]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
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 | /* | |
428870ff | 22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. |
36da08ef | 23 | * Copyright (c) 2012, Joyent, Inc. All rights reserved. |
c30e58c4 | 24 | * Copyright (c) 2011, 2017 by Delphix. All rights reserved. |
36da08ef | 25 | * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. |
d3c2ae1c | 26 | * Copyright 2015 Nexenta Systems, Inc. All rights reserved. |
34dc7c2f BB |
27 | */ |
28 | ||
34dc7c2f BB |
29 | /* |
30 | * DVA-based Adjustable Replacement Cache | |
31 | * | |
32 | * While much of the theory of operation used here is | |
33 | * based on the self-tuning, low overhead replacement cache | |
34 | * presented by Megiddo and Modha at FAST 2003, there are some | |
35 | * significant differences: | |
36 | * | |
37 | * 1. The Megiddo and Modha model assumes any page is evictable. | |
38 | * Pages in its cache cannot be "locked" into memory. This makes | |
39 | * the eviction algorithm simple: evict the last page in the list. | |
40 | * This also make the performance characteristics easy to reason | |
41 | * about. Our cache is not so simple. At any given moment, some | |
42 | * subset of the blocks in the cache are un-evictable because we | |
43 | * have handed out a reference to them. Blocks are only evictable | |
44 | * when there are no external references active. This makes | |
45 | * eviction far more problematic: we choose to evict the evictable | |
46 | * blocks that are the "lowest" in the list. | |
47 | * | |
48 | * There are times when it is not possible to evict the requested | |
49 | * space. In these circumstances we are unable to adjust the cache | |
50 | * size. To prevent the cache growing unbounded at these times we | |
51 | * implement a "cache throttle" that slows the flow of new data | |
52 | * into the cache until we can make space available. | |
53 | * | |
54 | * 2. The Megiddo and Modha model assumes a fixed cache size. | |
55 | * Pages are evicted when the cache is full and there is a cache | |
56 | * miss. Our model has a variable sized cache. It grows with | |
57 | * high use, but also tries to react to memory pressure from the | |
58 | * operating system: decreasing its size when system memory is | |
59 | * tight. | |
60 | * | |
61 | * 3. The Megiddo and Modha model assumes a fixed page size. All | |
d3cc8b15 | 62 | * elements of the cache are therefore exactly the same size. So |
34dc7c2f BB |
63 | * when adjusting the cache size following a cache miss, its simply |
64 | * a matter of choosing a single page to evict. In our model, we | |
65 | * have variable sized cache blocks (rangeing from 512 bytes to | |
d3cc8b15 | 66 | * 128K bytes). We therefore choose a set of blocks to evict to make |
34dc7c2f BB |
67 | * space for a cache miss that approximates as closely as possible |
68 | * the space used by the new block. | |
69 | * | |
70 | * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" | |
71 | * by N. Megiddo & D. Modha, FAST 2003 | |
72 | */ | |
73 | ||
74 | /* | |
75 | * The locking model: | |
76 | * | |
77 | * A new reference to a cache buffer can be obtained in two | |
78 | * ways: 1) via a hash table lookup using the DVA as a key, | |
79 | * or 2) via one of the ARC lists. The arc_read() interface | |
2aa34383 | 80 | * uses method 1, while the internal ARC algorithms for |
d3cc8b15 | 81 | * adjusting the cache use method 2. We therefore provide two |
34dc7c2f | 82 | * types of locks: 1) the hash table lock array, and 2) the |
2aa34383 | 83 | * ARC list locks. |
34dc7c2f | 84 | * |
5c839890 BC |
85 | * Buffers do not have their own mutexes, rather they rely on the |
86 | * hash table mutexes for the bulk of their protection (i.e. most | |
87 | * fields in the arc_buf_hdr_t are protected by these mutexes). | |
34dc7c2f BB |
88 | * |
89 | * buf_hash_find() returns the appropriate mutex (held) when it | |
90 | * locates the requested buffer in the hash table. It returns | |
91 | * NULL for the mutex if the buffer was not in the table. | |
92 | * | |
93 | * buf_hash_remove() expects the appropriate hash mutex to be | |
94 | * already held before it is invoked. | |
95 | * | |
2aa34383 | 96 | * Each ARC state also has a mutex which is used to protect the |
34dc7c2f | 97 | * buffer list associated with the state. When attempting to |
2aa34383 | 98 | * obtain a hash table lock while holding an ARC list lock you |
34dc7c2f BB |
99 | * must use: mutex_tryenter() to avoid deadlock. Also note that |
100 | * the active state mutex must be held before the ghost state mutex. | |
101 | * | |
ab26409d BB |
102 | * It as also possible to register a callback which is run when the |
103 | * arc_meta_limit is reached and no buffers can be safely evicted. In | |
104 | * this case the arc user should drop a reference on some arc buffers so | |
105 | * they can be reclaimed and the arc_meta_limit honored. For example, | |
106 | * when using the ZPL each dentry holds a references on a znode. These | |
107 | * dentries must be pruned before the arc buffer holding the znode can | |
108 | * be safely evicted. | |
109 | * | |
34dc7c2f BB |
110 | * Note that the majority of the performance stats are manipulated |
111 | * with atomic operations. | |
112 | * | |
b9541d6b | 113 | * The L2ARC uses the l2ad_mtx on each vdev for the following: |
34dc7c2f BB |
114 | * |
115 | * - L2ARC buflist creation | |
116 | * - L2ARC buflist eviction | |
117 | * - L2ARC write completion, which walks L2ARC buflists | |
118 | * - ARC header destruction, as it removes from L2ARC buflists | |
119 | * - ARC header release, as it removes from L2ARC buflists | |
120 | */ | |
121 | ||
d3c2ae1c GW |
122 | /* |
123 | * ARC operation: | |
124 | * | |
125 | * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure. | |
126 | * This structure can point either to a block that is still in the cache or to | |
127 | * one that is only accessible in an L2 ARC device, or it can provide | |
128 | * information about a block that was recently evicted. If a block is | |
129 | * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough | |
130 | * information to retrieve it from the L2ARC device. This information is | |
131 | * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block | |
132 | * that is in this state cannot access the data directly. | |
133 | * | |
134 | * Blocks that are actively being referenced or have not been evicted | |
135 | * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within | |
136 | * the arc_buf_hdr_t that will point to the data block in memory. A block can | |
137 | * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC | |
2aa34383 | 138 | * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and |
a6255b7f | 139 | * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd). |
2aa34383 DK |
140 | * |
141 | * The L1ARC's data pointer may or may not be uncompressed. The ARC has the | |
a6255b7f DQ |
142 | * ability to store the physical data (b_pabd) associated with the DVA of the |
143 | * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block, | |
2aa34383 DK |
144 | * it will match its on-disk compression characteristics. This behavior can be |
145 | * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the | |
a6255b7f | 146 | * compressed ARC functionality is disabled, the b_pabd will point to an |
2aa34383 DK |
147 | * uncompressed version of the on-disk data. |
148 | * | |
149 | * Data in the L1ARC is not accessed by consumers of the ARC directly. Each | |
150 | * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it. | |
151 | * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC | |
152 | * consumer. The ARC will provide references to this data and will keep it | |
153 | * cached until it is no longer in use. The ARC caches only the L1ARC's physical | |
154 | * data block and will evict any arc_buf_t that is no longer referenced. The | |
155 | * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the | |
d3c2ae1c GW |
156 | * "overhead_size" kstat. |
157 | * | |
2aa34383 DK |
158 | * Depending on the consumer, an arc_buf_t can be requested in uncompressed or |
159 | * compressed form. The typical case is that consumers will want uncompressed | |
160 | * data, and when that happens a new data buffer is allocated where the data is | |
161 | * decompressed for them to use. Currently the only consumer who wants | |
162 | * compressed arc_buf_t's is "zfs send", when it streams data exactly as it | |
163 | * exists on disk. When this happens, the arc_buf_t's data buffer is shared | |
164 | * with the arc_buf_hdr_t. | |
d3c2ae1c | 165 | * |
2aa34383 DK |
166 | * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The |
167 | * first one is owned by a compressed send consumer (and therefore references | |
168 | * the same compressed data buffer as the arc_buf_hdr_t) and the second could be | |
169 | * used by any other consumer (and has its own uncompressed copy of the data | |
170 | * buffer). | |
d3c2ae1c | 171 | * |
2aa34383 DK |
172 | * arc_buf_hdr_t |
173 | * +-----------+ | |
174 | * | fields | | |
175 | * | common to | | |
176 | * | L1- and | | |
177 | * | L2ARC | | |
178 | * +-----------+ | |
179 | * | l2arc_buf_hdr_t | |
180 | * | | | |
181 | * +-----------+ | |
182 | * | l1arc_buf_hdr_t | |
183 | * | | arc_buf_t | |
184 | * | b_buf +------------>+-----------+ arc_buf_t | |
a6255b7f | 185 | * | b_pabd +-+ |b_next +---->+-----------+ |
2aa34383 DK |
186 | * +-----------+ | |-----------| |b_next +-->NULL |
187 | * | |b_comp = T | +-----------+ | |
188 | * | |b_data +-+ |b_comp = F | | |
189 | * | +-----------+ | |b_data +-+ | |
190 | * +->+------+ | +-----------+ | | |
191 | * compressed | | | | | |
192 | * data | |<--------------+ | uncompressed | |
193 | * +------+ compressed, | data | |
194 | * shared +-->+------+ | |
195 | * data | | | |
196 | * | | | |
197 | * +------+ | |
d3c2ae1c GW |
198 | * |
199 | * When a consumer reads a block, the ARC must first look to see if the | |
2aa34383 DK |
200 | * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new |
201 | * arc_buf_t and either copies uncompressed data into a new data buffer from an | |
a6255b7f DQ |
202 | * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a |
203 | * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the | |
2aa34383 DK |
204 | * hdr is compressed and the desired compression characteristics of the |
205 | * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the | |
206 | * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be | |
207 | * the last buffer in the hdr's b_buf list, however a shared compressed buf can | |
208 | * be anywhere in the hdr's list. | |
d3c2ae1c GW |
209 | * |
210 | * The diagram below shows an example of an uncompressed ARC hdr that is | |
2aa34383 DK |
211 | * sharing its data with an arc_buf_t (note that the shared uncompressed buf is |
212 | * the last element in the buf list): | |
d3c2ae1c GW |
213 | * |
214 | * arc_buf_hdr_t | |
215 | * +-----------+ | |
216 | * | | | |
217 | * | | | |
218 | * | | | |
219 | * +-----------+ | |
220 | * l2arc_buf_hdr_t| | | |
221 | * | | | |
222 | * +-----------+ | |
223 | * l1arc_buf_hdr_t| | | |
224 | * | | arc_buf_t (shared) | |
225 | * | b_buf +------------>+---------+ arc_buf_t | |
226 | * | | |b_next +---->+---------+ | |
a6255b7f | 227 | * | b_pabd +-+ |---------| |b_next +-->NULL |
d3c2ae1c GW |
228 | * +-----------+ | | | +---------+ |
229 | * | |b_data +-+ | | | |
230 | * | +---------+ | |b_data +-+ | |
231 | * +->+------+ | +---------+ | | |
232 | * | | | | | |
233 | * uncompressed | | | | | |
234 | * data +------+ | | | |
235 | * ^ +->+------+ | | |
236 | * | uncompressed | | | | |
237 | * | data | | | | |
238 | * | +------+ | | |
239 | * +---------------------------------+ | |
240 | * | |
a6255b7f | 241 | * Writing to the ARC requires that the ARC first discard the hdr's b_pabd |
d3c2ae1c | 242 | * since the physical block is about to be rewritten. The new data contents |
2aa34383 DK |
243 | * will be contained in the arc_buf_t. As the I/O pipeline performs the write, |
244 | * it may compress the data before writing it to disk. The ARC will be called | |
245 | * with the transformed data and will bcopy the transformed on-disk block into | |
a6255b7f | 246 | * a newly allocated b_pabd. Writes are always done into buffers which have |
2aa34383 DK |
247 | * either been loaned (and hence are new and don't have other readers) or |
248 | * buffers which have been released (and hence have their own hdr, if there | |
249 | * were originally other readers of the buf's original hdr). This ensures that | |
250 | * the ARC only needs to update a single buf and its hdr after a write occurs. | |
d3c2ae1c | 251 | * |
a6255b7f DQ |
252 | * When the L2ARC is in use, it will also take advantage of the b_pabd. The |
253 | * L2ARC will always write the contents of b_pabd to the L2ARC. This means | |
2aa34383 | 254 | * that when compressed ARC is enabled that the L2ARC blocks are identical |
d3c2ae1c GW |
255 | * to the on-disk block in the main data pool. This provides a significant |
256 | * advantage since the ARC can leverage the bp's checksum when reading from the | |
257 | * L2ARC to determine if the contents are valid. However, if the compressed | |
2aa34383 | 258 | * ARC is disabled, then the L2ARC's block must be transformed to look |
d3c2ae1c GW |
259 | * like the physical block in the main data pool before comparing the |
260 | * checksum and determining its validity. | |
b5256303 TC |
261 | * |
262 | * The L1ARC has a slightly different system for storing encrypted data. | |
263 | * Raw (encrypted + possibly compressed) data has a few subtle differences from | |
264 | * data that is just compressed. The biggest difference is that it is not | |
265 | * possible to decrypt encrypted data (or visa versa) if the keys aren't loaded. | |
266 | * The other difference is that encryption cannot be treated as a suggestion. | |
267 | * If a caller would prefer compressed data, but they actually wind up with | |
268 | * uncompressed data the worst thing that could happen is there might be a | |
269 | * performance hit. If the caller requests encrypted data, however, we must be | |
270 | * sure they actually get it or else secret information could be leaked. Raw | |
271 | * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore, | |
272 | * may have both an encrypted version and a decrypted version of its data at | |
273 | * once. When a caller needs a raw arc_buf_t, it is allocated and the data is | |
274 | * copied out of this header. To avoid complications with b_pabd, raw buffers | |
275 | * cannot be shared. | |
d3c2ae1c GW |
276 | */ |
277 | ||
34dc7c2f BB |
278 | #include <sys/spa.h> |
279 | #include <sys/zio.h> | |
d3c2ae1c | 280 | #include <sys/spa_impl.h> |
3a17a7a9 | 281 | #include <sys/zio_compress.h> |
d3c2ae1c | 282 | #include <sys/zio_checksum.h> |
34dc7c2f BB |
283 | #include <sys/zfs_context.h> |
284 | #include <sys/arc.h> | |
36da08ef | 285 | #include <sys/refcount.h> |
b128c09f | 286 | #include <sys/vdev.h> |
9babb374 | 287 | #include <sys/vdev_impl.h> |
e8b96c60 | 288 | #include <sys/dsl_pool.h> |
a6255b7f | 289 | #include <sys/zio_checksum.h> |
ca0bf58d | 290 | #include <sys/multilist.h> |
a6255b7f | 291 | #include <sys/abd.h> |
b5256303 TC |
292 | #include <sys/zil.h> |
293 | #include <sys/fm/fs/zfs.h> | |
34dc7c2f | 294 | #ifdef _KERNEL |
93ce2b4c | 295 | #include <sys/shrinker.h> |
34dc7c2f | 296 | #include <sys/vmsystm.h> |
ab26409d | 297 | #include <sys/zpl.h> |
e9a77290 | 298 | #include <linux/page_compat.h> |
34dc7c2f BB |
299 | #endif |
300 | #include <sys/callb.h> | |
301 | #include <sys/kstat.h> | |
570827e1 | 302 | #include <sys/dmu_tx.h> |
428870ff | 303 | #include <zfs_fletcher.h> |
59ec819a | 304 | #include <sys/arc_impl.h> |
49ee64e5 | 305 | #include <sys/trace_arc.h> |
34dc7c2f | 306 | |
498877ba MA |
307 | #ifndef _KERNEL |
308 | /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ | |
309 | boolean_t arc_watch = B_FALSE; | |
310 | #endif | |
311 | ||
ca0bf58d PS |
312 | static kmutex_t arc_reclaim_lock; |
313 | static kcondvar_t arc_reclaim_thread_cv; | |
314 | static boolean_t arc_reclaim_thread_exit; | |
315 | static kcondvar_t arc_reclaim_waiters_cv; | |
316 | ||
e8b96c60 | 317 | /* |
ca0bf58d PS |
318 | * The number of headers to evict in arc_evict_state_impl() before |
319 | * dropping the sublist lock and evicting from another sublist. A lower | |
320 | * value means we're more likely to evict the "correct" header (i.e. the | |
321 | * oldest header in the arc state), but comes with higher overhead | |
322 | * (i.e. more invocations of arc_evict_state_impl()). | |
323 | */ | |
324 | int zfs_arc_evict_batch_limit = 10; | |
325 | ||
34dc7c2f | 326 | /* number of seconds before growing cache again */ |
ca67b33a | 327 | static int arc_grow_retry = 5; |
34dc7c2f | 328 | |
a6255b7f | 329 | /* shift of arc_c for calculating overflow limit in arc_get_data_impl */ |
ca67b33a | 330 | int zfs_arc_overflow_shift = 8; |
62422785 | 331 | |
728d6ae9 BB |
332 | /* shift of arc_c for calculating both min and max arc_p */ |
333 | static int arc_p_min_shift = 4; | |
334 | ||
d164b209 | 335 | /* log2(fraction of arc to reclaim) */ |
ca67b33a | 336 | static int arc_shrink_shift = 7; |
d164b209 | 337 | |
03b60eee DB |
338 | /* percent of pagecache to reclaim arc to */ |
339 | #ifdef _KERNEL | |
340 | static uint_t zfs_arc_pc_percent = 0; | |
341 | #endif | |
342 | ||
34dc7c2f | 343 | /* |
ca67b33a MA |
344 | * log2(fraction of ARC which must be free to allow growing). |
345 | * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, | |
346 | * when reading a new block into the ARC, we will evict an equal-sized block | |
347 | * from the ARC. | |
348 | * | |
349 | * This must be less than arc_shrink_shift, so that when we shrink the ARC, | |
350 | * we will still not allow it to grow. | |
34dc7c2f | 351 | */ |
ca67b33a | 352 | int arc_no_grow_shift = 5; |
bce45ec9 | 353 | |
49ddb315 | 354 | |
ca0bf58d PS |
355 | /* |
356 | * minimum lifespan of a prefetch block in clock ticks | |
357 | * (initialized in arc_init()) | |
358 | */ | |
d4a72f23 TC |
359 | static int arc_min_prefetch_ms; |
360 | static int arc_min_prescient_prefetch_ms; | |
ca0bf58d | 361 | |
e8b96c60 MA |
362 | /* |
363 | * If this percent of memory is free, don't throttle. | |
364 | */ | |
365 | int arc_lotsfree_percent = 10; | |
366 | ||
34dc7c2f BB |
367 | static int arc_dead; |
368 | ||
b128c09f BB |
369 | /* |
370 | * The arc has filled available memory and has now warmed up. | |
371 | */ | |
372 | static boolean_t arc_warm; | |
373 | ||
d3c2ae1c GW |
374 | /* |
375 | * log2 fraction of the zio arena to keep free. | |
376 | */ | |
377 | int arc_zio_arena_free_shift = 2; | |
378 | ||
34dc7c2f BB |
379 | /* |
380 | * These tunables are for performance analysis. | |
381 | */ | |
c28b2279 BB |
382 | unsigned long zfs_arc_max = 0; |
383 | unsigned long zfs_arc_min = 0; | |
384 | unsigned long zfs_arc_meta_limit = 0; | |
ca0bf58d | 385 | unsigned long zfs_arc_meta_min = 0; |
25458cbe TC |
386 | unsigned long zfs_arc_dnode_limit = 0; |
387 | unsigned long zfs_arc_dnode_reduce_percent = 10; | |
ca67b33a MA |
388 | int zfs_arc_grow_retry = 0; |
389 | int zfs_arc_shrink_shift = 0; | |
728d6ae9 | 390 | int zfs_arc_p_min_shift = 0; |
ca67b33a | 391 | int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ |
34dc7c2f | 392 | |
d3c2ae1c GW |
393 | int zfs_compressed_arc_enabled = B_TRUE; |
394 | ||
9907cc1c G |
395 | /* |
396 | * ARC will evict meta buffers that exceed arc_meta_limit. This | |
397 | * tunable make arc_meta_limit adjustable for different workloads. | |
398 | */ | |
399 | unsigned long zfs_arc_meta_limit_percent = 75; | |
400 | ||
401 | /* | |
402 | * Percentage that can be consumed by dnodes of ARC meta buffers. | |
403 | */ | |
404 | unsigned long zfs_arc_dnode_limit_percent = 10; | |
405 | ||
bc888666 | 406 | /* |
ca67b33a | 407 | * These tunables are Linux specific |
bc888666 | 408 | */ |
11f552fa | 409 | unsigned long zfs_arc_sys_free = 0; |
d4a72f23 TC |
410 | int zfs_arc_min_prefetch_ms = 0; |
411 | int zfs_arc_min_prescient_prefetch_ms = 0; | |
ca67b33a MA |
412 | int zfs_arc_p_dampener_disable = 1; |
413 | int zfs_arc_meta_prune = 10000; | |
414 | int zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED; | |
415 | int zfs_arc_meta_adjust_restarts = 4096; | |
7e8bddd0 | 416 | int zfs_arc_lotsfree_percent = 10; |
bc888666 | 417 | |
34dc7c2f BB |
418 | /* The 6 states: */ |
419 | static arc_state_t ARC_anon; | |
420 | static arc_state_t ARC_mru; | |
421 | static arc_state_t ARC_mru_ghost; | |
422 | static arc_state_t ARC_mfu; | |
423 | static arc_state_t ARC_mfu_ghost; | |
424 | static arc_state_t ARC_l2c_only; | |
425 | ||
426 | typedef struct arc_stats { | |
427 | kstat_named_t arcstat_hits; | |
428 | kstat_named_t arcstat_misses; | |
429 | kstat_named_t arcstat_demand_data_hits; | |
430 | kstat_named_t arcstat_demand_data_misses; | |
431 | kstat_named_t arcstat_demand_metadata_hits; | |
432 | kstat_named_t arcstat_demand_metadata_misses; | |
433 | kstat_named_t arcstat_prefetch_data_hits; | |
434 | kstat_named_t arcstat_prefetch_data_misses; | |
435 | kstat_named_t arcstat_prefetch_metadata_hits; | |
436 | kstat_named_t arcstat_prefetch_metadata_misses; | |
437 | kstat_named_t arcstat_mru_hits; | |
438 | kstat_named_t arcstat_mru_ghost_hits; | |
439 | kstat_named_t arcstat_mfu_hits; | |
440 | kstat_named_t arcstat_mfu_ghost_hits; | |
441 | kstat_named_t arcstat_deleted; | |
e49f1e20 WA |
442 | /* |
443 | * Number of buffers that could not be evicted because the hash lock | |
444 | * was held by another thread. The lock may not necessarily be held | |
445 | * by something using the same buffer, since hash locks are shared | |
446 | * by multiple buffers. | |
447 | */ | |
34dc7c2f | 448 | kstat_named_t arcstat_mutex_miss; |
0873bb63 BB |
449 | /* |
450 | * Number of buffers skipped when updating the access state due to the | |
451 | * header having already been released after acquiring the hash lock. | |
452 | */ | |
453 | kstat_named_t arcstat_access_skip; | |
e49f1e20 WA |
454 | /* |
455 | * Number of buffers skipped because they have I/O in progress, are | |
0873bb63 | 456 | * indirect prefetch buffers that have not lived long enough, or are |
e49f1e20 WA |
457 | * not from the spa we're trying to evict from. |
458 | */ | |
34dc7c2f | 459 | kstat_named_t arcstat_evict_skip; |
ca0bf58d PS |
460 | /* |
461 | * Number of times arc_evict_state() was unable to evict enough | |
462 | * buffers to reach its target amount. | |
463 | */ | |
464 | kstat_named_t arcstat_evict_not_enough; | |
428870ff BB |
465 | kstat_named_t arcstat_evict_l2_cached; |
466 | kstat_named_t arcstat_evict_l2_eligible; | |
467 | kstat_named_t arcstat_evict_l2_ineligible; | |
ca0bf58d | 468 | kstat_named_t arcstat_evict_l2_skip; |
34dc7c2f BB |
469 | kstat_named_t arcstat_hash_elements; |
470 | kstat_named_t arcstat_hash_elements_max; | |
471 | kstat_named_t arcstat_hash_collisions; | |
472 | kstat_named_t arcstat_hash_chains; | |
473 | kstat_named_t arcstat_hash_chain_max; | |
474 | kstat_named_t arcstat_p; | |
475 | kstat_named_t arcstat_c; | |
476 | kstat_named_t arcstat_c_min; | |
477 | kstat_named_t arcstat_c_max; | |
478 | kstat_named_t arcstat_size; | |
d3c2ae1c | 479 | /* |
a6255b7f | 480 | * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd. |
d3c2ae1c GW |
481 | * Note that the compressed bytes may match the uncompressed bytes |
482 | * if the block is either not compressed or compressed arc is disabled. | |
483 | */ | |
484 | kstat_named_t arcstat_compressed_size; | |
485 | /* | |
a6255b7f | 486 | * Uncompressed size of the data stored in b_pabd. If compressed |
d3c2ae1c GW |
487 | * arc is disabled then this value will be identical to the stat |
488 | * above. | |
489 | */ | |
490 | kstat_named_t arcstat_uncompressed_size; | |
491 | /* | |
492 | * Number of bytes stored in all the arc_buf_t's. This is classified | |
493 | * as "overhead" since this data is typically short-lived and will | |
494 | * be evicted from the arc when it becomes unreferenced unless the | |
495 | * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level | |
496 | * values have been set (see comment in dbuf.c for more information). | |
497 | */ | |
498 | kstat_named_t arcstat_overhead_size; | |
500445c0 PS |
499 | /* |
500 | * Number of bytes consumed by internal ARC structures necessary | |
501 | * for tracking purposes; these structures are not actually | |
502 | * backed by ARC buffers. This includes arc_buf_hdr_t structures | |
503 | * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only | |
504 | * caches), and arc_buf_t structures (allocated via arc_buf_t | |
505 | * cache). | |
506 | */ | |
34dc7c2f | 507 | kstat_named_t arcstat_hdr_size; |
500445c0 PS |
508 | /* |
509 | * Number of bytes consumed by ARC buffers of type equal to | |
510 | * ARC_BUFC_DATA. This is generally consumed by buffers backing | |
511 | * on disk user data (e.g. plain file contents). | |
512 | */ | |
d164b209 | 513 | kstat_named_t arcstat_data_size; |
500445c0 PS |
514 | /* |
515 | * Number of bytes consumed by ARC buffers of type equal to | |
516 | * ARC_BUFC_METADATA. This is generally consumed by buffers | |
517 | * backing on disk data that is used for internal ZFS | |
518 | * structures (e.g. ZAP, dnode, indirect blocks, etc). | |
519 | */ | |
520 | kstat_named_t arcstat_metadata_size; | |
521 | /* | |
25458cbe | 522 | * Number of bytes consumed by dmu_buf_impl_t objects. |
500445c0 | 523 | */ |
25458cbe TC |
524 | kstat_named_t arcstat_dbuf_size; |
525 | /* | |
526 | * Number of bytes consumed by dnode_t objects. | |
527 | */ | |
528 | kstat_named_t arcstat_dnode_size; | |
529 | /* | |
530 | * Number of bytes consumed by bonus buffers. | |
531 | */ | |
532 | kstat_named_t arcstat_bonus_size; | |
500445c0 PS |
533 | /* |
534 | * Total number of bytes consumed by ARC buffers residing in the | |
535 | * arc_anon state. This includes *all* buffers in the arc_anon | |
536 | * state; e.g. data, metadata, evictable, and unevictable buffers | |
537 | * are all included in this value. | |
538 | */ | |
13be560d | 539 | kstat_named_t arcstat_anon_size; |
500445c0 PS |
540 | /* |
541 | * Number of bytes consumed by ARC buffers that meet the | |
542 | * following criteria: backing buffers of type ARC_BUFC_DATA, | |
543 | * residing in the arc_anon state, and are eligible for eviction | |
544 | * (e.g. have no outstanding holds on the buffer). | |
545 | */ | |
546 | kstat_named_t arcstat_anon_evictable_data; | |
547 | /* | |
548 | * Number of bytes consumed by ARC buffers that meet the | |
549 | * following criteria: backing buffers of type ARC_BUFC_METADATA, | |
550 | * residing in the arc_anon state, and are eligible for eviction | |
551 | * (e.g. have no outstanding holds on the buffer). | |
552 | */ | |
553 | kstat_named_t arcstat_anon_evictable_metadata; | |
554 | /* | |
555 | * Total number of bytes consumed by ARC buffers residing in the | |
556 | * arc_mru state. This includes *all* buffers in the arc_mru | |
557 | * state; e.g. data, metadata, evictable, and unevictable buffers | |
558 | * are all included in this value. | |
559 | */ | |
13be560d | 560 | kstat_named_t arcstat_mru_size; |
500445c0 PS |
561 | /* |
562 | * Number of bytes consumed by ARC buffers that meet the | |
563 | * following criteria: backing buffers of type ARC_BUFC_DATA, | |
564 | * residing in the arc_mru state, and are eligible for eviction | |
565 | * (e.g. have no outstanding holds on the buffer). | |
566 | */ | |
567 | kstat_named_t arcstat_mru_evictable_data; | |
568 | /* | |
569 | * Number of bytes consumed by ARC buffers that meet the | |
570 | * following criteria: backing buffers of type ARC_BUFC_METADATA, | |
571 | * residing in the arc_mru state, and are eligible for eviction | |
572 | * (e.g. have no outstanding holds on the buffer). | |
573 | */ | |
574 | kstat_named_t arcstat_mru_evictable_metadata; | |
575 | /* | |
576 | * Total number of bytes that *would have been* consumed by ARC | |
577 | * buffers in the arc_mru_ghost state. The key thing to note | |
578 | * here, is the fact that this size doesn't actually indicate | |
579 | * RAM consumption. The ghost lists only consist of headers and | |
580 | * don't actually have ARC buffers linked off of these headers. | |
581 | * Thus, *if* the headers had associated ARC buffers, these | |
582 | * buffers *would have* consumed this number of bytes. | |
583 | */ | |
13be560d | 584 | kstat_named_t arcstat_mru_ghost_size; |
500445c0 PS |
585 | /* |
586 | * Number of bytes that *would have been* consumed by ARC | |
587 | * buffers that are eligible for eviction, of type | |
588 | * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. | |
589 | */ | |
590 | kstat_named_t arcstat_mru_ghost_evictable_data; | |
591 | /* | |
592 | * Number of bytes that *would have been* consumed by ARC | |
593 | * buffers that are eligible for eviction, of type | |
594 | * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. | |
595 | */ | |
596 | kstat_named_t arcstat_mru_ghost_evictable_metadata; | |
597 | /* | |
598 | * Total number of bytes consumed by ARC buffers residing in the | |
599 | * arc_mfu state. This includes *all* buffers in the arc_mfu | |
600 | * state; e.g. data, metadata, evictable, and unevictable buffers | |
601 | * are all included in this value. | |
602 | */ | |
13be560d | 603 | kstat_named_t arcstat_mfu_size; |
500445c0 PS |
604 | /* |
605 | * Number of bytes consumed by ARC buffers that are eligible for | |
606 | * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu | |
607 | * state. | |
608 | */ | |
609 | kstat_named_t arcstat_mfu_evictable_data; | |
610 | /* | |
611 | * Number of bytes consumed by ARC buffers that are eligible for | |
612 | * eviction, of type ARC_BUFC_METADATA, and reside in the | |
613 | * arc_mfu state. | |
614 | */ | |
615 | kstat_named_t arcstat_mfu_evictable_metadata; | |
616 | /* | |
617 | * Total number of bytes that *would have been* consumed by ARC | |
618 | * buffers in the arc_mfu_ghost state. See the comment above | |
619 | * arcstat_mru_ghost_size for more details. | |
620 | */ | |
13be560d | 621 | kstat_named_t arcstat_mfu_ghost_size; |
500445c0 PS |
622 | /* |
623 | * Number of bytes that *would have been* consumed by ARC | |
624 | * buffers that are eligible for eviction, of type | |
625 | * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. | |
626 | */ | |
627 | kstat_named_t arcstat_mfu_ghost_evictable_data; | |
628 | /* | |
629 | * Number of bytes that *would have been* consumed by ARC | |
630 | * buffers that are eligible for eviction, of type | |
631 | * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. | |
632 | */ | |
633 | kstat_named_t arcstat_mfu_ghost_evictable_metadata; | |
34dc7c2f BB |
634 | kstat_named_t arcstat_l2_hits; |
635 | kstat_named_t arcstat_l2_misses; | |
636 | kstat_named_t arcstat_l2_feeds; | |
637 | kstat_named_t arcstat_l2_rw_clash; | |
d164b209 BB |
638 | kstat_named_t arcstat_l2_read_bytes; |
639 | kstat_named_t arcstat_l2_write_bytes; | |
34dc7c2f BB |
640 | kstat_named_t arcstat_l2_writes_sent; |
641 | kstat_named_t arcstat_l2_writes_done; | |
642 | kstat_named_t arcstat_l2_writes_error; | |
ca0bf58d | 643 | kstat_named_t arcstat_l2_writes_lock_retry; |
34dc7c2f BB |
644 | kstat_named_t arcstat_l2_evict_lock_retry; |
645 | kstat_named_t arcstat_l2_evict_reading; | |
b9541d6b | 646 | kstat_named_t arcstat_l2_evict_l1cached; |
34dc7c2f BB |
647 | kstat_named_t arcstat_l2_free_on_write; |
648 | kstat_named_t arcstat_l2_abort_lowmem; | |
649 | kstat_named_t arcstat_l2_cksum_bad; | |
650 | kstat_named_t arcstat_l2_io_error; | |
01850391 AG |
651 | kstat_named_t arcstat_l2_lsize; |
652 | kstat_named_t arcstat_l2_psize; | |
34dc7c2f BB |
653 | kstat_named_t arcstat_l2_hdr_size; |
654 | kstat_named_t arcstat_memory_throttle_count; | |
7cb67b45 BB |
655 | kstat_named_t arcstat_memory_direct_count; |
656 | kstat_named_t arcstat_memory_indirect_count; | |
70f02287 BB |
657 | kstat_named_t arcstat_memory_all_bytes; |
658 | kstat_named_t arcstat_memory_free_bytes; | |
659 | kstat_named_t arcstat_memory_available_bytes; | |
1834f2d8 BB |
660 | kstat_named_t arcstat_no_grow; |
661 | kstat_named_t arcstat_tempreserve; | |
662 | kstat_named_t arcstat_loaned_bytes; | |
ab26409d | 663 | kstat_named_t arcstat_prune; |
1834f2d8 BB |
664 | kstat_named_t arcstat_meta_used; |
665 | kstat_named_t arcstat_meta_limit; | |
25458cbe | 666 | kstat_named_t arcstat_dnode_limit; |
1834f2d8 | 667 | kstat_named_t arcstat_meta_max; |
ca0bf58d | 668 | kstat_named_t arcstat_meta_min; |
a8b2e306 | 669 | kstat_named_t arcstat_async_upgrade_sync; |
7f60329a | 670 | kstat_named_t arcstat_demand_hit_predictive_prefetch; |
d4a72f23 | 671 | kstat_named_t arcstat_demand_hit_prescient_prefetch; |
11f552fa BB |
672 | kstat_named_t arcstat_need_free; |
673 | kstat_named_t arcstat_sys_free; | |
b5256303 | 674 | kstat_named_t arcstat_raw_size; |
34dc7c2f BB |
675 | } arc_stats_t; |
676 | ||
677 | static arc_stats_t arc_stats = { | |
678 | { "hits", KSTAT_DATA_UINT64 }, | |
679 | { "misses", KSTAT_DATA_UINT64 }, | |
680 | { "demand_data_hits", KSTAT_DATA_UINT64 }, | |
681 | { "demand_data_misses", KSTAT_DATA_UINT64 }, | |
682 | { "demand_metadata_hits", KSTAT_DATA_UINT64 }, | |
683 | { "demand_metadata_misses", KSTAT_DATA_UINT64 }, | |
684 | { "prefetch_data_hits", KSTAT_DATA_UINT64 }, | |
685 | { "prefetch_data_misses", KSTAT_DATA_UINT64 }, | |
686 | { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, | |
687 | { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, | |
688 | { "mru_hits", KSTAT_DATA_UINT64 }, | |
689 | { "mru_ghost_hits", KSTAT_DATA_UINT64 }, | |
690 | { "mfu_hits", KSTAT_DATA_UINT64 }, | |
691 | { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, | |
692 | { "deleted", KSTAT_DATA_UINT64 }, | |
34dc7c2f | 693 | { "mutex_miss", KSTAT_DATA_UINT64 }, |
0873bb63 | 694 | { "access_skip", KSTAT_DATA_UINT64 }, |
34dc7c2f | 695 | { "evict_skip", KSTAT_DATA_UINT64 }, |
ca0bf58d | 696 | { "evict_not_enough", KSTAT_DATA_UINT64 }, |
428870ff BB |
697 | { "evict_l2_cached", KSTAT_DATA_UINT64 }, |
698 | { "evict_l2_eligible", KSTAT_DATA_UINT64 }, | |
699 | { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, | |
ca0bf58d | 700 | { "evict_l2_skip", KSTAT_DATA_UINT64 }, |
34dc7c2f BB |
701 | { "hash_elements", KSTAT_DATA_UINT64 }, |
702 | { "hash_elements_max", KSTAT_DATA_UINT64 }, | |
703 | { "hash_collisions", KSTAT_DATA_UINT64 }, | |
704 | { "hash_chains", KSTAT_DATA_UINT64 }, | |
705 | { "hash_chain_max", KSTAT_DATA_UINT64 }, | |
706 | { "p", KSTAT_DATA_UINT64 }, | |
707 | { "c", KSTAT_DATA_UINT64 }, | |
708 | { "c_min", KSTAT_DATA_UINT64 }, | |
709 | { "c_max", KSTAT_DATA_UINT64 }, | |
710 | { "size", KSTAT_DATA_UINT64 }, | |
d3c2ae1c GW |
711 | { "compressed_size", KSTAT_DATA_UINT64 }, |
712 | { "uncompressed_size", KSTAT_DATA_UINT64 }, | |
713 | { "overhead_size", KSTAT_DATA_UINT64 }, | |
34dc7c2f | 714 | { "hdr_size", KSTAT_DATA_UINT64 }, |
d164b209 | 715 | { "data_size", KSTAT_DATA_UINT64 }, |
500445c0 | 716 | { "metadata_size", KSTAT_DATA_UINT64 }, |
25458cbe TC |
717 | { "dbuf_size", KSTAT_DATA_UINT64 }, |
718 | { "dnode_size", KSTAT_DATA_UINT64 }, | |
719 | { "bonus_size", KSTAT_DATA_UINT64 }, | |
13be560d | 720 | { "anon_size", KSTAT_DATA_UINT64 }, |
500445c0 PS |
721 | { "anon_evictable_data", KSTAT_DATA_UINT64 }, |
722 | { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, | |
13be560d | 723 | { "mru_size", KSTAT_DATA_UINT64 }, |
500445c0 PS |
724 | { "mru_evictable_data", KSTAT_DATA_UINT64 }, |
725 | { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, | |
13be560d | 726 | { "mru_ghost_size", KSTAT_DATA_UINT64 }, |
500445c0 PS |
727 | { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, |
728 | { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, | |
13be560d | 729 | { "mfu_size", KSTAT_DATA_UINT64 }, |
500445c0 PS |
730 | { "mfu_evictable_data", KSTAT_DATA_UINT64 }, |
731 | { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, | |
13be560d | 732 | { "mfu_ghost_size", KSTAT_DATA_UINT64 }, |
500445c0 PS |
733 | { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, |
734 | { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, | |
34dc7c2f BB |
735 | { "l2_hits", KSTAT_DATA_UINT64 }, |
736 | { "l2_misses", KSTAT_DATA_UINT64 }, | |
737 | { "l2_feeds", KSTAT_DATA_UINT64 }, | |
738 | { "l2_rw_clash", KSTAT_DATA_UINT64 }, | |
d164b209 BB |
739 | { "l2_read_bytes", KSTAT_DATA_UINT64 }, |
740 | { "l2_write_bytes", KSTAT_DATA_UINT64 }, | |
34dc7c2f BB |
741 | { "l2_writes_sent", KSTAT_DATA_UINT64 }, |
742 | { "l2_writes_done", KSTAT_DATA_UINT64 }, | |
743 | { "l2_writes_error", KSTAT_DATA_UINT64 }, | |
ca0bf58d | 744 | { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, |
34dc7c2f BB |
745 | { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, |
746 | { "l2_evict_reading", KSTAT_DATA_UINT64 }, | |
b9541d6b | 747 | { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, |
34dc7c2f BB |
748 | { "l2_free_on_write", KSTAT_DATA_UINT64 }, |
749 | { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, | |
750 | { "l2_cksum_bad", KSTAT_DATA_UINT64 }, | |
751 | { "l2_io_error", KSTAT_DATA_UINT64 }, | |
752 | { "l2_size", KSTAT_DATA_UINT64 }, | |
3a17a7a9 | 753 | { "l2_asize", KSTAT_DATA_UINT64 }, |
34dc7c2f | 754 | { "l2_hdr_size", KSTAT_DATA_UINT64 }, |
1834f2d8 | 755 | { "memory_throttle_count", KSTAT_DATA_UINT64 }, |
7cb67b45 BB |
756 | { "memory_direct_count", KSTAT_DATA_UINT64 }, |
757 | { "memory_indirect_count", KSTAT_DATA_UINT64 }, | |
70f02287 BB |
758 | { "memory_all_bytes", KSTAT_DATA_UINT64 }, |
759 | { "memory_free_bytes", KSTAT_DATA_UINT64 }, | |
760 | { "memory_available_bytes", KSTAT_DATA_INT64 }, | |
1834f2d8 BB |
761 | { "arc_no_grow", KSTAT_DATA_UINT64 }, |
762 | { "arc_tempreserve", KSTAT_DATA_UINT64 }, | |
763 | { "arc_loaned_bytes", KSTAT_DATA_UINT64 }, | |
ab26409d | 764 | { "arc_prune", KSTAT_DATA_UINT64 }, |
1834f2d8 BB |
765 | { "arc_meta_used", KSTAT_DATA_UINT64 }, |
766 | { "arc_meta_limit", KSTAT_DATA_UINT64 }, | |
25458cbe | 767 | { "arc_dnode_limit", KSTAT_DATA_UINT64 }, |
1834f2d8 | 768 | { "arc_meta_max", KSTAT_DATA_UINT64 }, |
11f552fa | 769 | { "arc_meta_min", KSTAT_DATA_UINT64 }, |
a8b2e306 | 770 | { "async_upgrade_sync", KSTAT_DATA_UINT64 }, |
7f60329a | 771 | { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, |
d4a72f23 | 772 | { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 }, |
11f552fa | 773 | { "arc_need_free", KSTAT_DATA_UINT64 }, |
b5256303 TC |
774 | { "arc_sys_free", KSTAT_DATA_UINT64 }, |
775 | { "arc_raw_size", KSTAT_DATA_UINT64 } | |
34dc7c2f BB |
776 | }; |
777 | ||
778 | #define ARCSTAT(stat) (arc_stats.stat.value.ui64) | |
779 | ||
780 | #define ARCSTAT_INCR(stat, val) \ | |
d3cc8b15 | 781 | atomic_add_64(&arc_stats.stat.value.ui64, (val)) |
34dc7c2f | 782 | |
428870ff | 783 | #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) |
34dc7c2f BB |
784 | #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) |
785 | ||
786 | #define ARCSTAT_MAX(stat, val) { \ | |
787 | uint64_t m; \ | |
788 | while ((val) > (m = arc_stats.stat.value.ui64) && \ | |
789 | (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ | |
790 | continue; \ | |
791 | } | |
792 | ||
793 | #define ARCSTAT_MAXSTAT(stat) \ | |
794 | ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) | |
795 | ||
796 | /* | |
797 | * We define a macro to allow ARC hits/misses to be easily broken down by | |
798 | * two separate conditions, giving a total of four different subtypes for | |
799 | * each of hits and misses (so eight statistics total). | |
800 | */ | |
801 | #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ | |
802 | if (cond1) { \ | |
803 | if (cond2) { \ | |
804 | ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ | |
805 | } else { \ | |
806 | ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ | |
807 | } \ | |
808 | } else { \ | |
809 | if (cond2) { \ | |
810 | ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ | |
811 | } else { \ | |
812 | ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ | |
813 | } \ | |
814 | } | |
815 | ||
816 | kstat_t *arc_ksp; | |
428870ff | 817 | static arc_state_t *arc_anon; |
34dc7c2f BB |
818 | static arc_state_t *arc_mru; |
819 | static arc_state_t *arc_mru_ghost; | |
820 | static arc_state_t *arc_mfu; | |
821 | static arc_state_t *arc_mfu_ghost; | |
822 | static arc_state_t *arc_l2c_only; | |
823 | ||
824 | /* | |
825 | * There are several ARC variables that are critical to export as kstats -- | |
826 | * but we don't want to have to grovel around in the kstat whenever we wish to | |
827 | * manipulate them. For these variables, we therefore define them to be in | |
828 | * terms of the statistic variable. This assures that we are not introducing | |
829 | * the possibility of inconsistency by having shadow copies of the variables, | |
830 | * while still allowing the code to be readable. | |
831 | */ | |
832 | #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ | |
833 | #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ | |
834 | #define arc_c ARCSTAT(arcstat_c) /* target size of cache */ | |
835 | #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ | |
836 | #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ | |
d3c2ae1c | 837 | #define arc_no_grow ARCSTAT(arcstat_no_grow) /* do not grow cache size */ |
1834f2d8 BB |
838 | #define arc_tempreserve ARCSTAT(arcstat_tempreserve) |
839 | #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes) | |
23c0a133 | 840 | #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ |
25458cbe | 841 | #define arc_dnode_limit ARCSTAT(arcstat_dnode_limit) /* max size for dnodes */ |
ca0bf58d | 842 | #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ |
23c0a133 GW |
843 | #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ |
844 | #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ | |
25458cbe TC |
845 | #define arc_dbuf_size ARCSTAT(arcstat_dbuf_size) /* dbuf metadata */ |
846 | #define arc_dnode_size ARCSTAT(arcstat_dnode_size) /* dnode metadata */ | |
847 | #define arc_bonus_size ARCSTAT(arcstat_bonus_size) /* bonus buffer metadata */ | |
11f552fa BB |
848 | #define arc_need_free ARCSTAT(arcstat_need_free) /* bytes to be freed */ |
849 | #define arc_sys_free ARCSTAT(arcstat_sys_free) /* target system free bytes */ | |
34dc7c2f | 850 | |
b5256303 TC |
851 | /* size of all b_rabd's in entire arc */ |
852 | #define arc_raw_size ARCSTAT(arcstat_raw_size) | |
d3c2ae1c GW |
853 | /* compressed size of entire arc */ |
854 | #define arc_compressed_size ARCSTAT(arcstat_compressed_size) | |
855 | /* uncompressed size of entire arc */ | |
856 | #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) | |
857 | /* number of bytes in the arc from arc_buf_t's */ | |
858 | #define arc_overhead_size ARCSTAT(arcstat_overhead_size) | |
3a17a7a9 | 859 | |
ab26409d BB |
860 | static list_t arc_prune_list; |
861 | static kmutex_t arc_prune_mtx; | |
f6046738 | 862 | static taskq_t *arc_prune_taskq; |
428870ff | 863 | |
34dc7c2f BB |
864 | #define GHOST_STATE(state) \ |
