841 * bottom -- Lower end of the range to check (written to earlier).
9842 * top -- Upper end of the range to check (written to later).
9843 * check -- The address for which we want to determine if it sits in
9844 * between the top and bottom.
9845 *
9846 * The 3-way conditional below represents the following cases:
9847 *
9848 * bottom < top : Sequentially ordered case:
9849 * <check>--------+-------------------+
9850 * | (overlap here?) |
9851 * L2ARC dev V V
9852 * |---------------<bottom>============<top>--------------|
9853 *
9854 * bottom > top: Looped-around case:
9855 * <check>--------+------------------+
9856 * | (overlap here?) |
9857 * L2ARC dev V V
9858 * |===============<top>---------------<bottom>===========|
9859 * ^ ^
9860 * | (or here?) |
9861 * +---------------+---------<check>
9862 *
9863 * top == bottom : Just a single address comparison.
9864 */
9865 boolean_t
9866 l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check)
9867 {
9868 if (bottom < top)
9869 return (bottom <= check && check <= top);
9870 else if (bottom > top)
9871 return (check <= top || bottom <= check);
9872 else
9873 return (check == top);
9874 }