865 | ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ | |
866 | (state) == arc_l2c_only) | |
867 | ||
2a432414 GW |
868 | #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) |
869 | #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) | |
870 | #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) | |
871 | #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) | |
d4a72f23 TC |
872 | #define HDR_PRESCIENT_PREFETCH(hdr) \ |
873 | ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) | |
d3c2ae1c GW |
874 | #define HDR_COMPRESSION_ENABLED(hdr) \ |
875 | ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) | |
b9541d6b | 876 | |
2a432414 GW |
877 | #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) |
878 | #define HDR_L2_READING(hdr) \ | |
d3c2ae1c GW |
879 | (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ |
880 | ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) | |
2a432414 GW |
881 | #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) |
882 | #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) | |
883 | #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) | |
b5256303 TC |
884 | #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED) |
885 | #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH) | |
d3c2ae1c | 886 | #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) |
34dc7c2f | 887 | |
b9541d6b | 888 | #define HDR_ISTYPE_METADATA(hdr) \ |
d3c2ae1c | 889 | ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) |
b9541d6b CW |
890 | #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) |
891 | ||
892 | #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) | |
893 | #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) | |
b5256303 TC |
894 | #define HDR_HAS_RABD(hdr) \ |
895 | (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \ | |
896 | (hdr)->b_crypt_hdr.b_rabd != NULL) | |
897 | #define HDR_ENCRYPTED(hdr) \ | |
898 | (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot)) | |
899 | #define HDR_AUTHENTICATED(hdr) \ | |
900 | (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot)) | |
b9541d6b | 901 | |
d3c2ae1c GW |
902 | /* For storing compression mode in b_flags */ |
903 | #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) | |
904 | ||
905 | #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ | |
906 | HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) | |
907 | #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ | |
908 | HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); | |
909 | ||
910 | #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) | |
524b4217 DK |
911 | #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) |
912 | #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) | |
b5256303 | 913 | #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED) |
d3c2ae1c | 914 | |
34dc7c2f BB |
915 | /* |
916 | * Other sizes | |
917 | */ | |
918 | ||
b5256303 TC |
919 | #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) |
920 | #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr)) | |
b9541d6b | 921 | #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) |
34dc7c2f BB |
922 | |
923 | /* | |
924 | * Hash table routines | |
925 | */ | |
926 | ||
00b46022 BB |
927 | #define HT_LOCK_ALIGN 64 |
928 | #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN))) | |
34dc7c2f BB |
929 | |
930 | struct ht_lock { | |
931 | kmutex_t ht_lock; | |
932 | #ifdef _KERNEL | |
00b46022 | 933 | unsigned char pad[HT_LOCK_PAD]; |
34dc7c2f BB |
934 | #endif |
935 | }; | |
936 | ||
b31d8ea7 | 937 | #define BUF_LOCKS 8192 |
34dc7c2f BB |
938 | typedef struct buf_hash_table { |
939 | uint64_t ht_mask; | |
940 | arc_buf_hdr_t **ht_table; | |
941 | struct ht_lock ht_locks[BUF_LOCKS]; | |
942 | } buf_hash_table_t; | |
943 | ||
944 | static buf_hash_table_t buf_hash_table; | |
945 | ||
946 | #define BUF_HASH_INDEX(spa, dva, birth) \ | |
947 | (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) | |
948 | #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) | |
949 | #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) | |
428870ff BB |
950 | #define HDR_LOCK(hdr) \ |
951 | (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) | |
34dc7c2f BB |
952 | |
953 | uint64_t zfs_crc64_table[256]; | |
954 | ||
955 | /* | |
956 | * Level 2 ARC | |
957 | */ | |
958 | ||
959 | #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ | |
3a17a7a9 | 960 | #define L2ARC_HEADROOM 2 /* num of writes */ |
8a09d5fd | 961 | |
3a17a7a9 SK |
962 | /* |
963 | * If we discover during ARC scan any buffers to be compressed, we boost | |
964 | * our headroom for the next scanning cycle by this percentage multiple. | |
965 | */ | |
966 | #define L2ARC_HEADROOM_BOOST 200 | |
d164b209 BB |
967 | #define L2ARC_FEED_SECS 1 /* caching interval secs */ |
968 | #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ | |
34dc7c2f | 969 | |
4aafab91 G |
970 | /* |
971 | * We can feed L2ARC from two states of ARC buffers, mru and mfu, | |
972 | * and each of the state has two types: data and metadata. | |
973 | */ | |
974 | #define L2ARC_FEED_TYPES 4 | |
975 | ||
34dc7c2f BB |
976 | #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) |
977 | #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) | |
978 | ||
d3cc8b15 | 979 | /* L2ARC Performance Tunables */ |
abd8610c BB |
980 | unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */ |
981 | unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */ | |
982 | unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */ | |
3a17a7a9 | 983 | unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; |
abd8610c BB |
984 | unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ |
985 | unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */ | |
986 | int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ | |
987 | int l2arc_feed_again = B_TRUE; /* turbo warmup */ | |
c93504f0 | 988 | int l2arc_norw = B_FALSE; /* no reads during writes */ |
34dc7c2f BB |
989 | |
990 | /* | |
991 | * L2ARC Internals | |
992 | */ | |
34dc7c2f BB |
993 | static list_t L2ARC_dev_list; /* device list */ |
994 | static list_t *l2arc_dev_list; /* device list pointer */ | |
995 | static kmutex_t l2arc_dev_mtx; /* device list mutex */ | |
996 | static l2arc_dev_t *l2arc_dev_last; /* last device used */ | |
34dc7c2f BB |
997 | static list_t L2ARC_free_on_write; /* free after write buf list */ |
998 | static list_t *l2arc_free_on_write; /* free after write list ptr */ | |
999 | static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ | |
1000 | static uint64_t l2arc_ndev; /* number of devices */ | |
1001 | ||
1002 | typedef struct l2arc_read_callback { | |
2aa34383 | 1003 | arc_buf_hdr_t *l2rcb_hdr; /* read header */ |
3a17a7a9 | 1004 | blkptr_t l2rcb_bp; /* original blkptr */ |
5dbd68a3 | 1005 | zbookmark_phys_t l2rcb_zb; /* original bookmark */ |
3a17a7a9 | 1006 | int l2rcb_flags; /* original flags */ |
82710e99 | 1007 | abd_t *l2rcb_abd; /* temporary buffer */ |
34dc7c2f BB |
1008 | } l2arc_read_callback_t; |
1009 | ||
34dc7c2f BB |
1010 | typedef struct l2arc_data_free { |
1011 | /* protected by l2arc_free_on_write_mtx */ | |
a6255b7f | 1012 | abd_t *l2df_abd; |
34dc7c2f | 1013 | size_t l2df_size; |
d3c2ae1c | 1014 | arc_buf_contents_t l2df_type; |
34dc7c2f BB |
1015 | list_node_t l2df_list_node; |
1016 | } l2arc_data_free_t; | |
1017 | ||
b5256303 TC |
1018 | typedef enum arc_fill_flags { |
1019 | ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */ | |
1020 | ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */ | |
1021 | ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */ | |
1022 | ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */ | |
1023 | ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */ | |
1024 | } arc_fill_flags_t; | |
1025 | ||
34dc7c2f BB |
1026 | static kmutex_t l2arc_feed_thr_lock; |
1027 | static kcondvar_t l2arc_feed_thr_cv; | |
1028 | static uint8_t l2arc_thread_exit; | |
1029 | ||
a6255b7f | 1030 | static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *); |
d3c2ae1c | 1031 | static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); |
a6255b7f DQ |
1032 | static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *); |
1033 | static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *); | |
d3c2ae1c | 1034 | static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); |
a6255b7f | 1035 | static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag); |
b5256303 TC |
1036 | static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t); |
1037 | static void arc_hdr_alloc_abd(arc_buf_hdr_t *, boolean_t); | |
2a432414 | 1038 | static void arc_access(arc_buf_hdr_t *, kmutex_t *); |
ca0bf58d | 1039 | static boolean_t arc_is_overflowing(void); |
2a432414 | 1040 | static void arc_buf_watch(arc_buf_t *); |
ca67b33a | 1041 | static void arc_tuning_update(void); |
25458cbe | 1042 | static void arc_prune_async(int64_t); |
9edb3695 | 1043 | static uint64_t arc_all_memory(void); |
2a432414 | 1044 | |
b9541d6b CW |
1045 | static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); |
1046 | static uint32_t arc_bufc_to_flags(arc_buf_contents_t); | |
d3c2ae1c GW |
1047 | static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); |
1048 | static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); | |
b9541d6b | 1049 | |
2a432414 GW |
1050 | static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); |
1051 | static void l2arc_read_done(zio_t *); | |
34dc7c2f BB |
1052 | |
1053 | static uint64_t | |
d164b209 | 1054 | buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) |
34dc7c2f | 1055 | { |
34dc7c2f BB |
1056 | uint8_t *vdva = (uint8_t *)dva; |
1057 | uint64_t crc = -1ULL; | |
1058 | int i; | |
1059 | ||
1060 | ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); | |
1061 | ||
1062 | for (i = 0; i < sizeof (dva_t); i++) | |
1063 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; | |
1064 | ||
d164b209 | 1065 | crc ^= (spa>>8) ^ birth; |
34dc7c2f BB |
1066 | |
1067 | return (crc); | |
1068 | } | |
1069 | ||
d3c2ae1c GW |
1070 | #define HDR_EMPTY(hdr) \ |
1071 | ((hdr)->b_dva.dva_word[0] == 0 && \ | |
1072 | (hdr)->b_dva.dva_word[1] == 0) | |
34dc7c2f | 1073 | |
d3c2ae1c GW |
1074 | #define HDR_EQUAL(spa, dva, birth, hdr) \ |
1075 | ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ | |
1076 | ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ | |
1077 | ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) | |
34dc7c2f | 1078 | |
428870ff BB |
1079 | static void |
1080 | buf_discard_identity(arc_buf_hdr_t *hdr) | |
1081 | { | |
1082 | hdr->b_dva.dva_word[0] = 0; | |
1083 | hdr->b_dva.dva_word[1] = 0; | |
1084 | hdr->b_birth = 0; | |
428870ff BB |
1085 | } |
1086 | ||
34dc7c2f | 1087 | static arc_buf_hdr_t * |
9b67f605 | 1088 | buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) |
34dc7c2f | 1089 | { |
9b67f605 MA |
1090 | const dva_t *dva = BP_IDENTITY(bp); |
1091 | uint64_t birth = BP_PHYSICAL_BIRTH(bp); | |
34dc7c2f BB |
1092 | uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); |
1093 | kmutex_t *hash_lock = BUF_HASH_LOCK(idx); | |
2a432414 | 1094 | arc_buf_hdr_t *hdr; |
34dc7c2f BB |
1095 | |
1096 | mutex_enter(hash_lock); | |
2a432414 GW |
1097 | for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; |
1098 | hdr = hdr->b_hash_next) { | |
d3c2ae1c | 1099 | if (HDR_EQUAL(spa, dva, birth, hdr)) { |
34dc7c2f | 1100 | *lockp = hash_lock; |
2a432414 | 1101 | return (hdr); |
34dc7c2f BB |
1102 | } |
1103 | } | |
1104 | mutex_exit(hash_lock); | |
1105 | *lockp = NULL; | |
1106 | return (NULL); | |
1107 | } | |
1108 | ||
1109 | /* | |
1110 | * Insert an entry into the hash table. If there is already an element | |
1111 | * equal to elem in the hash table, then the already existing element | |
1112 | * will be returned and the new element will not be inserted. | |
1113 | * Otherwise returns NULL. | |
b9541d6b | 1114 | * If lockp == NULL, the caller is assumed to already hold the hash lock. |
34dc7c2f BB |
1115 | */ |
1116 | static arc_buf_hdr_t * | |
2a432414 | 1117 | buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) |
34dc7c2f | 1118 | { |
2a432414 | 1119 | uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); |
34dc7c2f | 1120 | kmutex_t *hash_lock = BUF_HASH_LOCK(idx); |
2a432414 | 1121 | arc_buf_hdr_t *fhdr; |
34dc7c2f BB |
1122 | uint32_t i; |
1123 | ||
2a432414 GW |
1124 | ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); |
1125 | ASSERT(hdr->b_birth != 0); | |
1126 | ASSERT(!HDR_IN_HASH_TABLE(hdr)); | |
b9541d6b CW |
1127 | |
1128 | if (lockp != NULL) { | |
1129 | *lockp = hash_lock; | |
1130 | mutex_enter(hash_lock); | |
1131 | } else { | |
1132 | ASSERT(MUTEX_HELD(hash_lock)); | |
1133 | } | |
1134 | ||
2a432414 GW |
1135 | for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; |
1136 | fhdr = fhdr->b_hash_next, i++) { | |
d3c2ae1c | 1137 | if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) |
2a432414 | 1138 | return (fhdr); |
34dc7c2f BB |
1139 | } |
1140 | ||
2a432414 GW |
1141 | hdr->b_hash_next = buf_hash_table.ht_table[idx]; |
1142 | buf_hash_table.ht_table[idx] = hdr; | |
d3c2ae1c | 1143 | arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); |
34dc7c2f BB |
1144 | |
1145 | /* collect some hash table performance data */ | |
1146 | if (i > 0) { | |
1147 | ARCSTAT_BUMP(arcstat_hash_collisions); | |
1148 | if (i == 1) | |
1149 | ARCSTAT_BUMP(arcstat_hash_chains); | |
1150 | ||
1151 | ARCSTAT_MAX(arcstat_hash_chain_max, i); | |
1152 | } | |
1153 | ||
1154 | ARCSTAT_BUMP(arcstat_hash_elements); | |
1155 | ARCSTAT_MAXSTAT(arcstat_hash_elements); | |
1156 | ||
1157 | return (NULL); | |
1158 | } | |
1159 | ||
1160 | static void | |
2a432414 | 1161 | buf_hash_remove(arc_buf_hdr_t *hdr) |
34dc7c2f | 1162 | { |
2a432414 GW |
1163 | arc_buf_hdr_t *fhdr, **hdrp; |
1164 | uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); | |
34dc7c2f BB |
1165 | |
1166 | ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); | |
2a432414 | 1167 | ASSERT(HDR_IN_HASH_TABLE(hdr)); |
34dc7c2f | 1168 | |
2a432414 GW |
1169 | hdrp = &buf_hash_table.ht_table[idx]; |
1170 | while ((fhdr = *hdrp) != hdr) { | |
d3c2ae1c | 1171 | ASSERT3P(fhdr, !=, NULL); |
2a432414 | 1172 | hdrp = &fhdr->b_hash_next; |
34dc7c2f | 1173 | } |
2a432414 GW |
1174 | *hdrp = hdr->b_hash_next; |
1175 | hdr->b_hash_next = NULL; | |
d3c2ae1c | 1176 | arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); |
34dc7c2f BB |
1177 | |
1178 | /* collect some hash table performance data */ | |
1179 | ARCSTAT_BUMPDOWN(arcstat_hash_elements); | |
1180 | ||
1181 | if (buf_hash_table.ht_table[idx] && | |
1182 | buf_hash_table.ht_table[idx]->b_hash_next == NULL) | |
1183 | ARCSTAT_BUMPDOWN(arcstat_hash_chains); | |
1184 | } | |
1185 | ||
1186 | /* | |
1187 | * Global data structures and functions for the buf kmem cache. | |
1188 | */ | |
b5256303 | 1189 | |
b9541d6b | 1190 | static kmem_cache_t *hdr_full_cache; |
b5256303 | 1191 | static kmem_cache_t *hdr_full_crypt_cache; |
b9541d6b | 1192 | static kmem_cache_t *hdr_l2only_cache; |
34dc7c2f BB |
1193 | static kmem_cache_t *buf_cache; |
1194 | ||
1195 | static void | |
1196 | buf_fini(void) | |
1197 | { | |
1198 | int i; | |
1199 | ||
93ce2b4c | 1200 | #if defined(_KERNEL) |
d1d7e268 MK |
1201 | /* |
1202 | * Large allocations which do not require contiguous pages | |
1203 | * should be using vmem_free() in the linux kernel\ | |
1204 | */ | |
00b46022 BB |
1205 | vmem_free(buf_hash_table.ht_table, |
1206 | (buf_hash_table.ht_mask + 1) * sizeof (void *)); | |
1207 | #else | |
34dc7c2f BB |
1208 | kmem_free(buf_hash_table.ht_table, |
1209 | (buf_hash_table.ht_mask + 1) * sizeof (void *)); | |
00b46022 | 1210 | #endif |
34dc7c2f BB |
1211 | for (i = 0; i < BUF_LOCKS; i++) |
1212 | mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); | |
b9541d6b | 1213 | kmem_cache_destroy(hdr_full_cache); |
b5256303 | 1214 | kmem_cache_destroy(hdr_full_crypt_cache); |
b9541d6b | 1215 | kmem_cache_destroy(hdr_l2only_cache); |
34dc7c2f BB |
1216 | kmem_cache_destroy(buf_cache); |
1217 | } | |
1218 | ||
1219 | /* | |
1220 | * Constructor callback - called when the cache is empty | |
1221 | * and a new buf is requested. | |
1222 | */ | |
1223 | /* ARGSUSED */ | |
1224 | static int | |
b9541d6b CW |
1225 | hdr_full_cons(void *vbuf, void *unused, int kmflag) |
1226 | { | |
1227 | arc_buf_hdr_t *hdr = vbuf; | |
1228 | ||
1229 | bzero(hdr, HDR_FULL_SIZE); | |
ae76f45c | 1230 | hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; |
b9541d6b CW |
1231 | cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); |
1232 | refcount_create(&hdr->b_l1hdr.b_refcnt); | |
1233 | mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); | |
1234 | list_link_init(&hdr->b_l1hdr.b_arc_node); | |
1235 | list_link_init(&hdr->b_l2hdr.b_l2node); | |
ca0bf58d | 1236 | multilist_link_init(&hdr->b_l1hdr.b_arc_node); |
b9541d6b CW |
1237 | arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); |
1238 | ||
1239 | return (0); | |
1240 | } | |
1241 | ||
b5256303 TC |
1242 | /* ARGSUSED */ |
1243 | static int | |
1244 | hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag) | |
1245 | { | |
1246 | arc_buf_hdr_t *hdr = vbuf; | |
1247 | ||
1248 | hdr_full_cons(vbuf, unused, kmflag); | |
1249 | bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr)); | |
1250 | arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS); | |
1251 | ||
1252 | return (0); | |
1253 | } | |
1254 | ||
b9541d6b CW |
1255 | /* ARGSUSED */ |
1256 | static int | |
1257 | hdr_l2only_cons(void *vbuf, void *unused, int kmflag) | |
34dc7c2f | 1258 | { |
2a432414 GW |
1259 | arc_buf_hdr_t *hdr = vbuf; |
1260 | ||
b9541d6b CW |
1261 | bzero(hdr, HDR_L2ONLY_SIZE); |
1262 | arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); | |
34dc7c2f | 1263 | |
34dc7c2f BB |
1264 | return (0); |
1265 | } | |
1266 | ||
b128c09f BB |
1267 | /* ARGSUSED */ |
1268 | static int | |
1269 | buf_cons(void *vbuf, void *unused, int kmflag) | |
1270 | { | |
1271 | arc_buf_t *buf = vbuf; | |
1272 | ||
1273 | bzero(buf, sizeof (arc_buf_t)); | |
428870ff | 1274 | mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); |
d164b209 BB |
1275 | arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); |
1276 | ||
b128c09f BB |
1277 | return (0); |
1278 | } | |
1279 | ||
34dc7c2f BB |
1280 | /* |
1281 | * Destructor callback - called when a cached buf is | |
1282 | * no longer required. | |
1283 | */ | |
1284 | /* ARGSUSED */ | |
1285 | static void | |
b9541d6b | 1286 | hdr_full_dest(void *vbuf, void *unused) |
34dc7c2f | 1287 | { |
2a432414 | 1288 | arc_buf_hdr_t *hdr = vbuf; |
34dc7c2f | 1289 | |
d3c2ae1c | 1290 | ASSERT(HDR_EMPTY(hdr)); |
b9541d6b CW |
1291 | cv_destroy(&hdr->b_l1hdr.b_cv); |
1292 | refcount_destroy(&hdr->b_l1hdr.b_refcnt); | |
1293 | mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); | |
ca0bf58d | 1294 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); |
b9541d6b CW |
1295 | arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); |
1296 | } | |
1297 | ||
b5256303 TC |
1298 | /* ARGSUSED */ |
1299 | static void | |
1300 | hdr_full_crypt_dest(void *vbuf, void *unused) | |
1301 | { | |
1302 | arc_buf_hdr_t *hdr = vbuf; | |
1303 | ||
1304 | hdr_full_dest(vbuf, unused); | |
1305 | arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS); | |
1306 | } | |
1307 | ||
b9541d6b CW |
1308 | /* ARGSUSED */ |
1309 | static void | |
1310 | hdr_l2only_dest(void *vbuf, void *unused) | |
1311 | { | |
1312 | ASSERTV(arc_buf_hdr_t *hdr = vbuf); | |
1313 | ||
d3c2ae1c | 1314 | ASSERT(HDR_EMPTY(hdr)); |
b9541d6b | 1315 | arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); |
34dc7c2f BB |
1316 | } |
1317 | ||
b128c09f BB |
1318 | /* ARGSUSED */ |
1319 | static void | |
1320 | buf_dest(void *vbuf, void *unused) | |
1321 | { | |
1322 | arc_buf_t *buf = vbuf; | |
1323 | ||
428870ff | 1324 | mutex_destroy(&buf->b_evict_lock); |
d164b209 | 1325 | arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); |
b128c09f BB |
1326 | } |
1327 | ||
8c8af9d8 BB |
1328 | /* |
1329 | * Reclaim callback -- invoked when memory is low. | |
1330 | */ | |
1331 | /* ARGSUSED */ | |
1332 | static void | |
1333 | hdr_recl(void *unused) | |
1334 | { | |
1335 | dprintf("hdr_recl called\n"); | |
1336 | /* | |
1337 | * umem calls the reclaim func when we destroy the buf cache, | |
1338 | * which is after we do arc_fini(). | |
1339 | */ | |
1340 | if (!arc_dead) | |
1341 | cv_signal(&arc_reclaim_thread_cv); | |
1342 | } | |
1343 | ||
34dc7c2f BB |
1344 | static void |
1345 | buf_init(void) | |
1346 | { | |
2db28197 | 1347 | uint64_t *ct = NULL; |
34dc7c2f BB |
1348 | uint64_t hsize = 1ULL << 12; |
1349 | int i, j; | |
1350 | ||
1351 | /* | |
1352 | * The hash table is big enough to fill all of physical memory | |
49ddb315 MA |
1353 | * with an average block size of zfs_arc_average_blocksize (default 8K). |
1354 | * By default, the table will take up | |
1355 | * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). | |
34dc7c2f | 1356 | */ |
9edb3695 | 1357 | while (hsize * zfs_arc_average_blocksize < arc_all_memory()) |
34dc7c2f BB |
1358 | hsize <<= 1; |
1359 | retry: | |
1360 | buf_hash_table.ht_mask = hsize - 1; | |
93ce2b4c | 1361 | #if defined(_KERNEL) |
d1d7e268 MK |
1362 | /* |
1363 | * Large allocations which do not require contiguous pages | |
1364 | * should be using vmem_alloc() in the linux kernel | |
1365 | */ | |
00b46022 BB |
1366 | buf_hash_table.ht_table = |
1367 | vmem_zalloc(hsize * sizeof (void*), KM_SLEEP); | |
1368 | #else | |
34dc7c2f BB |
1369 | buf_hash_table.ht_table = |
1370 | kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); | |
00b46022 | 1371 | #endif |
34dc7c2f BB |
1372 | if (buf_hash_table.ht_table == NULL) { |
1373 | ASSERT(hsize > (1ULL << 8)); | |
1374 | hsize >>= 1; | |
1375 | goto retry; | |
1376 | } | |
1377 | ||
b9541d6b | 1378 | hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, |
8c8af9d8 | 1379 | 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); |
b5256303 TC |
1380 | hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt", |
1381 | HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest, | |
1382 | hdr_recl, NULL, NULL, 0); | |
b9541d6b | 1383 | hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", |
8c8af9d8 | 1384 | HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, |
b9541d6b | 1385 | NULL, NULL, 0); |
34dc7c2f | 1386 | buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), |
b128c09f | 1387 | 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); |
34dc7c2f BB |
1388 | |
1389 | for (i = 0; i < 256; i++) | |
1390 | for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) | |
1391 | *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); | |
1392 | ||
1393 | for (i = 0; i < BUF_LOCKS; i++) { | |
1394 | mutex_init(&buf_hash_table.ht_locks[i].ht_lock, | |
40d06e3c | 1395 | NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f BB |
1396 | } |
1397 | } | |
1398 | ||
d3c2ae1c | 1399 | #define ARC_MINTIME (hz>>4) /* 62 ms */ |
ca0bf58d | 1400 | |
2aa34383 DK |
1401 | /* |
1402 | * This is the size that the buf occupies in memory. If the buf is compressed, | |
1403 | * it will correspond to the compressed size. You should use this method of | |
1404 | * getting the buf size unless you explicitly need the logical size. | |
1405 | */ | |
1406 | uint64_t | |
1407 | arc_buf_size(arc_buf_t *buf) | |
1408 | { | |
1409 | return (ARC_BUF_COMPRESSED(buf) ? | |
1410 | HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); | |
1411 | } | |
1412 | ||
1413 | uint64_t | |
1414 | arc_buf_lsize(arc_buf_t *buf) | |
1415 | { | |
1416 | return (HDR_GET_LSIZE(buf->b_hdr)); | |
1417 | } | |
1418 | ||
b5256303 TC |
1419 | /* |
1420 | * This function will return B_TRUE if the buffer is encrypted in memory. | |
1421 | * This buffer can be decrypted by calling arc_untransform(). | |
1422 | */ | |
1423 | boolean_t | |
1424 | arc_is_encrypted(arc_buf_t *buf) | |
1425 | { | |
1426 | return (ARC_BUF_ENCRYPTED(buf) != 0); | |
1427 | } | |
1428 | ||
1429 | /* | |
1430 | * Returns B_TRUE if the buffer represents data that has not had its MAC | |
1431 | * verified yet. | |
1432 | */ | |
1433 | boolean_t | |
1434 | arc_is_unauthenticated(arc_buf_t *buf) | |
1435 | { | |
1436 | return (HDR_NOAUTH(buf->b_hdr) != 0); | |
1437 | } | |
1438 | ||
1439 | void | |
1440 | arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt, | |
1441 | uint8_t *iv, uint8_t *mac) | |
1442 | { | |
1443 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
1444 | ||
1445 | ASSERT(HDR_PROTECTED(hdr)); | |
1446 | ||
1447 | bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN); | |
1448 | bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN); | |
1449 | bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN); | |
1450 | *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ? | |
1451 | ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER; | |
1452 | } | |
1453 | ||
1454 | /* | |
1455 | * Indicates how this buffer is compressed in memory. If it is not compressed | |
1456 | * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with | |
1457 | * arc_untransform() as long as it is also unencrypted. | |
1458 | */ | |
2aa34383 DK |
1459 | enum zio_compress |
1460 | arc_get_compression(arc_buf_t *buf) | |
1461 | { | |
1462 | return (ARC_BUF_COMPRESSED(buf) ? | |
1463 | HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); | |
1464 | } | |
1465 | ||
b5256303 TC |
1466 | /* |
1467 | * Return the compression algorithm used to store this data in the ARC. If ARC | |
1468 | * compression is enabled or this is an encrypted block, this will be the same | |
1469 | * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF. | |
1470 | */ | |
1471 | static inline enum zio_compress | |
1472 | arc_hdr_get_compress(arc_buf_hdr_t *hdr) | |
1473 | { | |
1474 | return (HDR_COMPRESSION_ENABLED(hdr) ? | |
1475 | HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF); | |
1476 | } | |
1477 | ||
d3c2ae1c GW |
1478 | static inline boolean_t |
1479 | arc_buf_is_shared(arc_buf_t *buf) | |
1480 | { | |
1481 | boolean_t shared = (buf->b_data != NULL && | |
a6255b7f DQ |
1482 | buf->b_hdr->b_l1hdr.b_pabd != NULL && |
1483 | abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) && | |
1484 | buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd)); | |
d3c2ae1c | 1485 | IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); |
2aa34383 DK |
1486 | IMPLY(shared, ARC_BUF_SHARED(buf)); |
1487 | IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); | |
524b4217 DK |
1488 | |
1489 | /* | |
1490 | * It would be nice to assert arc_can_share() too, but the "hdr isn't | |
1491 | * already being shared" requirement prevents us from doing that. | |
1492 | */ | |
1493 | ||
d3c2ae1c GW |
1494 | return (shared); |
1495 | } | |
ca0bf58d | 1496 | |
a7004725 DK |
1497 | /* |
1498 | * Free the checksum associated with this header. If there is no checksum, this | |
1499 | * is a no-op. | |
1500 | */ | |
d3c2ae1c GW |
1501 | static inline void |
1502 | arc_cksum_free(arc_buf_hdr_t *hdr) | |
1503 | { | |
1504 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
b5256303 | 1505 | |
d3c2ae1c GW |
1506 | mutex_enter(&hdr->b_l1hdr.b_freeze_lock); |
1507 | if (hdr->b_l1hdr.b_freeze_cksum != NULL) { | |
1508 | kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); | |
1509 | hdr->b_l1hdr.b_freeze_cksum = NULL; | |
b9541d6b | 1510 | } |
d3c2ae1c | 1511 | mutex_exit(&hdr->b_l1hdr.b_freeze_lock); |
b9541d6b CW |
1512 | } |
1513 | ||
a7004725 DK |
1514 | /* |
1515 | * Return true iff at least one of the bufs on hdr is not compressed. | |
b5256303 | 1516 | * Encrypted buffers count as compressed. |
a7004725 DK |
1517 | */ |
1518 | static boolean_t | |
1519 | arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) | |
1520 | { | |
1521 | for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { | |
1522 | if (!ARC_BUF_COMPRESSED(b)) { | |
1523 | return (B_TRUE); | |
1524 | } | |
1525 | } | |
1526 | return (B_FALSE); | |
1527 | } | |
1528 | ||
1529 | ||
524b4217 DK |
1530 | /* |
1531 | * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data | |
1532 | * matches the checksum that is stored in the hdr. If there is no checksum, | |
1533 | * or if the buf is compressed, this is a no-op. | |
1534 | */ | |
34dc7c2f BB |
1535 | static void |
1536 | arc_cksum_verify(arc_buf_t *buf) | |
1537 | { | |
d3c2ae1c | 1538 | arc_buf_hdr_t *hdr = buf->b_hdr; |
34dc7c2f BB |
1539 | zio_cksum_t zc; |
1540 | ||
1541 | if (!(zfs_flags & ZFS_DEBUG_MODIFY)) | |
1542 | return; | |
1543 | ||
524b4217 | 1544 | if (ARC_BUF_COMPRESSED(buf)) { |
a7004725 DK |
1545 | ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || |
1546 | arc_hdr_has_uncompressed_buf(hdr)); | |
524b4217 DK |
1547 | return; |
1548 | } | |
1549 | ||
d3c2ae1c GW |
1550 | ASSERT(HDR_HAS_L1HDR(hdr)); |
1551 | ||
1552 | mutex_enter(&hdr->b_l1hdr.b_freeze_lock); | |
1553 | if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { | |
1554 | mutex_exit(&hdr->b_l1hdr.b_freeze_lock); | |
34dc7c2f BB |
1555 | return; |
1556 | } | |
2aa34383 | 1557 | |
3c67d83a | 1558 | fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); |
d3c2ae1c | 1559 | if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) |
34dc7c2f | 1560 | panic("buffer modified while frozen!"); |
d3c2ae1c | 1561 | mutex_exit(&hdr->b_l1hdr.b_freeze_lock); |
34dc7c2f BB |
1562 | } |
1563 | ||
b5256303 TC |
1564 | /* |
1565 | * This function makes the assumption that data stored in the L2ARC | |
1566 | * will be transformed exactly as it is in the main pool. Because of | |
1567 | * this we can verify the checksum against the reading process's bp. | |
1568 | */ | |
d3c2ae1c GW |
1569 | static boolean_t |
1570 | arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) | |
34dc7c2f | 1571 | { |
d3c2ae1c GW |
1572 | ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); |
1573 | VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); | |
34dc7c2f | 1574 | |
d3c2ae1c GW |
1575 | /* |
1576 | * Block pointers always store the checksum for the logical data. | |
1577 | * If the block pointer has the gang bit set, then the checksum | |
1578 | * it represents is for the reconstituted data and not for an | |
1579 | * individual gang member. The zio pipeline, however, must be able to | |
1580 | * determine the checksum of each of the gang constituents so it | |
1581 | * treats the checksum comparison differently than what we need | |
1582 | * for l2arc blocks. This prevents us from using the | |
1583 | * zio_checksum_error() interface directly. Instead we must call the | |
1584 | * zio_checksum_error_impl() so that we can ensure the checksum is | |
1585 | * generated using the correct checksum algorithm and accounts for the | |
1586 | * logical I/O size and not just a gang fragment. | |
1587 | */ | |
b5256303 | 1588 | return (zio_checksum_error_impl(zio->io_spa, zio->io_bp, |
a6255b7f | 1589 | BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size, |
d3c2ae1c | 1590 | zio->io_offset, NULL) == 0); |
34dc7c2f BB |
1591 | } |
1592 | ||
524b4217 DK |
1593 | /* |
1594 | * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a | |
1595 | * checksum and attaches it to the buf's hdr so that we can ensure that the buf | |
1596 | * isn't modified later on. If buf is compressed or there is already a checksum | |
1597 | * on the hdr, this is a no-op (we only checksum uncompressed bufs). | |
1598 | */ | |
34dc7c2f | 1599 | static void |
d3c2ae1c | 1600 | arc_cksum_compute(arc_buf_t *buf) |
34dc7c2f | 1601 | { |
d3c2ae1c GW |
1602 | arc_buf_hdr_t *hdr = buf->b_hdr; |
1603 | ||
1604 | if (!(zfs_flags & ZFS_DEBUG_MODIFY)) | |
34dc7c2f BB |
1605 | return; |
1606 | ||
d3c2ae1c | 1607 | ASSERT(HDR_HAS_L1HDR(hdr)); |
2aa34383 | 1608 | |
b9541d6b | 1609 | mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); |
d3c2ae1c | 1610 | if (hdr->b_l1hdr.b_freeze_cksum != NULL) { |
a7004725 | 1611 | ASSERT(arc_hdr_has_uncompressed_buf(hdr)); |
2aa34383 DK |
1612 | mutex_exit(&hdr->b_l1hdr.b_freeze_lock); |
1613 | return; | |
1614 | } else if (ARC_BUF_COMPRESSED(buf)) { | |
d3c2ae1c | 1615 | mutex_exit(&hdr->b_l1hdr.b_freeze_lock); |
34dc7c2f BB |
1616 | return; |
1617 | } | |
2aa34383 | 1618 | |
b5256303 | 1619 | ASSERT(!ARC_BUF_ENCRYPTED(buf)); |
2aa34383 | 1620 | ASSERT(!ARC_BUF_COMPRESSED(buf)); |
d3c2ae1c GW |
1621 | hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), |
1622 | KM_SLEEP); | |
3c67d83a | 1623 | fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, |
d3c2ae1c GW |
1624 | hdr->b_l1hdr.b_freeze_cksum); |
1625 | mutex_exit(&hdr->b_l1hdr.b_freeze_lock); | |
498877ba MA |
1626 | arc_buf_watch(buf); |
1627 | } | |
1628 | ||
1629 | #ifndef _KERNEL | |
1630 | void | |
1631 | arc_buf_sigsegv(int sig, siginfo_t *si, void *unused) | |
1632 | { | |
02730c33 | 1633 | panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr); |
498877ba MA |
1634 | } |
1635 | #endif | |
1636 | ||
1637 | /* ARGSUSED */ | |
1638 | static void | |
1639 | arc_buf_unwatch(arc_buf_t *buf) | |
1640 | { | |
1641 | #ifndef _KERNEL | |
1642 | if (arc_watch) { | |
a7004725 | 1643 | ASSERT0(mprotect(buf->b_data, arc_buf_size(buf), |
498877ba MA |
1644 | PROT_READ | PROT_WRITE)); |
1645 | } | |
1646 | #endif | |
1647 | } | |
1648 | ||
1649 | /* ARGSUSED */ | |
1650 | static void | |
1651 | arc_buf_watch(arc_buf_t *buf) | |
1652 | { | |
1653 | #ifndef _KERNEL | |
1654 | if (arc_watch) | |
2aa34383 | 1655 | ASSERT0(mprotect(buf->b_data, arc_buf_size(buf), |
d3c2ae1c | 1656 | PROT_READ)); |
498877ba | 1657 | #endif |
34dc7c2f BB |
1658 | } |
1659 | ||
b9541d6b CW |
1660 | static arc_buf_contents_t |
1661 | arc_buf_type(arc_buf_hdr_t *hdr) | |
1662 | { | |
d3c2ae1c | 1663 | arc_buf_contents_t type; |
b9541d6b | 1664 | if (HDR_ISTYPE_METADATA(hdr)) { |
d3c2ae1c | 1665 | type = ARC_BUFC_METADATA; |
b9541d6b | 1666 | } else { |
d3c2ae1c | 1667 | type = ARC_BUFC_DATA; |
b9541d6b | 1668 | } |
d3c2ae1c GW |
1669 | VERIFY3U(hdr->b_type, ==, type); |
1670 | return (type); | |
b9541d6b CW |
1671 | } |
1672 | ||
2aa34383 DK |
1673 | boolean_t |
1674 | arc_is_metadata(arc_buf_t *buf) | |
1675 | { | |
1676 | return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); | |
1677 | } | |
1678 | ||
b9541d6b CW |
1679 | static uint32_t |
1680 | arc_bufc_to_flags(arc_buf_contents_t type) | |
1681 | { | |
1682 | switch (type) { | |
1683 | case ARC_BUFC_DATA: | |
1684 | /* metadata field is 0 if buffer contains normal data */ | |
1685 | return (0); | |
1686 | case ARC_BUFC_METADATA: | |
1687 | return (ARC_FLAG_BUFC_METADATA); | |
1688 | default: | |
1689 | break; | |
1690 | } | |
1691 | panic("undefined ARC buffer type!"); | |
1692 | return ((uint32_t)-1); | |
1693 | } | |
1694 | ||
34dc7c2f BB |
1695 | void |
1696 | arc_buf_thaw(arc_buf_t *buf) | |
1697 | { | |
d3c2ae1c GW |
1698 | arc_buf_hdr_t *hdr = buf->b_hdr; |
1699 | ||
2aa34383 DK |
1700 | ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); |
1701 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
1702 | ||
524b4217 | 1703 | arc_cksum_verify(buf); |
34dc7c2f | 1704 | |
2aa34383 DK |
1705 | /* |
1706 | * Compressed buffers do not manipulate the b_freeze_cksum or | |
1707 | * allocate b_thawed. | |
1708 | */ | |
1709 | if (ARC_BUF_COMPRESSED(buf)) { | |
a7004725 DK |
1710 | ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || |
1711 | arc_hdr_has_uncompressed_buf(hdr)); | |
2aa34383 DK |
1712 | return; |
1713 | } | |
1714 | ||
d3c2ae1c GW |
1715 | ASSERT(HDR_HAS_L1HDR(hdr)); |
1716 | arc_cksum_free(hdr); | |
498877ba | 1717 | arc_buf_unwatch(buf); |
34dc7c2f BB |
1718 | } |
1719 | ||
1720 | void | |
1721 | arc_buf_freeze(arc_buf_t *buf) | |
1722 | { | |
d3c2ae1c | 1723 | arc_buf_hdr_t *hdr = buf->b_hdr; |
428870ff BB |
1724 | kmutex_t *hash_lock; |
1725 | ||
34dc7c2f BB |
1726 | if (!(zfs_flags & ZFS_DEBUG_MODIFY)) |
1727 | return; | |
1728 | ||
2aa34383 | 1729 | if (ARC_BUF_COMPRESSED(buf)) { |
a7004725 DK |
1730 | ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || |
1731 | arc_hdr_has_uncompressed_buf(hdr)); | |
2aa34383 DK |
1732 | return; |
1733 | } | |
1734 | ||
d3c2ae1c | 1735 | hash_lock = HDR_LOCK(hdr); |
428870ff BB |
1736 | mutex_enter(hash_lock); |
1737 | ||
d3c2ae1c GW |
1738 | ASSERT(HDR_HAS_L1HDR(hdr)); |
1739 | ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || | |
1740 | hdr->b_l1hdr.b_state == arc_anon); | |
1741 | arc_cksum_compute(buf); | |
428870ff | 1742 | mutex_exit(hash_lock); |
34dc7c2f BB |
1743 | } |
1744 | ||
d3c2ae1c GW |
1745 | /* |
1746 | * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, | |
1747 | * the following functions should be used to ensure that the flags are | |
1748 | * updated in a thread-safe way. When manipulating the flags either | |
1749 | * the hash_lock must be held or the hdr must be undiscoverable. This | |
1750 | * ensures that we're not racing with any other threads when updating | |
1751 | * the flags. | |
1752 | */ | |
1753 | static inline void | |
1754 | arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) | |
1755 | { | |
1756 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); | |
1757 | hdr->b_flags |= flags; | |
1758 | } | |
1759 | ||
1760 | static inline void | |
1761 | arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) | |
1762 | { | |
1763 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); | |
1764 | hdr->b_flags &= ~flags; | |
1765 | } | |
1766 | ||
1767 | /* | |
1768 | * Setting the compression bits in the arc_buf_hdr_t's b_flags is | |
1769 | * done in a special way since we have to clear and set bits | |
1770 | * at the same time. Consumers that wish to set the compression bits | |
1771 | * must use this function to ensure that the flags are updated in | |
1772 | * thread-safe manner. | |
1773 | */ | |
1774 | static void | |
1775 | arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) | |
1776 | { | |
1777 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); | |
1778 | ||
1779 | /* | |
1780 | * Holes and embedded blocks will always have a psize = 0 so | |
1781 | * we ignore the compression of the blkptr and set the | |
d3c2ae1c GW |
1782 | * want to uncompress them. Mark them as uncompressed. |
1783 | */ | |
1784 | if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { | |
1785 | arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); | |
d3c2ae1c | 1786 | ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); |
d3c2ae1c GW |
1787 | } else { |
1788 | arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); | |
d3c2ae1c GW |
1789 | ASSERT(HDR_COMPRESSION_ENABLED(hdr)); |
1790 | } | |
b5256303 TC |
1791 | |
1792 | HDR_SET_COMPRESS(hdr, cmp); | |
1793 | ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); | |
d3c2ae1c GW |
1794 | } |
1795 | ||
524b4217 DK |
1796 | /* |
1797 | * Looks for another buf on the same hdr which has the data decompressed, copies | |
1798 | * from it, and returns true. If no such buf exists, returns false. | |
1799 | */ | |
1800 | static boolean_t | |
1801 | arc_buf_try_copy_decompressed_data(arc_buf_t *buf) | |
1802 | { | |
1803 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
524b4217 DK |
1804 | boolean_t copied = B_FALSE; |
1805 | ||
1806 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
1807 | ASSERT3P(buf->b_data, !=, NULL); | |
1808 | ASSERT(!ARC_BUF_COMPRESSED(buf)); | |
1809 | ||
a7004725 | 1810 | for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; |
524b4217 DK |
1811 | from = from->b_next) { |
1812 | /* can't use our own data buffer */ | |
1813 | if (from == buf) { | |
1814 | continue; | |
1815 | } | |
1816 | ||
1817 | if (!ARC_BUF_COMPRESSED(from)) { | |
1818 | bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); | |
1819 | copied = B_TRUE; | |
1820 | break; | |
1821 | } | |
1822 | } | |
1823 | ||
1824 | /* | |
1825 | * There were no decompressed bufs, so there should not be a | |
1826 | * checksum on the hdr either. | |
1827 | */ | |
1828 | EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); | |
1829 | ||
1830 | return (copied); | |
1831 | } | |
1832 | ||
b5256303 TC |
1833 | /* |
1834 | * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t. | |
1835 | */ | |
1836 | static uint64_t | |
1837 | arc_hdr_size(arc_buf_hdr_t *hdr) | |
1838 | { | |
1839 | uint64_t size; | |
1840 | ||
1841 | if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF && | |
1842 | HDR_GET_PSIZE(hdr) > 0) { | |
1843 | size = HDR_GET_PSIZE(hdr); | |
1844 | } else { | |
1845 | ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); | |
1846 | size = HDR_GET_LSIZE(hdr); | |
1847 | } | |
1848 | return (size); | |
1849 | } | |
1850 | ||
1851 | static int | |
1852 | arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj) | |
1853 | { | |
1854 | int ret; | |
1855 | uint64_t csize; | |
1856 | uint64_t lsize = HDR_GET_LSIZE(hdr); | |
1857 | uint64_t psize = HDR_GET_PSIZE(hdr); | |
1858 | void *tmpbuf = NULL; | |
1859 | abd_t *abd = hdr->b_l1hdr.b_pabd; | |
1860 | ||
1861 | ASSERT(HDR_LOCK(hdr) == NULL || MUTEX_HELD(HDR_LOCK(hdr))); | |
1862 | ASSERT(HDR_AUTHENTICATED(hdr)); | |
1863 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); | |
1864 | ||
1865 | /* | |
1866 | * The MAC is calculated on the compressed data that is stored on disk. | |
1867 | * However, if compressed arc is disabled we will only have the | |
1868 | * decompressed data available to us now. Compress it into a temporary | |
1869 | * abd so we can verify the MAC. The performance overhead of this will | |
1870 | * be relatively low, since most objects in an encrypted objset will | |
1871 | * be encrypted (instead of authenticated) anyway. | |
1872 | */ | |
1873 | if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && | |
1874 | !HDR_COMPRESSION_ENABLED(hdr)) { | |
1875 | tmpbuf = zio_buf_alloc(lsize); | |
1876 | abd = abd_get_from_buf(tmpbuf, lsize); | |
1877 | abd_take_ownership_of_buf(abd, B_TRUE); | |
1878 | ||
1879 | csize = zio_compress_data(HDR_GET_COMPRESS(hdr), | |
1880 | hdr->b_l1hdr.b_pabd, tmpbuf, lsize); | |
1881 | ASSERT3U(csize, <=, psize); | |
1882 | abd_zero_off(abd, csize, psize - csize); | |
1883 | } | |
1884 | ||
1885 | /* | |
1886 | * Authentication is best effort. We authenticate whenever the key is | |
1887 | * available. If we succeed we clear ARC_FLAG_NOAUTH. | |
1888 | */ | |
1889 | if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) { | |
1890 | ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); | |
1891 | ASSERT3U(lsize, ==, psize); | |
1892 | ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd, | |
1893 | psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS); | |
1894 | } else { | |
1895 | ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize, | |
1896 | hdr->b_crypt_hdr.b_mac); | |
1897 | } | |
1898 | ||
1899 | if (ret == 0) | |
1900 | arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH); | |
1901 | else if (ret != ENOENT) | |
1902 | goto error; | |
1903 | ||
1904 | if (tmpbuf != NULL) | |
1905 | abd_free(abd); | |
1906 | ||
1907 | return (0); | |
1908 | ||
1909 | error: | |
1910 | if (tmpbuf != NULL) | |
1911 | abd_free(abd); | |
1912 | ||
1913 | return (ret); | |
1914 | } | |
1915 | ||
1916 | /* | |
1917 | * This function will take a header that only has raw encrypted data in | |
1918 | * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in | |
1919 | * b_l1hdr.b_pabd. If designated in the header flags, this function will | |
1920 | * also decompress the data. | |
1921 | */ | |
1922 | static int | |
be9a5c35 | 1923 | arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb) |
b5256303 TC |
1924 | { |
1925 | int ret; | |
b5256303 TC |
1926 | abd_t *cabd = NULL; |
1927 | void *tmp = NULL; | |
1928 | boolean_t no_crypt = B_FALSE; | |
1929 | boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS); | |
1930 | ||
1931 | ASSERT(HDR_LOCK(hdr) == NULL || MUTEX_HELD(HDR_LOCK(hdr))); | |
1932 | ASSERT(HDR_ENCRYPTED(hdr)); | |
1933 | ||
1934 | arc_hdr_alloc_abd(hdr, B_FALSE); | |
1935 | ||
be9a5c35 TC |
1936 | ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot, |
1937 | B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv, | |
1938 | hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd, | |
b5256303 TC |
1939 | hdr->b_crypt_hdr.b_rabd, &no_crypt); |
1940 | if (ret != 0) | |
1941 | goto error; | |
1942 | ||
1943 | if (no_crypt) { | |
1944 | abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd, | |
1945 | HDR_GET_PSIZE(hdr)); | |
1946 | } | |
1947 | ||
1948 | /* | |
1949 | * If this header has disabled arc compression but the b_pabd is | |
1950 | * compressed after decrypting it, we need to decompress the newly | |
1951 | * decrypted data. | |
1952 | */ | |
1953 | if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && | |
1954 | !HDR_COMPRESSION_ENABLED(hdr)) { | |
1955 | /* | |
1956 | * We want to make sure that we are correctly honoring the | |
1957 | * zfs_abd_scatter_enabled setting, so we allocate an abd here | |
1958 | * and then loan a buffer from it, rather than allocating a | |
1959 | * linear buffer and wrapping it in an abd later. | |
1960 | */ | |
1961 | cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); | |
1962 | tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr)); | |
1963 | ||
1964 | ret = zio_decompress_data(HDR_GET_COMPRESS(hdr), | |
1965 | hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr), | |
1966 | HDR_GET_LSIZE(hdr)); | |
1967 | if (ret != 0) { | |
1968 | abd_return_buf(cabd, tmp, arc_hdr_size(hdr)); | |
1969 | goto error; | |
1970 | } | |
1971 | ||
1972 | abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr)); | |
1973 | arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, | |
1974 | arc_hdr_size(hdr), hdr); | |
1975 | hdr->b_l1hdr.b_pabd = cabd; | |
1976 | } | |
1977 | ||
b5256303 TC |
1978 | return (0); |
1979 | ||
1980 | error: | |
1981 | arc_hdr_free_abd(hdr, B_FALSE); | |
b5256303 TC |
1982 | if (cabd != NULL) |
1983 | arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr); | |
1984 | ||
1985 | return (ret); | |
1986 | } | |
1987 | ||
1988 | /* | |
1989 | * This function is called during arc_buf_fill() to prepare the header's | |
1990 | * abd plaintext pointer for use. This involves authenticated protected | |
1991 | * data and decrypting encrypted data into the plaintext abd. | |
1992 | */ | |
1993 | static int | |
1994 | arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa, | |
be9a5c35 | 1995 | const zbookmark_phys_t *zb, boolean_t noauth) |
b5256303 TC |
1996 | { |
1997 | int ret; | |
1998 | ||
1999 | ASSERT(HDR_PROTECTED(hdr)); | |
2000 | ||
2001 | if (hash_lock != NULL) | |
2002 | mutex_enter(hash_lock); | |
2003 | ||
2004 | if (HDR_NOAUTH(hdr) && !noauth) { | |
2005 | /* | |
2006 | * The caller requested authenticated data but our data has | |
2007 | * not been authenticated yet. Verify the MAC now if we can. | |
2008 | */ | |
be9a5c35 | 2009 | ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset); |
b5256303 TC |
2010 | if (ret != 0) |
2011 | goto error; | |
2012 | } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) { | |
2013 | /* | |
2014 | * If we only have the encrypted version of the data, but the | |
2015 | * unencrypted version was requested we take this opportunity | |
2016 | * to store the decrypted version in the header for future use. | |
2017 | */ | |
be9a5c35 | 2018 | ret = arc_hdr_decrypt(hdr, spa, zb); |
b5256303 TC |
2019 | if (ret != 0) |
2020 | goto error; | |
2021 | } | |
2022 | ||
2023 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); | |
2024 | ||
2025 | if (hash_lock != NULL) | |
2026 | mutex_exit(hash_lock); | |
2027 | ||
2028 | return (0); | |
2029 | ||
2030 | error: | |
2031 | if (hash_lock != NULL) | |
2032 | mutex_exit(hash_lock); | |
2033 | ||
2034 | return (ret); | |
2035 | } | |
2036 | ||
2037 | /* | |
2038 | * This function is used by the dbuf code to decrypt bonus buffers in place. | |
2039 | * The dbuf code itself doesn't have any locking for decrypting a shared dnode | |
2040 | * block, so we use the hash lock here to protect against concurrent calls to | |
2041 | * arc_buf_fill(). | |
2042 | */ | |
2043 | static void | |
2044 | arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock) | |
2045 | { | |
2046 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
2047 | ||
2048 | ASSERT(HDR_ENCRYPTED(hdr)); | |
2049 | ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE); | |
2050 | ASSERT(HDR_LOCK(hdr) == NULL || MUTEX_HELD(HDR_LOCK(hdr))); | |
2051 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); | |
2052 | ||
2053 | zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data, | |
2054 | arc_buf_size(buf)); | |
2055 | buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED; | |
2056 | buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; | |
2057 | hdr->b_crypt_hdr.b_ebufcnt -= 1; | |
2058 | } | |
2059 | ||
524b4217 DK |
2060 | /* |
2061 | * Given a buf that has a data buffer attached to it, this function will | |
2062 | * efficiently fill the buf with data of the specified compression setting from | |
2063 | * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr | |
2064 | * are already sharing a data buf, no copy is performed. | |
2065 | * | |
2066 | * If the buf is marked as compressed but uncompressed data was requested, this | |
2067 | * will allocate a new data buffer for the buf, remove that flag, and fill the | |
2068 | * buf with uncompressed data. You can't request a compressed buf on a hdr with | |
2069 | * uncompressed data, and (since we haven't added support for it yet) if you | |
2070 | * want compressed data your buf must already be marked as compressed and have | |
2071 | * the correct-sized data buffer. | |
2072 | */ | |
2073 | static int | |
be9a5c35 TC |
2074 | arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb, |
2075 | arc_fill_flags_t flags) | |
d3c2ae1c | 2076 | { |
b5256303 | 2077 | int error = 0; |
d3c2ae1c | 2078 | arc_buf_hdr_t *hdr = buf->b_hdr; |
b5256303 TC |
2079 | boolean_t hdr_compressed = |
2080 | (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF); | |
2081 | boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0; | |
2082 | boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0; | |
d3c2ae1c | 2083 | dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; |
b5256303 | 2084 | kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr); |
d3c2ae1c | 2085 | |
524b4217 | 2086 | ASSERT3P(buf->b_data, !=, NULL); |
b5256303 | 2087 | IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf)); |
524b4217 | 2088 | IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); |
b5256303 TC |
2089 | IMPLY(encrypted, HDR_ENCRYPTED(hdr)); |
2090 | IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf)); | |
2091 | IMPLY(encrypted, ARC_BUF_COMPRESSED(buf)); | |
2092 | IMPLY(encrypted, !ARC_BUF_SHARED(buf)); | |
2093 | ||
2094 | /* | |
2095 | * If the caller wanted encrypted data we just need to copy it from | |
2096 | * b_rabd and potentially byteswap it. We won't be able to do any | |
2097 | * further transforms on it. | |
2098 | */ | |
2099 | if (encrypted) { | |
2100 | ASSERT(HDR_HAS_RABD(hdr)); | |
2101 | abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd, | |
2102 | HDR_GET_PSIZE(hdr)); | |
2103 | goto byteswap; | |
2104 | } | |
2105 | ||
2106 | /* | |
2107 | * Adjust encrypted and authenticated headers to accomodate the | |
2108 | * request if needed. | |
2109 | */ | |
2110 | if (HDR_PROTECTED(hdr)) { | |
2111 | error = arc_fill_hdr_crypt(hdr, hash_lock, spa, | |
be9a5c35 | 2112 | zb, !!(flags & ARC_FILL_NOAUTH)); |
2c24b5b1 TC |
2113 | if (error != 0) { |
2114 | arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); | |
b5256303 | 2115 | return (error); |
2c24b5b1 | 2116 | } |
b5256303 TC |
2117 | } |
2118 | ||
2119 | /* | |
2120 | * There is a special case here for dnode blocks which are | |
2121 | * decrypting their bonus buffers. These blocks may request to | |
2122 | * be decrypted in-place. This is necessary because there may | |
2123 | * be many dnodes pointing into this buffer and there is | |
2124 | * currently no method to synchronize replacing the backing | |
2125 | * b_data buffer and updating all of the pointers. Here we use | |
2126 | * the hash lock to ensure there are no races. If the need | |
2127 | * arises for other types to be decrypted in-place, they must | |
2128 | * add handling here as well. | |
2129 | */ | |
2130 | if ((flags & ARC_FILL_IN_PLACE) != 0) { | |
2131 | ASSERT(!hdr_compressed); | |
2132 | ASSERT(!compressed); | |
2133 | ASSERT(!encrypted); | |
2134 | ||
2135 | if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) { | |
2136 | ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE); | |
2137 | ||
2138 | if (hash_lock != NULL) | |
2139 | mutex_enter(hash_lock); | |
2140 | arc_buf_untransform_in_place(buf, hash_lock); | |
2141 | if (hash_lock != NULL) | |
2142 | mutex_exit(hash_lock); | |
2143 | ||
2144 | /* Compute the hdr's checksum if necessary */ | |
2145 | arc_cksum_compute(buf); | |
2146 | } | |
2147 | ||
2148 | return (0); | |
2149 | } | |
524b4217 DK |
2150 | |
2151 | if (hdr_compressed == compressed) { | |
2aa34383 | 2152 | if (!arc_buf_is_shared(buf)) { |
a6255b7f | 2153 | abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd, |
524b4217 | 2154 | arc_buf_size(buf)); |
2aa34383 | 2155 | } |
d3c2ae1c | 2156 | } else { |
524b4217 DK |
2157 | ASSERT(hdr_compressed); |
2158 | ASSERT(!compressed); | |
d3c2ae1c | 2159 | ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); |
2aa34383 DK |
2160 | |
2161 | /* | |
524b4217 DK |
2162 | * If the buf is sharing its data with the hdr, unlink it and |
2163 | * allocate a new data buffer for the buf. | |
2aa34383 | 2164 | */ |
524b4217 DK |
2165 | if (arc_buf_is_shared(buf)) { |
2166 | ASSERT(ARC_BUF_COMPRESSED(buf)); | |
2167 | ||
2168 | /* We need to give the buf it's own b_data */ | |
2169 | buf->b_flags &= ~ARC_BUF_FLAG_SHARED; | |
2aa34383 DK |
2170 | buf->b_data = |
2171 | arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); | |
2172 | arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); | |
2173 | ||
524b4217 | 2174 | /* Previously overhead was 0; just add new overhead */ |
2aa34383 | 2175 | ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); |
524b4217 DK |
2176 | } else if (ARC_BUF_COMPRESSED(buf)) { |
2177 | /* We need to reallocate the buf's b_data */ | |
2178 | arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), | |
2179 | buf); | |
2180 | buf->b_data = | |
2181 | arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); | |
2182 | ||
2183 | /* We increased the size of b_data; update overhead */ | |
2184 | ARCSTAT_INCR(arcstat_overhead_size, | |
2185 | HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); | |
2aa34383 DK |
2186 | } |
2187 | ||
524b4217 DK |
2188 | /* |
2189 | * Regardless of the buf's previous compression settings, it | |
2190 | * should not be compressed at the end of this function. | |
2191 | */ | |
2192 | buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; | |
2193 | ||
2194 | /* | |
2195 | * Try copying the data from another buf which already has a | |
2196 | * decompressed version. If that's not possible, it's time to | |
2197 | * bite the bullet and decompress the data from the hdr. | |
2198 | */ | |
2199 | if (arc_buf_try_copy_decompressed_data(buf)) { | |
2200 | /* Skip byteswapping and checksumming (already done) */ | |
2201 | ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL); | |
2202 | return (0); | |
2203 | } else { | |
b5256303 | 2204 | error = zio_decompress_data(HDR_GET_COMPRESS(hdr), |
a6255b7f | 2205 | hdr->b_l1hdr.b_pabd, buf->b_data, |
524b4217 DK |
2206 | HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); |
2207 | ||
2208 | /* | |
2209 | * Absent hardware errors or software bugs, this should | |
2210 | * be impossible, but log it anyway so we can debug it. | |
2211 | */ | |
2212 | if (error != 0) { | |
2213 | zfs_dbgmsg( | |
2214 | "hdr %p, compress %d, psize %d, lsize %d", | |
b5256303 | 2215 | hdr, arc_hdr_get_compress(hdr), |
524b4217 | 2216 | HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); |
2c24b5b1 | 2217 | arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); |
524b4217 DK |
2218 | return (SET_ERROR(EIO)); |
2219 | } | |
d3c2ae1c GW |
2220 | } |
2221 | } | |
524b4217 | 2222 | |
b5256303 | 2223 | byteswap: |
524b4217 | 2224 | /* Byteswap the buf's data if necessary */ |
d3c2ae1c GW |
2225 | if (bswap != DMU_BSWAP_NUMFUNCS) { |
2226 | ASSERT(!HDR_SHARED_DATA(hdr)); | |
2227 | ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); | |
2228 | dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); | |
2229 | } | |
524b4217 DK |
2230 | |
2231 | /* Compute the hdr's checksum if necessary */ | |
d3c2ae1c | 2232 | arc_cksum_compute(buf); |
524b4217 | 2233 | |
d3c2ae1c GW |
2234 | return (0); |
2235 | } | |
2236 | ||
2237 | /* | |
b5256303 TC |
2238 | * If this function is being called to decrypt an encrypted buffer or verify an |
2239 | * authenticated one, the key must be loaded and a mapping must be made | |
2240 | * available in the keystore via spa_keystore_create_mapping() or one of its | |
2241 | * callers. | |
d3c2ae1c | 2242 | */ |
b5256303 | 2243 | int |
a2c2ed1b TC |
2244 | arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb, |
2245 | boolean_t in_place) | |
d3c2ae1c | 2246 | { |
a2c2ed1b | 2247 | int ret; |
b5256303 | 2248 | arc_fill_flags_t flags = 0; |
d3c2ae1c | 2249 | |
b5256303 TC |
2250 | if (in_place) |
2251 | flags |= ARC_FILL_IN_PLACE; | |
2252 | ||
be9a5c35 | 2253 | ret = arc_buf_fill(buf, spa, zb, flags); |
a2c2ed1b TC |
2254 | if (ret == ECKSUM) { |
2255 | /* | |
2256 | * Convert authentication and decryption errors to EIO | |
2257 | * (and generate an ereport) before leaving the ARC. | |
2258 | */ | |
2259 | ret = SET_ERROR(EIO); | |
be9a5c35 | 2260 | spa_log_error(spa, zb); |
a2c2ed1b TC |
2261 | zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, |
2262 | spa, NULL, zb, NULL, 0, 0); | |
2263 | } | |
2264 | ||
2265 | return (ret); | |
d3c2ae1c GW |
2266 | } |
2267 | ||
2268 | /* | |
2269 | * Increment the amount of evictable space in the arc_state_t's refcount. | |
2270 | * We account for the space used by the hdr and the arc buf individually | |
2271 | * so that we can add and remove them from the refcount individually. | |
2272 | */ | |
34dc7c2f | 2273 | static void |
d3c2ae1c GW |
2274 | arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) |
2275 | { | |
2276 | arc_buf_contents_t type = arc_buf_type(hdr); | |
d3c2ae1c GW |
2277 | |
2278 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
2279 | ||
2280 | if (GHOST_STATE(state)) { | |
2281 | ASSERT0(hdr->b_l1hdr.b_bufcnt); | |
2282 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
a6255b7f | 2283 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 2284 | ASSERT(!HDR_HAS_RABD(hdr)); |
2aa34383 DK |
2285 | (void) refcount_add_many(&state->arcs_esize[type], |
2286 | HDR_GET_LSIZE(hdr), hdr); | |
d3c2ae1c GW |
2287 | return; |
2288 | } | |
2289 | ||
2290 | ASSERT(!GHOST_STATE(state)); | |
a6255b7f | 2291 | if (hdr->b_l1hdr.b_pabd != NULL) { |
d3c2ae1c GW |
2292 | (void) refcount_add_many(&state->arcs_esize[type], |
2293 | arc_hdr_size(hdr), hdr); | |
2294 | } | |
b5256303 TC |
2295 | if (HDR_HAS_RABD(hdr)) { |
2296 | (void) refcount_add_many(&state->arcs_esize[type], | |
2297 | HDR_GET_PSIZE(hdr), hdr); | |
2298 | } | |
2299 | ||
1c27024e DB |
2300 | for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; |
2301 | buf = buf->b_next) { | |
2aa34383 | 2302 | if (arc_buf_is_shared(buf)) |
d3c2ae1c | 2303 | continue; |
2aa34383 DK |
2304 | (void) refcount_add_many(&state->arcs_esize[type], |
2305 | arc_buf_size(buf), buf); | |
d3c2ae1c GW |
2306 | } |
2307 | } | |
2308 | ||
2309 | /* | |
2310 | * Decrement the amount of evictable space in the arc_state_t's refcount. | |
2311 | * We account for the space used by the hdr and the arc buf individually | |
2312 | * so that we can add and remove them from the refcount individually. | |
2313 | */ | |
2314 | static void | |
2aa34383 | 2315 | arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) |
d3c2ae1c GW |
2316 | { |
2317 | arc_buf_contents_t type = arc_buf_type(hdr); | |
d3c2ae1c GW |
2318 | |
2319 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
2320 | ||
2321 | if (GHOST_STATE(state)) { | |
2322 | ASSERT0(hdr->b_l1hdr.b_bufcnt); | |
2323 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
a6255b7f | 2324 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 2325 | ASSERT(!HDR_HAS_RABD(hdr)); |
d3c2ae1c | 2326 | (void) refcount_remove_many(&state->arcs_esize[type], |
2aa34383 | 2327 | HDR_GET_LSIZE(hdr), hdr); |
d3c2ae1c GW |
2328 | return; |
2329 | } | |
2330 | ||
2331 | ASSERT(!GHOST_STATE(state)); | |
a6255b7f | 2332 | if (hdr->b_l1hdr.b_pabd != NULL) { |
d3c2ae1c GW |
2333 | (void) refcount_remove_many(&state->arcs_esize[type], |
2334 | arc_hdr_size(hdr), hdr); | |
2335 | } | |
b5256303 TC |
2336 | if (HDR_HAS_RABD(hdr)) { |
2337 | (void) refcount_remove_many(&state->arcs_esize[type], | |
2338 | HDR_GET_PSIZE(hdr), hdr); | |
2339 | } | |
2340 | ||
1c27024e DB |
2341 | for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; |
2342 | buf = buf->b_next) { | |
2aa34383 | 2343 | if (arc_buf_is_shared(buf)) |
d3c2ae1c | 2344 | continue; |
d3c2ae1c | 2345 | (void) refcount_remove_many(&state->arcs_esize[type], |
2aa34383 | 2346 | arc_buf_size(buf), buf); |
d3c2ae1c GW |
2347 | } |
2348 | } | |
2349 | ||
2350 | /* | |
2351 | * Add a reference to this hdr indicating that someone is actively | |
2352 | * referencing that memory. When the refcount transitions from 0 to 1, | |
2353 | * we remove it from the respective arc_state_t list to indicate that | |
2354 | * it is not evictable. | |
2355 | */ | |
2356 | static void | |
2357 | add_reference(arc_buf_hdr_t *hdr, void *tag) | |
34dc7c2f | 2358 | { |
b9541d6b CW |
2359 | arc_state_t *state; |
2360 | ||
2361 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
d3c2ae1c GW |
2362 | if (!MUTEX_HELD(HDR_LOCK(hdr))) { |
2363 | ASSERT(hdr->b_l1hdr.b_state == arc_anon); | |
2364 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); | |
2365 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
2366 | } | |
34dc7c2f | 2367 | |
b9541d6b CW |
2368 | state = hdr->b_l1hdr.b_state; |
2369 | ||
2370 | if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && | |
2371 | (state != arc_anon)) { | |
2372 | /* We don't use the L2-only state list. */ | |
2373 | if (state != arc_l2c_only) { | |
64fc7762 | 2374 | multilist_remove(state->arcs_list[arc_buf_type(hdr)], |
d3c2ae1c | 2375 | hdr); |
2aa34383 | 2376 | arc_evictable_space_decrement(hdr, state); |
34dc7c2f | 2377 | } |
b128c09f | 2378 | /* remove the prefetch flag if we get a reference */ |
d3c2ae1c | 2379 | arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); |
34dc7c2f BB |
2380 | } |
2381 | } | |
2382 | ||
d3c2ae1c GW |
2383 | /* |
2384 | * Remove a reference from this hdr. When the reference transitions from | |
2385 | * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's | |
2386 | * list making it eligible for eviction. | |
2387 | */ | |
34dc7c2f | 2388 | static int |
2a432414 | 2389 | remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) |
34dc7c2f BB |
2390 | { |
2391 | int cnt; | |
b9541d6b | 2392 | arc_state_t *state = hdr->b_l1hdr.b_state; |
34dc7c2f | 2393 | |
b9541d6b | 2394 | ASSERT(HDR_HAS_L1HDR(hdr)); |
34dc7c2f BB |
2395 | ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); |
2396 | ASSERT(!GHOST_STATE(state)); | |
2397 | ||
b9541d6b CW |
2398 | /* |
2399 | * arc_l2c_only counts as a ghost state so we don't need to explicitly | |
2400 | * check to prevent usage of the arc_l2c_only list. | |
2401 | */ | |
2402 | if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && | |
34dc7c2f | 2403 | (state != arc_anon)) { |
64fc7762 | 2404 | multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr); |
d3c2ae1c GW |
2405 | ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); |
2406 | arc_evictable_space_increment(hdr, state); | |
34dc7c2f BB |
2407 | } |
2408 | return (cnt); | |
2409 | } | |
2410 | ||
e0b0ca98 BB |
2411 | /* |
2412 | * Returns detailed information about a specific arc buffer. When the | |
2413 | * state_index argument is set the function will calculate the arc header | |
2414 | * list position for its arc state. Since this requires a linear traversal | |
2415 | * callers are strongly encourage not to do this. However, it can be helpful | |
2416 | * for targeted analysis so the functionality is provided. | |
2417 | */ | |
2418 | void | |
2419 | arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index) | |
2420 | { | |
2421 | arc_buf_hdr_t *hdr = ab->b_hdr; | |
b9541d6b CW |
2422 | l1arc_buf_hdr_t *l1hdr = NULL; |
2423 | l2arc_buf_hdr_t *l2hdr = NULL; | |
2424 | arc_state_t *state = NULL; | |
2425 | ||
8887c7d7 TC |
2426 | memset(abi, 0, sizeof (arc_buf_info_t)); |
2427 | ||
2428 | if (hdr == NULL) | |
2429 | return; | |
2430 | ||
2431 | abi->abi_flags = hdr->b_flags; | |
2432 | ||
b9541d6b CW |
2433 | if (HDR_HAS_L1HDR(hdr)) { |
2434 | l1hdr = &hdr->b_l1hdr; | |
2435 | state = l1hdr->b_state; | |
2436 | } | |
2437 | if (HDR_HAS_L2HDR(hdr)) | |
2438 | l2hdr = &hdr->b_l2hdr; | |
e0b0ca98 | 2439 | |
b9541d6b | 2440 | if (l1hdr) { |
d3c2ae1c | 2441 | abi->abi_bufcnt = l1hdr->b_bufcnt; |
b9541d6b CW |
2442 | abi->abi_access = l1hdr->b_arc_access; |
2443 | abi->abi_mru_hits = l1hdr->b_mru_hits; | |
2444 | abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits; | |
2445 | abi->abi_mfu_hits = l1hdr->b_mfu_hits; | |
2446 | abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits; | |
2447 | abi->abi_holds = refcount_count(&l1hdr->b_refcnt); | |
2448 | } | |
2449 | ||
2450 | if (l2hdr) { | |
2451 | abi->abi_l2arc_dattr = l2hdr->b_daddr; | |
b9541d6b CW |
2452 | abi->abi_l2arc_hits = l2hdr->b_hits; |
2453 | } | |
2454 | ||
e0b0ca98 | 2455 | abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON; |
b9541d6b | 2456 | abi->abi_state_contents = arc_buf_type(hdr); |
d3c2ae1c | 2457 | abi->abi_size = arc_hdr_size(hdr); |
e0b0ca98 BB |
2458 | } |
2459 | ||
34dc7c2f | 2460 | /* |
ca0bf58d | 2461 | * Move the supplied buffer to the indicated state. The hash lock |
34dc7c2f BB |
2462 | * for the buffer must be held by the caller. |
2463 | */ | |
2464 | static void | |
2a432414 GW |
2465 | arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, |
2466 | kmutex_t *hash_lock) | |
34dc7c2f | 2467 | { |
b9541d6b CW |
2468 | arc_state_t *old_state; |
2469 | int64_t refcnt; | |
d3c2ae1c GW |
2470 | uint32_t bufcnt; |
2471 | boolean_t update_old, update_new; | |
b9541d6b CW |
2472 | arc_buf_contents_t buftype = arc_buf_type(hdr); |
2473 | ||
2474 | /* | |
2475 | * We almost always have an L1 hdr here, since we call arc_hdr_realloc() | |
2476 | * in arc_read() when bringing a buffer out of the L2ARC. However, the | |
2477 | * L1 hdr doesn't always exist when we change state to arc_anon before | |
2478 | * destroying a header, in which case reallocating to add the L1 hdr is | |
2479 | * pointless. | |
2480 | */ | |
2481 | if (HDR_HAS_L1HDR(hdr)) { | |
2482 | old_state = hdr->b_l1hdr.b_state; | |
2483 | refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); | |
d3c2ae1c | 2484 | bufcnt = hdr->b_l1hdr.b_bufcnt; |
b5256303 TC |
2485 | update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL || |
2486 | HDR_HAS_RABD(hdr)); | |
b9541d6b CW |
2487 | } else { |
2488 | old_state = arc_l2c_only; | |
2489 | refcnt = 0; | |
d3c2ae1c GW |
2490 | bufcnt = 0; |
2491 | update_old = B_FALSE; | |
b9541d6b | 2492 | } |
d3c2ae1c | 2493 | update_new = update_old; |
34dc7c2f BB |
2494 | |
2495 | ASSERT(MUTEX_HELD(hash_lock)); | |
e8b96c60 | 2496 | ASSERT3P(new_state, !=, old_state); |
d3c2ae1c GW |
2497 | ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); |
2498 | ASSERT(old_state != arc_anon || bufcnt <= 1); | |
34dc7c2f BB |
2499 | |
2500 | /* | |
2501 | * If this buffer is evictable, transfer it from the | |
2502 | * old state list to the new state list. | |
2503 | */ | |
2504 | if (refcnt == 0) { | |
b9541d6b | 2505 | if (old_state != arc_anon && old_state != arc_l2c_only) { |
b9541d6b | 2506 | ASSERT(HDR_HAS_L1HDR(hdr)); |
64fc7762 | 2507 | multilist_remove(old_state->arcs_list[buftype], hdr); |
34dc7c2f | 2508 | |
d3c2ae1c GW |
2509 | if (GHOST_STATE(old_state)) { |
2510 | ASSERT0(bufcnt); | |
2511 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
2512 | update_old = B_TRUE; | |
34dc7c2f | 2513 | } |
2aa34383 | 2514 | arc_evictable_space_decrement(hdr, old_state); |
34dc7c2f | 2515 | } |
b9541d6b | 2516 | if (new_state != arc_anon && new_state != arc_l2c_only) { |
b9541d6b CW |
2517 | /* |
2518 | * An L1 header always exists here, since if we're | |
2519 | * moving to some L1-cached state (i.e. not l2c_only or | |
2520 | * anonymous), we realloc the header to add an L1hdr | |
2521 | * beforehand. | |
2522 | */ | |
2523 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
64fc7762 | 2524 | multilist_insert(new_state->arcs_list[buftype], hdr); |
34dc7c2f | 2525 | |
34dc7c2f | 2526 | if (GHOST_STATE(new_state)) { |
d3c2ae1c GW |
2527 | ASSERT0(bufcnt); |
2528 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
2529 | update_new = B_TRUE; | |
34dc7c2f | 2530 | } |
d3c2ae1c | 2531 | arc_evictable_space_increment(hdr, new_state); |
34dc7c2f BB |
2532 | } |
2533 | } | |
2534 | ||
d3c2ae1c | 2535 | ASSERT(!HDR_EMPTY(hdr)); |
2a432414 GW |
2536 | if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) |
2537 | buf_hash_remove(hdr); | |
34dc7c2f | 2538 | |
b9541d6b | 2539 | /* adjust state sizes (ignore arc_l2c_only) */ |
36da08ef | 2540 | |
d3c2ae1c | 2541 | if (update_new && new_state != arc_l2c_only) { |
36da08ef PS |
2542 | ASSERT(HDR_HAS_L1HDR(hdr)); |
2543 | if (GHOST_STATE(new_state)) { | |
d3c2ae1c | 2544 | ASSERT0(bufcnt); |
36da08ef PS |
2545 | |
2546 | /* | |
d3c2ae1c | 2547 | * When moving a header to a ghost state, we first |
36da08ef | 2548 | * remove all arc buffers. Thus, we'll have a |
d3c2ae1c | 2549 | * bufcnt of zero, and no arc buffer to use for |
36da08ef PS |
2550 | * the reference. As a result, we use the arc |
2551 | * header pointer for the reference. | |
2552 | */ | |
2553 | (void) refcount_add_many(&new_state->arcs_size, | |
d3c2ae1c | 2554 | HDR_GET_LSIZE(hdr), hdr); |
a6255b7f | 2555 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 2556 | ASSERT(!HDR_HAS_RABD(hdr)); |
36da08ef | 2557 | } else { |
d3c2ae1c | 2558 | uint32_t buffers = 0; |
36da08ef PS |
2559 | |
2560 | /* | |
2561 | * Each individual buffer holds a unique reference, | |
2562 | * thus we must remove each of these references one | |
2563 | * at a time. | |
2564 | */ | |
1c27024e | 2565 | for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; |
36da08ef | 2566 | buf = buf->b_next) { |
d3c2ae1c GW |
2567 | ASSERT3U(bufcnt, !=, 0); |
2568 | buffers++; | |
2569 | ||
2570 | /* | |
2571 | * When the arc_buf_t is sharing the data | |
2572 | * block with the hdr, the owner of the | |
2573 | * reference belongs to the hdr. Only | |
2574 | * add to the refcount if the arc_buf_t is | |
2575 | * not shared. | |
2576 | */ | |
2aa34383 | 2577 | if (arc_buf_is_shared(buf)) |
d3c2ae1c | 2578 | continue; |
d3c2ae1c | 2579 | |
36da08ef | 2580 | (void) refcount_add_many(&new_state->arcs_size, |
2aa34383 | 2581 | arc_buf_size(buf), buf); |
d3c2ae1c GW |
2582 | } |
2583 | ASSERT3U(bufcnt, ==, buffers); | |
2584 | ||
a6255b7f | 2585 | if (hdr->b_l1hdr.b_pabd != NULL) { |
d3c2ae1c GW |
2586 | (void) refcount_add_many(&new_state->arcs_size, |
2587 | arc_hdr_size(hdr), hdr); | |
b5256303 TC |
2588 | } |
2589 | ||
2590 | if (HDR_HAS_RABD(hdr)) { | |
2591 | (void) refcount_add_many(&new_state->arcs_size, | |
2592 | HDR_GET_PSIZE(hdr), hdr); | |
36da08ef PS |
2593 | } |
2594 | } | |
2595 | } | |
2596 | ||
d3c2ae1c | 2597 | if (update_old && old_state != arc_l2c_only) { |
36da08ef PS |
2598 | ASSERT(HDR_HAS_L1HDR(hdr)); |
2599 | if (GHOST_STATE(old_state)) { | |
d3c2ae1c | 2600 | ASSERT0(bufcnt); |
a6255b7f | 2601 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 2602 | ASSERT(!HDR_HAS_RABD(hdr)); |
d3c2ae1c | 2603 | |
36da08ef PS |
2604 | /* |
2605 | * When moving a header off of a ghost state, | |
d3c2ae1c GW |
2606 | * the header will not contain any arc buffers. |
2607 | * We use the arc header pointer for the reference | |
2608 | * which is exactly what we did when we put the | |
2609 | * header on the ghost state. | |
36da08ef PS |
2610 | */ |
2611 | ||
36da08ef | 2612 | (void) refcount_remove_many(&old_state->arcs_size, |
d3c2ae1c | 2613 | HDR_GET_LSIZE(hdr), hdr); |
36da08ef | 2614 | } else { |
d3c2ae1c | 2615 | uint32_t buffers = 0; |
36da08ef PS |
2616 | |
2617 | /* | |
2618 | * Each individual buffer holds a unique reference, | |
2619 | * thus we must remove each of these references one | |
2620 | * at a time. | |
2621 | */ | |
1c27024e | 2622 | for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; |
36da08ef | 2623 | buf = buf->b_next) { |
d3c2ae1c GW |
2624 | ASSERT3U(bufcnt, !=, 0); |
2625 | buffers++; | |
2626 | ||
2627 | /* | |
2628 | * When the arc_buf_t is sharing the data | |
2629 | * block with the hdr, the owner of the | |
2630 | * reference belongs to the hdr. Only | |
2631 | * add to the refcount if the arc_buf_t is | |
2632 | * not shared. | |
2633 | */ | |
2aa34383 | 2634 | if (arc_buf_is_shared(buf)) |
d3c2ae1c | 2635 | continue; |
d3c2ae1c | 2636 | |
36da08ef | 2637 | (void) refcount_remove_many( |
2aa34383 | 2638 | &old_state->arcs_size, arc_buf_size(buf), |
d3c2ae1c | 2639 | buf); |
36da08ef | 2640 | } |
d3c2ae1c | 2641 | ASSERT3U(bufcnt, ==, buffers); |
b5256303 TC |
2642 | ASSERT(hdr->b_l1hdr.b_pabd != NULL || |
2643 | HDR_HAS_RABD(hdr)); | |
2644 | ||
2645 | if (hdr->b_l1hdr.b_pabd != NULL) { | |
2646 | (void) refcount_remove_many( | |
2647 | &old_state->arcs_size, arc_hdr_size(hdr), | |
2648 | hdr); | |
2649 | } | |
2650 | ||
2651 | if (HDR_HAS_RABD(hdr)) { | |
2652 | (void) refcount_remove_many( | |
2653 | &old_state->arcs_size, HDR_GET_PSIZE(hdr), | |
2654 | hdr); | |
2655 | } | |
36da08ef | 2656 | } |
34dc7c2f | 2657 | } |
36da08ef | 2658 | |
b9541d6b CW |
2659 | if (HDR_HAS_L1HDR(hdr)) |
2660 | hdr->b_l1hdr.b_state = new_state; | |
34dc7c2f | 2661 | |
b9541d6b CW |
2662 | /* |
2663 | * L2 headers should never be on the L2 state list since they don't | |
2664 | * have L1 headers allocated. | |
2665 | */ | |
64fc7762 MA |
2666 | ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && |
2667 | multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); | |
34dc7c2f BB |
2668 | } |
2669 | ||
2670 | void | |
d164b209 | 2671 | arc_space_consume(uint64_t space, arc_space_type_t type) |
34dc7c2f | 2672 | { |
d164b209 BB |
2673 | ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); |
2674 | ||
2675 | switch (type) { | |
e75c13c3 BB |
2676 | default: |
2677 | break; | |
d164b209 BB |
2678 | case ARC_SPACE_DATA: |
2679 | ARCSTAT_INCR(arcstat_data_size, space); | |
2680 | break; | |
cc7f677c | 2681 | case ARC_SPACE_META: |
500445c0 | 2682 | ARCSTAT_INCR(arcstat_metadata_size, space); |
cc7f677c | 2683 | break; |
25458cbe TC |
2684 | case ARC_SPACE_BONUS: |
2685 | ARCSTAT_INCR(arcstat_bonus_size, space); | |
2686 | break; | |
2687 | case ARC_SPACE_DNODE: | |
2688 | ARCSTAT_INCR(arcstat_dnode_size, space); | |
2689 | break; | |
2690 | case ARC_SPACE_DBUF: | |
2691 | ARCSTAT_INCR(arcstat_dbuf_size, space); | |
d164b209 BB |
2692 | break; |
2693 | case ARC_SPACE_HDRS: | |
2694 | ARCSTAT_INCR(arcstat_hdr_size, space); | |
2695 | break; | |
2696 | case ARC_SPACE_L2HDRS: | |
2697 | ARCSTAT_INCR(arcstat_l2_hdr_size, space); | |
2698 | break; | |
2699 | } | |
2700 | ||
500445c0 | 2701 | if (type != ARC_SPACE_DATA) |
cc7f677c PS |
2702 | ARCSTAT_INCR(arcstat_meta_used, space); |
2703 | ||
34dc7c2f BB |
2704 | atomic_add_64(&arc_size, space); |
2705 | } | |
2706 | ||
2707 | void | |
d164b209 | 2708 | arc_space_return(uint64_t space, arc_space_type_t type) |
34dc7c2f | 2709 | { |
d164b209 BB |
2710 | ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); |
2711 | ||
2712 | switch (type) { | |
e75c13c3 BB |
2713 | default: |
2714 | break; | |
d164b209 BB |
2715 | case ARC_SPACE_DATA: |
2716 | ARCSTAT_INCR(arcstat_data_size, -space); | |
2717 | break; | |
cc7f677c | 2718 | case ARC_SPACE_META: |
500445c0 | 2719 | ARCSTAT_INCR(arcstat_metadata_size, -space); |
cc7f677c | 2720 | break; |
25458cbe TC |
2721 | case ARC_SPACE_BONUS: |
2722 | ARCSTAT_INCR(arcstat_bonus_size, -space); | |
2723 | break; | |
2724 | case ARC_SPACE_DNODE: | |
2725 | ARCSTAT_INCR(arcstat_dnode_size, -space); | |
2726 | break; | |
2727 | case ARC_SPACE_DBUF: | |
2728 | ARCSTAT_INCR(arcstat_dbuf_size, -space); | |
d164b209 BB |
2729 | break; |
2730 | case ARC_SPACE_HDRS: | |
2731 | ARCSTAT_INCR(arcstat_hdr_size, -space); | |
2732 | break; | |
2733 | case ARC_SPACE_L2HDRS: | |
2734 | ARCSTAT_INCR(arcstat_l2_hdr_size, -space); | |
2735 | break; | |
2736 | } | |
2737 | ||
cc7f677c PS |
2738 | if (type != ARC_SPACE_DATA) { |
2739 | ASSERT(arc_meta_used >= space); | |
500445c0 PS |
2740 | if (arc_meta_max < arc_meta_used) |
2741 | arc_meta_max = arc_meta_used; | |
cc7f677c PS |
2742 | ARCSTAT_INCR(arcstat_meta_used, -space); |
2743 | } | |
2744 | ||
34dc7c2f BB |
2745 | ASSERT(arc_size >= space); |
2746 | atomic_add_64(&arc_size, -space); | |
2747 | } | |
2748 | ||
d3c2ae1c | 2749 | /* |
524b4217 | 2750 | * Given a hdr and a buf, returns whether that buf can share its b_data buffer |
a6255b7f | 2751 | * with the hdr's b_pabd. |
d3c2ae1c | 2752 | */ |
524b4217 DK |
2753 | static boolean_t |
2754 | arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) | |
2755 | { | |
524b4217 DK |
2756 | /* |
2757 | * The criteria for sharing a hdr's data are: | |
b5256303 TC |
2758 | * 1. the buffer is not encrypted |
2759 | * 2. the hdr's compression matches the buf's compression | |
2760 | * 3. the hdr doesn't need to be byteswapped | |
2761 | * 4. the hdr isn't already being shared | |
2762 | * 5. the buf is either compressed or it is the last buf in the hdr list | |
524b4217 | 2763 | * |
b5256303 | 2764 | * Criterion #5 maintains the invariant that shared uncompressed |
524b4217 DK |
2765 | * bufs must be the final buf in the hdr's b_buf list. Reading this, you |
2766 | * might ask, "if a compressed buf is allocated first, won't that be the | |
2767 | * last thing in the list?", but in that case it's impossible to create | |
2768 | * a shared uncompressed buf anyway (because the hdr must be compressed | |
2769 | * to have the compressed buf). You might also think that #3 is | |
2770 | * sufficient to make this guarantee, however it's possible | |
2771 | * (specifically in the rare L2ARC write race mentioned in | |
2772 | * arc_buf_alloc_impl()) there will be an existing uncompressed buf that | |
2773 | * is sharable, but wasn't at the time of its allocation. Rather than | |
2774 | * allow a new shared uncompressed buf to be created and then shuffle | |
2775 | * the list around to make it the last element, this simply disallows | |
2776 | * sharing if the new buf isn't the first to be added. | |
2777 | */ | |
2778 | ASSERT3P(buf->b_hdr, ==, hdr); | |
b5256303 TC |
2779 | boolean_t hdr_compressed = |
2780 | arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF; | |
a7004725 | 2781 | boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; |
b5256303 TC |
2782 | return (!ARC_BUF_ENCRYPTED(buf) && |
2783 | buf_compressed == hdr_compressed && | |
524b4217 DK |
2784 | hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && |
2785 | !HDR_SHARED_DATA(hdr) && | |
2786 | (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); | |
2787 | } | |
2788 | ||
2789 | /* | |
2790 | * Allocate a buf for this hdr. If you care about the data that's in the hdr, | |
2791 | * or if you want a compressed buffer, pass those flags in. Returns 0 if the | |
2792 | * copy was made successfully, or an error code otherwise. | |
2793 | */ | |
2794 | static int | |
be9a5c35 TC |
2795 | arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb, |
2796 | void *tag, boolean_t encrypted, boolean_t compressed, boolean_t noauth, | |
524b4217 | 2797 | boolean_t fill, arc_buf_t **ret) |
34dc7c2f | 2798 | { |
34dc7c2f | 2799 | arc_buf_t *buf; |
b5256303 | 2800 | arc_fill_flags_t flags = ARC_FILL_LOCKED; |
34dc7c2f | 2801 | |
d3c2ae1c GW |
2802 | ASSERT(HDR_HAS_L1HDR(hdr)); |
2803 | ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); | |
2804 | VERIFY(hdr->b_type == ARC_BUFC_DATA || | |
2805 | hdr->b_type == ARC_BUFC_METADATA); | |
524b4217 DK |
2806 | ASSERT3P(ret, !=, NULL); |
2807 | ASSERT3P(*ret, ==, NULL); | |
b5256303 | 2808 | IMPLY(encrypted, compressed); |
d3c2ae1c | 2809 | |
b9541d6b CW |
2810 | hdr->b_l1hdr.b_mru_hits = 0; |
2811 | hdr->b_l1hdr.b_mru_ghost_hits = 0; | |
2812 | hdr->b_l1hdr.b_mfu_hits = 0; | |
2813 | hdr->b_l1hdr.b_mfu_ghost_hits = 0; | |
2814 | hdr->b_l1hdr.b_l2_hits = 0; | |
2815 | ||
524b4217 | 2816 | buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); |
34dc7c2f BB |
2817 | buf->b_hdr = hdr; |
2818 | buf->b_data = NULL; | |
2aa34383 | 2819 | buf->b_next = hdr->b_l1hdr.b_buf; |
524b4217 | 2820 | buf->b_flags = 0; |
b9541d6b | 2821 | |
d3c2ae1c GW |
2822 | add_reference(hdr, tag); |
2823 | ||
2824 | /* | |
2825 | * We're about to change the hdr's b_flags. We must either | |
2826 | * hold the hash_lock or be undiscoverable. | |
2827 | */ | |
2828 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); | |
2829 | ||
2830 | /* | |
524b4217 | 2831 | * Only honor requests for compressed bufs if the hdr is actually |
b5256303 TC |
2832 | * compressed. This must be overriden if the buffer is encrypted since |
2833 | * encrypted buffers cannot be decompressed. | |
524b4217 | 2834 | */ |
b5256303 TC |
2835 | if (encrypted) { |
2836 | buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; | |
2837 | buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED; | |
2838 | flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED; | |
2839 | } else if (compressed && | |
2840 | arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) { | |
524b4217 | 2841 | buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; |
b5256303 TC |
2842 | flags |= ARC_FILL_COMPRESSED; |
2843 | } | |
2844 | ||
2845 | if (noauth) { | |
2846 | ASSERT0(encrypted); | |
2847 | flags |= ARC_FILL_NOAUTH; | |
2848 | } | |
524b4217 | 2849 | |
524b4217 DK |
2850 | /* |
2851 | * If the hdr's data can be shared then we share the data buffer and | |
2852 | * set the appropriate bit in the hdr's b_flags to indicate the hdr is | |
2aa34383 | 2853 | * allocate a new buffer to store the buf's data. |
524b4217 | 2854 | * |
a6255b7f DQ |
2855 | * There are two additional restrictions here because we're sharing |
2856 | * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be | |
2857 | * actively involved in an L2ARC write, because if this buf is used by | |
2858 | * an arc_write() then the hdr's data buffer will be released when the | |
524b4217 | 2859 | * write completes, even though the L2ARC write might still be using it. |
a6255b7f DQ |
2860 | * Second, the hdr's ABD must be linear so that the buf's user doesn't |
2861 | * need to be ABD-aware. | |
d3c2ae1c | 2862 | */ |
a7004725 | 2863 | boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) && |
b5256303 | 2864 | hdr->b_l1hdr.b_pabd != NULL && abd_is_linear(hdr->b_l1hdr.b_pabd); |
524b4217 DK |
2865 | |
2866 | /* Set up b_data and sharing */ | |
2867 | if (can_share) { | |
a6255b7f | 2868 | buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd); |
524b4217 | 2869 | buf->b_flags |= ARC_BUF_FLAG_SHARED; |
d3c2ae1c GW |
2870 | arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); |
2871 | } else { | |
524b4217 DK |
2872 | buf->b_data = |
2873 | arc_get_data_buf(hdr, arc_buf_size(buf), buf); | |
2874 | ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); | |
d3c2ae1c GW |
2875 | } |
2876 | VERIFY3P(buf->b_data, !=, NULL); | |
b9541d6b CW |
2877 | |
2878 | hdr->b_l1hdr.b_buf = buf; | |
d3c2ae1c | 2879 | hdr->b_l1hdr.b_bufcnt += 1; |
b5256303 TC |
2880 | if (encrypted) |
2881 | hdr->b_crypt_hdr.b_ebufcnt += 1; | |
b9541d6b | 2882 | |
524b4217 DK |
2883 | /* |
2884 | * If the user wants the data from the hdr, we need to either copy or | |
2885 | * decompress the data. | |
2886 | */ | |
2887 | if (fill) { | |
be9a5c35 TC |
2888 | ASSERT3P(zb, !=, NULL); |
2889 | return (arc_buf_fill(buf, spa, zb, flags)); | |
524b4217 | 2890 | } |
d3c2ae1c | 2891 | |
524b4217 | 2892 | return (0); |
34dc7c2f BB |
2893 | } |
2894 | ||
9babb374 BB |
2895 | static char *arc_onloan_tag = "onloan"; |
2896 | ||
a7004725 DK |
2897 | static inline void |
2898 | arc_loaned_bytes_update(int64_t delta) | |
2899 | { | |
2900 | atomic_add_64(&arc_loaned_bytes, delta); | |
2901 | ||
2902 | /* assert that it did not wrap around */ | |
2903 | ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); | |
2904 | } | |
2905 | ||
9babb374 BB |
2906 | /* |
2907 | * Loan out an anonymous arc buffer. Loaned buffers are not counted as in | |
2908 | * flight data by arc_tempreserve_space() until they are "returned". Loaned | |
2909 | * buffers must be returned to the arc before they can be used by the DMU or | |
2910 | * freed. | |
2911 | */ | |
2912 | arc_buf_t * | |
2aa34383 | 2913 | arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) |
9babb374 | 2914 | { |
2aa34383 DK |
2915 | arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, |
2916 | is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); | |
9babb374 | 2917 | |
5152a740 | 2918 | arc_loaned_bytes_update(arc_buf_size(buf)); |
a7004725 | 2919 | |
9babb374 BB |
2920 | return (buf); |
2921 | } | |
2922 | ||
2aa34383 DK |
2923 | arc_buf_t * |
2924 | arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, | |
2925 | enum zio_compress compression_type) | |
2926 | { | |
2927 | arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, | |
2928 | psize, lsize, compression_type); | |
2929 | ||
5152a740 | 2930 | arc_loaned_bytes_update(arc_buf_size(buf)); |
a7004725 | 2931 | |
2aa34383 DK |
2932 | return (buf); |
2933 | } | |
2934 | ||
b5256303 TC |
2935 | arc_buf_t * |
2936 | arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder, | |
2937 | const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, | |
2938 | dmu_object_type_t ot, uint64_t psize, uint64_t lsize, | |
2939 | enum zio_compress compression_type) | |
2940 | { | |
2941 | arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj, | |
2942 | byteorder, salt, iv, mac, ot, psize, lsize, compression_type); | |
2943 | ||
2944 | atomic_add_64(&arc_loaned_bytes, psize); | |
2945 | return (buf); | |
2946 | } | |
2947 | ||
2aa34383 | 2948 | |
9babb374 BB |
2949 | /* |
2950 | * Return a loaned arc buffer to the arc. | |
2951 | */ | |
2952 | void | |
2953 | arc_return_buf(arc_buf_t *buf, void *tag) | |
2954 | { | |
2955 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
2956 | ||
d3c2ae1c | 2957 | ASSERT3P(buf->b_data, !=, NULL); |
b9541d6b CW |
2958 | ASSERT(HDR_HAS_L1HDR(hdr)); |
2959 | (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); | |
2960 | (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); | |
9babb374 | 2961 | |
a7004725 | 2962 | arc_loaned_bytes_update(-arc_buf_size(buf)); |
9babb374 BB |
2963 | } |
2964 | ||
428870ff BB |
2965 | /* Detach an arc_buf from a dbuf (tag) */ |
2966 | void | |
2967 | arc_loan_inuse_buf(arc_buf_t *buf, void *tag) | |
2968 | { | |
b9541d6b | 2969 | arc_buf_hdr_t *hdr = buf->b_hdr; |
428870ff | 2970 | |
d3c2ae1c | 2971 | ASSERT3P(buf->b_data, !=, NULL); |
b9541d6b CW |
2972 | ASSERT(HDR_HAS_L1HDR(hdr)); |
2973 | (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); | |
2974 | (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); | |
428870ff | 2975 | |
a7004725 | 2976 | arc_loaned_bytes_update(arc_buf_size(buf)); |
428870ff BB |
2977 | } |
2978 | ||
d3c2ae1c | 2979 | static void |
a6255b7f | 2980 | l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type) |
34dc7c2f | 2981 | { |
d3c2ae1c | 2982 | l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); |
34dc7c2f | 2983 | |
a6255b7f | 2984 | df->l2df_abd = abd; |
d3c2ae1c GW |
2985 | df->l2df_size = size; |
2986 | df->l2df_type = type; | |
2987 | mutex_enter(&l2arc_free_on_write_mtx); | |
2988 | list_insert_head(l2arc_free_on_write, df); | |
2989 | mutex_exit(&l2arc_free_on_write_mtx); | |
2990 | } | |
428870ff | 2991 | |
d3c2ae1c | 2992 | static void |
b5256303 | 2993 | arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata) |
d3c2ae1c GW |
2994 | { |
2995 | arc_state_t *state = hdr->b_l1hdr.b_state; | |
2996 | arc_buf_contents_t type = arc_buf_type(hdr); | |
b5256303 | 2997 | uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr); |
1eb5bfa3 | 2998 | |
d3c2ae1c GW |
2999 | /* protected by hash lock, if in the hash table */ |
3000 | if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { | |
3001 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); | |
3002 | ASSERT(state != arc_anon && state != arc_l2c_only); | |
3003 | ||
3004 | (void) refcount_remove_many(&state->arcs_esize[type], | |
3005 | size, hdr); | |
1eb5bfa3 | 3006 | } |
d3c2ae1c | 3007 | (void) refcount_remove_many(&state->arcs_size, size, hdr); |
423e7b62 AG |
3008 | if (type == ARC_BUFC_METADATA) { |
3009 | arc_space_return(size, ARC_SPACE_META); | |
3010 | } else { | |
3011 | ASSERT(type == ARC_BUFC_DATA); | |
3012 | arc_space_return(size, ARC_SPACE_DATA); | |
3013 | } | |
d3c2ae1c | 3014 | |
b5256303 TC |
3015 | if (free_rdata) { |
3016 | l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type); | |
3017 | } else { | |
3018 | l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type); | |
3019 | } | |
34dc7c2f BB |
3020 | } |
3021 | ||
d3c2ae1c GW |
3022 | /* |
3023 | * Share the arc_buf_t's data with the hdr. Whenever we are sharing the | |
3024 | * data buffer, we transfer the refcount ownership to the hdr and update | |
3025 | * the appropriate kstats. | |
3026 | */ | |
3027 | static void | |
3028 | arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) | |
34dc7c2f | 3029 | { |
524b4217 | 3030 | ASSERT(arc_can_share(hdr, buf)); |
a6255b7f | 3031 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 3032 | ASSERT(!ARC_BUF_ENCRYPTED(buf)); |
d3c2ae1c | 3033 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); |
34dc7c2f BB |
3034 | |
3035 | /* | |
d3c2ae1c GW |
3036 | * Start sharing the data buffer. We transfer the |
3037 | * refcount ownership to the hdr since it always owns | |
3038 | * the refcount whenever an arc_buf_t is shared. | |
34dc7c2f | 3039 | */ |
d3c2ae1c | 3040 | refcount_transfer_ownership(&hdr->b_l1hdr.b_state->arcs_size, buf, hdr); |
a6255b7f DQ |
3041 | hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf)); |
3042 | abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd, | |
3043 | HDR_ISTYPE_METADATA(hdr)); | |
d3c2ae1c | 3044 | arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); |
524b4217 | 3045 | buf->b_flags |= ARC_BUF_FLAG_SHARED; |
34dc7c2f | 3046 | |
d3c2ae1c GW |
3047 | /* |
3048 | * Since we've transferred ownership to the hdr we need | |
3049 | * to increment its compressed and uncompressed kstats and | |
3050 | * decrement the overhead size. | |
3051 | */ | |
3052 | ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); | |
3053 | ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); | |
2aa34383 | 3054 | ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); |
34dc7c2f BB |
3055 | } |
3056 | ||
ca0bf58d | 3057 | static void |
d3c2ae1c | 3058 | arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) |
ca0bf58d | 3059 | { |
d3c2ae1c | 3060 | ASSERT(arc_buf_is_shared(buf)); |
a6255b7f | 3061 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); |
d3c2ae1c | 3062 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); |
ca0bf58d | 3063 | |
d3c2ae1c GW |
3064 | /* |
3065 | * We are no longer sharing this buffer so we need | |
3066 | * to transfer its ownership to the rightful owner. | |
3067 | */ | |
3068 | refcount_transfer_ownership(&hdr->b_l1hdr.b_state->arcs_size, hdr, buf); | |
3069 | arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); | |
a6255b7f DQ |
3070 | abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd); |
3071 | abd_put(hdr->b_l1hdr.b_pabd); | |
3072 | hdr->b_l1hdr.b_pabd = NULL; | |
524b4217 | 3073 | buf->b_flags &= ~ARC_BUF_FLAG_SHARED; |
d3c2ae1c GW |
3074 | |
3075 | /* | |
3076 | * Since the buffer is no longer shared between | |
3077 | * the arc buf and the hdr, count it as overhead. | |
3078 | */ | |
3079 | ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); | |
3080 | ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); | |
2aa34383 | 3081 | ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); |
ca0bf58d PS |
3082 | } |
3083 | ||
34dc7c2f | 3084 | /* |
2aa34383 DK |
3085 | * Remove an arc_buf_t from the hdr's buf list and return the last |
3086 | * arc_buf_t on the list. If no buffers remain on the list then return | |
3087 | * NULL. | |
3088 | */ | |
3089 | static arc_buf_t * | |
3090 | arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) | |
3091 | { | |
2aa34383 DK |
3092 | ASSERT(HDR_HAS_L1HDR(hdr)); |
3093 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); | |
3094 | ||
a7004725 DK |
3095 | arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; |
3096 | arc_buf_t *lastbuf = NULL; | |
3097 | ||
2aa34383 DK |
3098 | /* |
3099 | * Remove the buf from the hdr list and locate the last | |
3100 | * remaining buffer on the list. | |
3101 | */ | |
3102 | while (*bufp != NULL) { | |
3103 | if (*bufp == buf) | |
3104 | *bufp = buf->b_next; | |
3105 | ||
3106 | /* | |
3107 | * If we've removed a buffer in the middle of | |
3108 | * the list then update the lastbuf and update | |
3109 | * bufp. | |
3110 | */ | |
3111 | if (*bufp != NULL) { | |
3112 | lastbuf = *bufp; | |
3113 | bufp = &(*bufp)->b_next; | |
3114 | } | |
3115 | } | |
3116 | buf->b_next = NULL; | |
3117 | ASSERT3P(lastbuf, !=, buf); | |
3118 | IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); | |
3119 | IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); | |
3120 | IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); | |
3121 | ||
3122 | return (lastbuf); | |
3123 | } | |
3124 | ||
3125 | /* | |
3126 | * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's | |
3127 | * list and free it. | |
34dc7c2f BB |
3128 | */ |
3129 | static void | |
2aa34383 | 3130 | arc_buf_destroy_impl(arc_buf_t *buf) |
34dc7c2f | 3131 | { |
498877ba | 3132 | arc_buf_hdr_t *hdr = buf->b_hdr; |
ca0bf58d PS |
3133 | |
3134 | /* | |
524b4217 DK |
3135 | * Free up the data associated with the buf but only if we're not |
3136 | * sharing this with the hdr. If we are sharing it with the hdr, the | |
3137 | * hdr is responsible for doing the free. | |
ca0bf58d | 3138 | */ |
d3c2ae1c GW |
3139 | if (buf->b_data != NULL) { |
3140 | /* | |
3141 | * We're about to change the hdr's b_flags. We must either | |
3142 | * hold the hash_lock or be undiscoverable. | |
3143 | */ | |
3144 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); | |
3145 | ||
524b4217 | 3146 | arc_cksum_verify(buf); |
d3c2ae1c GW |
3147 | arc_buf_unwatch(buf); |
3148 | ||
2aa34383 | 3149 | if (arc_buf_is_shared(buf)) { |
d3c2ae1c GW |
3150 | arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); |
3151 | } else { | |
2aa34383 | 3152 | uint64_t size = arc_buf_size(buf); |
d3c2ae1c GW |
3153 | arc_free_data_buf(hdr, buf->b_data, size, buf); |
3154 | ARCSTAT_INCR(arcstat_overhead_size, -size); | |
3155 | } | |
3156 | buf->b_data = NULL; | |
3157 | ||
3158 | ASSERT(hdr->b_l1hdr.b_bufcnt > 0); | |
3159 | hdr->b_l1hdr.b_bufcnt -= 1; | |
b5256303 | 3160 | |
da5d4697 | 3161 | if (ARC_BUF_ENCRYPTED(buf)) { |
b5256303 TC |
3162 | hdr->b_crypt_hdr.b_ebufcnt -= 1; |
3163 | ||
da5d4697 D |
3164 | /* |
3165 | * If we have no more encrypted buffers and we've | |
3166 | * already gotten a copy of the decrypted data we can | |
3167 | * free b_rabd to save some space. | |
3168 | */ | |
3169 | if (hdr->b_crypt_hdr.b_ebufcnt == 0 && | |
3170 | HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL && | |
3171 | !HDR_IO_IN_PROGRESS(hdr)) { | |
3172 | arc_hdr_free_abd(hdr, B_TRUE); | |
3173 | } | |
440a3eb9 | 3174 | } |
d3c2ae1c GW |
3175 | } |
3176 | ||
a7004725 | 3177 | arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); |
d3c2ae1c | 3178 | |
524b4217 | 3179 | if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { |
2aa34383 | 3180 | /* |
524b4217 | 3181 | * If the current arc_buf_t is sharing its data buffer with the |
a6255b7f | 3182 | * hdr, then reassign the hdr's b_pabd to share it with the new |
524b4217 DK |
3183 | * buffer at the end of the list. The shared buffer is always |
3184 | * the last one on the hdr's buffer list. | |
3185 | * | |
3186 | * There is an equivalent case for compressed bufs, but since | |
3187 | * they aren't guaranteed to be the last buf in the list and | |
3188 | * that is an exceedingly rare case, we just allow that space be | |
b5256303 TC |
3189 | * wasted temporarily. We must also be careful not to share |
3190 | * encrypted buffers, since they cannot be shared. | |
2aa34383 | 3191 | */ |
b5256303 | 3192 | if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) { |
524b4217 | 3193 | /* Only one buf can be shared at once */ |
2aa34383 | 3194 | VERIFY(!arc_buf_is_shared(lastbuf)); |
524b4217 DK |
3195 | /* hdr is uncompressed so can't have compressed buf */ |
3196 | VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); | |
d3c2ae1c | 3197 | |
a6255b7f | 3198 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); |
b5256303 | 3199 | arc_hdr_free_abd(hdr, B_FALSE); |
d3c2ae1c | 3200 | |
2aa34383 DK |
3201 | /* |
3202 | * We must setup a new shared block between the | |
3203 | * last buffer and the hdr. The data would have | |
3204 | * been allocated by the arc buf so we need to transfer | |
3205 | * ownership to the hdr since it's now being shared. | |
3206 | */ | |
3207 | arc_share_buf(hdr, lastbuf); | |
3208 | } | |
3209 | } else if (HDR_SHARED_DATA(hdr)) { | |
d3c2ae1c | 3210 | /* |
2aa34383 DK |
3211 | * Uncompressed shared buffers are always at the end |
3212 | * of the list. Compressed buffers don't have the | |
3213 | * same requirements. This makes it hard to | |
3214 | * simply assert that the lastbuf is shared so | |
3215 | * we rely on the hdr's compression flags to determine | |
3216 | * if we have a compressed, shared buffer. | |
d3c2ae1c | 3217 | */ |
2aa34383 DK |
3218 | ASSERT3P(lastbuf, !=, NULL); |
3219 | ASSERT(arc_buf_is_shared(lastbuf) || | |
b5256303 | 3220 | arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF); |
ca0bf58d PS |
3221 | } |
3222 | ||
a7004725 DK |
3223 | /* |
3224 | * Free the checksum if we're removing the last uncompressed buf from | |
3225 | * this hdr. | |
3226 | */ | |
3227 | if (!arc_hdr_has_uncompressed_buf(hdr)) { | |
d3c2ae1c | 3228 | arc_cksum_free(hdr); |
a7004725 | 3229 | } |
d3c2ae1c GW |
3230 | |
3231 | /* clean up the buf */ | |
3232 | buf->b_hdr = NULL; | |
3233 | kmem_cache_free(buf_cache, buf); | |
3234 | } | |
3235 | ||
3236 | static void | |
b5256303 | 3237 | arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, boolean_t alloc_rdata) |
d3c2ae1c | 3238 | { |
b5256303 TC |
3239 | uint64_t size; |
3240 | ||
d3c2ae1c GW |
3241 | ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); |
3242 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
b5256303 TC |
3243 | ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata); |
3244 | IMPLY(alloc_rdata, HDR_PROTECTED(hdr)); | |
d3c2ae1c | 3245 | |
b5256303 TC |
3246 | if (alloc_rdata) { |
3247 | size = HDR_GET_PSIZE(hdr); | |
3248 | ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL); | |
3249 | hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr); | |
3250 | ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL); | |
3251 | ARCSTAT_INCR(arcstat_raw_size, size); | |
3252 | } else { | |
3253 | size = arc_hdr_size(hdr); | |
3254 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); | |
3255 | hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr); | |
3256 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); | |
3257 | } | |
3258 | ||
3259 | ARCSTAT_INCR(arcstat_compressed_size, size); | |
d3c2ae1c GW |
3260 | ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); |
3261 | } | |
3262 | ||
3263 | static void | |
b5256303 | 3264 | arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata) |
d3c2ae1c | 3265 | { |
b5256303 TC |
3266 | uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr); |
3267 | ||
d3c2ae1c | 3268 | ASSERT(HDR_HAS_L1HDR(hdr)); |
b5256303 TC |
3269 | ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); |
3270 | IMPLY(free_rdata, HDR_HAS_RABD(hdr)); | |
d3c2ae1c | 3271 | |
ca0bf58d | 3272 | /* |
d3c2ae1c GW |
3273 | * If the hdr is currently being written to the l2arc then |
3274 | * we defer freeing the data by adding it to the l2arc_free_on_write | |
3275 | * list. The l2arc will free the data once it's finished | |
3276 | * writing it to the l2arc device. | |
ca0bf58d | 3277 | */ |
d3c2ae1c | 3278 | if (HDR_L2_WRITING(hdr)) { |
b5256303 | 3279 | arc_hdr_free_on_write(hdr, free_rdata); |
d3c2ae1c | 3280 | ARCSTAT_BUMP(arcstat_l2_free_on_write); |
b5256303 TC |
3281 | } else if (free_rdata) { |
3282 | arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr); | |
d3c2ae1c | 3283 | } else { |
b5256303 | 3284 | arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr); |
ca0bf58d PS |
3285 | } |
3286 | ||
b5256303 TC |
3287 | if (free_rdata) { |
3288 | hdr->b_crypt_hdr.b_rabd = NULL; | |
3289 | ARCSTAT_INCR(arcstat_raw_size, -size); | |
3290 | } else { | |
3291 | hdr->b_l1hdr.b_pabd = NULL; | |
3292 | } | |
3293 | ||
3294 | if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr)) | |
3295 | hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; | |
3296 | ||
3297 | ARCSTAT_INCR(arcstat_compressed_size, -size); | |
d3c2ae1c GW |
3298 | ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); |
3299 | } | |
3300 | ||
3301 | static arc_buf_hdr_t * | |
3302 | arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, | |
b5256303 TC |
3303 | boolean_t protected, enum zio_compress compression_type, |
3304 | arc_buf_contents_t type, boolean_t alloc_rdata) | |
d3c2ae1c GW |
3305 | { |
3306 | arc_buf_hdr_t *hdr; | |
3307 | ||
d3c2ae1c | 3308 | VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); |
b5256303 TC |
3309 | if (protected) { |
3310 | hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE); | |
3311 | } else { | |
3312 | hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); | |
3313 | } | |
d3c2ae1c | 3314 | |
d3c2ae1c GW |
3315 | ASSERT(HDR_EMPTY(hdr)); |
3316 | ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); | |
3317 | HDR_SET_PSIZE(hdr, psize); | |
3318 | HDR_SET_LSIZE(hdr, lsize); | |
3319 | hdr->b_spa = spa; | |
3320 | hdr->b_type = type; | |
3321 | hdr->b_flags = 0; | |
3322 | arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); | |
2aa34383 | 3323 | arc_hdr_set_compress(hdr, compression_type); |
b5256303 TC |
3324 | if (protected) |
3325 | arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED); | |
ca0bf58d | 3326 | |
d3c2ae1c GW |
3327 | hdr->b_l1hdr.b_state = arc_anon; |
3328 | hdr->b_l1hdr.b_arc_access = 0; | |
3329 | hdr->b_l1hdr.b_bufcnt = 0; | |
3330 | hdr->b_l1hdr.b_buf = NULL; | |
ca0bf58d | 3331 | |
d3c2ae1c GW |
3332 | /* |
3333 | * Allocate the hdr's buffer. This will contain either | |
3334 | * the compressed or uncompressed data depending on the block | |
3335 | * it references and compressed arc enablement. | |
3336 | */ | |
b5256303 | 3337 | arc_hdr_alloc_abd(hdr, alloc_rdata); |
d3c2ae1c | 3338 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
ca0bf58d | 3339 | |
d3c2ae1c | 3340 | return (hdr); |
ca0bf58d PS |
3341 | } |
3342 | ||
bd089c54 | 3343 | /* |
d3c2ae1c GW |
3344 | * Transition between the two allocation states for the arc_buf_hdr struct. |
3345 | * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without | |
3346 | * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller | |
3347 | * version is used when a cache buffer is only in the L2ARC in order to reduce | |
3348 | * memory usage. | |
bd089c54 | 3349 | */ |
d3c2ae1c GW |
3350 | static arc_buf_hdr_t * |
3351 | arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) | |
34dc7c2f | 3352 | { |
1c27024e DB |
3353 | ASSERT(HDR_HAS_L2HDR(hdr)); |
3354 | ||
d3c2ae1c GW |
3355 | arc_buf_hdr_t *nhdr; |
3356 | l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; | |
34dc7c2f | 3357 | |
d3c2ae1c GW |
3358 | ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || |
3359 | (old == hdr_l2only_cache && new == hdr_full_cache)); | |
34dc7c2f | 3360 | |
b5256303 TC |
3361 | /* |
3362 | * if the caller wanted a new full header and the header is to be | |
3363 | * encrypted we will actually allocate the header from the full crypt | |
3364 | * cache instead. The same applies to freeing from the old cache. | |
3365 | */ | |
3366 | if (HDR_PROTECTED(hdr) && new == hdr_full_cache) | |
3367 | new = hdr_full_crypt_cache; | |
3368 | if (HDR_PROTECTED(hdr) && old == hdr_full_cache) | |
3369 | old = hdr_full_crypt_cache; | |
3370 | ||
d3c2ae1c | 3371 | nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); |
428870ff | 3372 | |
d3c2ae1c GW |
3373 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); |
3374 | buf_hash_remove(hdr); | |
ca0bf58d | 3375 | |
d3c2ae1c | 3376 | bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); |
34dc7c2f | 3377 | |
b5256303 | 3378 | if (new == hdr_full_cache || new == hdr_full_crypt_cache) { |
d3c2ae1c GW |
3379 | arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); |
3380 | /* | |
3381 | * arc_access and arc_change_state need to be aware that a | |
3382 | * header has just come out of L2ARC, so we set its state to | |
3383 | * l2c_only even though it's about to change. | |
3384 | */ | |
3385 | nhdr->b_l1hdr.b_state = arc_l2c_only; | |
34dc7c2f | 3386 | |
d3c2ae1c | 3387 | /* Verify previous threads set to NULL before freeing */ |
a6255b7f | 3388 | ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 3389 | ASSERT(!HDR_HAS_RABD(hdr)); |
d3c2ae1c GW |
3390 | } else { |
3391 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
3392 | ASSERT0(hdr->b_l1hdr.b_bufcnt); | |
3393 | ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); | |
36da08ef | 3394 | |
d3c2ae1c GW |
3395 | /* |
3396 | * If we've reached here, We must have been called from | |
3397 | * arc_evict_hdr(), as such we should have already been | |
3398 | * removed from any ghost list we were previously on | |
3399 | * (which protects us from racing with arc_evict_state), | |
3400 | * thus no locking is needed during this check. | |
3401 | */ | |
3402 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); | |
1eb5bfa3 GW |
3403 | |
3404 | /* | |
d3c2ae1c GW |
3405 | * A buffer must not be moved into the arc_l2c_only |
3406 | * state if it's not finished being written out to the | |
a6255b7f | 3407 | * l2arc device. Otherwise, the b_l1hdr.b_pabd field |
d3c2ae1c | 3408 | * might try to be accessed, even though it was removed. |
1eb5bfa3 | 3409 | */ |
d3c2ae1c | 3410 | VERIFY(!HDR_L2_WRITING(hdr)); |
a6255b7f | 3411 | VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 3412 | ASSERT(!HDR_HAS_RABD(hdr)); |
d3c2ae1c GW |
3413 | |
3414 | arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); | |
34dc7c2f | 3415 | } |
d3c2ae1c GW |
3416 | /* |
3417 | * The header has been reallocated so we need to re-insert it into any | |
3418 | * lists it was on. | |
3419 | */ | |
3420 | (void) buf_hash_insert(nhdr, NULL); | |
34dc7c2f | 3421 | |
d3c2ae1c | 3422 | ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); |
34dc7c2f | 3423 | |
d3c2ae1c GW |
3424 | mutex_enter(&dev->l2ad_mtx); |
3425 | ||
3426 | /* | |
3427 | * We must place the realloc'ed header back into the list at | |
3428 | * the same spot. Otherwise, if it's placed earlier in the list, | |
3429 | * l2arc_write_buffers() could find it during the function's | |
3430 | * write phase, and try to write it out to the l2arc. | |
3431 | */ | |
3432 | list_insert_after(&dev->l2ad_buflist, hdr, nhdr); | |
3433 | list_remove(&dev->l2ad_buflist, hdr); | |
34dc7c2f | 3434 | |
d3c2ae1c | 3435 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f | 3436 | |
d3c2ae1c GW |
3437 | /* |
3438 | * Since we're using the pointer address as the tag when | |
3439 | * incrementing and decrementing the l2ad_alloc refcount, we | |
3440 | * must remove the old pointer (that we're about to destroy) and | |
3441 | * add the new pointer to the refcount. Otherwise we'd remove | |
3442 | * the wrong pointer address when calling arc_hdr_destroy() later. | |
3443 | */ | |
3444 | ||
3445 | (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); | |
3446 | (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); | |
3447 | ||
3448 | buf_discard_identity(hdr); | |
3449 | kmem_cache_free(old, hdr); | |
3450 | ||
3451 | return (nhdr); | |
3452 | } | |
3453 | ||
b5256303 TC |
3454 | /* |
3455 | * This function allows an L1 header to be reallocated as a crypt | |
3456 | * header and vice versa. If we are going to a crypt header, the | |
3457 | * new fields will be zeroed out. | |
3458 | */ | |
3459 | static arc_buf_hdr_t * | |
3460 | arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt) | |
3461 | { | |
3462 | arc_buf_hdr_t *nhdr; | |
3463 | arc_buf_t *buf; | |
3464 | kmem_cache_t *ncache, *ocache; | |
3465 | ||
3466 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
3467 | ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt); | |
3468 | ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); | |
3469 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); | |
3470 | ||
3471 | if (need_crypt) { | |
3472 | ncache = hdr_full_crypt_cache; | |
3473 | ocache = hdr_full_cache; | |
3474 | } else { | |
3475 | ncache = hdr_full_cache; | |
3476 | ocache = hdr_full_crypt_cache; | |
3477 | } | |
3478 | ||
3479 | nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE); | |
3480 | bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); | |
3481 | nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum; | |
3482 | nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt; | |
3483 | nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap; | |
3484 | nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state; | |
3485 | nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access; | |
3486 | nhdr->b_l1hdr.b_mru_hits = hdr->b_l1hdr.b_mru_hits; | |
3487 | nhdr->b_l1hdr.b_mru_ghost_hits = hdr->b_l1hdr.b_mru_ghost_hits; | |
3488 | nhdr->b_l1hdr.b_mfu_hits = hdr->b_l1hdr.b_mfu_hits; | |
3489 | nhdr->b_l1hdr.b_mfu_ghost_hits = hdr->b_l1hdr.b_mfu_ghost_hits; | |
3490 | nhdr->b_l1hdr.b_l2_hits = hdr->b_l1hdr.b_l2_hits; | |
3491 | nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb; | |
3492 | nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd; | |
3493 | nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf; | |
3494 | ||
3495 | /* | |
3496 | * This refcount_add() exists only to ensure that the individual | |
3497 | * arc buffers always point to a header that is referenced, avoiding | |
3498 | * a small race condition that could trigger ASSERTs. | |
3499 | */ | |
3500 | (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG); | |
3501 | ||
3502 | for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) { | |
3503 | mutex_enter(&buf->b_evict_lock); | |
3504 | buf->b_hdr = nhdr; | |
3505 | mutex_exit(&buf->b_evict_lock); | |
3506 | } | |
3507 | ||
3508 | refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt); | |
3509 | (void) refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG); | |
3510 | ||
3511 | if (need_crypt) { | |
3512 | arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED); | |
3513 | } else { | |
3514 | arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED); | |
3515 | } | |
3516 | ||
3517 | buf_discard_identity(hdr); | |
3518 | kmem_cache_free(ocache, hdr); | |
3519 | ||
3520 | return (nhdr); | |
3521 | } | |
3522 | ||
3523 | /* | |
3524 | * This function is used by the send / receive code to convert a newly | |
3525 | * allocated arc_buf_t to one that is suitable for a raw encrypted write. It | |
3526 | * is also used to allow the root objset block to be uupdated without altering | |
3527 | * its embedded MACs. Both block types will always be uncompressed so we do not | |
3528 | * have to worry about compression type or psize. | |
3529 | */ | |
3530 | void | |
3531 | arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder, | |
3532 | dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv, | |
3533 | const uint8_t *mac) | |
3534 | { | |
3535 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
3536 | ||
3537 | ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET); | |
3538 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
3539 | ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); | |
3540 | ||
3541 | buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED); | |
3542 | if (!HDR_PROTECTED(hdr)) | |
3543 | hdr = arc_hdr_realloc_crypt(hdr, B_TRUE); | |
3544 | hdr->b_crypt_hdr.b_dsobj = dsobj; | |
3545 | hdr->b_crypt_hdr.b_ot = ot; | |
3546 | hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ? | |
3547 | DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot); | |
3548 | if (!arc_hdr_has_uncompressed_buf(hdr)) | |
3549 | arc_cksum_free(hdr); | |
3550 | ||
3551 | if (salt != NULL) | |
3552 | bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN); | |
3553 | if (iv != NULL) | |
3554 | bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN); | |
3555 | if (mac != NULL) | |
3556 | bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN); | |
3557 | } | |
3558 | ||
d3c2ae1c GW |
3559 | /* |
3560 | * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. | |
3561 | * The buf is returned thawed since we expect the consumer to modify it. | |
3562 | */ | |
3563 | arc_buf_t * | |
2aa34383 | 3564 | arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) |
d3c2ae1c | 3565 | { |
d3c2ae1c | 3566 | arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, |
b5256303 | 3567 | B_FALSE, ZIO_COMPRESS_OFF, type, B_FALSE); |
d3c2ae1c | 3568 | ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); |
2aa34383 | 3569 | |
a7004725 | 3570 | arc_buf_t *buf = NULL; |
be9a5c35 | 3571 | VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE, |
b5256303 | 3572 | B_FALSE, B_FALSE, &buf)); |
d3c2ae1c | 3573 | arc_buf_thaw(buf); |
2aa34383 DK |
3574 | |
3575 | return (buf); | |
3576 | } | |
3577 | ||
3578 | /* | |
3579 | * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this | |
3580 | * for bufs containing metadata. | |
3581 | */ | |
3582 | arc_buf_t * | |
3583 | arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, | |
3584 | enum zio_compress compression_type) | |
3585 | { | |
2aa34383 DK |
3586 | ASSERT3U(lsize, >, 0); |
3587 | ASSERT3U(lsize, >=, psize); | |
b5256303 TC |
3588 | ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF); |
3589 | ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS); | |
2aa34383 | 3590 | |
a7004725 | 3591 | arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, |
b5256303 | 3592 | B_FALSE, compression_type, ARC_BUFC_DATA, B_FALSE); |
2aa34383 DK |
3593 | ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); |
3594 | ||
a7004725 | 3595 | arc_buf_t *buf = NULL; |
be9a5c35 | 3596 | VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, |
b5256303 | 3597 | B_TRUE, B_FALSE, B_FALSE, &buf)); |
2aa34383 DK |
3598 | arc_buf_thaw(buf); |
3599 | ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); | |
3600 | ||
a6255b7f DQ |
3601 | if (!arc_buf_is_shared(buf)) { |
3602 | /* | |
3603 | * To ensure that the hdr has the correct data in it if we call | |
b5256303 | 3604 | * arc_untransform() on this buf before it's been written to |
a6255b7f DQ |
3605 | * disk, it's easiest if we just set up sharing between the |
3606 | * buf and the hdr. | |
3607 | */ | |
3608 | ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd)); | |
b5256303 | 3609 | arc_hdr_free_abd(hdr, B_FALSE); |
a6255b7f DQ |
3610 | arc_share_buf(hdr, buf); |
3611 | } | |
3612 | ||
d3c2ae1c | 3613 | return (buf); |
34dc7c2f BB |
3614 | } |
3615 | ||
b5256303 TC |
3616 | arc_buf_t * |
3617 | arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder, | |
3618 | const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, | |
3619 | dmu_object_type_t ot, uint64_t psize, uint64_t lsize, | |
3620 | enum zio_compress compression_type) | |
3621 | { | |
3622 | arc_buf_hdr_t *hdr; | |
3623 | arc_buf_t *buf; | |
3624 | arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ? | |
3625 | ARC_BUFC_METADATA : ARC_BUFC_DATA; | |
3626 | ||
3627 | ASSERT3U(lsize, >, 0); | |
3628 | ASSERT3U(lsize, >=, psize); | |
3629 | ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF); | |
3630 | ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS); | |
3631 | ||
3632 | hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE, | |
3633 | compression_type, type, B_TRUE); | |
3634 | ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); | |
3635 | ||
3636 | hdr->b_crypt_hdr.b_dsobj = dsobj; | |
3637 | hdr->b_crypt_hdr.b_ot = ot; | |
3638 | hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ? | |
3639 | DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot); | |
3640 | bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN); | |
3641 | bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN); | |
3642 | bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN); | |
3643 | ||
3644 | /* | |
3645 | * This buffer will be considered encrypted even if the ot is not an | |
3646 | * encrypted type. It will become authenticated instead in | |
3647 | * arc_write_ready(). | |
3648 | */ | |
3649 | buf = NULL; | |
be9a5c35 | 3650 | VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE, |
b5256303 TC |
3651 | B_FALSE, B_FALSE, &buf)); |
3652 | arc_buf_thaw(buf); | |
3653 | ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); | |
3654 | ||
3655 | return (buf); | |
3656 | } | |
3657 | ||
d962d5da PS |
3658 | static void |
3659 | arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) | |
3660 | { | |
3661 | l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; | |
3662 | l2arc_dev_t *dev = l2hdr->b_dev; | |
01850391 | 3663 | uint64_t psize = arc_hdr_size(hdr); |
d962d5da PS |
3664 | |
3665 | ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); | |
3666 | ASSERT(HDR_HAS_L2HDR(hdr)); | |
3667 | ||
3668 | list_remove(&dev->l2ad_buflist, hdr); | |
3669 | ||
01850391 AG |
3670 | ARCSTAT_INCR(arcstat_l2_psize, -psize); |
3671 | ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); | |
d962d5da | 3672 | |
01850391 | 3673 | vdev_space_update(dev->l2ad_vdev, -psize, 0, 0); |
d962d5da | 3674 | |
01850391 | 3675 | (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr); |
d3c2ae1c | 3676 | arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); |
d962d5da PS |
3677 | } |
3678 | ||
34dc7c2f BB |
3679 | static void |
3680 | arc_hdr_destroy(arc_buf_hdr_t *hdr) | |
3681 | { | |
b9541d6b CW |
3682 | if (HDR_HAS_L1HDR(hdr)) { |
3683 | ASSERT(hdr->b_l1hdr.b_buf == NULL || | |
d3c2ae1c | 3684 | hdr->b_l1hdr.b_bufcnt > 0); |
b9541d6b CW |
3685 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
3686 | ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); | |
3687 | } | |
34dc7c2f | 3688 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); |
b9541d6b CW |
3689 | ASSERT(!HDR_IN_HASH_TABLE(hdr)); |
3690 | ||
d3c2ae1c GW |
3691 | if (!HDR_EMPTY(hdr)) |
3692 | buf_discard_identity(hdr); | |
3693 | ||
b9541d6b | 3694 | if (HDR_HAS_L2HDR(hdr)) { |
d962d5da PS |
3695 | l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; |
3696 | boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); | |
428870ff | 3697 | |
d962d5da PS |
3698 | if (!buflist_held) |
3699 | mutex_enter(&dev->l2ad_mtx); | |
b9541d6b | 3700 | |
ca0bf58d | 3701 | /* |
d962d5da PS |
3702 | * Even though we checked this conditional above, we |
3703 | * need to check this again now that we have the | |
3704 | * l2ad_mtx. This is because we could be racing with | |
3705 | * another thread calling l2arc_evict() which might have | |
3706 | * destroyed this header's L2 portion as we were waiting | |
3707 | * to acquire the l2ad_mtx. If that happens, we don't | |
3708 | * want to re-destroy the header's L2 portion. | |
ca0bf58d | 3709 | */ |
d962d5da PS |
3710 | if (HDR_HAS_L2HDR(hdr)) |
3711 | arc_hdr_l2hdr_destroy(hdr); | |
428870ff BB |
3712 | |
3713 | if (!buflist_held) | |
d962d5da | 3714 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f BB |
3715 | } |
3716 | ||
d3c2ae1c GW |
3717 | if (HDR_HAS_L1HDR(hdr)) { |
3718 | arc_cksum_free(hdr); | |
b9541d6b | 3719 | |
d3c2ae1c | 3720 | while (hdr->b_l1hdr.b_buf != NULL) |
2aa34383 | 3721 | arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); |
34dc7c2f | 3722 | |
b5256303 TC |
3723 | if (hdr->b_l1hdr.b_pabd != NULL) { |
3724 | arc_hdr_free_abd(hdr, B_FALSE); | |
3725 | } | |
3726 | ||
440a3eb9 | 3727 | if (HDR_HAS_RABD(hdr)) |
b5256303 | 3728 | arc_hdr_free_abd(hdr, B_TRUE); |
b9541d6b CW |
3729 | } |
3730 | ||
34dc7c2f | 3731 | ASSERT3P(hdr->b_hash_next, ==, NULL); |
b9541d6b | 3732 | if (HDR_HAS_L1HDR(hdr)) { |
ca0bf58d | 3733 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); |
b9541d6b | 3734 | ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); |
b5256303 TC |
3735 | |
3736 | if (!HDR_PROTECTED(hdr)) { | |
3737 | kmem_cache_free(hdr_full_cache, hdr); | |
3738 | } else { | |
3739 | kmem_cache_free(hdr_full_crypt_cache, hdr); | |
3740 | } | |
b9541d6b CW |
3741 | } else { |
3742 | kmem_cache_free(hdr_l2only_cache, hdr); | |
3743 | } | |
34dc7c2f BB |
3744 | } |
3745 | ||
3746 | void | |
d3c2ae1c | 3747 | arc_buf_destroy(arc_buf_t *buf, void* tag) |
34dc7c2f BB |
3748 | { |
3749 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
96c080cb | 3750 | kmutex_t *hash_lock = HDR_LOCK(hdr); |
34dc7c2f | 3751 | |
b9541d6b | 3752 | if (hdr->b_l1hdr.b_state == arc_anon) { |
d3c2ae1c GW |
3753 | ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); |
3754 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
3755 | VERIFY0(remove_reference(hdr, NULL, tag)); | |
3756 | arc_hdr_destroy(hdr); | |
3757 | return; | |
34dc7c2f BB |
3758 | } |
3759 | ||
3760 | mutex_enter(hash_lock); | |
d3c2ae1c GW |
3761 | ASSERT3P(hdr, ==, buf->b_hdr); |
3762 | ASSERT(hdr->b_l1hdr.b_bufcnt > 0); | |
428870ff | 3763 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); |
d3c2ae1c GW |
3764 | ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); |
3765 | ASSERT3P(buf->b_data, !=, NULL); | |
34dc7c2f BB |
3766 | |
3767 | (void) remove_reference(hdr, hash_lock, tag); | |
2aa34383 | 3768 | arc_buf_destroy_impl(buf); |
34dc7c2f | 3769 | mutex_exit(hash_lock); |
34dc7c2f BB |
3770 | } |
3771 | ||
34dc7c2f | 3772 | /* |
ca0bf58d PS |
3773 | * Evict the arc_buf_hdr that is provided as a parameter. The resultant |
3774 | * state of the header is dependent on its state prior to entering this | |
3775 | * function. The following transitions are possible: | |
34dc7c2f | 3776 | * |
ca0bf58d PS |
3777 | * - arc_mru -> arc_mru_ghost |
3778 | * - arc_mfu -> arc_mfu_ghost | |
3779 | * - arc_mru_ghost -> arc_l2c_only | |
3780 | * - arc_mru_ghost -> deleted | |
3781 | * - arc_mfu_ghost -> arc_l2c_only | |
3782 | * - arc_mfu_ghost -> deleted | |
34dc7c2f | 3783 | */ |
ca0bf58d PS |
3784 | static int64_t |
3785 | arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) | |
34dc7c2f | 3786 | { |
ca0bf58d PS |
3787 | arc_state_t *evicted_state, *state; |
3788 | int64_t bytes_evicted = 0; | |
d4a72f23 TC |
3789 | int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ? |
3790 | arc_min_prescient_prefetch_ms : arc_min_prefetch_ms; | |
34dc7c2f | 3791 | |
ca0bf58d PS |
3792 | ASSERT(MUTEX_HELD(hash_lock)); |
3793 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
e8b96c60 | 3794 | |
ca0bf58d PS |
3795 | state = hdr->b_l1hdr.b_state; |
3796 | if (GHOST_STATE(state)) { | |
3797 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
d3c2ae1c | 3798 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); |
e8b96c60 MA |
3799 | |
3800 | /* | |
ca0bf58d | 3801 | * l2arc_write_buffers() relies on a header's L1 portion |
a6255b7f | 3802 | * (i.e. its b_pabd field) during it's write phase. |
ca0bf58d PS |
3803 | * Thus, we cannot push a header onto the arc_l2c_only |
3804 | * state (removing its L1 piece) until the header is | |
3805 | * done being written to the l2arc. | |
e8b96c60 | 3806 | */ |
ca0bf58d PS |
3807 | if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { |
3808 | ARCSTAT_BUMP(arcstat_evict_l2_skip); | |
3809 | return (bytes_evicted); | |
e8b96c60 MA |
3810 | } |
3811 | ||
ca0bf58d | 3812 | ARCSTAT_BUMP(arcstat_deleted); |
d3c2ae1c | 3813 | bytes_evicted += HDR_GET_LSIZE(hdr); |
428870ff | 3814 | |
ca0bf58d | 3815 | DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); |
428870ff | 3816 | |
ca0bf58d | 3817 | if (HDR_HAS_L2HDR(hdr)) { |
a6255b7f | 3818 | ASSERT(hdr->b_l1hdr.b_pabd == NULL); |
b5256303 | 3819 | ASSERT(!HDR_HAS_RABD(hdr)); |
ca0bf58d PS |
3820 | /* |
3821 | * This buffer is cached on the 2nd Level ARC; | |
3822 | * don't destroy the header. | |
3823 | */ | |
3824 | arc_change_state(arc_l2c_only, hdr, hash_lock); | |
3825 | /* | |
3826 | * dropping from L1+L2 cached to L2-only, | |
3827 | * realloc to remove the L1 header. | |
3828 | */ | |
3829 | hdr = arc_hdr_realloc(hdr, hdr_full_cache, | |
3830 | hdr_l2only_cache); | |
34dc7c2f | 3831 | } else { |
ca0bf58d PS |
3832 | arc_change_state(arc_anon, hdr, hash_lock); |
3833 | arc_hdr_destroy(hdr); | |
34dc7c2f | 3834 | } |
ca0bf58d | 3835 | return (bytes_evicted); |
34dc7c2f BB |
3836 | } |
3837 | ||
ca0bf58d PS |
3838 | ASSERT(state == arc_mru || state == arc_mfu); |
3839 | evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; | |
34dc7c2f | 3840 | |
ca0bf58d PS |
3841 | /* prefetch buffers have a minimum lifespan */ |
3842 | if (HDR_IO_IN_PROGRESS(hdr) || | |
3843 | ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && | |
2b84817f TC |
3844 | ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < |
3845 | MSEC_TO_TICK(min_lifetime))) { | |
ca0bf58d PS |
3846 | ARCSTAT_BUMP(arcstat_evict_skip); |
3847 | return (bytes_evicted); | |
da8ccd0e PS |
3848 | } |
3849 | ||
ca0bf58d | 3850 | ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); |
ca0bf58d PS |
3851 | while (hdr->b_l1hdr.b_buf) { |
3852 | arc_buf_t *buf = hdr->b_l1hdr.b_buf; | |
3853 | if (!mutex_tryenter(&buf->b_evict_lock)) { | |
3854 | ARCSTAT_BUMP(arcstat_mutex_miss); | |
3855 | break; | |
3856 | } | |
3857 | if (buf->b_data != NULL) | |
d3c2ae1c GW |
3858 | bytes_evicted += HDR_GET_LSIZE(hdr); |
3859 | mutex_exit(&buf->b_evict_lock); | |
2aa34383 | 3860 | arc_buf_destroy_impl(buf); |
ca0bf58d | 3861 | } |
34dc7c2f | 3862 | |
ca0bf58d | 3863 | if (HDR_HAS_L2HDR(hdr)) { |
d3c2ae1c | 3864 | ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); |
ca0bf58d | 3865 | } else { |
d3c2ae1c GW |
3866 | if (l2arc_write_eligible(hdr->b_spa, hdr)) { |
3867 | ARCSTAT_INCR(arcstat_evict_l2_eligible, | |
3868 | HDR_GET_LSIZE(hdr)); | |
3869 | } else { | |
3870 | ARCSTAT_INCR(arcstat_evict_l2_ineligible, | |
3871 | HDR_GET_LSIZE(hdr)); | |
3872 | } | |
ca0bf58d | 3873 | } |
34dc7c2f | 3874 | |
d3c2ae1c GW |
3875 | if (hdr->b_l1hdr.b_bufcnt == 0) { |
3876 | arc_cksum_free(hdr); | |
3877 | ||
3878 | bytes_evicted += arc_hdr_size(hdr); | |
3879 | ||
3880 | /* | |
3881 | * If this hdr is being evicted and has a compressed | |
3882 | * buffer then we discard it here before we change states. | |
3883 | * This ensures that the accounting is updated correctly | |
a6255b7f | 3884 | * in arc_free_data_impl(). |
d3c2ae1c | 3885 | */ |
b5256303 TC |
3886 | if (hdr->b_l1hdr.b_pabd != NULL) |
3887 | arc_hdr_free_abd(hdr, B_FALSE); | |
3888 | ||
3889 | if (HDR_HAS_RABD(hdr)) | |
3890 | arc_hdr_free_abd(hdr, B_TRUE); | |
d3c2ae1c | 3891 | |
ca0bf58d PS |
3892 | arc_change_state(evicted_state, hdr, hash_lock); |
3893 | ASSERT(HDR_IN_HASH_TABLE(hdr)); | |
d3c2ae1c | 3894 | arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); |
ca0bf58d PS |
3895 | DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); |
3896 | } | |
34dc7c2f | 3897 | |
ca0bf58d | 3898 | return (bytes_evicted); |
34dc7c2f BB |
3899 | } |
3900 | ||
ca0bf58d PS |
3901 | static uint64_t |
3902 | arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, | |
3903 | uint64_t spa, int64_t bytes) | |
34dc7c2f | 3904 | { |
ca0bf58d PS |
3905 | multilist_sublist_t *mls; |
3906 | uint64_t bytes_evicted = 0; | |
3907 | arc_buf_hdr_t *hdr; | |
34dc7c2f | 3908 | kmutex_t *hash_lock; |
ca0bf58d | 3909 | int evict_count = 0; |
34dc7c2f | 3910 | |
ca0bf58d | 3911 | ASSERT3P(marker, !=, NULL); |
96c080cb | 3912 | IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); |
ca0bf58d PS |
3913 | |
3914 | mls = multilist_sublist_lock(ml, idx); | |
572e2857 | 3915 | |
ca0bf58d PS |
3916 | for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; |
3917 | hdr = multilist_sublist_prev(mls, marker)) { | |
3918 | if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || | |
3919 | (evict_count >= zfs_arc_evict_batch_limit)) | |
3920 | break; | |
3921 | ||
3922 | /* | |
3923 | * To keep our iteration location, move the marker | |
3924 | * forward. Since we're not holding hdr's hash lock, we | |
3925 | * must be very careful and not remove 'hdr' from the | |
3926 | * sublist. Otherwise, other consumers might mistake the | |
3927 | * 'hdr' as not being on a sublist when they call the | |
3928 | * multilist_link_active() function (they all rely on | |
3929 | * the hash lock protecting concurrent insertions and | |
3930 | * removals). multilist_sublist_move_forward() was | |
3931 | * specifically implemented to ensure this is the case | |
3932 | * (only 'marker' will be removed and re-inserted). | |
3933 | */ | |
3934 | multilist_sublist_move_forward(mls, marker); | |
3935 | ||
3936 | /* | |
3937 | * The only case where the b_spa field should ever be | |
3938 | * zero, is the marker headers inserted by | |
3939 | * arc_evict_state(). It's possible for multiple threads | |
3940 | * to be calling arc_evict_state() concurrently (e.g. | |
3941 | * dsl_pool_close() and zio_inject_fault()), so we must | |
3942 | * skip any markers we see from these other threads. | |
3943 | */ | |
2a432414 | 3944 | if (hdr->b_spa == 0) |
572e2857 BB |
3945 | continue; |
3946 | ||
ca0bf58d PS |
3947 | /* we're only interested in evicting buffers of a certain spa */ |
3948 | if (spa != 0 && hdr->b_spa != spa) { | |
3949 | ARCSTAT_BUMP(arcstat_evict_skip); | |
428870ff | 3950 | continue; |
ca0bf58d PS |
3951 | } |
3952 | ||
3953 | hash_lock = HDR_LOCK(hdr); | |
e8b96c60 MA |
3954 | |
3955 | /* | |
ca0bf58d PS |
3956 | * We aren't calling this function from any code path |
3957 | * that would already be holding a hash lock, so we're | |
3958 | * asserting on this assumption to be defensive in case | |
3959 | * this ever changes. Without this check, it would be | |
3960 | * possible to incorrectly increment arcstat_mutex_miss | |
3961 | * below (e.g. if the code changed such that we called | |
3962 | * this function with a hash lock held). | |
e8b96c60 | 3963 | */ |
ca0bf58d PS |
3964 | ASSERT(!MUTEX_HELD(hash_lock)); |
3965 | ||
34dc7c2f | 3966 | if (mutex_tryenter(hash_lock)) { |
ca0bf58d PS |
3967 | uint64_t evicted = arc_evict_hdr(hdr, hash_lock); |
3968 | mutex_exit(hash_lock); | |
34dc7c2f | 3969 | |
ca0bf58d | 3970 | bytes_evicted += evicted; |
34dc7c2f | 3971 | |
572e2857 | 3972 | /* |
ca0bf58d PS |
3973 | * If evicted is zero, arc_evict_hdr() must have |
3974 | * decided to skip this header, don't increment | |
3975 | * evict_count in this case. | |
572e2857 | 3976 | */ |
ca0bf58d PS |
3977 | if (evicted != 0) |
3978 | evict_count++; | |
3979 | ||
3980 | /* | |
3981 | * If arc_size isn't overflowing, signal any | |
3982 | * threads that might happen to be waiting. | |
3983 | * | |
3984 | * For each header evicted, we wake up a single | |
3985 | * thread. If we used cv_broadcast, we could | |
3986 | * wake up "too many" threads causing arc_size | |
3987 | * to significantly overflow arc_c; since | |
a6255b7f | 3988 | * arc_get_data_impl() doesn't check for overflow |
ca0bf58d PS |
3989 | * when it's woken up (it doesn't because it's |
3990 | * possible for the ARC to be overflowing while | |
3991 | * full of un-evictable buffers, and the | |
3992 | * function should proceed in this case). | |
3993 | * | |
3994 | * If threads are left sleeping, due to not | |
3995 | * using cv_broadcast, they will be woken up | |
3996 | * just before arc_reclaim_thread() sleeps. | |
3997 | */ | |
3998 | mutex_enter(&arc_reclaim_lock); | |
3999 | if (!arc_is_overflowing()) | |
4000 | cv_signal(&arc_reclaim_waiters_cv); | |
4001 | mutex_exit(&arc_reclaim_lock); | |
e8b96c60 | 4002 | } else { |
ca0bf58d | 4003 | ARCSTAT_BUMP(arcstat_mutex_miss); |
e8b96c60 | 4004 | } |
34dc7c2f | 4005 | } |
34dc7c2f | 4006 | |
ca0bf58d | 4007 | multilist_sublist_unlock(mls); |
34dc7c2f | 4008 | |
ca0bf58d | 4009 | return (bytes_evicted); |
34dc7c2f BB |
4010 | } |
4011 | ||
ca0bf58d PS |
4012 | /* |
4013 | * Evict buffers from the given arc state, until we've removed the | |
4014 | * specified number of bytes. Move the removed buffers to the | |
4015 | * appropriate evict state. | |
4016 | * | |
4017 | * This function makes a "best effort". It skips over any buffers | |
4018 | * it can't get a hash_lock on, and so, may not catch all candidates. | |
4019 | * It may also return without evicting as much space as requested. | |
4020 | * | |
4021 | * If bytes is specified using the special value ARC_EVICT_ALL, this | |
4022 | * will evict all available (i.e. unlocked and evictable) buffers from | |
4023 | * the given arc state; which is used by arc_flush(). | |
4024 | */ | |
4025 | static uint64_t | |
4026 | arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, | |
4027 | arc_buf_contents_t type) | |
34dc7c2f | 4028 | { |
ca0bf58d | 4029 | uint64_t total_evicted = 0; |
64fc7762 | 4030 | multilist_t *ml = state->arcs_list[type]; |
ca0bf58d PS |
4031 | int num_sublists; |
4032 | arc_buf_hdr_t **markers; | |
ca0bf58d | 4033 | |
96c080cb | 4034 | IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); |
ca0bf58d PS |
4035 | |
4036 | num_sublists = multilist_get_num_sublists(ml); | |
d164b209 BB |
4037 | |
4038 | /* | |
ca0bf58d PS |
4039 | * If we've tried to evict from each sublist, made some |
4040 | * progress, but still have not hit the target number of bytes | |
4041 | * to evict, we want to keep trying. The markers allow us to | |
4042 | * pick up where we left off for each individual sublist, rather | |
4043 | * than starting from the tail each time. | |
d164b209 | 4044 | */ |
ca0bf58d | 4045 | markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); |
1c27024e | 4046 | for (int i = 0; i < num_sublists; i++) { |
ca0bf58d | 4047 | multilist_sublist_t *mls; |
34dc7c2f | 4048 | |
ca0bf58d PS |
4049 | markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); |
4050 | ||
4051 | /* | |
4052 | * A b_spa of 0 is used to indicate that this header is | |
4053 | * a marker. This fact is used in arc_adjust_type() and | |
4054 | * arc_evict_state_impl(). | |
4055 | */ | |
4056 | markers[i]->b_spa = 0; | |
34dc7c2f | 4057 | |
ca0bf58d PS |
4058 | mls = multilist_sublist_lock(ml, i); |
4059 | multilist_sublist_insert_tail(mls, markers[i]); | |
4060 | multilist_sublist_unlock(mls); | |
34dc7c2f BB |
4061 | } |
4062 | ||
d164b209 | 4063 | /* |
ca0bf58d PS |
4064 | * While we haven't hit our target number of bytes to evict, or |
4065 | * we're evicting all available buffers. | |
d164b209 | 4066 | */ |
ca0bf58d | 4067 | while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { |
25458cbe TC |
4068 | int sublist_idx = multilist_get_random_index(ml); |
4069 | uint64_t scan_evicted = 0; | |
4070 | ||
4071 | /* | |
4072 | * Try to reduce pinned dnodes with a floor of arc_dnode_limit. | |
4073 | * Request that 10% of the LRUs be scanned by the superblock | |
4074 | * shrinker. | |
4075 | */ | |
4076 | if (type == ARC_BUFC_DATA && arc_dnode_size > arc_dnode_limit) | |
4077 | arc_prune_async((arc_dnode_size - arc_dnode_limit) / | |
4078 | sizeof (dnode_t) / zfs_arc_dnode_reduce_percent); | |
4079 | ||
ca0bf58d PS |
4080 | /* |
4081 | * Start eviction using a randomly selected sublist, | |
4082 | * this is to try and evenly balance eviction across all | |
4083 | * sublists. Always starting at the same sublist | |
4084 | * (e.g. index 0) would cause evictions to favor certain | |
4085 | * sublists over others. | |
4086 | */ | |
1c27024e | 4087 | for (int i = 0; i < num_sublists; i++) { |
ca0bf58d PS |
4088 | uint64_t bytes_remaining; |
4089 | uint64_t bytes_evicted; | |
d164b209 | 4090 | |
ca0bf58d PS |
4091 | if (bytes == ARC_EVICT_ALL) |
4092 | bytes_remaining = ARC_EVICT_ALL; | |
4093 | else if (total_evicted < bytes) | |
4094 | bytes_remaining = bytes - total_evicted; | |
4095 | else | |
4096 | break; | |
34dc7c2f | 4097 | |
ca0bf58d PS |
4098 | bytes_evicted = arc_evict_state_impl(ml, sublist_idx, |
4099 | markers[sublist_idx], spa, bytes_remaining); | |
4100 | ||
4101 | scan_evicted += bytes_evicted; | |
4102 | total_evicted += bytes_evicted; | |
4103 | ||
4104 | /* we've reached the end, wrap to the beginning */ | |
4105 | if (++sublist_idx >= num_sublists) | |
4106 | sublist_idx = 0; | |
4107 | } | |
4108 | ||
4109 | /* | |
4110 | * If we didn't evict anything during this scan, we have | |
4111 | * no reason to believe we'll evict more during another | |
4112 | * scan, so break the loop. | |
4113 | */ | |
4114 | if (scan_evicted == 0) { | |
4115 | /* This isn't possible, let's make that obvious */ | |
4116 | ASSERT3S(bytes, !=, 0); | |
34dc7c2f | 4117 | |
ca0bf58d PS |
4118 | /* |
4119 | * When bytes is ARC_EVICT_ALL, the only way to | |
4120 | * break the loop is when scan_evicted is zero. | |
4121 | * In that case, we actually have evicted enough, | |
4122 | * so we don't want to increment the kstat. | |
4123 | */ | |
4124 | if (bytes != ARC_EVICT_ALL) { | |
4125 | ASSERT3S(total_evicted, <, bytes); | |
4126 | ARCSTAT_BUMP(arcstat_evict_not_enough); | |
4127 | } | |
d164b209 | 4128 | |
ca0bf58d PS |
4129 | break; |
4130 | } | |
d164b209 | 4131 | } |
34dc7c2f | 4132 | |
1c27024e | 4133 | for (int i = 0; i < num_sublists; i++) { |
ca0bf58d PS |
4134 | multilist_sublist_t *mls = multilist_sublist_lock(ml, i); |
4135 | multilist_sublist_remove(mls, markers[i]); | |
4136 | multilist_sublist_unlock(mls); | |
34dc7c2f | 4137 | |
ca0bf58d | 4138 | kmem_cache_free(hdr_full_cache, markers[i]); |
34dc7c2f | 4139 | } |
ca0bf58d PS |
4140 | kmem_free(markers, sizeof (*markers) * num_sublists); |
4141 | ||
4142 | return (total_evicted); | |
4143 | } | |
4144 | ||
4145 | /* | |
4146 | * Flush all "evictable" data of the given type from the arc state | |
4147 | * specified. This will not evict any "active" buffers (i.e. referenced). | |
4148 | * | |
d3c2ae1c | 4149 | * When 'retry' is set to B_FALSE, the function will make a single pass |
ca0bf58d PS |
4150 | * over the state and evict any buffers that it can. Since it doesn't |
4151 | * continually retry the eviction, it might end up leaving some buffers | |
4152 | * in the ARC due to lock misses. | |
4153 | * | |
d3c2ae1c | 4154 | * When 'retry' is set to B_TRUE, the function will continually retry the |
ca0bf58d PS |
4155 | * eviction until *all* evictable buffers have been removed from the |
4156 | * state. As a result, if concurrent insertions into the state are | |
4157 | * allowed (e.g. if the ARC isn't shutting down), this function might | |
4158 | * wind up in an infinite loop, continually trying to evict buffers. | |
4159 | */ | |
4160 | static uint64_t | |
4161 | arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, | |
4162 | boolean_t retry) | |
4163 | { | |
4164 | uint64_t evicted = 0; | |
4165 | ||
d3c2ae1c | 4166 | while (refcount_count(&state->arcs_esize[type]) != 0) { |
ca0bf58d PS |
4167 | evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); |
4168 | ||
4169 | if (!retry) | |
4170 | break; | |
4171 | } | |
4172 | ||
4173 | return (evicted); | |
34dc7c2f BB |
4174 | } |
4175 | ||
ab26409d | 4176 | /* |
ef5b2e10 BB |
4177 | * Helper function for arc_prune_async() it is responsible for safely |
4178 | * handling the execution of a registered arc_prune_func_t. | |
ab26409d BB |
4179 | */ |
4180 | static void | |
f6046738 | 4181 | arc_prune_task(void *ptr) |
ab26409d | 4182 | { |
f6046738 BB |
4183 | arc_prune_t *ap = (arc_prune_t *)ptr; |
4184 | arc_prune_func_t *func = ap->p_pfunc; | |
ab26409d | 4185 | |
f6046738 BB |
4186 | if (func != NULL) |
4187 | func(ap->p_adjust, ap->p_private); | |
ab26409d | 4188 | |
4442f60d | 4189 | refcount_remove(&ap->p_refcnt, func); |
f6046738 | 4190 | } |
ab26409d | 4191 | |
f6046738 BB |
4192 | /* |
4193 | * Notify registered consumers they must drop holds on a portion of the ARC | |
4194 | * buffered they reference. This provides a mechanism to ensure the ARC can | |
4195 | * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This | |
4196 | * is analogous to dnlc_reduce_cache() but more generic. | |
4197 | * | |
ef5b2e10 | 4198 | * This operation is performed asynchronously so it may be safely called |
ca67b33a | 4199 | * in the context of the arc_reclaim_thread(). A reference is taken here |
f6046738 BB |
4200 | * for each registered arc_prune_t and the arc_prune_task() is responsible |
4201 | * for releasing it once the registered arc_prune_func_t has completed. | |
4202 | */ | |
4203 | static void | |
4204 | arc_prune_async(int64_t adjust) | |
4205 | { | |
4206 | arc_prune_t *ap; | |
ab26409d | 4207 | |
f6046738 BB |
4208 | mutex_enter(&arc_prune_mtx); |
4209 | for (ap = list_head(&arc_prune_list); ap != NULL; | |
4210 | ap = list_next(&arc_prune_list, ap)) { | |
ab26409d | 4211 | |
f6046738 BB |
4212 | if (refcount_count(&ap->p_refcnt) >= 2) |
4213 | continue; | |
ab26409d | 4214 | |
f6046738 BB |
4215 | refcount_add(&ap->p_refcnt, ap->p_pfunc); |
4216 | ap->p_adjust = adjust; | |
b60eac3d | 4217 | if (taskq_dispatch(arc_prune_taskq, arc_prune_task, |
48d3eb40 | 4218 | ap, TQ_SLEEP) == TASKQID_INVALID) { |
b60eac3d | 4219 | refcount_remove(&ap->p_refcnt, ap->p_pfunc); |
4220 | continue; | |
4221 | } | |
f6046738 | 4222 | ARCSTAT_BUMP(arcstat_prune); |
ab26409d | 4223 | } |
ab26409d BB |
4224 | mutex_exit(&arc_prune_mtx); |
4225 | } | |
4226 | ||
ca0bf58d PS |
4227 | /* |
4228 | * Evict the specified number of bytes from the state specified, | |
4229 | * restricting eviction to the spa and type given. This function | |
4230 | * prevents us from trying to evict more from a state's list than | |
4231 | * is "evictable", and to skip evicting altogether when passed a | |
4232 | * negative value for "bytes". In contrast, arc_evict_state() will | |
4233 | * evict everything it can, when passed a negative value for "bytes". | |
4234 | */ | |
4235 | static uint64_t | |
4236 | arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, | |
4237 | arc_buf_contents_t type) | |
4238 | { | |
4239 | int64_t delta; | |
4240 | ||
d3c2ae1c GW |
4241 | if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { |
4242 | delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); | |
ca0bf58d PS |
4243 | return (arc_evict_state(state, spa, delta, type)); |
4244 | } | |
4245 | ||
4246 | return (0); | |
4247 | } | |
4248 | ||
4249 | /* | |
4250 | * The goal of this function is to evict enough meta data buffers from the | |
4251 | * ARC in order to enforce the arc_meta_limit. Achieving this is slightly | |
4252 | * more complicated than it appears because it is common for data buffers | |
4253 | * to have holds on meta data buffers. In addition, dnode meta data buffers | |
4254 | * will be held by the dnodes in the block preventing them from being freed. | |
4255 | * This means we can't simply traverse the ARC and expect to always find | |
4256 | * enough unheld meta data buffer to release. | |
4257 | * | |
4258 | * Therefore, this function has been updated to make alternating passes | |
4259 | * over the ARC releasing data buffers and then newly unheld meta data | |
4260 | * buffers. This ensures forward progress is maintained and arc_meta_used | |
4261 | * will decrease. Normally this is sufficient, but if required the ARC | |
4262 | * will call the registered prune callbacks causing dentry and inodes to | |
4263 | * be dropped from the VFS cache. This will make dnode meta data buffers | |
4264 | * available for reclaim. | |
4265 | */ | |
4266 | static uint64_t | |
f6046738 | 4267 | arc_adjust_meta_balanced(void) |
ca0bf58d | 4268 | { |
25e2ab16 TC |
4269 | int64_t delta, prune = 0, adjustmnt; |
4270 | uint64_t total_evicted = 0; | |
ca0bf58d | 4271 | arc_buf_contents_t type = ARC_BUFC_DATA; |
ca67b33a | 4272 | int restarts = MAX(zfs_arc_meta_adjust_restarts, 0); |
ca0bf58d PS |
4273 | |
4274 | restart: | |
4275 | /* | |
4276 | * This slightly differs than the way we evict from the mru in | |
4277 | * arc_adjust because we don't have a "target" value (i.e. no | |
4278 | * "meta" arc_p). As a result, I think we can completely | |
4279 | * cannibalize the metadata in the MRU before we evict the | |
4280 | * metadata from the MFU. I think we probably need to implement a | |
4281 | * "metadata arc_p" value to do this properly. | |
4282 | */ | |
4283 | adjustmnt = arc_meta_used - arc_meta_limit; | |
4284 | ||
d3c2ae1c GW |
4285 | if (adjustmnt > 0 && refcount_count(&arc_mru->arcs_esize[type]) > 0) { |
4286 | delta = MIN(refcount_count(&arc_mru->arcs_esize[type]), | |
4287 | adjustmnt); | |
ca0bf58d PS |
4288 | total_evicted += arc_adjust_impl(arc_mru, 0, delta, type); |
4289 | adjustmnt -= delta; | |
4290 | } | |
4291 | ||
4292 | /* | |
4293 | * We can't afford to recalculate adjustmnt here. If we do, | |
4294 | * new metadata buffers can sneak into the MRU or ANON lists, | |
4295 | * thus penalize the MFU metadata. Although the fudge factor is | |
4296 | * small, it has been empirically shown to be significant for | |
4297 | * certain workloads (e.g. creating many empty directories). As | |
4298 | * such, we use the original calculation for adjustmnt, and | |
4299 | * simply decrement the amount of data evicted from the MRU. | |
4300 | */ | |
4301 | ||
d3c2ae1c GW |
4302 | if (adjustmnt > 0 && refcount_count(&arc_mfu->arcs_esize[type]) > 0) { |
4303 | delta = MIN(refcount_count(&arc_mfu->arcs_esize[type]), | |
4304 | adjustmnt); | |
ca0bf58d PS |
4305 | total_evicted += arc_adjust_impl(arc_mfu, 0, delta, type); |
4306 | } | |
4307 | ||
4308 | adjustmnt = arc_meta_used - arc_meta_limit; | |
4309 | ||
d3c2ae1c GW |
4310 | if (adjustmnt > 0 && |
4311 | refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) { | |
ca0bf58d | 4312 | delta = MIN(adjustmnt, |
d3c2ae1c | 4313 | refcount_count(&arc_mru_ghost->arcs_esize[type])); |
ca0bf58d PS |
4314 | total_evicted += arc_adjust_impl(arc_mru_ghost, 0, delta, type); |
4315 | adjustmnt -= delta; | |
4316 | } | |
4317 | ||
d3c2ae1c GW |
4318 | if (adjustmnt > 0 && |
4319 | refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) { | |
ca0bf58d | 4320 | delta = MIN(adjustmnt, |
d3c2ae1c | 4321 | refcount_count(&arc_mfu_ghost->arcs_esize[type])); |
ca0bf58d PS |
4322 | total_evicted += arc_adjust_impl(arc_mfu_ghost, 0, delta, type); |
4323 | } | |
4324 | ||
4325 | /* | |
4326 | * If after attempting to make the requested adjustment to the ARC | |
4327 | * the meta limit is still being exceeded then request that the | |
4328 | * higher layers drop some cached objects which have holds on ARC | |
4329 | * meta buffers. Requests to the upper layers will be made with | |
4330 | * increasingly large scan sizes until the ARC is below the limit. | |
4331 | */ | |
4332 | if (arc_meta_used > arc_meta_limit) { | |
4333 | if (type == ARC_BUFC_DATA) { | |
4334 | type = ARC_BUFC_METADATA; | |
4335 | } else { | |
4336 | type = ARC_BUFC_DATA; | |
4337 | ||
4338 | if (zfs_arc_meta_prune) { | |
4339 | prune += zfs_arc_meta_prune; | |
f6046738 | 4340 | arc_prune_async(prune); |
ca0bf58d PS |
4341 | } |
4342 | } | |
4343 | ||
4344 | if (restarts > 0) { | |
4345 | restarts--; | |
4346 | goto restart; | |
4347 | } | |
4348 | } | |
4349 | return (total_evicted); | |
4350 | } | |
4351 | ||
f6046738 BB |
4352 | /* |
4353 | * Evict metadata buffers from the cache, such that arc_meta_used is | |
4354 | * capped by the arc_meta_limit tunable. | |
4355 | */ | |
4356 | static uint64_t | |
4357 | arc_adjust_meta_only(void) | |
4358 | { | |
4359 | uint64_t total_evicted = 0; | |
4360 | int64_t target; | |
4361 | ||
4362 | /* | |
4363 | * If we're over the meta limit, we want to evict enough | |
4364 | * metadata to get back under the meta limit. We don't want to | |
4365 | * evict so much that we drop the MRU below arc_p, though. If | |
4366 | * we're over the meta limit more than we're over arc_p, we | |
4367 | * evict some from the MRU here, and some from the MFU below. | |
4368 | */ | |
4369 | target = MIN((int64_t)(arc_meta_used - arc_meta_limit), | |
36da08ef PS |
4370 | (int64_t)(refcount_count(&arc_anon->arcs_size) + |
4371 | refcount_count(&arc_mru->arcs_size) - arc_p)); | |
f6046738 BB |
4372 | |
4373 | total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); | |
4374 | ||
4375 | /* | |
4376 | * Similar to the above, we want to evict enough bytes to get us | |
4377 | * below the meta limit, but not so much as to drop us below the | |
2aa34383 | 4378 | * space allotted to the MFU (which is defined as arc_c - arc_p). |
f6046738 BB |
4379 | */ |
4380 | target = MIN((int64_t)(arc_meta_used - arc_meta_limit), | |
36da08ef | 4381 | (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); |
f6046738 BB |
4382 | |
4383 | total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); | |
4384 | ||
4385 | return (total_evicted); | |
4386 | } | |
4387 | ||
4388 | static uint64_t | |
4389 | arc_adjust_meta(void) | |
4390 | { | |
4391 | if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY) | |
4392 | return (arc_adjust_meta_only()); | |
4393 | else | |
4394 | return (arc_adjust_meta_balanced()); | |
4395 | } | |
4396 | ||
ca0bf58d PS |
4397 | /* |
4398 | * Return the type of the oldest buffer in the given arc state | |
4399 | * | |
4400 | * This function will select a random sublist of type ARC_BUFC_DATA and | |
4401 | * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist | |
4402 | * is compared, and the type which contains the "older" buffer will be | |
4403 | * returned. | |
4404 | */ | |
4405 | static arc_buf_contents_t | |
4406 | arc_adjust_type(arc_state_t *state) | |
4407 | { | |
64fc7762 MA |
4408 | multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA]; |
4409 | multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA]; | |
ca0bf58d PS |
4410 | int data_idx = multilist_get_random_index(data_ml); |
4411 | int meta_idx = multilist_get_random_index(meta_ml); | |
4412 | multilist_sublist_t *data_mls; | |
4413 | multilist_sublist_t *meta_mls; | |
4414 | arc_buf_contents_t type; | |
4415 | arc_buf_hdr_t *data_hdr; | |
4416 | arc_buf_hdr_t *meta_hdr; | |
4417 | ||
4418 | /* | |
4419 | * We keep the sublist lock until we're finished, to prevent | |
4420 | * the headers from being destroyed via arc_evict_state(). | |
4421 | */ | |
4422 | data_mls = multilist_sublist_lock(data_ml, data_idx); | |
4423 | meta_mls = multilist_sublist_lock(meta_ml, meta_idx); | |
4424 | ||
4425 | /* | |
4426 | * These two loops are to ensure we skip any markers that | |
4427 | * might be at the tail of the lists due to arc_evict_state(). | |
4428 | */ | |
4429 | ||
4430 | for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; | |
4431 | data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { | |
4432 | if (data_hdr->b_spa != 0) | |
4433 | break; | |
4434 | } | |
4435 | ||
4436 | for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; | |
4437 | meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { | |
4438 | if (meta_hdr->b_spa != 0) | |
4439 | break; | |
4440 | } | |
4441 | ||
4442 | if (data_hdr == NULL && meta_hdr == NULL) { | |
4443 | type = ARC_BUFC_DATA; | |
4444 | } else if (data_hdr == NULL) { | |
4445 | ASSERT3P(meta_hdr, !=, NULL); | |
4446 | type = ARC_BUFC_METADATA; | |
4447 | } else if (meta_hdr == NULL) { | |
4448 | ASSERT3P(data_hdr, !=, NULL); | |
4449 | type = ARC_BUFC_DATA; | |
4450 | } else { | |
4451 | ASSERT3P(data_hdr, !=, NULL); | |
4452 | ASSERT3P(meta_hdr, !=, NULL); | |
4453 | ||
4454 | /* The headers can't be on the sublist without an L1 header */ | |
4455 | ASSERT(HDR_HAS_L1HDR(data_hdr)); | |
4456 | ASSERT(HDR_HAS_L1HDR(meta_hdr)); | |
4457 | ||
4458 | if (data_hdr->b_l1hdr.b_arc_access < | |
4459 | meta_hdr->b_l1hdr.b_arc_access) { | |
4460 | type = ARC_BUFC_DATA; | |
4461 | } else { | |
4462 | type = ARC_BUFC_METADATA; | |
4463 | } | |
4464 | } | |
4465 | ||
4466 | multilist_sublist_unlock(meta_mls); | |
4467 | multilist_sublist_unlock(data_mls); | |
4468 | ||
4469 | return (type); | |
4470 | } | |
4471 | ||
4472 | /* | |
4473 | * Evict buffers from the cache, such that arc_size is capped by arc_c. | |
4474 | */ | |
4475 | static uint64_t | |
4476 | arc_adjust(void) | |
4477 | { | |
4478 | uint64_t total_evicted = 0; | |
4479 | uint64_t bytes; | |
4480 | int64_t target; | |
4481 | ||
4482 | /* | |
4483 | * If we're over arc_meta_limit, we want to correct that before | |
4484 | * potentially evicting data buffers below. | |
4485 | */ | |
4486 | total_evicted += arc_adjust_meta(); | |
4487 | ||
4488 | /* | |
4489 | * Adjust MRU size | |
4490 | * | |
4491 | * If we're over the target cache size, we want to evict enough | |
4492 | * from the list to get back to our target size. We don't want | |
4493 | * to evict too much from the MRU, such that it drops below | |
4494 | * arc_p. So, if we're over our target cache size more than | |
4495 | * the MRU is over arc_p, we'll evict enough to get back to | |
4496 | * arc_p here, and then evict more from the MFU below. | |
4497 | */ | |
4498 | target = MIN((int64_t)(arc_size - arc_c), | |
36da08ef PS |
4499 | (int64_t)(refcount_count(&arc_anon->arcs_size) + |
4500 | refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); | |
ca0bf58d PS |
4501 | |
4502 | /* | |
4503 | * If we're below arc_meta_min, always prefer to evict data. | |
4504 | * Otherwise, try to satisfy the requested number of bytes to | |
4505 | * evict from the type which contains older buffers; in an | |
4506 | * effort to keep newer buffers in the cache regardless of their | |
4507 | * type. If we cannot satisfy the number of bytes from this | |
4508 | * type, spill over into the next type. | |
4509 | */ | |
4510 | if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && | |
4511 | arc_meta_used > arc_meta_min) { | |
4512 | bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); | |
4513 | total_evicted += bytes; | |
4514 | ||
4515 | /* | |
4516 | * If we couldn't evict our target number of bytes from | |
4517 | * metadata, we try to get the rest from data. | |
4518 | */ | |
4519 | target -= bytes; | |
4520 | ||
4521 | total_evicted += | |
4522 | arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); | |
4523 | } else { | |
4524 | bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); | |
4525 | total_evicted += bytes; | |
4526 | ||
4527 | /* | |
4528 | * If we couldn't evict our target number of bytes from | |
4529 | * data, we try to get the rest from metadata. | |
4530 | */ | |
4531 | target -= bytes; | |
4532 | ||
4533 | total_evicted += | |
4534 | arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); | |
4535 | } | |
4536 | ||
4537 | /* | |
4538 | * Adjust MFU size | |
4539 | * | |
4540 | * Now that we've tried to evict enough from the MRU to get its | |
4541 | * size back to arc_p, if we're still above the target cache | |
4542 | * size, we evict the rest from the MFU. | |
4543 | */ | |
4544 | target = arc_size - arc_c; | |
4545 | ||
a7b10a93 | 4546 | if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && |
ca0bf58d PS |
4547 | arc_meta_used > arc_meta_min) { |
4548 | bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); | |
4549 | total_evicted += bytes; | |
4550 | ||
4551 | /* | |
4552 | * If we couldn't evict our target number of bytes from | |
4553 | * metadata, we try to get the rest from data. | |
4554 | */ | |
4555 | target -= bytes; | |
4556 | ||
4557 | total_evicted += | |
4558 | arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); | |
4559 | } else { | |
4560 | bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); | |
4561 | total_evicted += bytes; | |
4562 | ||
4563 | /* | |
4564 | * If we couldn't evict our target number of bytes from | |
4565 | * data, we try to get the rest from data. | |
4566 | */ | |
4567 | target -= bytes; | |
4568 | ||
4569 | total_evicted += | |
4570 | arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); | |
4571 | } | |
4572 | ||
4573 | /* | |
4574 | * Adjust ghost lists | |
4575 | * | |
4576 | * In addition to the above, the ARC also defines target values | |
4577 | * for the ghost lists. The sum of the mru list and mru ghost | |
4578 | * list should never exceed the target size of the cache, and | |
4579 | * the sum of the mru list, mfu list, mru ghost list, and mfu | |
4580 | * ghost list should never exceed twice the target size of the | |
4581 | * cache. The following logic enforces these limits on the ghost | |
4582 | * caches, and evicts from them as needed. | |
4583 | */ | |
36da08ef PS |
4584 | target = refcount_count(&arc_mru->arcs_size) + |
4585 | refcount_count(&arc_mru_ghost->arcs_size) - arc_c; | |
ca0bf58d PS |
4586 | |
4587 | bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); | |
4588 | total_evicted += bytes; | |
4589 | ||
4590 | target -= bytes; | |
4591 | ||
4592 | total_evicted += | |
4593 | arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); | |
4594 | ||
4595 | /* | |
4596 | * We assume the sum of the mru list and mfu list is less than | |
4597 | * or equal to arc_c (we enforced this above), which means we | |
4598 | * can use the simpler of the two equations below: | |
4599 | * | |
4600 | * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c | |
4601 | * mru ghost + mfu ghost <= arc_c | |
4602 | */ | |
36da08ef PS |
4603 | target = refcount_count(&arc_mru_ghost->arcs_size) + |
4604 | refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; | |
ca0bf58d PS |
4605 | |
4606 | bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); | |
4607 | total_evicted += bytes; | |
4608 | ||
4609 | target -= bytes; | |
4610 | ||
4611 | total_evicted += | |
4612 | arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); | |
4613 | ||
4614 | return (total_evicted); | |
4615 | } | |
4616 | ||
ca0bf58d PS |
4617 | void |
4618 | arc_flush(spa_t *spa, boolean_t retry) | |
ab26409d | 4619 | { |
ca0bf58d | 4620 | uint64_t guid = 0; |
94520ca4 | 4621 | |
bc888666 | 4622 | /* |
d3c2ae1c | 4623 | * If retry is B_TRUE, a spa must not be specified since we have |
ca0bf58d PS |
4624 | * no good way to determine if all of a spa's buffers have been |
4625 | * evicted from an arc state. | |
bc888666 | 4626 | */ |
ca0bf58d | 4627 | ASSERT(!retry || spa == 0); |
d164b209 | 4628 | |
b9541d6b | 4629 | if (spa != NULL) |
3541dc6d | 4630 | guid = spa_load_guid(spa); |
d164b209 | 4631 | |
ca0bf58d PS |
4632 | (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); |
4633 | (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); | |
4634 | ||
4635 | (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); | |
4636 | (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); | |
4637 | ||
4638 | (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); | |
4639 | (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); | |
34dc7c2f | 4640 | |
ca0bf58d PS |
4641 | (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); |
4642 | (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); | |
34dc7c2f BB |
4643 | } |
4644 | ||
34dc7c2f | 4645 | void |
ca67b33a | 4646 | arc_shrink(int64_t to_free) |
34dc7c2f | 4647 | { |
1b8951b3 | 4648 | uint64_t c = arc_c; |
34dc7c2f | 4649 | |
1b8951b3 TC |
4650 | if (c > to_free && c - to_free > arc_c_min) { |
4651 | arc_c = c - to_free; | |
ca67b33a | 4652 | atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); |
34dc7c2f BB |
4653 | if (arc_c > arc_size) |
4654 | arc_c = MAX(arc_size, arc_c_min); | |
4655 | if (arc_p > arc_c) | |
4656 | arc_p = (arc_c >> 1); | |
4657 | ASSERT(arc_c >= arc_c_min); | |
4658 | ASSERT((int64_t)arc_p >= 0); | |
1b8951b3 TC |
4659 | } else { |
4660 | arc_c = arc_c_min; | |
34dc7c2f BB |
4661 | } |
4662 | ||
4663 | if (arc_size > arc_c) | |
ca0bf58d | 4664 | (void) arc_adjust(); |
34dc7c2f BB |
4665 | } |
4666 | ||
9edb3695 BB |
4667 | /* |
4668 | * Return maximum amount of memory that we could possibly use. Reduced | |
4669 | * to half of all memory in user space which is primarily used for testing. | |
4670 | */ | |
4671 | static uint64_t | |
4672 | arc_all_memory(void) | |
4673 | { | |
4674 | #ifdef _KERNEL | |
70f02287 BB |
4675 | #ifdef CONFIG_HIGHMEM |
4676 | return (ptob(totalram_pages - totalhigh_pages)); | |
4677 | #else | |
4678 | return (ptob(totalram_pages)); | |
4679 | #endif /* CONFIG_HIGHMEM */ | |
9edb3695 BB |
4680 | #else |
4681 | return (ptob(physmem) / 2); | |
70f02287 | 4682 | #endif /* _KERNEL */ |
9edb3695 BB |
4683 | } |
4684 | ||
70f02287 BB |
4685 | /* |
4686 | * Return the amount of memory that is considered free. In user space | |
4687 | * which is primarily used for testing we pretend that free memory ranges | |
4688 | * from 0-20% of all memory. | |
4689 | */ | |
787acae0 GDN |
4690 | static uint64_t |
4691 | arc_free_memory(void) | |
4692 | { | |
70f02287 BB |
4693 | #ifdef _KERNEL |
4694 | #ifdef CONFIG_HIGHMEM | |
4695 | struct sysinfo si; | |
4696 | si_meminfo(&si); | |
4697 | return (ptob(si.freeram - si.freehigh)); | |
4698 | #else | |
70f02287 | 4699 | return (ptob(nr_free_pages() + |
e9a77290 | 4700 | nr_inactive_file_pages() + |
4701 | nr_inactive_anon_pages() + | |
4702 | nr_slab_reclaimable_pages())); | |
4703 | ||
70f02287 BB |
4704 | #endif /* CONFIG_HIGHMEM */ |
4705 | #else | |
4706 | return (spa_get_random(arc_all_memory() * 20 / 100)); | |
4707 | #endif /* _KERNEL */ | |
787acae0 | 4708 | } |
787acae0 | 4709 | |
ca67b33a MA |
4710 | typedef enum free_memory_reason_t { |
4711 | FMR_UNKNOWN, | |
4712 | FMR_NEEDFREE, | |
4713 | FMR_LOTSFREE, | |
4714 | FMR_SWAPFS_MINFREE, | |
4715 | FMR_PAGES_PP_MAXIMUM, | |
4716 | FMR_HEAP_ARENA, | |
4717 | FMR_ZIO_ARENA, | |
4718 | } free_memory_reason_t; | |
4719 | ||
4720 | int64_t last_free_memory; | |
4721 | free_memory_reason_t last_free_reason; | |
4722 | ||
4723 | #ifdef _KERNEL | |
ca67b33a MA |
4724 | /* |
4725 | * Additional reserve of pages for pp_reserve. | |
4726 | */ | |
4727 | int64_t arc_pages_pp_reserve = 64; | |
4728 | ||
4729 | /* | |
4730 | * Additional reserve of pages for swapfs. | |
4731 | */ | |
4732 | int64_t arc_swapfs_reserve = 64; | |
ca67b33a MA |
4733 | #endif /* _KERNEL */ |
4734 | ||
4735 | /* | |
4736 | * Return the amount of memory that can be consumed before reclaim will be | |
4737 | * needed. Positive if there is sufficient free memory, negative indicates | |
4738 | * the amount of memory that needs to be freed up. | |
4739 | */ | |
4740 | static int64_t | |
4741 | arc_available_memory(void) | |
4742 | { | |
4743 | int64_t lowest = INT64_MAX; | |
4744 | free_memory_reason_t r = FMR_UNKNOWN; | |
ca67b33a | 4745 | #ifdef _KERNEL |
ca67b33a | 4746 | int64_t n; |
11f552fa | 4747 | #ifdef __linux__ |
70f02287 BB |
4748 | #ifdef freemem |
4749 | #undef freemem | |
4750 | #endif | |
11f552fa BB |
4751 | pgcnt_t needfree = btop(arc_need_free); |
4752 | pgcnt_t lotsfree = btop(arc_sys_free); | |
4753 | pgcnt_t desfree = 0; | |
70f02287 | 4754 | pgcnt_t freemem = btop(arc_free_memory()); |
9edb3695 BB |
4755 | #endif |
4756 | ||
ca67b33a MA |
4757 | if (needfree > 0) { |
4758 | n = PAGESIZE * (-needfree); | |
4759 | if (n < lowest) { | |
4760 | lowest = n; | |
4761 | r = FMR_NEEDFREE; | |
4762 | } | |
4763 | } | |
4764 | ||
4765 | /* | |
4766 | * check that we're out of range of the pageout scanner. It starts to | |
4767 | * schedule paging if freemem is less than lotsfree and needfree. | |
4768 | * lotsfree is the high-water mark for pageout, and needfree is the | |
4769 | * number of needed free pages. We add extra pages here to make sure | |
4770 | * the scanner doesn't start up while we're freeing memory. | |
4771 | */ | |
70f02287 | 4772 | n = PAGESIZE * (freemem - lotsfree - needfree - desfree); |
ca67b33a MA |
4773 | if (n < lowest) { |
4774 | lowest = n; | |
4775 | r = FMR_LOTSFREE; | |
4776 | } | |
4777 | ||
11f552fa | 4778 | #ifndef __linux__ |
ca67b33a MA |
4779 | /* |
4780 | * check to make sure that swapfs has enough space so that anon | |
4781 | * reservations can still succeed. anon_resvmem() checks that the | |
4782 | * availrmem is greater than swapfs_minfree, and the number of reserved | |
4783 | * swap pages. We also add a bit of extra here just to prevent | |
4784 | * circumstances from getting really dire. | |
4785 | */ | |
4786 | n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - | |
4787 | desfree - arc_swapfs_reserve); | |
4788 | if (n < lowest) { | |
4789 | lowest = n; | |
4790 | r = FMR_SWAPFS_MINFREE; | |
4791 | } | |
4792 | ||
ca67b33a MA |
4793 | /* |
4794 | * Check that we have enough availrmem that memory locking (e.g., via | |
4795 | * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum | |
4796 | * stores the number of pages that cannot be locked; when availrmem | |
4797 | * drops below pages_pp_maximum, page locking mechanisms such as | |
4798 | * page_pp_lock() will fail.) | |
4799 | */ | |
4800 | n = PAGESIZE * (availrmem - pages_pp_maximum - | |
4801 | arc_pages_pp_reserve); | |
4802 | if (n < lowest) { | |
4803 | lowest = n; | |
4804 | r = FMR_PAGES_PP_MAXIMUM; | |
4805 | } | |
11f552fa | 4806 | #endif |
ca67b33a | 4807 | |
70f02287 | 4808 | #if defined(_ILP32) |
ca67b33a | 4809 | /* |
70f02287 | 4810 | * If we're on a 32-bit platform, it's possible that we'll exhaust the |
ca67b33a MA |
4811 | * kernel heap space before we ever run out of available physical |
4812 | * memory. Most checks of the size of the heap_area compare against | |
4813 | * tune.t_minarmem, which is the minimum available real memory that we | |
4814 | * can have in the system. However, this is generally fixed at 25 pages | |
4815 | * which is so low that it's useless. In this comparison, we seek to | |
4816 | * calculate the total heap-size, and reclaim if more than 3/4ths of the | |
4817 | * heap is allocated. (Or, in the calculation, if less than 1/4th is | |
4818 | * free) | |
4819 | */ | |
4820 | n = vmem_size(heap_arena, VMEM_FREE) - | |
4821 | (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); | |
4822 | if (n < lowest) { | |
4823 | lowest = n; | |
4824 | r = FMR_HEAP_ARENA; | |
4825 | } | |
4826 | #endif | |
4827 | ||
4828 | /* | |
4829 | * If zio data pages are being allocated out of a separate heap segment, | |
4830 | * then enforce that the size of available vmem for this arena remains | |
d3c2ae1c | 4831 | * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free. |
ca67b33a | 4832 | * |
d3c2ae1c GW |
4833 | * Note that reducing the arc_zio_arena_free_shift keeps more virtual |
4834 | * memory (in the zio_arena) free, which can avoid memory | |
4835 | * fragmentation issues. | |
ca67b33a MA |
4836 | */ |
4837 | if (zio_arena != NULL) { | |
9edb3695 BB |
4838 | n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - |
4839 | (vmem_size(zio_arena, VMEM_ALLOC) >> | |
4840 | arc_zio_arena_free_shift); | |
ca67b33a MA |
4841 | if (n < lowest) { |
4842 | lowest = n; | |
4843 | r = FMR_ZIO_ARENA; | |
4844 | } | |
4845 | } | |
11f552fa | 4846 | #else /* _KERNEL */ |
ca67b33a MA |
4847 | /* Every 100 calls, free a small amount */ |
4848 | if (spa_get_random(100) == 0) | |
4849 | lowest = -1024; | |
11f552fa | 4850 | #endif /* _KERNEL */ |
ca67b33a MA |
4851 | |
4852 | last_free_memory = lowest; | |
4853 | last_free_reason = r; | |
4854 | ||
4855 | return (lowest); | |
4856 | } | |
4857 | ||
4858 | /* | |
4859 | * Determine if the system is under memory pressure and is asking | |
d3c2ae1c | 4860 | * to reclaim memory. A return value of B_TRUE indicates that the system |
ca67b33a MA |
4861 | * is under memory pressure and that the arc should adjust accordingly. |
4862 | */ | |
4863 | static boolean_t | |
4864 | arc_reclaim_needed(void) | |
4865 | { | |
4866 | return (arc_available_memory() < 0); | |
4867 | } | |
4868 | ||
34dc7c2f | 4869 | static void |
ca67b33a | 4870 | arc_kmem_reap_now(void) |
34dc7c2f BB |
4871 | { |
4872 | size_t i; | |
4873 | kmem_cache_t *prev_cache = NULL; | |
4874 | kmem_cache_t *prev_data_cache = NULL; | |
4875 | extern kmem_cache_t *zio_buf_cache[]; | |
4876 | extern kmem_cache_t *zio_data_buf_cache[]; | |
669dedb3 | 4877 | extern kmem_cache_t *range_seg_cache; |
34dc7c2f | 4878 | |
70f02287 | 4879 | #ifdef _KERNEL |
f6046738 BB |
4880 | if ((arc_meta_used >= arc_meta_limit) && zfs_arc_meta_prune) { |
4881 | /* | |
4882 | * We are exceeding our meta-data cache limit. | |
4883 | * Prune some entries to release holds on meta-data. | |
4884 | */ | |
ef5b2e10 | 4885 | arc_prune_async(zfs_arc_meta_prune); |
f6046738 | 4886 | } |
70f02287 BB |
4887 | #if defined(_ILP32) |
4888 | /* | |
4889 | * Reclaim unused memory from all kmem caches. | |
4890 | */ | |
4891 | kmem_reap(); | |
4892 | #endif | |
4893 | #endif | |
f6046738 | 4894 | |
34dc7c2f | 4895 | for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { |
70f02287 | 4896 | #if defined(_ILP32) |
d0c614ec | 4897 | /* reach upper limit of cache size on 32-bit */ |
4898 | if (zio_buf_cache[i] == NULL) | |
4899 | break; | |
4900 | #endif | |
34dc7c2f BB |
4901 | if (zio_buf_cache[i] != prev_cache) { |
4902 | prev_cache = zio_buf_cache[i]; | |
4903 | kmem_cache_reap_now(zio_buf_cache[i]); | |
4904 | } | |
4905 | if (zio_data_buf_cache[i] != prev_data_cache) { | |
4906 | prev_data_cache = zio_data_buf_cache[i]; | |
4907 | kmem_cache_reap_now(zio_data_buf_cache[i]); | |
4908 | } | |
4909 | } | |
ca0bf58d | 4910 | kmem_cache_reap_now(buf_cache); |
b9541d6b CW |
4911 | kmem_cache_reap_now(hdr_full_cache); |
4912 | kmem_cache_reap_now(hdr_l2only_cache); | |
669dedb3 | 4913 | kmem_cache_reap_now(range_seg_cache); |
ca67b33a MA |
4914 | |
4915 | if (zio_arena != NULL) { | |
4916 | /* | |
4917 | * Ask the vmem arena to reclaim unused memory from its | |
4918 | * quantum caches. | |
4919 | */ | |
4920 | vmem_qcache_reap(zio_arena); | |
4921 | } | |
34dc7c2f BB |
4922 | } |
4923 | ||
302f753f | 4924 | /* |
a6255b7f | 4925 | * Threads can block in arc_get_data_impl() waiting for this thread to evict |
ca0bf58d | 4926 | * enough data and signal them to proceed. When this happens, the threads in |
a6255b7f | 4927 | * arc_get_data_impl() are sleeping while holding the hash lock for their |
ca0bf58d PS |
4928 | * particular arc header. Thus, we must be careful to never sleep on a |
4929 | * hash lock in this thread. This is to prevent the following deadlock: | |
4930 | * | |
a6255b7f | 4931 | * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L", |
ca0bf58d PS |
4932 | * waiting for the reclaim thread to signal it. |
4933 | * | |
4934 | * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, | |
4935 | * fails, and goes to sleep forever. | |
4936 | * | |
4937 | * This possible deadlock is avoided by always acquiring a hash lock | |
4938 | * using mutex_tryenter() from arc_reclaim_thread(). | |
302f753f | 4939 | */ |
867959b5 | 4940 | /* ARGSUSED */ |
34dc7c2f | 4941 | static void |
c25b8f99 | 4942 | arc_reclaim_thread(void *unused) |
34dc7c2f | 4943 | { |
ca67b33a | 4944 | fstrans_cookie_t cookie = spl_fstrans_mark(); |
ae6d0c60 | 4945 | hrtime_t growtime = 0; |
34dc7c2f BB |
4946 | callb_cpr_t cpr; |
4947 | ||
ca0bf58d | 4948 | CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); |
34dc7c2f | 4949 | |
ca0bf58d | 4950 | mutex_enter(&arc_reclaim_lock); |
ca67b33a | 4951 | while (!arc_reclaim_thread_exit) { |
ca67b33a | 4952 | uint64_t evicted = 0; |
30fffb90 | 4953 | uint64_t need_free = arc_need_free; |
ca67b33a | 4954 | arc_tuning_update(); |
34dc7c2f | 4955 | |
d3c2ae1c GW |
4956 | /* |
4957 | * This is necessary in order for the mdb ::arc dcmd to | |
4958 | * show up to date information. Since the ::arc command | |
4959 | * does not call the kstat's update function, without | |
4960 | * this call, the command may show stale stats for the | |
4961 | * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even | |
4962 | * with this change, the data might be up to 1 second | |
4963 | * out of date; but that should suffice. The arc_state_t | |
4964 | * structures can be queried directly if more accurate | |
4965 | * information is needed. | |
4966 | */ | |
4967 | #ifndef __linux__ | |
4968 | if (arc_ksp != NULL) | |
4969 | arc_ksp->ks_update(arc_ksp, KSTAT_READ); | |
4970 | #endif | |
ca67b33a | 4971 | mutex_exit(&arc_reclaim_lock); |
34dc7c2f | 4972 | |
0a252dae GM |
4973 | /* |
4974 | * We call arc_adjust() before (possibly) calling | |
4975 | * arc_kmem_reap_now(), so that we can wake up | |
4976 | * arc_get_data_buf() sooner. | |
4977 | */ | |
4978 | evicted = arc_adjust(); | |
4979 | ||
4980 | int64_t free_memory = arc_available_memory(); | |
ca67b33a | 4981 | if (free_memory < 0) { |
34dc7c2f | 4982 | |
ca67b33a | 4983 | arc_no_grow = B_TRUE; |
b128c09f | 4984 | arc_warm = B_TRUE; |
34dc7c2f | 4985 | |
ca67b33a MA |
4986 | /* |
4987 | * Wait at least zfs_grow_retry (default 5) seconds | |
4988 | * before considering growing. | |
4989 | */ | |
ae6d0c60 | 4990 | growtime = gethrtime() + SEC2NSEC(arc_grow_retry); |
6a8f9b6b | 4991 | |
ca67b33a | 4992 | arc_kmem_reap_now(); |
34dc7c2f | 4993 | |
ca67b33a MA |
4994 | /* |
4995 | * If we are still low on memory, shrink the ARC | |
4996 | * so that we have arc_shrink_min free space. | |
4997 | */ | |
4998 | free_memory = arc_available_memory(); | |
34dc7c2f | 4999 | |
1c27024e DB |
5000 | int64_t to_free = |
5001 | (arc_c >> arc_shrink_shift) - free_memory; | |
ca67b33a MA |
5002 | if (to_free > 0) { |
5003 | #ifdef _KERNEL | |
30fffb90 | 5004 | to_free = MAX(to_free, need_free); |
ca67b33a MA |
5005 | #endif |
5006 | arc_shrink(to_free); | |
5007 | } | |
5008 | } else if (free_memory < arc_c >> arc_no_grow_shift) { | |
5009 | arc_no_grow = B_TRUE; | |
ae6d0c60 | 5010 | } else if (gethrtime() >= growtime) { |
ca67b33a MA |
5011 | arc_no_grow = B_FALSE; |
5012 | } | |
bce45ec9 | 5013 | |
ca67b33a | 5014 | mutex_enter(&arc_reclaim_lock); |
bce45ec9 | 5015 | |
ca67b33a MA |
5016 | /* |
5017 | * If evicted is zero, we couldn't evict anything via | |
5018 | * arc_adjust(). This could be due to hash lock | |
5019 | * collisions, but more likely due to the majority of | |
5020 | * arc buffers being unevictable. Therefore, even if | |
5021 | * arc_size is above arc_c, another pass is unlikely to | |
5022 | * be helpful and could potentially cause us to enter an | |
5023 | * infinite loop. | |
5024 | */ | |
5025 | if (arc_size <= arc_c || evicted == 0) { | |
5026 | /* | |
5027 | * We're either no longer overflowing, or we | |
5028 | * can't evict anything more, so we should wake | |
30fffb90 DB |
5029 | * up any threads before we go to sleep and remove |
5030 | * the bytes we were working on from arc_need_free | |
5031 | * since nothing more will be done here. | |
ca67b33a MA |
5032 | */ |
5033 | cv_broadcast(&arc_reclaim_waiters_cv); | |
30fffb90 | 5034 | ARCSTAT_INCR(arcstat_need_free, -need_free); |
bce45ec9 | 5035 | |
ca67b33a MA |
5036 | /* |
5037 | * Block until signaled, or after one second (we | |
5038 | * might need to perform arc_kmem_reap_now() | |
5039 | * even if we aren't being signalled) | |
5040 | */ | |
5041 | CALLB_CPR_SAFE_BEGIN(&cpr); | |
a9bb2b68 | 5042 | (void) cv_timedwait_sig_hires(&arc_reclaim_thread_cv, |
ae6d0c60 | 5043 | &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); |
ca67b33a MA |
5044 | CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); |
5045 | } | |
ca0bf58d | 5046 | } |
bce45ec9 | 5047 | |
d3c2ae1c | 5048 | arc_reclaim_thread_exit = B_FALSE; |
ca0bf58d PS |
5049 | cv_broadcast(&arc_reclaim_thread_cv); |
5050 | CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ | |
5051 | spl_fstrans_unmark(cookie); | |
5052 | thread_exit(); | |
5053 | } | |
5054 | ||
7cb67b45 BB |
5055 | #ifdef _KERNEL |
5056 | /* | |
302f753f BB |
5057 | * Determine the amount of memory eligible for eviction contained in the |
5058 | * ARC. All clean data reported by the ghost lists can always be safely | |
5059 | * evicted. Due to arc_c_min, the same does not hold for all clean data | |
5060 | * contained by the regular mru and mfu lists. | |
5061 | * | |
5062 | * In the case of the regular mru and mfu lists, we need to report as | |
5063 | * much clean data as possible, such that evicting that same reported | |
5064 | * data will not bring arc_size below arc_c_min. Thus, in certain | |
5065 | * circumstances, the total amount of clean data in the mru and mfu | |
5066 | * lists might not actually be evictable. | |
5067 | * | |
5068 | * The following two distinct cases are accounted for: | |
5069 | * | |
5070 | * 1. The sum of the amount of dirty data contained by both the mru and | |
5071 | * mfu lists, plus the ARC's other accounting (e.g. the anon list), | |
5072 | * is greater than or equal to arc_c_min. | |
5073 | * (i.e. amount of dirty data >= arc_c_min) | |
5074 | * | |
5075 | * This is the easy case; all clean data contained by the mru and mfu | |
5076 | * lists is evictable. Evicting all clean data can only drop arc_size | |
5077 | * to the amount of dirty data, which is greater than arc_c_min. | |
5078 | * | |
5079 | * 2. The sum of the amount of dirty data contained by both the mru and | |
5080 | * mfu lists, plus the ARC's other accounting (e.g. the anon list), | |
5081 | * is less than arc_c_min. | |
5082 | * (i.e. arc_c_min > amount of dirty data) | |
5083 | * | |
5084 | * 2.1. arc_size is greater than or equal arc_c_min. | |
5085 | * (i.e. arc_size >= arc_c_min > amount of dirty data) | |
5086 | * | |
5087 | * In this case, not all clean data from the regular mru and mfu | |
5088 | * lists is actually evictable; we must leave enough clean data | |
5089 | * to keep arc_size above arc_c_min. Thus, the maximum amount of | |
5090 | * evictable data from the two lists combined, is exactly the | |
5091 | * difference between arc_size and arc_c_min. | |
5092 | * | |
5093 | * 2.2. arc_size is less than arc_c_min | |
5094 | * (i.e. arc_c_min > arc_size > amount of dirty data) | |
5095 | * | |
5096 | * In this case, none of the data contained in the mru and mfu | |
5097 | * lists is evictable, even if it's clean. Since arc_size is | |
5098 | * already below arc_c_min, evicting any more would only | |
5099 | * increase this negative difference. | |
7cb67b45 | 5100 | */ |
302f753f | 5101 | static uint64_t |
4ea3f864 GM |
5102 | arc_evictable_memory(void) |
5103 | { | |
302f753f | 5104 | uint64_t arc_clean = |
d3c2ae1c GW |
5105 | refcount_count(&arc_mru->arcs_esize[ARC_BUFC_DATA]) + |
5106 | refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA]) + | |
5107 | refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_DATA]) + | |
5108 | refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); | |
302f753f BB |
5109 | uint64_t arc_dirty = MAX((int64_t)arc_size - (int64_t)arc_clean, 0); |
5110 | ||
03b60eee DB |
5111 | /* |
5112 | * Scale reported evictable memory in proportion to page cache, cap | |
5113 | * at specified min/max. | |
5114 | */ | |
e9a77290 | 5115 | uint64_t min = (ptob(nr_file_pages()) / 100) * zfs_arc_pc_percent; |
03b60eee DB |
5116 | min = MAX(arc_c_min, MIN(arc_c_max, min)); |
5117 | ||
5118 | if (arc_dirty >= min) | |
9b50146d | 5119 | return (arc_clean); |
302f753f | 5120 | |
03b60eee | 5121 | return (MAX((int64_t)arc_size - (int64_t)min, 0)); |
302f753f BB |
5122 | } |
5123 | ||
ed6e9cc2 TC |
5124 | /* |
5125 | * If sc->nr_to_scan is zero, the caller is requesting a query of the | |
5126 | * number of objects which can potentially be freed. If it is nonzero, | |
5127 | * the request is to free that many objects. | |
5128 | * | |
5129 | * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks | |
5130 | * in struct shrinker and also require the shrinker to return the number | |
5131 | * of objects freed. | |
5132 | * | |
5133 | * Older kernels require the shrinker to return the number of freeable | |
5134 | * objects following the freeing of nr_to_free. | |
5135 | */ | |
5136 | static spl_shrinker_t | |
7e7baeca | 5137 | __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc) |
7cb67b45 | 5138 | { |
ed6e9cc2 | 5139 | int64_t pages; |
7cb67b45 | 5140 | |
302f753f BB |
5141 | /* The arc is considered warm once reclaim has occurred */ |
5142 | if (unlikely(arc_warm == B_FALSE)) | |
5143 | arc_warm = B_TRUE; | |
7cb67b45 | 5144 | |
302f753f | 5145 | /* Return the potential number of reclaimable pages */ |
ed6e9cc2 | 5146 | pages = btop((int64_t)arc_evictable_memory()); |
302f753f BB |
5147 | if (sc->nr_to_scan == 0) |
5148 | return (pages); | |
3fd70ee6 BB |
5149 | |
5150 | /* Not allowed to perform filesystem reclaim */ | |
7e7baeca | 5151 | if (!(sc->gfp_mask & __GFP_FS)) |
ed6e9cc2 | 5152 | return (SHRINK_STOP); |
3fd70ee6 | 5153 | |
7cb67b45 | 5154 | /* Reclaim in progress */ |
b855550c DB |
5155 | if (mutex_tryenter(&arc_reclaim_lock) == 0) { |
5156 | ARCSTAT_INCR(arcstat_need_free, ptob(sc->nr_to_scan)); | |
2e91c2fb | 5157 | return (0); |
b855550c | 5158 | } |
7cb67b45 | 5159 | |
ca0bf58d PS |
5160 | mutex_exit(&arc_reclaim_lock); |
5161 | ||
302f753f BB |
5162 | /* |
5163 | * Evict the requested number of pages by shrinking arc_c the | |
44813aef | 5164 | * requested amount. |
302f753f BB |
5165 | */ |
5166 | if (pages > 0) { | |
ca67b33a | 5167 | arc_shrink(ptob(sc->nr_to_scan)); |
44813aef DB |
5168 | if (current_is_kswapd()) |
5169 | arc_kmem_reap_now(); | |
ed6e9cc2 | 5170 | #ifdef HAVE_SPLIT_SHRINKER_CALLBACK |
4149bf49 DB |
5171 | pages = MAX((int64_t)pages - |
5172 | (int64_t)btop(arc_evictable_memory()), 0); | |
ed6e9cc2 | 5173 | #else |
1e3cb67b | 5174 | pages = btop(arc_evictable_memory()); |
ed6e9cc2 | 5175 | #endif |
1a31dcf5 DB |
5176 | /* |
5177 | * We've shrunk what we can, wake up threads. | |
5178 | */ | |
5179 | cv_broadcast(&arc_reclaim_waiters_cv); | |
44813aef | 5180 | } else |
ed6e9cc2 | 5181 | pages = SHRINK_STOP; |
302f753f BB |
5182 | |
5183 | /* | |
5184 | * When direct reclaim is observed it usually indicates a rapid | |
5185 | * increase in memory pressure. This occurs because the kswapd | |
5186 | * threads were unable to asynchronously keep enough free memory | |
5187 | * available. In this case set arc_no_grow to briefly pause arc | |
5188 | * growth to avoid compounding the memory pressure. | |
5189 | */ | |
7cb67b45 | 5190 | if (current_is_kswapd()) { |
302f753f | 5191 | ARCSTAT_BUMP(arcstat_memory_indirect_count); |
7cb67b45 | 5192 | } else { |
302f753f | 5193 | arc_no_grow = B_TRUE; |
44813aef | 5194 | arc_kmem_reap_now(); |
302f753f | 5195 | ARCSTAT_BUMP(arcstat_memory_direct_count); |
7cb67b45 BB |
5196 | } |
5197 | ||
1e3cb67b | 5198 | return (pages); |
7cb67b45 | 5199 | } |
7e7baeca | 5200 | SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func); |
7cb67b45 BB |
5201 | |
5202 | SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS); | |
5203 | #endif /* _KERNEL */ | |
5204 | ||
34dc7c2f BB |
5205 | /* |
5206 | * Adapt arc info given the number of bytes we are trying to add and | |
4e33ba4c | 5207 | * the state that we are coming from. This function is only called |
34dc7c2f BB |
5208 | * when we are adding new content to the cache. |
5209 | */ | |
5210 | static void | |
5211 | arc_adapt(int bytes, arc_state_t *state) | |
5212 | { | |
5213 | int mult; | |
728d6ae9 | 5214 | uint64_t arc_p_min = (arc_c >> arc_p_min_shift); |
36da08ef PS |
5215 | int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); |
5216 | int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); | |
34dc7c2f BB |
5217 | |
5218 | if (state == arc_l2c_only) | |
5219 | return; | |
5220 | ||
5221 | ASSERT(bytes > 0); | |
5222 | /* | |
5223 | * Adapt the target size of the MRU list: | |
5224 | * - if we just hit in the MRU ghost list, then increase | |
5225 | * the target size of the MRU list. | |
5226 | * - if we just hit in the MFU ghost list, then increase | |
5227 | * the target size of the MFU list by decreasing the | |
5228 | * target size of the MRU list. | |
5229 | */ | |
5230 | if (state == arc_mru_ghost) { | |
36da08ef | 5231 | mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); |
62422785 PS |
5232 | if (!zfs_arc_p_dampener_disable) |
5233 | mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ | |
34dc7c2f | 5234 | |
728d6ae9 | 5235 | arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); |
34dc7c2f | 5236 | } else if (state == arc_mfu_ghost) { |
d164b209 BB |
5237 | uint64_t delta; |
5238 | ||
36da08ef | 5239 | mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); |
62422785 PS |
5240 | if (!zfs_arc_p_dampener_disable) |
5241 | mult = MIN(mult, 10); | |
34dc7c2f | 5242 | |
d164b209 | 5243 | delta = MIN(bytes * mult, arc_p); |
728d6ae9 | 5244 | arc_p = MAX(arc_p_min, arc_p - delta); |
34dc7c2f BB |
5245 | } |
5246 | ASSERT((int64_t)arc_p >= 0); | |
5247 | ||
ca67b33a MA |
5248 | if (arc_reclaim_needed()) { |
5249 | cv_signal(&arc_reclaim_thread_cv); | |
5250 | return; | |
5251 | } | |
5252 | ||
34dc7c2f BB |
5253 | if (arc_no_grow) |
5254 | return; | |
5255 | ||
5256 | if (arc_c >= arc_c_max) | |
5257 | return; | |
5258 | ||
5259 | /* | |
5260 | * If we're within (2 * maxblocksize) bytes of the target | |
5261 | * cache size, increment the target cache size | |
5262 | */ | |
935434ef | 5263 | ASSERT3U(arc_c, >=, 2ULL << SPA_MAXBLOCKSHIFT); |
121b3cae | 5264 | if (arc_size >= arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { |
34dc7c2f BB |
5265 | atomic_add_64(&arc_c, (int64_t)bytes); |
5266 | if (arc_c > arc_c_max) | |
5267 | arc_c = arc_c_max; | |
5268 | else if (state == arc_anon) | |
5269 | atomic_add_64(&arc_p, (int64_t)bytes); | |
5270 | if (arc_p > arc_c) | |
5271 | arc_p = arc_c; | |
5272 | } | |
5273 | ASSERT((int64_t)arc_p >= 0); | |
5274 | } | |
5275 | ||
5276 | /* | |
ca0bf58d PS |
5277 | * Check if arc_size has grown past our upper threshold, determined by |
5278 | * zfs_arc_overflow_shift. | |
34dc7c2f | 5279 | */ |
ca0bf58d PS |
5280 | static boolean_t |
5281 | arc_is_overflowing(void) | |
34dc7c2f | 5282 | { |
ca0bf58d PS |
5283 | /* Always allow at least one block of overflow */ |
5284 | uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, | |
5285 | arc_c >> zfs_arc_overflow_shift); | |
34dc7c2f | 5286 | |
ca0bf58d | 5287 | return (arc_size >= arc_c + overflow); |
34dc7c2f BB |
5288 | } |
5289 | ||
a6255b7f DQ |
5290 | static abd_t * |
5291 | arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag) | |
5292 | { | |
5293 | arc_buf_contents_t type = arc_buf_type(hdr); | |
5294 | ||
5295 | arc_get_data_impl(hdr, size, tag); | |
5296 | if (type == ARC_BUFC_METADATA) { | |
5297 | return (abd_alloc(size, B_TRUE)); | |
5298 | } else { | |
5299 | ASSERT(type == ARC_BUFC_DATA); | |
5300 | return (abd_alloc(size, B_FALSE)); | |
5301 | } | |
5302 | } | |
5303 | ||
5304 | static void * | |
5305 | arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) | |
5306 | { | |
5307 | arc_buf_contents_t type = arc_buf_type(hdr); | |
5308 | ||
5309 | arc_get_data_impl(hdr, size, tag); | |
5310 | if (type == ARC_BUFC_METADATA) { | |
5311 | return (zio_buf_alloc(size)); | |
5312 | } else { | |
5313 | ASSERT(type == ARC_BUFC_DATA); | |
5314 | return (zio_data_buf_alloc(size)); | |
5315 | } | |
5316 | } | |
5317 | ||
34dc7c2f | 5318 | /* |
d3c2ae1c GW |
5319 | * Allocate a block and return it to the caller. If we are hitting the |
5320 | * hard limit for the cache size, we must sleep, waiting for the eviction | |
5321 | * thread to catch up. If we're past the target size but below the hard | |
5322 | * limit, we'll only signal the reclaim thread and continue on. | |
34dc7c2f | 5323 | */ |
a6255b7f DQ |
5324 | static void |
5325 | arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) | |
34dc7c2f | 5326 | { |
a6255b7f DQ |
5327 | arc_state_t *state = hdr->b_l1hdr.b_state; |
5328 | arc_buf_contents_t type = arc_buf_type(hdr); | |
34dc7c2f BB |
5329 | |
5330 | arc_adapt(size, state); | |
5331 | ||
5332 | /* | |
ca0bf58d PS |
5333 | * If arc_size is currently overflowing, and has grown past our |
5334 | * upper limit, we must be adding data faster than the evict | |
5335 | * thread can evict. Thus, to ensure we don't compound the | |
5336 | * problem by adding more data and forcing arc_size to grow even | |
5337 | * further past it's target size, we halt and wait for the | |
5338 | * eviction thread to catch up. | |
5339 | * | |
5340 | * It's also possible that the reclaim thread is unable to evict | |
5341 | * enough buffers to get arc_size below the overflow limit (e.g. | |
5342 | * due to buffers being un-evictable, or hash lock collisions). | |
5343 | * In this case, we want to proceed regardless if we're | |
5344 | * overflowing; thus we don't use a while loop here. | |
34dc7c2f | 5345 | */ |
ca0bf58d PS |
5346 | if (arc_is_overflowing()) { |
5347 | mutex_enter(&arc_reclaim_lock); | |
5348 | ||
5349 | /* | |
5350 | * Now that we've acquired the lock, we may no longer be | |
5351 | * over the overflow limit, lets check. | |
5352 | * | |
5353 | * We're ignoring the case of spurious wake ups. If that | |
5354 | * were to happen, it'd let this thread consume an ARC | |
5355 | * buffer before it should have (i.e. before we're under | |
5356 | * the overflow limit and were signalled by the reclaim | |
5357 | * thread). As long as that is a rare occurrence, it | |
5358 | * shouldn't cause any harm. | |
5359 | */ | |
5360 | if (arc_is_overflowing()) { | |
5361 | cv_signal(&arc_reclaim_thread_cv); | |
5362 | cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); | |
34dc7c2f | 5363 | } |
34dc7c2f | 5364 | |
ca0bf58d | 5365 | mutex_exit(&arc_reclaim_lock); |
34dc7c2f | 5366 | } |
ab26409d | 5367 | |
d3c2ae1c | 5368 | VERIFY3U(hdr->b_type, ==, type); |
da8ccd0e | 5369 | if (type == ARC_BUFC_METADATA) { |
ca0bf58d PS |
5370 | arc_space_consume(size, ARC_SPACE_META); |
5371 | } else { | |
ca0bf58d | 5372 | arc_space_consume(size, ARC_SPACE_DATA); |
da8ccd0e PS |
5373 | } |
5374 | ||
34dc7c2f BB |
5375 | /* |
5376 | * Update the state size. Note that ghost states have a | |
5377 | * "ghost size" and so don't need to be updated. | |
5378 | */ | |
d3c2ae1c | 5379 | if (!GHOST_STATE(state)) { |
34dc7c2f | 5380 | |
d3c2ae1c | 5381 | (void) refcount_add_many(&state->arcs_size, size, tag); |
ca0bf58d PS |
5382 | |
5383 | /* | |
5384 | * If this is reached via arc_read, the link is | |
5385 | * protected by the hash lock. If reached via | |
5386 | * arc_buf_alloc, the header should not be accessed by | |
5387 | * any other thread. And, if reached via arc_read_done, | |
5388 | * the hash lock will protect it if it's found in the | |
5389 | * hash table; otherwise no other thread should be | |
5390 | * trying to [add|remove]_reference it. | |
5391 | */ | |
5392 | if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { | |
b9541d6b | 5393 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
d3c2ae1c GW |
5394 | (void) refcount_add_many(&state->arcs_esize[type], |
5395 | size, tag); | |
34dc7c2f | 5396 | } |
d3c2ae1c | 5397 | |
34dc7c2f BB |
5398 | /* |
5399 | * If we are growing the cache, and we are adding anonymous | |
5400 | * data, and we have outgrown arc_p, update arc_p | |
5401 | */ | |
ca0bf58d | 5402 | if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && |
36da08ef PS |
5403 | (refcount_count(&arc_anon->arcs_size) + |
5404 | refcount_count(&arc_mru->arcs_size) > arc_p)) | |
34dc7c2f BB |
5405 | arc_p = MIN(arc_c, arc_p + size); |
5406 | } | |
a6255b7f DQ |
5407 | } |
5408 | ||
5409 | static void | |
5410 | arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag) | |
5411 | { | |
5412 | arc_free_data_impl(hdr, size, tag); | |
5413 | abd_free(abd); | |
5414 | } | |
5415 | ||
5416 | static void | |
5417 | arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag) | |
5418 | { | |
5419 | arc_buf_contents_t type = arc_buf_type(hdr); | |
5420 | ||
5421 | arc_free_data_impl(hdr, size, tag); | |
5422 | if (type == ARC_BUFC_METADATA) { | |
5423 | zio_buf_free(buf, size); | |
5424 | } else { | |
5425 | ASSERT(type == ARC_BUFC_DATA); | |
5426 | zio_data_buf_free(buf, size); | |
5427 | } | |
d3c2ae1c GW |
5428 | } |
5429 | ||
5430 | /* | |
5431 | * Free the arc data buffer. | |
5432 | */ | |
5433 | static void | |
a6255b7f | 5434 | arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) |
d3c2ae1c GW |
5435 | { |
5436 | arc_state_t *state = hdr->b_l1hdr.b_state; | |
5437 | arc_buf_contents_t type = arc_buf_type(hdr); | |
5438 | ||
5439 | /* protected by hash lock, if in the hash table */ | |
5440 | if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { | |
5441 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); | |
5442 | ASSERT(state != arc_anon && state != arc_l2c_only); | |
5443 | ||
5444 | (void) refcount_remove_many(&state->arcs_esize[type], | |
5445 | size, tag); | |
5446 | } | |
5447 | (void) refcount_remove_many(&state->arcs_size, size, tag); | |
5448 | ||
5449 | VERIFY3U(hdr->b_type, ==, type); | |
5450 | if (type == ARC_BUFC_METADATA) { | |
d3c2ae1c GW |
5451 | arc_space_return(size, ARC_SPACE_META); |
5452 | } else { | |
5453 | ASSERT(type == ARC_BUFC_DATA); | |
d3c2ae1c GW |
5454 | arc_space_return(size, ARC_SPACE_DATA); |
5455 | } | |
34dc7c2f BB |
5456 | } |
5457 | ||
5458 | /* | |
5459 | * This routine is called whenever a buffer is accessed. | |
5460 | * NOTE: the hash lock is dropped in this function. | |
5461 | */ | |
5462 | static void | |
2a432414 | 5463 | arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) |
34dc7c2f | 5464 | { |
428870ff BB |
5465 | clock_t now; |
5466 | ||
34dc7c2f | 5467 | ASSERT(MUTEX_HELD(hash_lock)); |
b9541d6b | 5468 | ASSERT(HDR_HAS_L1HDR(hdr)); |
34dc7c2f | 5469 | |
b9541d6b | 5470 | if (hdr->b_l1hdr.b_state == arc_anon) { |
34dc7c2f BB |
5471 | /* |
5472 | * This buffer is not in the cache, and does not | |
5473 | * appear in our "ghost" list. Add the new buffer | |
5474 | * to the MRU state. | |
5475 | */ | |
5476 | ||
b9541d6b CW |
5477 | ASSERT0(hdr->b_l1hdr.b_arc_access); |
5478 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); | |
2a432414 GW |
5479 | DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); |
5480 | arc_change_state(arc_mru, hdr, hash_lock); | |
34dc7c2f | 5481 | |
b9541d6b | 5482 | } else if (hdr->b_l1hdr.b_state == arc_mru) { |
428870ff BB |
5483 | now = ddi_get_lbolt(); |
5484 | ||
34dc7c2f BB |
5485 | /* |
5486 | * If this buffer is here because of a prefetch, then either: | |
5487 | * - clear the flag if this is a "referencing" read | |
5488 | * (any subsequent access will bump this into the MFU state). | |
5489 | * or | |
5490 | * - move the buffer to the head of the list if this is | |
5491 | * another prefetch (to make it less likely to be evicted). | |
5492 | */ | |
d4a72f23 | 5493 | if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) { |
b9541d6b | 5494 | if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { |
ca0bf58d PS |
5495 | /* link protected by hash lock */ |
5496 | ASSERT(multilist_link_active( | |
b9541d6b | 5497 | &hdr->b_l1hdr.b_arc_node)); |
34dc7c2f | 5498 | } else { |
d4a72f23 TC |
5499 | arc_hdr_clear_flags(hdr, |
5500 | ARC_FLAG_PREFETCH | | |
5501 | ARC_FLAG_PRESCIENT_PREFETCH); | |
b9541d6b | 5502 | atomic_inc_32(&hdr->b_l1hdr.b_mru_hits); |
34dc7c2f BB |
5503 | ARCSTAT_BUMP(arcstat_mru_hits); |
5504 | } | |
b9541d6b | 5505 | hdr->b_l1hdr.b_arc_access = now; |
34dc7c2f BB |
5506 | return; |
5507 | } | |
5508 | ||
5509 | /* | |
5510 | * This buffer has been "accessed" only once so far, | |
5511 | * but it is still in the cache. Move it to the MFU | |
5512 | * state. | |
5513 | */ | |
b9541d6b CW |
5514 | if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access + |
5515 | ARC_MINTIME)) { | |
34dc7c2f BB |
5516 | /* |
5517 | * More than 125ms have passed since we | |
5518 | * instantiated this buffer. Move it to the | |
5519 | * most frequently used state. | |
5520 | */ | |
b9541d6b | 5521 | hdr->b_l1hdr.b_arc_access = now; |
2a432414 GW |
5522 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); |
5523 | arc_change_state(arc_mfu, hdr, hash_lock); | |
34dc7c2f | 5524 | } |
b9541d6b | 5525 | atomic_inc_32(&hdr->b_l1hdr.b_mru_hits); |
34dc7c2f | 5526 | ARCSTAT_BUMP(arcstat_mru_hits); |
b9541d6b | 5527 | } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { |
34dc7c2f BB |
5528 | arc_state_t *new_state; |
5529 | /* | |
5530 | * This buffer has been "accessed" recently, but | |
5531 | * was evicted from the cache. Move it to the | |
5532 | * MFU state. | |
5533 | */ | |
5534 | ||
d4a72f23 | 5535 | if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) { |
34dc7c2f | 5536 | new_state = arc_mru; |
d4a72f23 TC |
5537 | if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) { |
5538 | arc_hdr_clear_flags(hdr, | |
5539 | ARC_FLAG_PREFETCH | | |
5540 | ARC_FLAG_PRESCIENT_PREFETCH); | |
5541 | } | |
2a432414 | 5542 | DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); |
34dc7c2f BB |
5543 | } else { |
5544 | new_state = arc_mfu; | |
2a432414 | 5545 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); |
34dc7c2f BB |
5546 | } |
5547 | ||
b9541d6b | 5548 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); |
2a432414 | 5549 | arc_change_state(new_state, hdr, hash_lock); |
34dc7c2f | 5550 | |
b9541d6b | 5551 | atomic_inc_32(&hdr->b_l1hdr.b_mru_ghost_hits); |
34dc7c2f | 5552 | ARCSTAT_BUMP(arcstat_mru_ghost_hits); |
b9541d6b | 5553 | } else if (hdr->b_l1hdr.b_state == arc_mfu) { |
34dc7c2f BB |
5554 | /* |
5555 | * This buffer has been accessed more than once and is | |
5556 | * still in the cache. Keep it in the MFU state. | |
5557 | * | |
5558 | * NOTE: an add_reference() that occurred when we did | |
5559 | * the arc_read() will have kicked this off the list. | |
5560 | * If it was a prefetch, we will explicitly move it to | |
5561 | * the head of the list now. | |
5562 | */ | |
d4a72f23 | 5563 | |
b9541d6b | 5564 | atomic_inc_32(&hdr->b_l1hdr.b_mfu_hits); |
34dc7c2f | 5565 | ARCSTAT_BUMP(arcstat_mfu_hits); |
b9541d6b CW |
5566 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); |
5567 | } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { | |
34dc7c2f BB |
5568 | arc_state_t *new_state = arc_mfu; |
5569 | /* | |
5570 | * This buffer has been accessed more than once but has | |
5571 | * been evicted from the cache. Move it back to the | |
5572 | * MFU state. | |
5573 | */ | |
5574 | ||
d4a72f23 | 5575 | if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) { |
34dc7c2f BB |
5576 | /* |
5577 | * This is a prefetch access... | |
5578 | * move this block back to the MRU state. | |
5579 | */ | |
34dc7c2f BB |
5580 | new_state = arc_mru; |
5581 | } | |
5582 | ||
b9541d6b | 5583 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); |
2a432414 GW |
5584 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); |
5585 | arc_change_state(new_state, hdr, hash_lock); | |
34dc7c2f | 5586 | |
b9541d6b | 5587 | atomic_inc_32(&hdr->b_l1hdr.b_mfu_ghost_hits); |
34dc7c2f | 5588 | ARCSTAT_BUMP(arcstat_mfu_ghost_hits); |
b9541d6b | 5589 | } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { |
34dc7c2f BB |
5590 | /* |
5591 | * This buffer is on the 2nd Level ARC. | |
5592 | */ | |
5593 | ||
b9541d6b | 5594 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); |
2a432414 GW |
5595 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); |
5596 | arc_change_state(arc_mfu, hdr, hash_lock); | |
34dc7c2f | 5597 | } else { |
b9541d6b CW |
5598 | cmn_err(CE_PANIC, "invalid arc state 0x%p", |
5599 | hdr->b_l1hdr.b_state); | |
34dc7c2f BB |
5600 | } |
5601 | } | |
5602 | ||
0873bb63 BB |
5603 | /* |
5604 | * This routine is called by dbuf_hold() to update the arc_access() state | |
5605 | * which otherwise would be skipped for entries in the dbuf cache. | |
5606 | */ | |
5607 | void | |
5608 | arc_buf_access(arc_buf_t *buf) | |
5609 | { | |
5610 | mutex_enter(&buf->b_evict_lock); | |
5611 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
5612 | ||
5613 | /* | |
5614 | * Avoid taking the hash_lock when possible as an optimization. | |
5615 | * The header must be checked again under the hash_lock in order | |
5616 | * to handle the case where it is concurrently being released. | |
5617 | */ | |
5618 | if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { | |
5619 | mutex_exit(&buf->b_evict_lock); | |
5620 | return; | |
5621 | } | |
5622 | ||
5623 | kmutex_t *hash_lock = HDR_LOCK(hdr); | |
5624 | mutex_enter(hash_lock); | |
5625 | ||
5626 | if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { | |
5627 | mutex_exit(hash_lock); | |
5628 | mutex_exit(&buf->b_evict_lock); | |
5629 | ARCSTAT_BUMP(arcstat_access_skip); | |
5630 | return; | |
5631 | } | |
5632 | ||
5633 | mutex_exit(&buf->b_evict_lock); | |
5634 | ||
5635 | ASSERT(hdr->b_l1hdr.b_state == arc_mru || | |
5636 | hdr->b_l1hdr.b_state == arc_mfu); | |
5637 | ||
5638 | DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); | |
5639 | arc_access(hdr, hash_lock); | |
5640 | mutex_exit(hash_lock); | |
5641 | ||
5642 | ARCSTAT_BUMP(arcstat_hits); | |
5643 | ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr) && !HDR_PRESCIENT_PREFETCH(hdr), | |
5644 | demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits); | |
5645 | } | |
5646 | ||
b5256303 | 5647 | /* a generic arc_read_done_func_t which you can use */ |
34dc7c2f BB |
5648 | /* ARGSUSED */ |
5649 | void | |
d4a72f23 TC |
5650 | arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, |
5651 | arc_buf_t *buf, void *arg) | |
34dc7c2f | 5652 | { |
d4a72f23 TC |
5653 | if (buf == NULL) |
5654 | return; | |
5655 | ||
5656 | bcopy(buf->b_data, arg, arc_buf_size(buf)); | |
d3c2ae1c | 5657 | arc_buf_destroy(buf, arg); |
34dc7c2f BB |
5658 | } |
5659 | ||
b5256303 | 5660 | /* a generic arc_read_done_func_t */ |
d4a72f23 | 5661 | /* ARGSUSED */ |
34dc7c2f | 5662 | void |
d4a72f23 TC |
5663 | arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, |
5664 | arc_buf_t *buf, void *arg) | |
34dc7c2f BB |
5665 | { |
5666 | arc_buf_t **bufp = arg; | |
d4a72f23 TC |
5667 | |
5668 | if (buf == NULL) { | |
34dc7c2f BB |
5669 | *bufp = NULL; |
5670 | } else { | |
5671 | *bufp = buf; | |
428870ff | 5672 | ASSERT(buf->b_data); |
34dc7c2f BB |
5673 | } |
5674 | } | |
5675 | ||
d3c2ae1c GW |
5676 | static void |
5677 | arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) | |
5678 | { | |
5679 | if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { | |
5680 | ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); | |
b5256303 | 5681 | ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF); |
d3c2ae1c GW |
5682 | } else { |
5683 | if (HDR_COMPRESSION_ENABLED(hdr)) { | |
b5256303 | 5684 | ASSERT3U(arc_hdr_get_compress(hdr), ==, |
d3c2ae1c GW |
5685 | BP_GET_COMPRESS(bp)); |
5686 | } | |
5687 | ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); | |
5688 | ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); | |
b5256303 | 5689 | ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp)); |
d3c2ae1c GW |
5690 | } |
5691 | } | |
5692 | ||
34dc7c2f BB |
5693 | static void |
5694 | arc_read_done(zio_t *zio) | |
5695 | { | |
b5256303 | 5696 | blkptr_t *bp = zio->io_bp; |
d3c2ae1c | 5697 | arc_buf_hdr_t *hdr = zio->io_private; |
9b67f605 | 5698 | kmutex_t *hash_lock = NULL; |
524b4217 DK |
5699 | arc_callback_t *callback_list; |
5700 | arc_callback_t *acb; | |
2aa34383 | 5701 | boolean_t freeable = B_FALSE; |
a7004725 | 5702 | |
34dc7c2f BB |
5703 | /* |
5704 | * The hdr was inserted into hash-table and removed from lists | |
5705 | * prior to starting I/O. We should find this header, since | |
5706 | * it's in the hash table, and it should be legit since it's | |
5707 | * not possible to evict it during the I/O. The only possible | |
5708 | * reason for it not to be found is if we were freed during the | |
5709 | * read. | |
5710 | */ | |
9b67f605 | 5711 | if (HDR_IN_HASH_TABLE(hdr)) { |
31df97cd DB |
5712 | arc_buf_hdr_t *found; |
5713 | ||
9b67f605 MA |
5714 | ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); |
5715 | ASSERT3U(hdr->b_dva.dva_word[0], ==, | |
5716 | BP_IDENTITY(zio->io_bp)->dva_word[0]); | |
5717 | ASSERT3U(hdr->b_dva.dva_word[1], ==, | |
5718 | BP_IDENTITY(zio->io_bp)->dva_word[1]); | |
5719 | ||
31df97cd | 5720 | found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock); |
9b67f605 | 5721 | |
d3c2ae1c | 5722 | ASSERT((found == hdr && |
9b67f605 MA |
5723 | DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || |
5724 | (found == hdr && HDR_L2_READING(hdr))); | |
d3c2ae1c GW |
5725 | ASSERT3P(hash_lock, !=, NULL); |
5726 | } | |
5727 | ||
b5256303 TC |
5728 | if (BP_IS_PROTECTED(bp)) { |
5729 | hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp); | |
5730 | hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset; | |
5731 | zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt, | |
5732 | hdr->b_crypt_hdr.b_iv); | |
5733 | ||
5734 | if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) { | |
5735 | void *tmpbuf; | |
5736 | ||
5737 | tmpbuf = abd_borrow_buf_copy(zio->io_abd, | |
5738 | sizeof (zil_chain_t)); | |
5739 | zio_crypt_decode_mac_zil(tmpbuf, | |
5740 | hdr->b_crypt_hdr.b_mac); | |
5741 | abd_return_buf(zio->io_abd, tmpbuf, | |
5742 | sizeof (zil_chain_t)); | |
5743 | } else { | |
5744 | zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac); | |
5745 | } | |
5746 | } | |
5747 | ||
d4a72f23 | 5748 | if (zio->io_error == 0) { |
d3c2ae1c GW |
5749 | /* byteswap if necessary */ |
5750 | if (BP_SHOULD_BYTESWAP(zio->io_bp)) { | |
5751 | if (BP_GET_LEVEL(zio->io_bp) > 0) { | |
5752 | hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; | |
5753 | } else { | |
5754 | hdr->b_l1hdr.b_byteswap = | |
5755 | DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); | |
5756 | } | |
5757 | } else { | |
5758 | hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; | |
5759 | } | |
9b67f605 | 5760 | } |
34dc7c2f | 5761 | |
d3c2ae1c | 5762 | arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); |
b9541d6b | 5763 | if (l2arc_noprefetch && HDR_PREFETCH(hdr)) |
d3c2ae1c | 5764 | arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); |
34dc7c2f | 5765 | |
b9541d6b | 5766 | callback_list = hdr->b_l1hdr.b_acb; |
d3c2ae1c | 5767 | ASSERT3P(callback_list, !=, NULL); |
34dc7c2f | 5768 | |
d4a72f23 TC |
5769 | if (hash_lock && zio->io_error == 0 && |
5770 | hdr->b_l1hdr.b_state == arc_anon) { | |
428870ff BB |
5771 | /* |
5772 | * Only call arc_access on anonymous buffers. This is because | |
5773 | * if we've issued an I/O for an evicted buffer, we've already | |
5774 | * called arc_access (to prevent any simultaneous readers from | |
5775 | * getting confused). | |
5776 | */ | |
5777 | arc_access(hdr, hash_lock); | |
5778 | } | |
5779 | ||
524b4217 DK |
5780 | /* |
5781 | * If a read request has a callback (i.e. acb_done is not NULL), then we | |
5782 | * make a buf containing the data according to the parameters which were | |
5783 | * passed in. The implementation of arc_buf_alloc_impl() ensures that we | |
5784 | * aren't needlessly decompressing the data multiple times. | |
5785 | */ | |
a7004725 | 5786 | int callback_cnt = 0; |
2aa34383 DK |
5787 | for (acb = callback_list; acb != NULL; acb = acb->acb_next) { |
5788 | if (!acb->acb_done) | |
5789 | continue; | |
5790 | ||
2aa34383 | 5791 | callback_cnt++; |
524b4217 | 5792 | |
d4a72f23 TC |
5793 | if (zio->io_error != 0) |
5794 | continue; | |
5795 | ||
b5256303 | 5796 | int error = arc_buf_alloc_impl(hdr, zio->io_spa, |
be9a5c35 | 5797 | &acb->acb_zb, acb->acb_private, acb->acb_encrypted, |
d4a72f23 | 5798 | acb->acb_compressed, acb->acb_noauth, B_TRUE, |
440a3eb9 | 5799 | &acb->acb_buf); |
d4a72f23 | 5800 | if (error != 0) { |
2c24b5b1 TC |
5801 | (void) remove_reference(hdr, hash_lock, |
5802 | acb->acb_private); | |
5803 | arc_buf_destroy_impl(acb->acb_buf); | |
d4a72f23 TC |
5804 | acb->acb_buf = NULL; |
5805 | } | |
b5256303 TC |
5806 | |
5807 | /* | |
440a3eb9 | 5808 | * Assert non-speculative zios didn't fail because an |
b5256303 TC |
5809 | * encryption key wasn't loaded |
5810 | */ | |
a2c2ed1b | 5811 | ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) || |
be9a5c35 | 5812 | error != EACCES); |
b5256303 TC |
5813 | |
5814 | /* | |
5815 | * If we failed to decrypt, report an error now (as the zio | |
5816 | * layer would have done if it had done the transforms). | |
5817 | */ | |
5818 | if (error == ECKSUM) { | |
5819 | ASSERT(BP_IS_PROTECTED(bp)); | |
5820 | error = SET_ERROR(EIO); | |
b5256303 | 5821 | if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) { |
be9a5c35 | 5822 | spa_log_error(zio->io_spa, &acb->acb_zb); |
b5256303 | 5823 | zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, |
be9a5c35 | 5824 | zio->io_spa, NULL, &acb->acb_zb, zio, 0, 0); |
b5256303 TC |
5825 | } |
5826 | } | |
5827 | ||
d4a72f23 | 5828 | if (zio->io_error == 0) |
524b4217 | 5829 | zio->io_error = error; |
34dc7c2f | 5830 | } |
b9541d6b | 5831 | hdr->b_l1hdr.b_acb = NULL; |
d3c2ae1c | 5832 | arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); |
440a3eb9 | 5833 | if (callback_cnt == 0) |
b5256303 | 5834 | ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); |
34dc7c2f | 5835 | |
b9541d6b CW |
5836 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || |
5837 | callback_list != NULL); | |
34dc7c2f | 5838 | |
d4a72f23 | 5839 | if (zio->io_error == 0) { |
d3c2ae1c GW |
5840 | arc_hdr_verify(hdr, zio->io_bp); |
5841 | } else { | |
5842 | arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); | |
b9541d6b | 5843 | if (hdr->b_l1hdr.b_state != arc_anon) |
34dc7c2f BB |
5844 | arc_change_state(arc_anon, hdr, hash_lock); |
5845 | if (HDR_IN_HASH_TABLE(hdr)) | |
5846 | buf_hash_remove(hdr); | |
b9541d6b | 5847 | freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); |
34dc7c2f BB |
5848 | } |
5849 | ||
5850 | /* | |
5851 | * Broadcast before we drop the hash_lock to avoid the possibility | |
5852 | * that the hdr (and hence the cv) might be freed before we get to | |
5853 | * the cv_broadcast(). | |
5854 | */ | |
b9541d6b | 5855 | cv_broadcast(&hdr->b_l1hdr.b_cv); |
34dc7c2f | 5856 | |
b9541d6b | 5857 | if (hash_lock != NULL) { |
34dc7c2f BB |
5858 | mutex_exit(hash_lock); |
5859 | } else { | |
5860 | /* | |
5861 | * This block was freed while we waited for the read to | |
5862 | * complete. It has been removed from the hash table and | |
5863 | * moved to the anonymous state (so that it won't show up | |
5864 | * in the cache). | |
5865 | */ | |
b9541d6b CW |
5866 | ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); |
5867 | freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); | |
34dc7c2f BB |
5868 | } |
5869 | ||
5870 | /* execute each callback and free its structure */ | |
5871 | while ((acb = callback_list) != NULL) { | |
b5256303 | 5872 | if (acb->acb_done) { |
d4a72f23 TC |
5873 | acb->acb_done(zio, &zio->io_bookmark, zio->io_bp, |
5874 | acb->acb_buf, acb->acb_private); | |
b5256303 | 5875 | } |
34dc7c2f BB |
5876 | |
5877 | if (acb->acb_zio_dummy != NULL) { | |
5878 | acb->acb_zio_dummy->io_error = zio->io_error; | |
5879 | zio_nowait(acb->acb_zio_dummy); | |
5880 | } | |
5881 | ||
5882 | callback_list = acb->acb_next; | |
5883 | kmem_free(acb, sizeof (arc_callback_t)); | |
5884 | } | |
5885 | ||
5886 | if (freeable) | |
5887 | arc_hdr_destroy(hdr); | |
5888 | } | |
5889 | ||
5890 | /* | |
5c839890 | 5891 | * "Read" the block at the specified DVA (in bp) via the |
34dc7c2f BB |
5892 | * cache. If the block is found in the cache, invoke the provided |
5893 | * callback immediately and return. Note that the `zio' parameter | |
5894 | * in the callback will be NULL in this case, since no IO was | |
5895 | * required. If the block is not in the cache pass the read request | |
5896 | * on to the spa with a substitute callback function, so that the | |
5897 | * requested block will be added to the cache. | |
5898 | * | |
5899 | * If a read request arrives for a block that has a read in-progress, | |
5900 | * either wait for the in-progress read to complete (and return the | |
5901 | * results); or, if this is a read with a "done" func, add a record | |
5902 | * to the read to invoke the "done" func when the read completes, | |
5903 | * and return; or just return. | |
5904 | * | |
5905 | * arc_read_done() will invoke all the requested "done" functions | |
5906 | * for readers of this block. | |
5907 | */ | |
5908 | int | |
b5256303 TC |
5909 | arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, |
5910 | arc_read_done_func_t *done, void *private, zio_priority_t priority, | |
5911 | int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb) | |
34dc7c2f | 5912 | { |
9b67f605 | 5913 | arc_buf_hdr_t *hdr = NULL; |
9b67f605 | 5914 | kmutex_t *hash_lock = NULL; |
34dc7c2f | 5915 | zio_t *rzio; |
3541dc6d | 5916 | uint64_t guid = spa_load_guid(spa); |
b5256303 TC |
5917 | boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0; |
5918 | boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) && | |
5919 | (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0; | |
5920 | boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) && | |
5921 | (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0; | |
1421c891 | 5922 | int rc = 0; |
34dc7c2f | 5923 | |
9b67f605 MA |
5924 | ASSERT(!BP_IS_EMBEDDED(bp) || |
5925 | BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); | |
5926 | ||
34dc7c2f | 5927 | top: |
9b67f605 MA |
5928 | if (!BP_IS_EMBEDDED(bp)) { |
5929 | /* | |
5930 | * Embedded BP's have no DVA and require no I/O to "read". | |
5931 | * Create an anonymous arc buf to back it. | |
5932 | */ | |
5933 | hdr = buf_hash_find(guid, bp, &hash_lock); | |
5934 | } | |
5935 | ||
b5256303 TC |
5936 | /* |
5937 | * Determine if we have an L1 cache hit or a cache miss. For simplicity | |
5938 | * we maintain encrypted data seperately from compressed / uncompressed | |
5939 | * data. If the user is requesting raw encrypted data and we don't have | |
5940 | * that in the header we will read from disk to guarantee that we can | |
5941 | * get it even if the encryption keys aren't loaded. | |
5942 | */ | |
5943 | if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) || | |
5944 | (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) { | |
d3c2ae1c | 5945 | arc_buf_t *buf = NULL; |
2a432414 | 5946 | *arc_flags |= ARC_FLAG_CACHED; |
34dc7c2f BB |
5947 | |
5948 | if (HDR_IO_IN_PROGRESS(hdr)) { | |
a8b2e306 | 5949 | zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head; |
34dc7c2f | 5950 | |
a8b2e306 | 5951 | ASSERT3P(head_zio, !=, NULL); |
7f60329a MA |
5952 | if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && |
5953 | priority == ZIO_PRIORITY_SYNC_READ) { | |
5954 | /* | |
a8b2e306 TC |
5955 | * This is a sync read that needs to wait for |
5956 | * an in-flight async read. Request that the | |
5957 | * zio have its priority upgraded. | |
7f60329a | 5958 | */ |
a8b2e306 TC |
5959 | zio_change_priority(head_zio, priority); |
5960 | DTRACE_PROBE1(arc__async__upgrade__sync, | |
7f60329a | 5961 | arc_buf_hdr_t *, hdr); |
a8b2e306 | 5962 | ARCSTAT_BUMP(arcstat_async_upgrade_sync); |
7f60329a MA |
5963 | } |
5964 | if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { | |
d3c2ae1c GW |
5965 | arc_hdr_clear_flags(hdr, |
5966 | ARC_FLAG_PREDICTIVE_PREFETCH); | |
7f60329a MA |
5967 | } |
5968 | ||
2a432414 | 5969 | if (*arc_flags & ARC_FLAG_WAIT) { |
b9541d6b | 5970 | cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); |
34dc7c2f BB |
5971 | mutex_exit(hash_lock); |
5972 | goto top; | |
5973 | } | |
2a432414 | 5974 | ASSERT(*arc_flags & ARC_FLAG_NOWAIT); |
34dc7c2f BB |
5975 | |
5976 | if (done) { | |
7f60329a | 5977 | arc_callback_t *acb = NULL; |
34dc7c2f BB |
5978 | |
5979 | acb = kmem_zalloc(sizeof (arc_callback_t), | |
79c76d5b | 5980 | KM_SLEEP); |
34dc7c2f BB |
5981 | acb->acb_done = done; |
5982 | acb->acb_private = private; | |
a7004725 | 5983 | acb->acb_compressed = compressed_read; |
440a3eb9 TC |
5984 | acb->acb_encrypted = encrypted_read; |
5985 | acb->acb_noauth = noauth_read; | |
be9a5c35 | 5986 | acb->acb_zb = *zb; |
34dc7c2f BB |
5987 | if (pio != NULL) |
5988 | acb->acb_zio_dummy = zio_null(pio, | |
d164b209 | 5989 | spa, NULL, NULL, NULL, zio_flags); |
34dc7c2f | 5990 | |
d3c2ae1c | 5991 | ASSERT3P(acb->acb_done, !=, NULL); |
a8b2e306 | 5992 | acb->acb_zio_head = head_zio; |
b9541d6b CW |
5993 | acb->acb_next = hdr->b_l1hdr.b_acb; |
5994 | hdr->b_l1hdr.b_acb = acb; | |
34dc7c2f | 5995 | mutex_exit(hash_lock); |
1421c891 | 5996 | goto out; |
34dc7c2f BB |
5997 | } |
5998 | mutex_exit(hash_lock); | |
1421c891 | 5999 | goto out; |
34dc7c2f BB |
6000 | } |
6001 | ||
b9541d6b CW |
6002 | ASSERT(hdr->b_l1hdr.b_state == arc_mru || |
6003 | hdr->b_l1hdr.b_state == arc_mfu); | |
34dc7c2f BB |
6004 | |
6005 | if (done) { | |
7f60329a MA |
6006 | if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { |
6007 | /* | |
6008 | * This is a demand read which does not have to | |
6009 | * wait for i/o because we did a predictive | |
6010 | * prefetch i/o for it, which has completed. | |
6011 | */ | |
6012 | DTRACE_PROBE1( | |
6013 | arc__demand__hit__predictive__prefetch, | |
6014 | arc_buf_hdr_t *, hdr); | |
6015 | ARCSTAT_BUMP( | |
6016 | arcstat_demand_hit_predictive_prefetch); | |
d3c2ae1c GW |
6017 | arc_hdr_clear_flags(hdr, |
6018 | ARC_FLAG_PREDICTIVE_PREFETCH); | |
7f60329a | 6019 | } |
d4a72f23 TC |
6020 | |
6021 | if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) { | |
6022 | ARCSTAT_BUMP( | |
6023 | arcstat_demand_hit_prescient_prefetch); | |
6024 | arc_hdr_clear_flags(hdr, | |
6025 | ARC_FLAG_PRESCIENT_PREFETCH); | |
6026 | } | |
6027 | ||
d3c2ae1c GW |
6028 | ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); |
6029 | ||
524b4217 | 6030 | /* Get a buf with the desired data in it. */ |
be9a5c35 TC |
6031 | rc = arc_buf_alloc_impl(hdr, spa, zb, private, |
6032 | encrypted_read, compressed_read, noauth_read, | |
6033 | B_TRUE, &buf); | |
a2c2ed1b TC |
6034 | if (rc == ECKSUM) { |
6035 | /* | |
6036 | * Convert authentication and decryption errors | |
be9a5c35 TC |
6037 | * to EIO (and generate an ereport if needed) |
6038 | * before leaving the ARC. | |
a2c2ed1b TC |
6039 | */ |
6040 | rc = SET_ERROR(EIO); | |
be9a5c35 TC |
6041 | if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) { |
6042 | spa_log_error(spa, zb); | |
6043 | zfs_ereport_post( | |
6044 | FM_EREPORT_ZFS_AUTHENTICATION, | |
6045 | spa, NULL, zb, NULL, 0, 0); | |
6046 | } | |
a2c2ed1b | 6047 | } |
d4a72f23 | 6048 | if (rc != 0) { |
2c24b5b1 TC |
6049 | (void) remove_reference(hdr, hash_lock, |
6050 | private); | |
6051 | arc_buf_destroy_impl(buf); | |
d4a72f23 TC |
6052 | buf = NULL; |
6053 | } | |
6054 | ||
a2c2ed1b TC |
6055 | /* assert any errors weren't due to unloaded keys */ |
6056 | ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) || | |
be9a5c35 | 6057 | rc != EACCES); |
2a432414 | 6058 | } else if (*arc_flags & ARC_FLAG_PREFETCH && |
b9541d6b | 6059 | refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { |
d3c2ae1c | 6060 | arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); |
34dc7c2f BB |
6061 | } |
6062 | DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); | |
6063 | arc_access(hdr, hash_lock); | |
d4a72f23 TC |
6064 | if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH) |
6065 | arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH); | |
2a432414 | 6066 | if (*arc_flags & ARC_FLAG_L2CACHE) |
d3c2ae1c | 6067 | arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); |
34dc7c2f BB |
6068 | mutex_exit(hash_lock); |
6069 | ARCSTAT_BUMP(arcstat_hits); | |
b9541d6b CW |
6070 | ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), |
6071 | demand, prefetch, !HDR_ISTYPE_METADATA(hdr), | |
34dc7c2f BB |
6072 | data, metadata, hits); |
6073 | ||
6074 | if (done) | |
d4a72f23 | 6075 | done(NULL, zb, bp, buf, private); |
34dc7c2f | 6076 | } else { |
d3c2ae1c GW |
6077 | uint64_t lsize = BP_GET_LSIZE(bp); |
6078 | uint64_t psize = BP_GET_PSIZE(bp); | |
9b67f605 | 6079 | arc_callback_t *acb; |
b128c09f | 6080 | vdev_t *vd = NULL; |
a117a6d6 | 6081 | uint64_t addr = 0; |
d164b209 | 6082 | boolean_t devw = B_FALSE; |
d3c2ae1c | 6083 | uint64_t size; |
440a3eb9 | 6084 | abd_t *hdr_abd; |
34dc7c2f | 6085 | |
5f6d0b6f BB |
6086 | /* |
6087 | * Gracefully handle a damaged logical block size as a | |
1cdb86cb | 6088 | * checksum error. |
5f6d0b6f | 6089 | */ |
d3c2ae1c | 6090 | if (lsize > spa_maxblocksize(spa)) { |
1cdb86cb | 6091 | rc = SET_ERROR(ECKSUM); |
5f6d0b6f BB |
6092 | goto out; |
6093 | } | |
6094 | ||
34dc7c2f BB |
6095 | if (hdr == NULL) { |
6096 | /* this block is not in the cache */ | |
9b67f605 | 6097 | arc_buf_hdr_t *exists = NULL; |
34dc7c2f | 6098 | arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); |
d3c2ae1c | 6099 | hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, |
b5256303 TC |
6100 | BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), type, |
6101 | encrypted_read); | |
d3c2ae1c | 6102 | |
9b67f605 MA |
6103 | if (!BP_IS_EMBEDDED(bp)) { |
6104 | hdr->b_dva = *BP_IDENTITY(bp); | |
6105 | hdr->b_birth = BP_PHYSICAL_BIRTH(bp); | |
9b67f605 MA |
6106 | exists = buf_hash_insert(hdr, &hash_lock); |
6107 | } | |
6108 | if (exists != NULL) { | |
34dc7c2f BB |
6109 | /* somebody beat us to the hash insert */ |
6110 | mutex_exit(hash_lock); | |
428870ff | 6111 | buf_discard_identity(hdr); |
d3c2ae1c | 6112 | arc_hdr_destroy(hdr); |
34dc7c2f BB |
6113 | goto top; /* restart the IO request */ |
6114 | } | |
34dc7c2f | 6115 | } else { |
b9541d6b | 6116 | /* |
b5256303 TC |
6117 | * This block is in the ghost cache or encrypted data |
6118 | * was requested and we didn't have it. If it was | |
6119 | * L2-only (and thus didn't have an L1 hdr), | |
6120 | * we realloc the header to add an L1 hdr. | |
b9541d6b CW |
6121 | */ |
6122 | if (!HDR_HAS_L1HDR(hdr)) { | |
6123 | hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, | |
6124 | hdr_full_cache); | |
6125 | } | |
6126 | ||
b5256303 TC |
6127 | if (GHOST_STATE(hdr->b_l1hdr.b_state)) { |
6128 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); | |
6129 | ASSERT(!HDR_HAS_RABD(hdr)); | |
6130 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
6131 | ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); | |
6132 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
6133 | ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); | |
6134 | } else if (HDR_IO_IN_PROGRESS(hdr)) { | |
6135 | /* | |
6136 | * If this header already had an IO in progress | |
6137 | * and we are performing another IO to fetch | |
6138 | * encrypted data we must wait until the first | |
6139 | * IO completes so as not to confuse | |
6140 | * arc_read_done(). This should be very rare | |
6141 | * and so the performance impact shouldn't | |
6142 | * matter. | |
6143 | */ | |
6144 | cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); | |
6145 | mutex_exit(hash_lock); | |
6146 | goto top; | |
6147 | } | |
34dc7c2f | 6148 | |
7f60329a | 6149 | /* |
d3c2ae1c | 6150 | * This is a delicate dance that we play here. |
b5256303 TC |
6151 | * This hdr might be in the ghost list so we access |
6152 | * it to move it out of the ghost list before we | |
d3c2ae1c GW |
6153 | * initiate the read. If it's a prefetch then |
6154 | * it won't have a callback so we'll remove the | |
6155 | * reference that arc_buf_alloc_impl() created. We | |
6156 | * do this after we've called arc_access() to | |
6157 | * avoid hitting an assert in remove_reference(). | |
7f60329a | 6158 | */ |
428870ff | 6159 | arc_access(hdr, hash_lock); |
b5256303 | 6160 | arc_hdr_alloc_abd(hdr, encrypted_read); |
d3c2ae1c | 6161 | } |
d3c2ae1c | 6162 | |
b5256303 TC |
6163 | if (encrypted_read) { |
6164 | ASSERT(HDR_HAS_RABD(hdr)); | |
6165 | size = HDR_GET_PSIZE(hdr); | |
6166 | hdr_abd = hdr->b_crypt_hdr.b_rabd; | |
d3c2ae1c | 6167 | zio_flags |= ZIO_FLAG_RAW; |
b5256303 TC |
6168 | } else { |
6169 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); | |
6170 | size = arc_hdr_size(hdr); | |
6171 | hdr_abd = hdr->b_l1hdr.b_pabd; | |
6172 | ||
6173 | if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) { | |
6174 | zio_flags |= ZIO_FLAG_RAW_COMPRESS; | |
6175 | } | |
6176 | ||
6177 | /* | |
6178 | * For authenticated bp's, we do not ask the ZIO layer | |
6179 | * to authenticate them since this will cause the entire | |
6180 | * IO to fail if the key isn't loaded. Instead, we | |
6181 | * defer authentication until arc_buf_fill(), which will | |
6182 | * verify the data when the key is available. | |
6183 | */ | |
6184 | if (BP_IS_AUTHENTICATED(bp)) | |
6185 | zio_flags |= ZIO_FLAG_RAW_ENCRYPT; | |
34dc7c2f BB |
6186 | } |
6187 | ||
b5256303 TC |
6188 | if (*arc_flags & ARC_FLAG_PREFETCH && |
6189 | refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) | |
d3c2ae1c | 6190 | arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); |
d4a72f23 TC |
6191 | if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH) |
6192 | arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH); | |
d3c2ae1c GW |
6193 | if (*arc_flags & ARC_FLAG_L2CACHE) |
6194 | arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); | |
b5256303 TC |
6195 | if (BP_IS_AUTHENTICATED(bp)) |
6196 | arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH); | |
d3c2ae1c GW |
6197 | if (BP_GET_LEVEL(bp) > 0) |
6198 | arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); | |
7f60329a | 6199 | if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) |
d3c2ae1c | 6200 | arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); |
b9541d6b | 6201 | ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); |
428870ff | 6202 | |
79c76d5b | 6203 | acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); |
34dc7c2f BB |
6204 | acb->acb_done = done; |
6205 | acb->acb_private = private; | |
2aa34383 | 6206 | acb->acb_compressed = compressed_read; |
b5256303 TC |
6207 | acb->acb_encrypted = encrypted_read; |
6208 | acb->acb_noauth = noauth_read; | |
be9a5c35 | 6209 | acb->acb_zb = *zb; |
34dc7c2f | 6210 | |
d3c2ae1c | 6211 | ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); |
b9541d6b | 6212 | hdr->b_l1hdr.b_acb = acb; |
d3c2ae1c | 6213 | arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); |
34dc7c2f | 6214 | |
b9541d6b CW |
6215 | if (HDR_HAS_L2HDR(hdr) && |
6216 | (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { | |
6217 | devw = hdr->b_l2hdr.b_dev->l2ad_writing; | |
6218 | addr = hdr->b_l2hdr.b_daddr; | |
b128c09f | 6219 | /* |
a1d477c2 | 6220 | * Lock out L2ARC device removal. |
b128c09f BB |
6221 | */ |
6222 | if (vdev_is_dead(vd) || | |
6223 | !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) | |
6224 | vd = NULL; | |
6225 | } | |
6226 | ||
a8b2e306 TC |
6227 | /* |
6228 | * We count both async reads and scrub IOs as asynchronous so | |
6229 | * that both can be upgraded in the event of a cache hit while | |
6230 | * the read IO is still in-flight. | |
6231 | */ | |
6232 | if (priority == ZIO_PRIORITY_ASYNC_READ || | |
6233 | priority == ZIO_PRIORITY_SCRUB) | |
d3c2ae1c GW |
6234 | arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); |
6235 | else | |
6236 | arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); | |
6237 | ||
e49f1e20 WA |
6238 | /* |
6239 | * At this point, we have a level 1 cache miss. Try again in | |
6240 | * L2ARC if possible. | |
6241 | */ | |
d3c2ae1c GW |
6242 | ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); |
6243 | ||
428870ff | 6244 | DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, |
d3c2ae1c | 6245 | uint64_t, lsize, zbookmark_phys_t *, zb); |
34dc7c2f | 6246 | ARCSTAT_BUMP(arcstat_misses); |
b9541d6b CW |
6247 | ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), |
6248 | demand, prefetch, !HDR_ISTYPE_METADATA(hdr), | |
34dc7c2f BB |
6249 | data, metadata, misses); |
6250 | ||
d164b209 | 6251 | if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { |
34dc7c2f BB |
6252 | /* |
6253 | * Read from the L2ARC if the following are true: | |
b128c09f BB |
6254 | * 1. The L2ARC vdev was previously cached. |
6255 | * 2. This buffer still has L2ARC metadata. | |
6256 | * 3. This buffer isn't currently writing to the L2ARC. | |
6257 | * 4. The L2ARC entry wasn't evicted, which may | |
6258 | * also have invalidated the vdev. | |
d164b209 | 6259 | * 5. This isn't prefetch and l2arc_noprefetch is set. |
34dc7c2f | 6260 | */ |
b9541d6b | 6261 | if (HDR_HAS_L2HDR(hdr) && |
d164b209 BB |
6262 | !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && |
6263 | !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { | |
34dc7c2f | 6264 | l2arc_read_callback_t *cb; |
82710e99 GDN |
6265 | abd_t *abd; |
6266 | uint64_t asize; | |
34dc7c2f BB |
6267 | |
6268 | DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); | |
6269 | ARCSTAT_BUMP(arcstat_l2_hits); | |
b9541d6b | 6270 | atomic_inc_32(&hdr->b_l2hdr.b_hits); |
34dc7c2f | 6271 | |
34dc7c2f | 6272 | cb = kmem_zalloc(sizeof (l2arc_read_callback_t), |
79c76d5b | 6273 | KM_SLEEP); |
d3c2ae1c | 6274 | cb->l2rcb_hdr = hdr; |
34dc7c2f BB |
6275 | cb->l2rcb_bp = *bp; |
6276 | cb->l2rcb_zb = *zb; | |
b128c09f | 6277 | cb->l2rcb_flags = zio_flags; |
34dc7c2f | 6278 | |
82710e99 GDN |
6279 | asize = vdev_psize_to_asize(vd, size); |
6280 | if (asize != size) { | |
6281 | abd = abd_alloc_for_io(asize, | |
6282 | HDR_ISTYPE_METADATA(hdr)); | |
6283 | cb->l2rcb_abd = abd; | |
6284 | } else { | |
b5256303 | 6285 | abd = hdr_abd; |
82710e99 GDN |
6286 | } |
6287 | ||
a117a6d6 | 6288 | ASSERT(addr >= VDEV_LABEL_START_SIZE && |
82710e99 | 6289 | addr + asize <= vd->vdev_psize - |
a117a6d6 GW |
6290 | VDEV_LABEL_END_SIZE); |
6291 | ||
34dc7c2f | 6292 | /* |
b128c09f BB |
6293 | * l2arc read. The SCL_L2ARC lock will be |
6294 | * released by l2arc_read_done(). | |
3a17a7a9 SK |
6295 | * Issue a null zio if the underlying buffer |
6296 | * was squashed to zero size by compression. | |
34dc7c2f | 6297 | */ |
b5256303 | 6298 | ASSERT3U(arc_hdr_get_compress(hdr), !=, |
d3c2ae1c GW |
6299 | ZIO_COMPRESS_EMPTY); |
6300 | rzio = zio_read_phys(pio, vd, addr, | |
82710e99 | 6301 | asize, abd, |
d3c2ae1c GW |
6302 | ZIO_CHECKSUM_OFF, |
6303 | l2arc_read_done, cb, priority, | |
6304 | zio_flags | ZIO_FLAG_DONT_CACHE | | |
6305 | ZIO_FLAG_CANFAIL | | |
6306 | ZIO_FLAG_DONT_PROPAGATE | | |
6307 | ZIO_FLAG_DONT_RETRY, B_FALSE); | |
a8b2e306 TC |
6308 | acb->acb_zio_head = rzio; |
6309 | ||
6310 | if (hash_lock != NULL) | |
6311 | mutex_exit(hash_lock); | |
d3c2ae1c | 6312 | |
34dc7c2f BB |
6313 | DTRACE_PROBE2(l2arc__read, vdev_t *, vd, |
6314 | zio_t *, rzio); | |
b5256303 TC |
6315 | ARCSTAT_INCR(arcstat_l2_read_bytes, |
6316 | HDR_GET_PSIZE(hdr)); | |
34dc7c2f | 6317 | |
2a432414 | 6318 | if (*arc_flags & ARC_FLAG_NOWAIT) { |
b128c09f | 6319 | zio_nowait(rzio); |
1421c891 | 6320 | goto out; |
b128c09f | 6321 | } |
34dc7c2f | 6322 | |
2a432414 | 6323 | ASSERT(*arc_flags & ARC_FLAG_WAIT); |
b128c09f | 6324 | if (zio_wait(rzio) == 0) |
1421c891 | 6325 | goto out; |
b128c09f BB |
6326 | |
6327 | /* l2arc read error; goto zio_read() */ | |
a8b2e306 TC |
6328 | if (hash_lock != NULL) |
6329 | mutex_enter(hash_lock); | |
34dc7c2f BB |
6330 | } else { |
6331 | DTRACE_PROBE1(l2arc__miss, | |
6332 | arc_buf_hdr_t *, hdr); | |
6333 | ARCSTAT_BUMP(arcstat_l2_misses); | |
6334 | if (HDR_L2_WRITING(hdr)) | |
6335 | ARCSTAT_BUMP(arcstat_l2_rw_clash); | |
b128c09f | 6336 | spa_config_exit(spa, SCL_L2ARC, vd); |
34dc7c2f | 6337 | } |
d164b209 BB |
6338 | } else { |
6339 | if (vd != NULL) | |
6340 | spa_config_exit(spa, SCL_L2ARC, vd); | |
6341 | if (l2arc_ndev != 0) { | |
6342 | DTRACE_PROBE1(l2arc__miss, | |
6343 | arc_buf_hdr_t *, hdr); | |
6344 | ARCSTAT_BUMP(arcstat_l2_misses); | |
6345 | } | |
34dc7c2f | 6346 | } |
34dc7c2f | 6347 | |
b5256303 | 6348 | rzio = zio_read(pio, spa, bp, hdr_abd, size, |
d3c2ae1c | 6349 | arc_read_done, hdr, priority, zio_flags, zb); |
a8b2e306 TC |
6350 | acb->acb_zio_head = rzio; |
6351 | ||
6352 | if (hash_lock != NULL) | |
6353 | mutex_exit(hash_lock); | |
34dc7c2f | 6354 | |
2a432414 | 6355 | if (*arc_flags & ARC_FLAG_WAIT) { |
1421c891 PS |
6356 | rc = zio_wait(rzio); |
6357 | goto out; | |
6358 | } | |
34dc7c2f | 6359 | |
2a432414 | 6360 | ASSERT(*arc_flags & ARC_FLAG_NOWAIT); |
34dc7c2f BB |
6361 | zio_nowait(rzio); |
6362 | } | |
1421c891 PS |
6363 | |
6364 | out: | |
157ef7f6 TC |
6365 | /* embedded bps don't actually go to disk */ |
6366 | if (!BP_IS_EMBEDDED(bp)) | |
6367 | spa_read_history_add(spa, zb, *arc_flags); | |
1421c891 | 6368 | return (rc); |
34dc7c2f BB |
6369 | } |
6370 | ||
ab26409d BB |
6371 | arc_prune_t * |
6372 | arc_add_prune_callback(arc_prune_func_t *func, void *private) | |
6373 | { | |
6374 | arc_prune_t *p; | |
6375 | ||
d1d7e268 | 6376 | p = kmem_alloc(sizeof (*p), KM_SLEEP); |
ab26409d BB |
6377 | p->p_pfunc = func; |
6378 | p->p_private = private; | |
6379 | list_link_init(&p->p_node); | |
6380 | refcount_create(&p->p_refcnt); | |
6381 | ||
6382 | mutex_enter(&arc_prune_mtx); | |
6383 | refcount_add(&p->p_refcnt, &arc_prune_list); | |
6384 | list_insert_head(&arc_prune_list, p); | |
6385 | mutex_exit(&arc_prune_mtx); | |
6386 | ||
6387 | return (p); | |
6388 | } | |
6389 | ||
6390 | void | |
6391 | arc_remove_prune_callback(arc_prune_t *p) | |
6392 | { | |
4442f60d | 6393 | boolean_t wait = B_FALSE; |
ab26409d BB |
6394 | mutex_enter(&arc_prune_mtx); |
6395 | list_remove(&arc_prune_list, p); | |
4442f60d CC |
6396 | if (refcount_remove(&p->p_refcnt, &arc_prune_list) > 0) |
6397 | wait = B_TRUE; | |
ab26409d | 6398 | mutex_exit(&arc_prune_mtx); |
4442f60d CC |
6399 | |
6400 | /* wait for arc_prune_task to finish */ | |
6401 | if (wait) | |
6402 | taskq_wait_outstanding(arc_prune_taskq, 0); | |
6403 | ASSERT0(refcount_count(&p->p_refcnt)); | |
6404 | refcount_destroy(&p->p_refcnt); | |
6405 | kmem_free(p, sizeof (*p)); | |
ab26409d BB |
6406 | } |
6407 | ||
df4474f9 MA |
6408 | /* |
6409 | * Notify the arc that a block was freed, and thus will never be used again. | |
6410 | */ | |
6411 | void | |
6412 | arc_freed(spa_t *spa, const blkptr_t *bp) | |
6413 | { | |
6414 | arc_buf_hdr_t *hdr; | |
6415 | kmutex_t *hash_lock; | |
6416 | uint64_t guid = spa_load_guid(spa); | |
6417 | ||
9b67f605 MA |
6418 | ASSERT(!BP_IS_EMBEDDED(bp)); |
6419 | ||
6420 | hdr = buf_hash_find(guid, bp, &hash_lock); | |
df4474f9 MA |
6421 | if (hdr == NULL) |
6422 | return; | |
df4474f9 | 6423 | |
d3c2ae1c GW |
6424 | /* |
6425 | * We might be trying to free a block that is still doing I/O | |
6426 | * (i.e. prefetch) or has a reference (i.e. a dedup-ed, | |
6427 | * dmu_sync-ed block). If this block is being prefetched, then it | |
6428 | * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr | |
6429 | * until the I/O completes. A block may also have a reference if it is | |
6430 | * part of a dedup-ed, dmu_synced write. The dmu_sync() function would | |
6431 | * have written the new block to its final resting place on disk but | |
6432 | * without the dedup flag set. This would have left the hdr in the MRU | |
6433 | * state and discoverable. When the txg finally syncs it detects that | |
6434 | * the block was overridden in open context and issues an override I/O. | |
6435 | * Since this is a dedup block, the override I/O will determine if the | |
6436 | * block is already in the DDT. If so, then it will replace the io_bp | |
6437 | * with the bp from the DDT and allow the I/O to finish. When the I/O | |
6438 | * reaches the done callback, dbuf_write_override_done, it will | |
6439 | * check to see if the io_bp and io_bp_override are identical. | |
6440 | * If they are not, then it indicates that the bp was replaced with | |
6441 | * the bp in the DDT and the override bp is freed. This allows | |
6442 | * us to arrive here with a reference on a block that is being | |
6443 | * freed. So if we have an I/O in progress, or a reference to | |
6444 | * this hdr, then we don't destroy the hdr. | |
6445 | */ | |
6446 | if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && | |
6447 | refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { | |
6448 | arc_change_state(arc_anon, hdr, hash_lock); | |
6449 | arc_hdr_destroy(hdr); | |
df4474f9 | 6450 | mutex_exit(hash_lock); |
bd089c54 | 6451 | } else { |
d3c2ae1c | 6452 | mutex_exit(hash_lock); |
34dc7c2f | 6453 | } |
34dc7c2f | 6454 | |
34dc7c2f BB |
6455 | } |
6456 | ||
6457 | /* | |
e49f1e20 WA |
6458 | * Release this buffer from the cache, making it an anonymous buffer. This |
6459 | * must be done after a read and prior to modifying the buffer contents. | |
34dc7c2f | 6460 | * If the buffer has more than one reference, we must make |
b128c09f | 6461 | * a new hdr for the buffer. |
34dc7c2f BB |
6462 | */ |
6463 | void | |
6464 | arc_release(arc_buf_t *buf, void *tag) | |
6465 | { | |
b9541d6b | 6466 | arc_buf_hdr_t *hdr = buf->b_hdr; |
34dc7c2f | 6467 | |
428870ff | 6468 | /* |
ca0bf58d | 6469 | * It would be nice to assert that if its DMU metadata (level > |
428870ff BB |
6470 | * 0 || it's the dnode file), then it must be syncing context. |
6471 | * But we don't know that information at this level. | |
6472 | */ | |
6473 | ||
6474 | mutex_enter(&buf->b_evict_lock); | |
b128c09f | 6475 | |
ca0bf58d PS |
6476 | ASSERT(HDR_HAS_L1HDR(hdr)); |
6477 | ||
b9541d6b CW |
6478 | /* |
6479 | * We don't grab the hash lock prior to this check, because if | |
6480 | * the buffer's header is in the arc_anon state, it won't be | |
6481 | * linked into the hash table. | |
6482 | */ | |
6483 | if (hdr->b_l1hdr.b_state == arc_anon) { | |
6484 | mutex_exit(&buf->b_evict_lock); | |
6485 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
6486 | ASSERT(!HDR_IN_HASH_TABLE(hdr)); | |
6487 | ASSERT(!HDR_HAS_L2HDR(hdr)); | |
d3c2ae1c | 6488 | ASSERT(HDR_EMPTY(hdr)); |
34dc7c2f | 6489 | |
d3c2ae1c | 6490 | ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); |
b9541d6b CW |
6491 | ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); |
6492 | ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); | |
6493 | ||
b9541d6b | 6494 | hdr->b_l1hdr.b_arc_access = 0; |
d3c2ae1c GW |
6495 | |
6496 | /* | |
6497 | * If the buf is being overridden then it may already | |
6498 | * have a hdr that is not empty. | |
6499 | */ | |
6500 | buf_discard_identity(hdr); | |
b9541d6b CW |
6501 | arc_buf_thaw(buf); |
6502 | ||
6503 | return; | |
34dc7c2f BB |
6504 | } |
6505 | ||
1c27024e | 6506 | kmutex_t *hash_lock = HDR_LOCK(hdr); |
b9541d6b CW |
6507 | mutex_enter(hash_lock); |
6508 | ||
6509 | /* | |
6510 | * This assignment is only valid as long as the hash_lock is | |
6511 | * held, we must be careful not to reference state or the | |
6512 | * b_state field after dropping the lock. | |
6513 | */ | |
1c27024e | 6514 | arc_state_t *state = hdr->b_l1hdr.b_state; |
b9541d6b CW |
6515 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); |
6516 | ASSERT3P(state, !=, arc_anon); | |
6517 | ||
6518 | /* this buffer is not on any list */ | |
2aa34383 | 6519 | ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); |
b9541d6b CW |
6520 | |
6521 | if (HDR_HAS_L2HDR(hdr)) { | |
b9541d6b | 6522 | mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); |
ca0bf58d PS |
6523 | |
6524 | /* | |
d962d5da PS |
6525 | * We have to recheck this conditional again now that |
6526 | * we're holding the l2ad_mtx to prevent a race with | |
6527 | * another thread which might be concurrently calling | |
6528 | * l2arc_evict(). In that case, l2arc_evict() might have | |
6529 | * destroyed the header's L2 portion as we were waiting | |
6530 | * to acquire the l2ad_mtx. | |
ca0bf58d | 6531 | */ |
d962d5da PS |
6532 | if (HDR_HAS_L2HDR(hdr)) |
6533 | arc_hdr_l2hdr_destroy(hdr); | |
ca0bf58d | 6534 | |
b9541d6b | 6535 | mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); |
b128c09f BB |
6536 | } |
6537 | ||
34dc7c2f BB |
6538 | /* |
6539 | * Do we have more than one buf? | |
6540 | */ | |
d3c2ae1c | 6541 | if (hdr->b_l1hdr.b_bufcnt > 1) { |
34dc7c2f | 6542 | arc_buf_hdr_t *nhdr; |
d164b209 | 6543 | uint64_t spa = hdr->b_spa; |
d3c2ae1c GW |
6544 | uint64_t psize = HDR_GET_PSIZE(hdr); |
6545 | uint64_t lsize = HDR_GET_LSIZE(hdr); | |
b5256303 TC |
6546 | boolean_t protected = HDR_PROTECTED(hdr); |
6547 | enum zio_compress compress = arc_hdr_get_compress(hdr); | |
b9541d6b | 6548 | arc_buf_contents_t type = arc_buf_type(hdr); |
d3c2ae1c | 6549 | VERIFY3U(hdr->b_type, ==, type); |
34dc7c2f | 6550 | |
b9541d6b | 6551 | ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); |
d3c2ae1c GW |
6552 | (void) remove_reference(hdr, hash_lock, tag); |
6553 | ||
524b4217 | 6554 | if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { |
d3c2ae1c | 6555 | ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); |
524b4217 DK |
6556 | ASSERT(ARC_BUF_LAST(buf)); |
6557 | } | |
d3c2ae1c | 6558 | |
34dc7c2f | 6559 | /* |
428870ff | 6560 | * Pull the data off of this hdr and attach it to |
d3c2ae1c GW |
6561 | * a new anonymous hdr. Also find the last buffer |
6562 | * in the hdr's buffer list. | |
34dc7c2f | 6563 | */ |
a7004725 | 6564 | arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); |
d3c2ae1c | 6565 | ASSERT3P(lastbuf, !=, NULL); |
34dc7c2f | 6566 | |
d3c2ae1c GW |
6567 | /* |
6568 | * If the current arc_buf_t and the hdr are sharing their data | |
524b4217 | 6569 | * buffer, then we must stop sharing that block. |
d3c2ae1c GW |
6570 | */ |
6571 | if (arc_buf_is_shared(buf)) { | |
6572 | ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); | |
d3c2ae1c GW |
6573 | VERIFY(!arc_buf_is_shared(lastbuf)); |
6574 | ||
6575 | /* | |
6576 | * First, sever the block sharing relationship between | |
a7004725 | 6577 | * buf and the arc_buf_hdr_t. |
d3c2ae1c GW |
6578 | */ |
6579 | arc_unshare_buf(hdr, buf); | |
2aa34383 DK |
6580 | |
6581 | /* | |
a6255b7f | 6582 | * Now we need to recreate the hdr's b_pabd. Since we |
524b4217 | 6583 | * have lastbuf handy, we try to share with it, but if |
a6255b7f | 6584 | * we can't then we allocate a new b_pabd and copy the |
524b4217 | 6585 | * data from buf into it. |
2aa34383 | 6586 | */ |
524b4217 DK |
6587 | if (arc_can_share(hdr, lastbuf)) { |
6588 | arc_share_buf(hdr, lastbuf); | |
6589 | } else { | |
b5256303 | 6590 | arc_hdr_alloc_abd(hdr, B_FALSE); |
a6255b7f DQ |
6591 | abd_copy_from_buf(hdr->b_l1hdr.b_pabd, |
6592 | buf->b_data, psize); | |
2aa34383 | 6593 | } |
d3c2ae1c GW |
6594 | VERIFY3P(lastbuf->b_data, !=, NULL); |
6595 | } else if (HDR_SHARED_DATA(hdr)) { | |
2aa34383 DK |
6596 | /* |
6597 | * Uncompressed shared buffers are always at the end | |
6598 | * of the list. Compressed buffers don't have the | |
6599 | * same requirements. This makes it hard to | |
6600 | * simply assert that the lastbuf is shared so | |
6601 | * we rely on the hdr's compression flags to determine | |
6602 | * if we have a compressed, shared buffer. | |
6603 | */ | |
6604 | ASSERT(arc_buf_is_shared(lastbuf) || | |
b5256303 | 6605 | arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF); |
2aa34383 | 6606 | ASSERT(!ARC_BUF_SHARED(buf)); |
d3c2ae1c | 6607 | } |
b5256303 TC |
6608 | |
6609 | ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); | |
b9541d6b | 6610 | ASSERT3P(state, !=, arc_l2c_only); |
36da08ef | 6611 | |
d3c2ae1c | 6612 | (void) refcount_remove_many(&state->arcs_size, |
2aa34383 | 6613 | arc_buf_size(buf), buf); |
36da08ef | 6614 | |
b9541d6b | 6615 | if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { |
b9541d6b | 6616 | ASSERT3P(state, !=, arc_l2c_only); |
d3c2ae1c | 6617 | (void) refcount_remove_many(&state->arcs_esize[type], |
2aa34383 | 6618 | arc_buf_size(buf), buf); |
34dc7c2f | 6619 | } |
1eb5bfa3 | 6620 | |
d3c2ae1c | 6621 | hdr->b_l1hdr.b_bufcnt -= 1; |
b5256303 TC |
6622 | if (ARC_BUF_ENCRYPTED(buf)) |
6623 | hdr->b_crypt_hdr.b_ebufcnt -= 1; | |
6624 | ||
34dc7c2f | 6625 | arc_cksum_verify(buf); |
498877ba | 6626 | arc_buf_unwatch(buf); |
34dc7c2f | 6627 | |
f486f584 TC |
6628 | /* if this is the last uncompressed buf free the checksum */ |
6629 | if (!arc_hdr_has_uncompressed_buf(hdr)) | |
6630 | arc_cksum_free(hdr); | |
6631 | ||
34dc7c2f BB |
6632 | mutex_exit(hash_lock); |
6633 | ||
d3c2ae1c | 6634 | /* |
a6255b7f | 6635 | * Allocate a new hdr. The new hdr will contain a b_pabd |
d3c2ae1c GW |
6636 | * buffer which will be freed in arc_write(). |
6637 | */ | |
b5256303 TC |
6638 | nhdr = arc_hdr_alloc(spa, psize, lsize, protected, |
6639 | compress, type, HDR_HAS_RABD(hdr)); | |
d3c2ae1c GW |
6640 | ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); |
6641 | ASSERT0(nhdr->b_l1hdr.b_bufcnt); | |
6642 | ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); | |
6643 | VERIFY3U(nhdr->b_type, ==, type); | |
6644 | ASSERT(!HDR_SHARED_DATA(nhdr)); | |
b9541d6b | 6645 | |
d3c2ae1c GW |
6646 | nhdr->b_l1hdr.b_buf = buf; |
6647 | nhdr->b_l1hdr.b_bufcnt = 1; | |
b5256303 TC |
6648 | if (ARC_BUF_ENCRYPTED(buf)) |
6649 | nhdr->b_crypt_hdr.b_ebufcnt = 1; | |
b9541d6b CW |
6650 | nhdr->b_l1hdr.b_mru_hits = 0; |
6651 | nhdr->b_l1hdr.b_mru_ghost_hits = 0; | |
6652 | nhdr->b_l1hdr.b_mfu_hits = 0; | |
6653 | nhdr->b_l1hdr.b_mfu_ghost_hits = 0; | |
6654 | nhdr->b_l1hdr.b_l2_hits = 0; | |
b9541d6b | 6655 | (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); |
34dc7c2f | 6656 | buf->b_hdr = nhdr; |
d3c2ae1c | 6657 | |
428870ff | 6658 | mutex_exit(&buf->b_evict_lock); |
d3c2ae1c GW |
6659 | (void) refcount_add_many(&arc_anon->arcs_size, |
6660 | HDR_GET_LSIZE(nhdr), buf); | |
34dc7c2f | 6661 | } else { |
428870ff | 6662 | mutex_exit(&buf->b_evict_lock); |
b9541d6b | 6663 | ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); |
ca0bf58d PS |
6664 | /* protected by hash lock, or hdr is on arc_anon */ |
6665 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); | |
34dc7c2f | 6666 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); |
b9541d6b CW |
6667 | hdr->b_l1hdr.b_mru_hits = 0; |
6668 | hdr->b_l1hdr.b_mru_ghost_hits = 0; | |
6669 | hdr->b_l1hdr.b_mfu_hits = 0; | |
6670 | hdr->b_l1hdr.b_mfu_ghost_hits = 0; | |
6671 | hdr->b_l1hdr.b_l2_hits = 0; | |
6672 | arc_change_state(arc_anon, hdr, hash_lock); | |
6673 | hdr->b_l1hdr.b_arc_access = 0; | |
34dc7c2f | 6674 | |
b5256303 | 6675 | mutex_exit(hash_lock); |
428870ff | 6676 | buf_discard_identity(hdr); |
34dc7c2f BB |
6677 | arc_buf_thaw(buf); |
6678 | } | |
34dc7c2f BB |
6679 | } |
6680 | ||
6681 | int | |
6682 | arc_released(arc_buf_t *buf) | |
6683 | { | |
b128c09f BB |
6684 | int released; |
6685 | ||
428870ff | 6686 | mutex_enter(&buf->b_evict_lock); |
b9541d6b CW |
6687 | released = (buf->b_data != NULL && |
6688 | buf->b_hdr->b_l1hdr.b_state == arc_anon); | |
428870ff | 6689 | mutex_exit(&buf->b_evict_lock); |
b128c09f | 6690 | return (released); |
34dc7c2f BB |
6691 | } |
6692 | ||
34dc7c2f BB |
6693 | #ifdef ZFS_DEBUG |
6694 | int | |
6695 | arc_referenced(arc_buf_t *buf) | |
6696 | { | |
b128c09f BB |
6697 | int referenced; |
6698 | ||
428870ff | 6699 | mutex_enter(&buf->b_evict_lock); |
b9541d6b | 6700 | referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); |
428870ff | 6701 | mutex_exit(&buf->b_evict_lock); |
b128c09f | 6702 | return (referenced); |
34dc7c2f BB |
6703 | } |
6704 | #endif | |
6705 | ||
6706 | static void | |
6707 | arc_write_ready(zio_t *zio) | |
6708 | { | |
6709 | arc_write_callback_t *callback = zio->io_private; | |
6710 | arc_buf_t *buf = callback->awcb_buf; | |
6711 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
b5256303 TC |
6712 | blkptr_t *bp = zio->io_bp; |
6713 | uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp); | |
a6255b7f | 6714 | fstrans_cookie_t cookie = spl_fstrans_mark(); |
34dc7c2f | 6715 | |
b9541d6b CW |
6716 | ASSERT(HDR_HAS_L1HDR(hdr)); |
6717 | ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); | |
d3c2ae1c | 6718 | ASSERT(hdr->b_l1hdr.b_bufcnt > 0); |
b128c09f | 6719 | |
34dc7c2f | 6720 | /* |
d3c2ae1c GW |
6721 | * If we're reexecuting this zio because the pool suspended, then |
6722 | * cleanup any state that was previously set the first time the | |
2aa34383 | 6723 | * callback was invoked. |
34dc7c2f | 6724 | */ |
d3c2ae1c GW |
6725 | if (zio->io_flags & ZIO_FLAG_REEXECUTED) { |
6726 | arc_cksum_free(hdr); | |
6727 | arc_buf_unwatch(buf); | |
a6255b7f | 6728 | if (hdr->b_l1hdr.b_pabd != NULL) { |
d3c2ae1c | 6729 | if (arc_buf_is_shared(buf)) { |
d3c2ae1c GW |
6730 | arc_unshare_buf(hdr, buf); |
6731 | } else { | |
b5256303 | 6732 | arc_hdr_free_abd(hdr, B_FALSE); |
d3c2ae1c | 6733 | } |
34dc7c2f | 6734 | } |
b5256303 TC |
6735 | |
6736 | if (HDR_HAS_RABD(hdr)) | |
6737 | arc_hdr_free_abd(hdr, B_TRUE); | |
34dc7c2f | 6738 | } |
a6255b7f | 6739 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 6740 | ASSERT(!HDR_HAS_RABD(hdr)); |
d3c2ae1c GW |
6741 | ASSERT(!HDR_SHARED_DATA(hdr)); |
6742 | ASSERT(!arc_buf_is_shared(buf)); | |
6743 | ||
6744 | callback->awcb_ready(zio, buf, callback->awcb_private); | |
6745 | ||
6746 | if (HDR_IO_IN_PROGRESS(hdr)) | |
6747 | ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); | |
6748 | ||
d3c2ae1c GW |
6749 | arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); |
6750 | ||
b5256303 TC |
6751 | if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr)) |
6752 | hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp)); | |
6753 | ||
6754 | if (BP_IS_PROTECTED(bp)) { | |
6755 | /* ZIL blocks are written through zio_rewrite */ | |
6756 | ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG); | |
6757 | ASSERT(HDR_PROTECTED(hdr)); | |
6758 | ||
ae76f45c TC |
6759 | if (BP_SHOULD_BYTESWAP(bp)) { |
6760 | if (BP_GET_LEVEL(bp) > 0) { | |
6761 | hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; | |
6762 | } else { | |
6763 | hdr->b_l1hdr.b_byteswap = | |
6764 | DMU_OT_BYTESWAP(BP_GET_TYPE(bp)); | |
6765 | } | |
6766 | } else { | |
6767 | hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; | |
6768 | } | |
6769 | ||
b5256303 TC |
6770 | hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp); |
6771 | hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset; | |
6772 | zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt, | |
6773 | hdr->b_crypt_hdr.b_iv); | |
6774 | zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac); | |
6775 | } | |
6776 | ||
6777 | /* | |
6778 | * If this block was written for raw encryption but the zio layer | |
6779 | * ended up only authenticating it, adjust the buffer flags now. | |
6780 | */ | |
6781 | if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) { | |
6782 | arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH); | |
6783 | buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED; | |
6784 | if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF) | |
6785 | buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; | |
b1d21733 TC |
6786 | } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) { |
6787 | buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED; | |
6788 | buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; | |
b5256303 TC |
6789 | } |
6790 | ||
6791 | /* this must be done after the buffer flags are adjusted */ | |
6792 | arc_cksum_compute(buf); | |
6793 | ||
1c27024e | 6794 | enum zio_compress compress; |
b5256303 | 6795 | if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { |
d3c2ae1c GW |
6796 | compress = ZIO_COMPRESS_OFF; |
6797 | } else { | |
b5256303 TC |
6798 | ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); |
6799 | compress = BP_GET_COMPRESS(bp); | |
d3c2ae1c GW |
6800 | } |
6801 | HDR_SET_PSIZE(hdr, psize); | |
6802 | arc_hdr_set_compress(hdr, compress); | |
6803 | ||
4807c0ba TC |
6804 | if (zio->io_error != 0 || psize == 0) |
6805 | goto out; | |
6806 | ||
d3c2ae1c | 6807 | /* |
b5256303 TC |
6808 | * Fill the hdr with data. If the buffer is encrypted we have no choice |
6809 | * but to copy the data into b_radb. If the hdr is compressed, the data | |
6810 | * we want is available from the zio, otherwise we can take it from | |
6811 | * the buf. | |
a6255b7f DQ |
6812 | * |
6813 | * We might be able to share the buf's data with the hdr here. However, | |
6814 | * doing so would cause the ARC to be full of linear ABDs if we write a | |
6815 | * lot of shareable data. As a compromise, we check whether scattered | |
6816 | * ABDs are allowed, and assume that if they are then the user wants | |
6817 | * the ARC to be primarily filled with them regardless of the data being | |
6818 | * written. Therefore, if they're allowed then we allocate one and copy | |
6819 | * the data into it; otherwise, we share the data directly if we can. | |
d3c2ae1c | 6820 | */ |
b5256303 | 6821 | if (ARC_BUF_ENCRYPTED(buf)) { |
4807c0ba | 6822 | ASSERT3U(psize, >, 0); |
b5256303 TC |
6823 | ASSERT(ARC_BUF_COMPRESSED(buf)); |
6824 | arc_hdr_alloc_abd(hdr, B_TRUE); | |
6825 | abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize); | |
6826 | } else if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) { | |
a6255b7f DQ |
6827 | /* |
6828 | * Ideally, we would always copy the io_abd into b_pabd, but the | |
6829 | * user may have disabled compressed ARC, thus we must check the | |
6830 | * hdr's compression setting rather than the io_bp's. | |
6831 | */ | |
b5256303 | 6832 | if (BP_IS_ENCRYPTED(bp)) { |
a6255b7f | 6833 | ASSERT3U(psize, >, 0); |
b5256303 TC |
6834 | arc_hdr_alloc_abd(hdr, B_TRUE); |
6835 | abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize); | |
6836 | } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF && | |
6837 | !ARC_BUF_COMPRESSED(buf)) { | |
6838 | ASSERT3U(psize, >, 0); | |
6839 | arc_hdr_alloc_abd(hdr, B_FALSE); | |
a6255b7f DQ |
6840 | abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize); |
6841 | } else { | |
6842 | ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr)); | |
b5256303 | 6843 | arc_hdr_alloc_abd(hdr, B_FALSE); |
a6255b7f DQ |
6844 | abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, |
6845 | arc_buf_size(buf)); | |
6846 | } | |
d3c2ae1c | 6847 | } else { |
a6255b7f | 6848 | ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd)); |
2aa34383 | 6849 | ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); |
d3c2ae1c | 6850 | ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); |
d3c2ae1c | 6851 | |
d3c2ae1c | 6852 | arc_share_buf(hdr, buf); |
d3c2ae1c | 6853 | } |
a6255b7f | 6854 | |
4807c0ba | 6855 | out: |
b5256303 | 6856 | arc_hdr_verify(hdr, bp); |
a6255b7f | 6857 | spl_fstrans_unmark(cookie); |
34dc7c2f BB |
6858 | } |
6859 | ||
bc77ba73 PD |
6860 | static void |
6861 | arc_write_children_ready(zio_t *zio) | |
6862 | { | |
6863 | arc_write_callback_t *callback = zio->io_private; | |
6864 | arc_buf_t *buf = callback->awcb_buf; | |
6865 | ||
6866 | callback->awcb_children_ready(zio, buf, callback->awcb_private); | |
6867 | } | |
6868 | ||
e8b96c60 MA |
6869 | /* |
6870 | * The SPA calls this callback for each physical write that happens on behalf | |
6871 | * of a logical write. See the comment in dbuf_write_physdone() for details. | |
6872 | */ | |
6873 | static void | |
6874 | arc_write_physdone(zio_t *zio) | |
6875 | { | |
6876 | arc_write_callback_t *cb = zio->io_private; | |
6877 | if (cb->awcb_physdone != NULL) | |
6878 | cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); | |
6879 | } | |
6880 | ||
34dc7c2f BB |
6881 | static void |
6882 | arc_write_done(zio_t *zio) | |
6883 | { | |
6884 | arc_write_callback_t *callback = zio->io_private; | |
6885 | arc_buf_t *buf = callback->awcb_buf; | |
6886 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
6887 | ||
d3c2ae1c | 6888 | ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); |
428870ff BB |
6889 | |
6890 | if (zio->io_error == 0) { | |
d3c2ae1c GW |
6891 | arc_hdr_verify(hdr, zio->io_bp); |
6892 | ||
9b67f605 | 6893 | if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { |
b0bc7a84 MG |
6894 | buf_discard_identity(hdr); |
6895 | } else { | |
6896 | hdr->b_dva = *BP_IDENTITY(zio->io_bp); | |
6897 | hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); | |
b0bc7a84 | 6898 | } |
428870ff | 6899 | } else { |
d3c2ae1c | 6900 | ASSERT(HDR_EMPTY(hdr)); |
428870ff | 6901 | } |
34dc7c2f | 6902 | |
34dc7c2f | 6903 | /* |
9b67f605 MA |
6904 | * If the block to be written was all-zero or compressed enough to be |
6905 | * embedded in the BP, no write was performed so there will be no | |
6906 | * dva/birth/checksum. The buffer must therefore remain anonymous | |
6907 | * (and uncached). | |
34dc7c2f | 6908 | */ |
d3c2ae1c | 6909 | if (!HDR_EMPTY(hdr)) { |
34dc7c2f BB |
6910 | arc_buf_hdr_t *exists; |
6911 | kmutex_t *hash_lock; | |
6912 | ||
524b4217 | 6913 | ASSERT3U(zio->io_error, ==, 0); |
428870ff | 6914 | |
34dc7c2f BB |
6915 | arc_cksum_verify(buf); |
6916 | ||
6917 | exists = buf_hash_insert(hdr, &hash_lock); | |
b9541d6b | 6918 | if (exists != NULL) { |
34dc7c2f BB |
6919 | /* |
6920 | * This can only happen if we overwrite for | |
6921 | * sync-to-convergence, because we remove | |
6922 | * buffers from the hash table when we arc_free(). | |
6923 | */ | |
428870ff BB |
6924 | if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { |
6925 | if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) | |
6926 | panic("bad overwrite, hdr=%p exists=%p", | |
6927 | (void *)hdr, (void *)exists); | |
b9541d6b CW |
6928 | ASSERT(refcount_is_zero( |
6929 | &exists->b_l1hdr.b_refcnt)); | |
428870ff BB |
6930 | arc_change_state(arc_anon, exists, hash_lock); |
6931 | mutex_exit(hash_lock); | |
6932 | arc_hdr_destroy(exists); | |
6933 | exists = buf_hash_insert(hdr, &hash_lock); | |
6934 | ASSERT3P(exists, ==, NULL); | |
03c6040b GW |
6935 | } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { |
6936 | /* nopwrite */ | |
6937 | ASSERT(zio->io_prop.zp_nopwrite); | |
6938 | if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) | |
6939 | panic("bad nopwrite, hdr=%p exists=%p", | |
6940 | (void *)hdr, (void *)exists); | |
428870ff BB |
6941 | } else { |
6942 | /* Dedup */ | |
d3c2ae1c | 6943 | ASSERT(hdr->b_l1hdr.b_bufcnt == 1); |
b9541d6b | 6944 | ASSERT(hdr->b_l1hdr.b_state == arc_anon); |
428870ff BB |
6945 | ASSERT(BP_GET_DEDUP(zio->io_bp)); |
6946 | ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); | |
6947 | } | |
34dc7c2f | 6948 | } |
d3c2ae1c | 6949 | arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); |
b128c09f | 6950 | /* if it's not anon, we are doing a scrub */ |
b9541d6b | 6951 | if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) |
b128c09f | 6952 | arc_access(hdr, hash_lock); |
34dc7c2f | 6953 | mutex_exit(hash_lock); |
34dc7c2f | 6954 | } else { |
d3c2ae1c | 6955 | arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); |
34dc7c2f BB |
6956 | } |
6957 | ||
b9541d6b | 6958 | ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
428870ff | 6959 | callback->awcb_done(zio, buf, callback->awcb_private); |
34dc7c2f | 6960 | |
a6255b7f | 6961 | abd_put(zio->io_abd); |
34dc7c2f BB |
6962 | kmem_free(callback, sizeof (arc_write_callback_t)); |
6963 | } | |
6964 | ||
6965 | zio_t * | |
428870ff | 6966 | arc_write(zio_t *pio, spa_t *spa, uint64_t txg, |
d3c2ae1c | 6967 | blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, |
b5256303 TC |
6968 | const zio_prop_t *zp, arc_write_done_func_t *ready, |
6969 | arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone, | |
6970 | arc_write_done_func_t *done, void *private, zio_priority_t priority, | |
5dbd68a3 | 6971 | int zio_flags, const zbookmark_phys_t *zb) |
34dc7c2f BB |
6972 | { |
6973 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
6974 | arc_write_callback_t *callback; | |
b128c09f | 6975 | zio_t *zio; |
82644107 | 6976 | zio_prop_t localprop = *zp; |
34dc7c2f | 6977 | |
d3c2ae1c GW |
6978 | ASSERT3P(ready, !=, NULL); |
6979 | ASSERT3P(done, !=, NULL); | |
34dc7c2f | 6980 | ASSERT(!HDR_IO_ERROR(hdr)); |
b9541d6b | 6981 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); |
d3c2ae1c GW |
6982 | ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); |
6983 | ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); | |
b128c09f | 6984 | if (l2arc) |
d3c2ae1c | 6985 | arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); |
82644107 | 6986 | |
b5256303 TC |
6987 | if (ARC_BUF_ENCRYPTED(buf)) { |
6988 | ASSERT(ARC_BUF_COMPRESSED(buf)); | |
6989 | localprop.zp_encrypt = B_TRUE; | |
6990 | localprop.zp_compress = HDR_GET_COMPRESS(hdr); | |
6991 | localprop.zp_byteorder = | |
6992 | (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ? | |
6993 | ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER; | |
6994 | bcopy(hdr->b_crypt_hdr.b_salt, localprop.zp_salt, | |
6995 | ZIO_DATA_SALT_LEN); | |
6996 | bcopy(hdr->b_crypt_hdr.b_iv, localprop.zp_iv, | |
6997 | ZIO_DATA_IV_LEN); | |
6998 | bcopy(hdr->b_crypt_hdr.b_mac, localprop.zp_mac, | |
6999 | ZIO_DATA_MAC_LEN); | |
7000 | if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) { | |
7001 | localprop.zp_nopwrite = B_FALSE; | |
7002 | localprop.zp_copies = | |
7003 | MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1); | |
7004 | } | |
2aa34383 | 7005 | zio_flags |= ZIO_FLAG_RAW; |
b5256303 TC |
7006 | } else if (ARC_BUF_COMPRESSED(buf)) { |
7007 | ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); | |
7008 | localprop.zp_compress = HDR_GET_COMPRESS(hdr); | |
7009 | zio_flags |= ZIO_FLAG_RAW_COMPRESS; | |
2aa34383 | 7010 | } |
79c76d5b | 7011 | callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); |
34dc7c2f | 7012 | callback->awcb_ready = ready; |
bc77ba73 | 7013 | callback->awcb_children_ready = children_ready; |
e8b96c60 | 7014 | callback->awcb_physdone = physdone; |
34dc7c2f BB |
7015 | callback->awcb_done = done; |
7016 | callback->awcb_private = private; | |
7017 | callback->awcb_buf = buf; | |
b128c09f | 7018 | |
d3c2ae1c | 7019 | /* |
a6255b7f | 7020 | * The hdr's b_pabd is now stale, free it now. A new data block |
d3c2ae1c GW |
7021 | * will be allocated when the zio pipeline calls arc_write_ready(). |
7022 | */ | |
a6255b7f | 7023 | if (hdr->b_l1hdr.b_pabd != NULL) { |
d3c2ae1c GW |
7024 | /* |
7025 | * If the buf is currently sharing the data block with | |
7026 | * the hdr then we need to break that relationship here. | |
7027 | * The hdr will remain with a NULL data pointer and the | |
7028 | * buf will take sole ownership of the block. | |
7029 | */ | |
7030 | if (arc_buf_is_shared(buf)) { | |
d3c2ae1c GW |
7031 | arc_unshare_buf(hdr, buf); |
7032 | } else { | |
b5256303 | 7033 | arc_hdr_free_abd(hdr, B_FALSE); |
d3c2ae1c GW |
7034 | } |
7035 | VERIFY3P(buf->b_data, !=, NULL); | |
d3c2ae1c | 7036 | } |
b5256303 TC |
7037 | |
7038 | if (HDR_HAS_RABD(hdr)) | |
7039 | arc_hdr_free_abd(hdr, B_TRUE); | |
7040 | ||
71a24c3c TC |
7041 | if (!(zio_flags & ZIO_FLAG_RAW)) |
7042 | arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); | |
b5256303 | 7043 | |
d3c2ae1c | 7044 | ASSERT(!arc_buf_is_shared(buf)); |
a6255b7f | 7045 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
d3c2ae1c | 7046 | |
a6255b7f DQ |
7047 | zio = zio_write(pio, spa, txg, bp, |
7048 | abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)), | |
82644107 | 7049 | HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready, |
bc77ba73 PD |
7050 | (children_ready != NULL) ? arc_write_children_ready : NULL, |
7051 | arc_write_physdone, arc_write_done, callback, | |
e8b96c60 | 7052 | priority, zio_flags, zb); |
34dc7c2f BB |
7053 | |
7054 | return (zio); | |
7055 | } | |
7056 | ||
34dc7c2f | 7057 | static int |
e8b96c60 | 7058 | arc_memory_throttle(uint64_t reserve, uint64_t txg) |
34dc7c2f BB |
7059 | { |
7060 | #ifdef _KERNEL | |
70f02287 | 7061 | uint64_t available_memory = arc_free_memory(); |
7e8bddd0 BB |
7062 | static uint64_t page_load = 0; |
7063 | static uint64_t last_txg = 0; | |
0c5493d4 | 7064 | |
70f02287 | 7065 | #if defined(_ILP32) |
9edb3695 BB |
7066 | available_memory = |
7067 | MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); | |
7068 | #endif | |
7069 | ||
7070 | if (available_memory > arc_all_memory() * arc_lotsfree_percent / 100) | |
ca67b33a MA |
7071 | return (0); |
7072 | ||
7e8bddd0 BB |
7073 | if (txg > last_txg) { |
7074 | last_txg = txg; | |
7075 | page_load = 0; | |
7076 | } | |
7e8bddd0 BB |
7077 | /* |
7078 | * If we are in pageout, we know that memory is already tight, | |
7079 | * the arc is already going to be evicting, so we just want to | |
7080 | * continue to let page writes occur as quickly as possible. | |
7081 | */ | |
7082 | if (current_is_kswapd()) { | |
70f02287 | 7083 | if (page_load > MAX(arc_sys_free / 4, available_memory) / 4) { |
7e8bddd0 BB |
7084 | DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim); |
7085 | return (SET_ERROR(ERESTART)); | |
7086 | } | |
7087 | /* Note: reserve is inflated, so we deflate */ | |
7088 | page_load += reserve / 8; | |
7089 | return (0); | |
7090 | } else if (page_load > 0 && arc_reclaim_needed()) { | |
ca67b33a | 7091 | /* memory is low, delay before restarting */ |
34dc7c2f | 7092 | ARCSTAT_INCR(arcstat_memory_throttle_count, 1); |
570827e1 | 7093 | DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim); |
2e528b49 | 7094 | return (SET_ERROR(EAGAIN)); |
34dc7c2f | 7095 | } |
7e8bddd0 | 7096 | page_load = 0; |
34dc7c2f BB |
7097 | #endif |
7098 | return (0); | |
7099 | } | |
7100 | ||
7101 | void | |
7102 | arc_tempreserve_clear(uint64_t reserve) | |
7103 | { | |
7104 | atomic_add_64(&arc_tempreserve, -reserve); | |
7105 | ASSERT((int64_t)arc_tempreserve >= 0); | |
7106 | } | |
7107 | ||
7108 | int | |
7109 | arc_tempreserve_space(uint64_t reserve, uint64_t txg) | |
7110 | { | |
7111 | int error; | |
9babb374 | 7112 | uint64_t anon_size; |
34dc7c2f | 7113 | |
1b8951b3 TC |
7114 | if (!arc_no_grow && |
7115 | reserve > arc_c/4 && | |
7116 | reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT)) | |
34dc7c2f | 7117 | arc_c = MIN(arc_c_max, reserve * 4); |
12f9a6a3 BB |
7118 | |
7119 | /* | |
7120 | * Throttle when the calculated memory footprint for the TXG | |
7121 | * exceeds the target ARC size. | |
7122 | */ | |
570827e1 BB |
7123 | if (reserve > arc_c) { |
7124 | DMU_TX_STAT_BUMP(dmu_tx_memory_reserve); | |
12f9a6a3 | 7125 | return (SET_ERROR(ERESTART)); |
570827e1 | 7126 | } |
34dc7c2f | 7127 | |
9babb374 BB |
7128 | /* |
7129 | * Don't count loaned bufs as in flight dirty data to prevent long | |
7130 | * network delays from blocking transactions that are ready to be | |
7131 | * assigned to a txg. | |
7132 | */ | |
a7004725 DK |
7133 | |
7134 | /* assert that it has not wrapped around */ | |
7135 | ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); | |
7136 | ||
36da08ef PS |
7137 | anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - |
7138 | arc_loaned_bytes), 0); | |
9babb374 | 7139 | |
34dc7c2f BB |
7140 | /* |
7141 | * Writes will, almost always, require additional memory allocations | |
d3cc8b15 | 7142 | * in order to compress/encrypt/etc the data. We therefore need to |
34dc7c2f BB |
7143 | * make sure that there is sufficient available memory for this. |
7144 | */ | |
e8b96c60 MA |
7145 | error = arc_memory_throttle(reserve, txg); |
7146 | if (error != 0) | |
34dc7c2f BB |
7147 | return (error); |
7148 | ||
7149 | /* | |
7150 | * Throttle writes when the amount of dirty data in the cache | |
7151 | * gets too large. We try to keep the cache less than half full | |
7152 | * of dirty blocks so that our sync times don't grow too large. | |
7153 | * Note: if two requests come in concurrently, we might let them | |
7154 | * both succeed, when one of them should fail. Not a huge deal. | |
7155 | */ | |
9babb374 BB |
7156 | |
7157 | if (reserve + arc_tempreserve + anon_size > arc_c / 2 && | |
7158 | anon_size > arc_c / 4) { | |
2fd92c3d | 7159 | #ifdef ZFS_DEBUG |
d3c2ae1c GW |
7160 | uint64_t meta_esize = |
7161 | refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); | |
7162 | uint64_t data_esize = | |
7163 | refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); | |
34dc7c2f BB |
7164 | dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " |
7165 | "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", | |
d3c2ae1c GW |
7166 | arc_tempreserve >> 10, meta_esize >> 10, |
7167 | data_esize >> 10, reserve >> 10, arc_c >> 10); | |
2fd92c3d | 7168 | #endif |
570827e1 | 7169 | DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle); |
2e528b49 | 7170 | return (SET_ERROR(ERESTART)); |
34dc7c2f BB |
7171 | } |
7172 | atomic_add_64(&arc_tempreserve, reserve); | |
7173 | return (0); | |
7174 | } | |
7175 | ||
13be560d BB |
7176 | static void |
7177 | arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, | |
7178 | kstat_named_t *evict_data, kstat_named_t *evict_metadata) | |
7179 | { | |
36da08ef | 7180 | size->value.ui64 = refcount_count(&state->arcs_size); |
d3c2ae1c GW |
7181 | evict_data->value.ui64 = |
7182 | refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); | |
7183 | evict_metadata->value.ui64 = | |
7184 | refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); | |
13be560d BB |
7185 | } |
7186 | ||
7187 | static int | |
7188 | arc_kstat_update(kstat_t *ksp, int rw) | |
7189 | { | |
7190 | arc_stats_t *as = ksp->ks_data; | |
7191 | ||
7192 | if (rw == KSTAT_WRITE) { | |
ecb2b7dc | 7193 | return (SET_ERROR(EACCES)); |
13be560d BB |
7194 | } else { |
7195 | arc_kstat_update_state(arc_anon, | |
7196 | &as->arcstat_anon_size, | |
500445c0 PS |
7197 | &as->arcstat_anon_evictable_data, |
7198 | &as->arcstat_anon_evictable_metadata); | |
13be560d BB |
7199 | arc_kstat_update_state(arc_mru, |
7200 | &as->arcstat_mru_size, | |
500445c0 PS |
7201 | &as->arcstat_mru_evictable_data, |
7202 | &as->arcstat_mru_evictable_metadata); | |
13be560d BB |
7203 | arc_kstat_update_state(arc_mru_ghost, |
7204 | &as->arcstat_mru_ghost_size, | |
500445c0 PS |
7205 | &as->arcstat_mru_ghost_evictable_data, |
7206 | &as->arcstat_mru_ghost_evictable_metadata); | |
13be560d BB |
7207 | arc_kstat_update_state(arc_mfu, |
7208 | &as->arcstat_mfu_size, | |
500445c0 PS |
7209 | &as->arcstat_mfu_evictable_data, |
7210 | &as->arcstat_mfu_evictable_metadata); | |
fc41c640 | 7211 | arc_kstat_update_state(arc_mfu_ghost, |
13be560d | 7212 | &as->arcstat_mfu_ghost_size, |
500445c0 PS |
7213 | &as->arcstat_mfu_ghost_evictable_data, |
7214 | &as->arcstat_mfu_ghost_evictable_metadata); | |
70f02287 BB |
7215 | |
7216 | as->arcstat_memory_all_bytes.value.ui64 = | |
7217 | arc_all_memory(); | |
7218 | as->arcstat_memory_free_bytes.value.ui64 = | |
7219 | arc_free_memory(); | |
7220 | as->arcstat_memory_available_bytes.value.i64 = | |
7221 | arc_available_memory(); | |
13be560d BB |
7222 | } |
7223 | ||
7224 | return (0); | |
7225 | } | |
7226 | ||
ca0bf58d PS |
7227 | /* |
7228 | * This function *must* return indices evenly distributed between all | |
7229 | * sublists of the multilist. This is needed due to how the ARC eviction | |
7230 | * code is laid out; arc_evict_state() assumes ARC buffers are evenly | |
7231 | * distributed between all sublists and uses this assumption when | |
7232 | * deciding which sublist to evict from and how much to evict from it. | |
7233 | */ | |
7234 | unsigned int | |
7235 | arc_state_multilist_index_func(multilist_t *ml, void *obj) | |
7236 | { | |
7237 | arc_buf_hdr_t *hdr = obj; | |
7238 | ||
7239 | /* | |
7240 | * We rely on b_dva to generate evenly distributed index | |
7241 | * numbers using buf_hash below. So, as an added precaution, | |
7242 | * let's make sure we never add empty buffers to the arc lists. | |
7243 | */ | |
d3c2ae1c | 7244 | ASSERT(!HDR_EMPTY(hdr)); |
ca0bf58d PS |
7245 | |
7246 | /* | |
7247 | * The assumption here, is the hash value for a given | |
7248 | * arc_buf_hdr_t will remain constant throughout its lifetime | |
7249 | * (i.e. its b_spa, b_dva, and b_birth fields don't change). | |
7250 | * Thus, we don't need to store the header's sublist index | |
7251 | * on insertion, as this index can be recalculated on removal. | |
7252 | * | |
7253 | * Also, the low order bits of the hash value are thought to be | |
7254 | * distributed evenly. Otherwise, in the case that the multilist | |
7255 | * has a power of two number of sublists, each sublists' usage | |
7256 | * would not be evenly distributed. | |
7257 | */ | |
7258 | return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % | |
7259 | multilist_get_num_sublists(ml)); | |
7260 | } | |
7261 | ||
ca67b33a MA |
7262 | /* |
7263 | * Called during module initialization and periodically thereafter to | |
7264 | * apply reasonable changes to the exposed performance tunings. Non-zero | |
7265 | * zfs_* values which differ from the currently set values will be applied. | |
7266 | */ | |
7267 | static void | |
7268 | arc_tuning_update(void) | |
7269 | { | |
b8a97fb1 | 7270 | uint64_t allmem = arc_all_memory(); |
7271 | unsigned long limit; | |
9edb3695 | 7272 | |
ca67b33a MA |
7273 | /* Valid range: 64M - <all physical memory> */ |
7274 | if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) && | |
7403d074 | 7275 | (zfs_arc_max >= 64 << 20) && (zfs_arc_max < allmem) && |
ca67b33a MA |
7276 | (zfs_arc_max > arc_c_min)) { |
7277 | arc_c_max = zfs_arc_max; | |
7278 | arc_c = arc_c_max; | |
7279 | arc_p = (arc_c >> 1); | |
b8a97fb1 | 7280 | if (arc_meta_limit > arc_c_max) |
7281 | arc_meta_limit = arc_c_max; | |
7282 | if (arc_dnode_limit > arc_meta_limit) | |
7283 | arc_dnode_limit = arc_meta_limit; | |
ca67b33a MA |
7284 | } |
7285 | ||
7286 | /* Valid range: 32M - <arc_c_max> */ | |
7287 | if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) && | |
7288 | (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) && | |
7289 | (zfs_arc_min <= arc_c_max)) { | |
7290 | arc_c_min = zfs_arc_min; | |
7291 | arc_c = MAX(arc_c, arc_c_min); | |
7292 | } | |
7293 | ||
7294 | /* Valid range: 16M - <arc_c_max> */ | |
7295 | if ((zfs_arc_meta_min) && (zfs_arc_meta_min != arc_meta_min) && | |
7296 | (zfs_arc_meta_min >= 1ULL << SPA_MAXBLOCKSHIFT) && | |
7297 | (zfs_arc_meta_min <= arc_c_max)) { | |
7298 | arc_meta_min = zfs_arc_meta_min; | |
b8a97fb1 | 7299 | if (arc_meta_limit < arc_meta_min) |
7300 | arc_meta_limit = arc_meta_min; | |
7301 | if (arc_dnode_limit < arc_meta_min) | |
7302 | arc_dnode_limit = arc_meta_min; | |
ca67b33a MA |
7303 | } |
7304 | ||
7305 | /* Valid range: <arc_meta_min> - <arc_c_max> */ | |
b8a97fb1 | 7306 | limit = zfs_arc_meta_limit ? zfs_arc_meta_limit : |
7307 | MIN(zfs_arc_meta_limit_percent, 100) * arc_c_max / 100; | |
7308 | if ((limit != arc_meta_limit) && | |
7309 | (limit >= arc_meta_min) && | |
7310 | (limit <= arc_c_max)) | |
7311 | arc_meta_limit = limit; | |
7312 | ||
7313 | /* Valid range: <arc_meta_min> - <arc_meta_limit> */ | |
7314 | limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit : | |
7315 | MIN(zfs_arc_dnode_limit_percent, 100) * arc_meta_limit / 100; | |
7316 | if ((limit != arc_dnode_limit) && | |
7317 | (limit >= arc_meta_min) && | |
7318 | (limit <= arc_meta_limit)) | |
7319 | arc_dnode_limit = limit; | |
25458cbe | 7320 | |
ca67b33a MA |
7321 | /* Valid range: 1 - N */ |
7322 | if (zfs_arc_grow_retry) | |
7323 | arc_grow_retry = zfs_arc_grow_retry; | |
7324 | ||
7325 | /* Valid range: 1 - N */ | |
7326 | if (zfs_arc_shrink_shift) { | |
7327 | arc_shrink_shift = zfs_arc_shrink_shift; | |
7328 | arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1); | |
7329 | } | |
7330 | ||
728d6ae9 BB |
7331 | /* Valid range: 1 - N */ |
7332 | if (zfs_arc_p_min_shift) | |
7333 | arc_p_min_shift = zfs_arc_p_min_shift; | |
7334 | ||
d4a72f23 TC |
7335 | /* Valid range: 1 - N ms */ |
7336 | if (zfs_arc_min_prefetch_ms) | |
7337 | arc_min_prefetch_ms = zfs_arc_min_prefetch_ms; | |
7338 | ||
7339 | /* Valid range: 1 - N ms */ | |
7340 | if (zfs_arc_min_prescient_prefetch_ms) { | |
7341 | arc_min_prescient_prefetch_ms = | |
7342 | zfs_arc_min_prescient_prefetch_ms; | |
7343 | } | |
11f552fa | 7344 | |
7e8bddd0 BB |
7345 | /* Valid range: 0 - 100 */ |
7346 | if ((zfs_arc_lotsfree_percent >= 0) && | |
7347 | (zfs_arc_lotsfree_percent <= 100)) | |
7348 | arc_lotsfree_percent = zfs_arc_lotsfree_percent; | |
7349 | ||
11f552fa BB |
7350 | /* Valid range: 0 - <all physical memory> */ |
7351 | if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free)) | |
9edb3695 | 7352 | arc_sys_free = MIN(MAX(zfs_arc_sys_free, 0), allmem); |
7e8bddd0 | 7353 | |
ca67b33a MA |
7354 | } |
7355 | ||
d3c2ae1c GW |
7356 | static void |
7357 | arc_state_init(void) | |
7358 | { | |
7359 | arc_anon = &ARC_anon; | |
7360 | arc_mru = &ARC_mru; | |
7361 | arc_mru_ghost = &ARC_mru_ghost; | |
7362 | arc_mfu = &ARC_mfu; | |
7363 | arc_mfu_ghost = &ARC_mfu_ghost; | |
7364 | arc_l2c_only = &ARC_l2c_only; | |
7365 | ||
64fc7762 MA |
7366 | arc_mru->arcs_list[ARC_BUFC_METADATA] = |
7367 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7368 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7369 | arc_state_multilist_index_func); |
64fc7762 MA |
7370 | arc_mru->arcs_list[ARC_BUFC_DATA] = |
7371 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7372 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7373 | arc_state_multilist_index_func); |
64fc7762 MA |
7374 | arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] = |
7375 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7376 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7377 | arc_state_multilist_index_func); |
64fc7762 MA |
7378 | arc_mru_ghost->arcs_list[ARC_BUFC_DATA] = |
7379 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7380 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7381 | arc_state_multilist_index_func); |
64fc7762 MA |
7382 | arc_mfu->arcs_list[ARC_BUFC_METADATA] = |
7383 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7384 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7385 | arc_state_multilist_index_func); |
64fc7762 MA |
7386 | arc_mfu->arcs_list[ARC_BUFC_DATA] = |
7387 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7388 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7389 | arc_state_multilist_index_func); |
64fc7762 MA |
7390 | arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] = |
7391 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7392 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7393 | arc_state_multilist_index_func); |
64fc7762 MA |
7394 | arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] = |
7395 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7396 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7397 | arc_state_multilist_index_func); |
64fc7762 MA |
7398 | arc_l2c_only->arcs_list[ARC_BUFC_METADATA] = |
7399 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7400 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7401 | arc_state_multilist_index_func); |
64fc7762 MA |
7402 | arc_l2c_only->arcs_list[ARC_BUFC_DATA] = |
7403 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7404 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7405 | arc_state_multilist_index_func); |
d3c2ae1c GW |
7406 | |
7407 | refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); | |
7408 | refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); | |
7409 | refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); | |
7410 | refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); | |
7411 | refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); | |
7412 | refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); | |
7413 | refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); | |
7414 | refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); | |
7415 | refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); | |
7416 | refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); | |
7417 | refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); | |
7418 | refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); | |
7419 | ||
7420 | refcount_create(&arc_anon->arcs_size); | |
7421 | refcount_create(&arc_mru->arcs_size); | |
7422 | refcount_create(&arc_mru_ghost->arcs_size); | |
7423 | refcount_create(&arc_mfu->arcs_size); | |
7424 | refcount_create(&arc_mfu_ghost->arcs_size); | |
7425 | refcount_create(&arc_l2c_only->arcs_size); | |
7426 | ||
7427 | arc_anon->arcs_state = ARC_STATE_ANON; | |
7428 | arc_mru->arcs_state = ARC_STATE_MRU; | |
7429 | arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST; | |
7430 | arc_mfu->arcs_state = ARC_STATE_MFU; | |
7431 | arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST; | |
7432 | arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY; | |
7433 | } | |
7434 | ||
7435 | static void | |
7436 | arc_state_fini(void) | |
7437 | { | |
7438 | refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); | |
7439 | refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); | |
7440 | refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); | |
7441 | refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); | |
7442 | refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); | |
7443 | refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); | |
7444 | refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); | |
7445 | refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); | |
7446 | refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); | |
7447 | refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); | |
7448 | refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); | |
7449 | refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); | |
7450 | ||
7451 | refcount_destroy(&arc_anon->arcs_size); | |
7452 | refcount_destroy(&arc_mru->arcs_size); | |
7453 | refcount_destroy(&arc_mru_ghost->arcs_size); | |
7454 | refcount_destroy(&arc_mfu->arcs_size); | |
7455 | refcount_destroy(&arc_mfu_ghost->arcs_size); | |
7456 | refcount_destroy(&arc_l2c_only->arcs_size); | |
7457 | ||
64fc7762 MA |
7458 | multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]); |
7459 | multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); | |
7460 | multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]); | |
7461 | multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); | |
7462 | multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]); | |
7463 | multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); | |
7464 | multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]); | |
7465 | multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); | |
7466 | multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]); | |
7467 | multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_DATA]); | |
d3c2ae1c GW |
7468 | } |
7469 | ||
7470 | uint64_t | |
e71cade6 | 7471 | arc_target_bytes(void) |
d3c2ae1c | 7472 | { |
e71cade6 | 7473 | return (arc_c); |
d3c2ae1c GW |
7474 | } |
7475 | ||
34dc7c2f BB |
7476 | void |
7477 | arc_init(void) | |
7478 | { | |
9edb3695 | 7479 | uint64_t percent, allmem = arc_all_memory(); |
ca67b33a | 7480 | |
ca0bf58d PS |
7481 | mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); |
7482 | cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); | |
7483 | cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); | |
7484 | ||
2b84817f TC |
7485 | arc_min_prefetch_ms = 1000; |
7486 | arc_min_prescient_prefetch_ms = 6000; | |
34dc7c2f | 7487 | |
34dc7c2f | 7488 | #ifdef _KERNEL |
7cb67b45 BB |
7489 | /* |
7490 | * Register a shrinker to support synchronous (direct) memory | |
7491 | * reclaim from the arc. This is done to prevent kswapd from | |
7492 | * swapping out pages when it is preferable to shrink the arc. | |
7493 | */ | |
7494 | spl_register_shrinker(&arc_shrinker); | |
11f552fa BB |
7495 | |
7496 | /* Set to 1/64 of all memory or a minimum of 512K */ | |
9edb3695 | 7497 | arc_sys_free = MAX(allmem / 64, (512 * 1024)); |
11f552fa | 7498 | arc_need_free = 0; |
34dc7c2f BB |
7499 | #endif |
7500 | ||
0a1f8cd9 TC |
7501 | /* Set max to 1/2 of all memory */ |
7502 | arc_c_max = allmem / 2; | |
7503 | ||
4ce3c45a BB |
7504 | #ifdef _KERNEL |
7505 | /* Set min cache to 1/32 of all memory, or 32MB, whichever is more */ | |
7506 | arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT); | |
7507 | #else | |
ab5cbbd1 BB |
7508 | /* |
7509 | * In userland, there's only the memory pressure that we artificially | |
7510 | * create (see arc_available_memory()). Don't let arc_c get too | |
7511 | * small, because it can cause transactions to be larger than | |
7512 | * arc_c, causing arc_tempreserve_space() to fail. | |
7513 | */ | |
0a1f8cd9 | 7514 | arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT); |
ab5cbbd1 BB |
7515 | #endif |
7516 | ||
34dc7c2f BB |
7517 | arc_c = arc_c_max; |
7518 | arc_p = (arc_c >> 1); | |
d3c2ae1c | 7519 | arc_size = 0; |
34dc7c2f | 7520 | |
ca67b33a MA |
7521 | /* Set min to 1/2 of arc_c_min */ |
7522 | arc_meta_min = 1ULL << SPA_MAXBLOCKSHIFT; | |
7523 | /* Initialize maximum observed usage to zero */ | |
1834f2d8 | 7524 | arc_meta_max = 0; |
9907cc1c G |
7525 | /* |
7526 | * Set arc_meta_limit to a percent of arc_c_max with a floor of | |
7527 | * arc_meta_min, and a ceiling of arc_c_max. | |
7528 | */ | |
7529 | percent = MIN(zfs_arc_meta_limit_percent, 100); | |
7530 | arc_meta_limit = MAX(arc_meta_min, (percent * arc_c_max) / 100); | |
7531 | percent = MIN(zfs_arc_dnode_limit_percent, 100); | |
7532 | arc_dnode_limit = (percent * arc_meta_limit) / 100; | |
34dc7c2f | 7533 | |
ca67b33a MA |
7534 | /* Apply user specified tunings */ |
7535 | arc_tuning_update(); | |
c52fca13 | 7536 | |
34dc7c2f BB |
7537 | /* if kmem_flags are set, lets try to use less memory */ |
7538 | if (kmem_debugging()) | |
7539 | arc_c = arc_c / 2; | |
7540 | if (arc_c < arc_c_min) | |
7541 | arc_c = arc_c_min; | |
7542 | ||
d3c2ae1c | 7543 | arc_state_init(); |
34dc7c2f BB |
7544 | buf_init(); |
7545 | ||
ab26409d BB |
7546 | list_create(&arc_prune_list, sizeof (arc_prune_t), |
7547 | offsetof(arc_prune_t, p_node)); | |
ab26409d | 7548 | mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f | 7549 | |
1229323d | 7550 | arc_prune_taskq = taskq_create("arc_prune", max_ncpus, defclsyspri, |
aa9af22c | 7551 | max_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC); |
f6046738 | 7552 | |
d3c2ae1c GW |
7553 | arc_reclaim_thread_exit = B_FALSE; |
7554 | ||
34dc7c2f BB |
7555 | arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, |
7556 | sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); | |
7557 | ||
7558 | if (arc_ksp != NULL) { | |
7559 | arc_ksp->ks_data = &arc_stats; | |
13be560d | 7560 | arc_ksp->ks_update = arc_kstat_update; |
34dc7c2f BB |
7561 | kstat_install(arc_ksp); |
7562 | } | |
7563 | ||
ca67b33a | 7564 | (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, |
1229323d | 7565 | TS_RUN, defclsyspri); |
34dc7c2f | 7566 | |
d3c2ae1c | 7567 | arc_dead = B_FALSE; |
b128c09f | 7568 | arc_warm = B_FALSE; |
34dc7c2f | 7569 | |
e8b96c60 MA |
7570 | /* |
7571 | * Calculate maximum amount of dirty data per pool. | |
7572 | * | |
7573 | * If it has been set by a module parameter, take that. | |
7574 | * Otherwise, use a percentage of physical memory defined by | |
7575 | * zfs_dirty_data_max_percent (default 10%) with a cap at | |
e99932f7 | 7576 | * zfs_dirty_data_max_max (default 4G or 25% of physical memory). |
e8b96c60 MA |
7577 | */ |
7578 | if (zfs_dirty_data_max_max == 0) | |
e99932f7 BB |
7579 | zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024, |
7580 | allmem * zfs_dirty_data_max_max_percent / 100); | |
e8b96c60 MA |
7581 | |
7582 | if (zfs_dirty_data_max == 0) { | |
9edb3695 | 7583 | zfs_dirty_data_max = allmem * |
e8b96c60 MA |
7584 | zfs_dirty_data_max_percent / 100; |
7585 | zfs_dirty_data_max = MIN(zfs_dirty_data_max, | |
7586 | zfs_dirty_data_max_max); | |
7587 | } | |
34dc7c2f BB |
7588 | } |
7589 | ||
7590 | void | |
7591 | arc_fini(void) | |
7592 | { | |
ab26409d BB |
7593 | arc_prune_t *p; |
7594 | ||
7cb67b45 BB |
7595 | #ifdef _KERNEL |
7596 | spl_unregister_shrinker(&arc_shrinker); | |
7597 | #endif /* _KERNEL */ | |
7598 | ||
ca0bf58d | 7599 | mutex_enter(&arc_reclaim_lock); |
d3c2ae1c | 7600 | arc_reclaim_thread_exit = B_TRUE; |
ca0bf58d PS |
7601 | /* |
7602 | * The reclaim thread will set arc_reclaim_thread_exit back to | |
d3c2ae1c | 7603 | * B_FALSE when it is finished exiting; we're waiting for that. |
ca0bf58d PS |
7604 | */ |
7605 | while (arc_reclaim_thread_exit) { | |
7606 | cv_signal(&arc_reclaim_thread_cv); | |
7607 | cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); | |
7608 | } | |
7609 | mutex_exit(&arc_reclaim_lock); | |
7610 | ||
d3c2ae1c GW |
7611 | /* Use B_TRUE to ensure *all* buffers are evicted */ |
7612 | arc_flush(NULL, B_TRUE); | |
34dc7c2f | 7613 | |
d3c2ae1c | 7614 | arc_dead = B_TRUE; |
34dc7c2f BB |
7615 | |
7616 | if (arc_ksp != NULL) { | |
7617 | kstat_delete(arc_ksp); | |
7618 | arc_ksp = NULL; | |
7619 | } | |
7620 | ||
f6046738 BB |
7621 | taskq_wait(arc_prune_taskq); |
7622 | taskq_destroy(arc_prune_taskq); | |
7623 | ||
ab26409d BB |
7624 | mutex_enter(&arc_prune_mtx); |
7625 | while ((p = list_head(&arc_prune_list)) != NULL) { | |
7626 | list_remove(&arc_prune_list, p); | |
7627 | refcount_remove(&p->p_refcnt, &arc_prune_list); | |
7628 | refcount_destroy(&p->p_refcnt); | |
7629 | kmem_free(p, sizeof (*p)); | |
7630 | } | |
7631 | mutex_exit(&arc_prune_mtx); | |
7632 | ||
7633 | list_destroy(&arc_prune_list); | |
7634 | mutex_destroy(&arc_prune_mtx); | |
ca0bf58d PS |
7635 | mutex_destroy(&arc_reclaim_lock); |
7636 | cv_destroy(&arc_reclaim_thread_cv); | |
7637 | cv_destroy(&arc_reclaim_waiters_cv); | |
7638 | ||
d3c2ae1c | 7639 | arc_state_fini(); |
34dc7c2f | 7640 | buf_fini(); |
9babb374 | 7641 | |
b9541d6b | 7642 | ASSERT0(arc_loaned_bytes); |
34dc7c2f BB |
7643 | } |
7644 | ||
7645 | /* | |
7646 | * Level 2 ARC | |
7647 | * | |
7648 | * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. | |
7649 | * It uses dedicated storage devices to hold cached data, which are populated | |
7650 | * using large infrequent writes. The main role of this cache is to boost | |
7651 | * the performance of random read workloads. The intended L2ARC devices | |
7652 | * include short-stroked disks, solid state disks, and other media with | |
7653 | * substantially faster read latency than disk. | |
7654 | * | |
7655 | * +-----------------------+ | |
7656 | * | ARC | | |
7657 | * +-----------------------+ | |
7658 | * | ^ ^ | |
7659 | * | | | | |
7660 | * l2arc_feed_thread() arc_read() | |
7661 | * | | | | |
7662 | * | l2arc read | | |
7663 | * V | | | |
7664 | * +---------------+ | | |
7665 | * | L2ARC | | | |
7666 | * +---------------+ | | |
7667 | * | ^ | | |
7668 | * l2arc_write() | | | |
7669 | * | | | | |
7670 | * V | | | |
7671 | * +-------+ +-------+ | |
7672 | * | vdev | | vdev | | |
7673 | * | cache | | cache | | |
7674 | * +-------+ +-------+ | |
7675 | * +=========+ .-----. | |
7676 | * : L2ARC : |-_____-| | |
7677 | * : devices : | Disks | | |
7678 | * +=========+ `-_____-' | |
7679 | * | |
7680 | * Read requests are satisfied from the following sources, in order: | |
7681 | * | |
7682 | * 1) ARC | |
7683 | * 2) vdev cache of L2ARC devices | |
7684 | * 3) L2ARC devices | |
7685 | * 4) vdev cache of disks | |
7686 | * 5) disks | |
7687 | * | |
7688 | * Some L2ARC device types exhibit extremely slow write performance. | |
7689 | * To accommodate for this there are some significant differences between | |
7690 | * the L2ARC and traditional cache design: | |
7691 | * | |
7692 | * 1. There is no eviction path from the ARC to the L2ARC. Evictions from | |
7693 | * the ARC behave as usual, freeing buffers and placing headers on ghost | |
7694 | * lists. The ARC does not send buffers to the L2ARC during eviction as | |
7695 | * this would add inflated write latencies for all ARC memory pressure. | |
7696 | * | |
7697 | * 2. The L2ARC attempts to cache data from the ARC before it is evicted. | |
7698 | * It does this by periodically scanning buffers from the eviction-end of | |
7699 | * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are | |
3a17a7a9 SK |
7700 | * not already there. It scans until a headroom of buffers is satisfied, |
7701 | * which itself is a buffer for ARC eviction. If a compressible buffer is | |
7702 | * found during scanning and selected for writing to an L2ARC device, we | |
7703 | * temporarily boost scanning headroom during the next scan cycle to make | |
7704 | * sure we adapt to compression effects (which might significantly reduce | |
7705 | * the data volume we write to L2ARC). The thread that does this is | |
34dc7c2f BB |
7706 | * l2arc_feed_thread(), illustrated below; example sizes are included to |
7707 | * provide a better sense of ratio than this diagram: | |
7708 | * | |
7709 | * head --> tail | |
7710 | * +---------------------+----------+ | |
7711 | * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC | |
7712 | * +---------------------+----------+ | o L2ARC eligible | |
7713 | * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer | |
7714 | * +---------------------+----------+ | | |
7715 | * 15.9 Gbytes ^ 32 Mbytes | | |
7716 | * headroom | | |
7717 | * l2arc_feed_thread() | |
7718 | * | | |
7719 | * l2arc write hand <--[oooo]--' | |
7720 | * | 8 Mbyte | |
7721 | * | write max | |
7722 | * V | |
7723 | * +==============================+ | |
7724 | * L2ARC dev |####|#|###|###| |####| ... | | |
7725 | * +==============================+ | |
7726 | * 32 Gbytes | |
7727 | * | |
7728 | * 3. If an ARC buffer is copied to the L2ARC but then hit instead of | |
7729 | * evicted, then the L2ARC has cached a buffer much sooner than it probably | |
7730 | * needed to, potentially wasting L2ARC device bandwidth and storage. It is | |
7731 | * safe to say that this is an uncommon case, since buffers at the end of | |
7732 | * the ARC lists have moved there due to inactivity. | |
7733 | * | |
7734 | * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, | |
7735 | * then the L2ARC simply misses copying some buffers. This serves as a | |
7736 | * pressure valve to prevent heavy read workloads from both stalling the ARC | |
7737 | * with waits and clogging the L2ARC with writes. This also helps prevent | |
7738 | * the potential for the L2ARC to churn if it attempts to cache content too | |
7739 | * quickly, such as during backups of the entire pool. | |
7740 | * | |
b128c09f BB |
7741 | * 5. After system boot and before the ARC has filled main memory, there are |
7742 | * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru | |
7743 | * lists can remain mostly static. Instead of searching from tail of these | |
7744 | * lists as pictured, the l2arc_feed_thread() will search from the list heads | |
7745 | * for eligible buffers, greatly increasing its chance of finding them. | |
7746 | * | |
7747 | * The L2ARC device write speed is also boosted during this time so that | |
7748 | * the L2ARC warms up faster. Since there have been no ARC evictions yet, | |
7749 | * there are no L2ARC reads, and no fear of degrading read performance | |
7750 | * through increased writes. | |
7751 | * | |
7752 | * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that | |
34dc7c2f BB |
7753 | * the vdev queue can aggregate them into larger and fewer writes. Each |
7754 | * device is written to in a rotor fashion, sweeping writes through | |
7755 | * available space then repeating. | |
7756 | * | |
b128c09f | 7757 | * 7. The L2ARC does not store dirty content. It never needs to flush |
34dc7c2f BB |
7758 | * write buffers back to disk based storage. |
7759 | * | |
b128c09f | 7760 | * 8. If an ARC buffer is written (and dirtied) which also exists in the |
34dc7c2f BB |
7761 | * L2ARC, the now stale L2ARC buffer is immediately dropped. |
7762 | * | |
7763 | * The performance of the L2ARC can be tweaked by a number of tunables, which | |
7764 | * may be necessary for different workloads: | |
7765 | * | |
7766 | * l2arc_write_max max write bytes per interval | |
b128c09f | 7767 | * l2arc_write_boost extra write bytes during device warmup |
34dc7c2f BB |
7768 | * l2arc_noprefetch skip caching prefetched buffers |
7769 | * l2arc_headroom number of max device writes to precache | |
3a17a7a9 SK |
7770 | * l2arc_headroom_boost when we find compressed buffers during ARC |
7771 | * scanning, we multiply headroom by this | |
7772 | * percentage factor for the next scan cycle, | |
7773 | * since more compressed buffers are likely to | |
7774 | * be present | |
34dc7c2f BB |
7775 | * l2arc_feed_secs seconds between L2ARC writing |
7776 | * | |
7777 | * Tunables may be removed or added as future performance improvements are | |
7778 | * integrated, and also may become zpool properties. | |
d164b209 BB |
7779 | * |
7780 | * There are three key functions that control how the L2ARC warms up: | |
7781 | * | |
7782 | * l2arc_write_eligible() check if a buffer is eligible to cache | |
7783 | * l2arc_write_size() calculate how much to write | |
7784 | * l2arc_write_interval() calculate sleep delay between writes | |
7785 | * | |
7786 | * These three functions determine what to write, how much, and how quickly | |
7787 | * to send writes. | |
34dc7c2f BB |
7788 | */ |
7789 | ||
d164b209 | 7790 | static boolean_t |
2a432414 | 7791 | l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) |
d164b209 BB |
7792 | { |
7793 | /* | |
7794 | * A buffer is *not* eligible for the L2ARC if it: | |
7795 | * 1. belongs to a different spa. | |
428870ff BB |
7796 | * 2. is already cached on the L2ARC. |
7797 | * 3. has an I/O in progress (it may be an incomplete read). | |
7798 | * 4. is flagged not eligible (zfs property). | |
d164b209 | 7799 | */ |
b9541d6b | 7800 | if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) || |
2a432414 | 7801 | HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr)) |
d164b209 BB |
7802 | return (B_FALSE); |
7803 | ||
7804 | return (B_TRUE); | |
7805 | } | |
7806 | ||
7807 | static uint64_t | |
3a17a7a9 | 7808 | l2arc_write_size(void) |
d164b209 BB |
7809 | { |
7810 | uint64_t size; | |
7811 | ||
3a17a7a9 SK |
7812 | /* |
7813 | * Make sure our globals have meaningful values in case the user | |
7814 | * altered them. | |
7815 | */ | |
7816 | size = l2arc_write_max; | |
7817 | if (size == 0) { | |
7818 | cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " | |
7819 | "be greater than zero, resetting it to the default (%d)", | |
7820 | L2ARC_WRITE_SIZE); | |
7821 | size = l2arc_write_max = L2ARC_WRITE_SIZE; | |
7822 | } | |
d164b209 BB |
7823 | |
7824 | if (arc_warm == B_FALSE) | |
3a17a7a9 | 7825 | size += l2arc_write_boost; |
d164b209 BB |
7826 | |
7827 | return (size); | |
7828 | ||
7829 | } | |
7830 | ||
7831 | static clock_t | |
7832 | l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) | |
7833 | { | |
428870ff | 7834 | clock_t interval, next, now; |
d164b209 BB |
7835 | |
7836 | /* | |
7837 | * If the ARC lists are busy, increase our write rate; if the | |
7838 | * lists are stale, idle back. This is achieved by checking | |
7839 | * how much we previously wrote - if it was more than half of | |
7840 | * what we wanted, schedule the next write much sooner. | |
7841 | */ | |
7842 | if (l2arc_feed_again && wrote > (wanted / 2)) | |
7843 | interval = (hz * l2arc_feed_min_ms) / 1000; | |
7844 | else | |
7845 | interval = hz * l2arc_feed_secs; | |
7846 | ||
428870ff BB |
7847 | now = ddi_get_lbolt(); |
7848 | next = MAX(now, MIN(now + interval, began + interval)); | |
d164b209 BB |
7849 | |
7850 | return (next); | |
7851 | } | |
7852 | ||
34dc7c2f BB |
7853 | /* |
7854 | * Cycle through L2ARC devices. This is how L2ARC load balances. | |
b128c09f | 7855 | * If a device is returned, this also returns holding the spa config lock. |
34dc7c2f BB |
7856 | */ |
7857 | static l2arc_dev_t * | |
7858 | l2arc_dev_get_next(void) | |
7859 | { | |
b128c09f | 7860 | l2arc_dev_t *first, *next = NULL; |
34dc7c2f | 7861 | |
b128c09f BB |
7862 | /* |
7863 | * Lock out the removal of spas (spa_namespace_lock), then removal | |
7864 | * of cache devices (l2arc_dev_mtx). Once a device has been selected, | |
7865 | * both locks will be dropped and a spa config lock held instead. | |
7866 | */ | |
7867 | mutex_enter(&spa_namespace_lock); | |
7868 | mutex_enter(&l2arc_dev_mtx); | |
7869 | ||
7870 | /* if there are no vdevs, there is nothing to do */ | |
7871 | if (l2arc_ndev == 0) | |
7872 | goto out; | |
7873 | ||
7874 | first = NULL; | |
7875 | next = l2arc_dev_last; | |
7876 | do { | |
7877 | /* loop around the list looking for a non-faulted vdev */ | |
7878 | if (next == NULL) { | |
34dc7c2f | 7879 | next = list_head(l2arc_dev_list); |
b128c09f BB |
7880 | } else { |
7881 | next = list_next(l2arc_dev_list, next); | |
7882 | if (next == NULL) | |
7883 | next = list_head(l2arc_dev_list); | |
7884 | } | |
7885 | ||
7886 | /* if we have come back to the start, bail out */ | |
7887 | if (first == NULL) | |
7888 | first = next; | |
7889 | else if (next == first) | |
7890 | break; | |
7891 | ||
7892 | } while (vdev_is_dead(next->l2ad_vdev)); | |
7893 | ||
7894 | /* if we were unable to find any usable vdevs, return NULL */ | |
7895 | if (vdev_is_dead(next->l2ad_vdev)) | |
7896 | next = NULL; | |
34dc7c2f BB |
7897 | |
7898 | l2arc_dev_last = next; | |
7899 | ||
b128c09f BB |
7900 | out: |
7901 | mutex_exit(&l2arc_dev_mtx); | |
7902 | ||
7903 | /* | |
7904 | * Grab the config lock to prevent the 'next' device from being | |
7905 | * removed while we are writing to it. | |
7906 | */ | |
7907 | if (next != NULL) | |
7908 | spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); | |
7909 | mutex_exit(&spa_namespace_lock); | |
7910 | ||
34dc7c2f BB |
7911 | return (next); |
7912 | } | |
7913 | ||
b128c09f BB |
7914 | /* |
7915 | * Free buffers that were tagged for destruction. | |
7916 | */ | |
7917 | static void | |
0bc8fd78 | 7918 | l2arc_do_free_on_write(void) |
b128c09f BB |
7919 | { |
7920 | list_t *buflist; | |
7921 | l2arc_data_free_t *df, *df_prev; | |
7922 | ||
7923 | mutex_enter(&l2arc_free_on_write_mtx); | |
7924 | buflist = l2arc_free_on_write; | |
7925 | ||
7926 | for (df = list_tail(buflist); df; df = df_prev) { | |
7927 | df_prev = list_prev(buflist, df); | |
a6255b7f DQ |
7928 | ASSERT3P(df->l2df_abd, !=, NULL); |
7929 | abd_free(df->l2df_abd); | |
b128c09f BB |
7930 | list_remove(buflist, df); |
7931 | kmem_free(df, sizeof (l2arc_data_free_t)); | |
7932 | } | |
7933 | ||
7934 | mutex_exit(&l2arc_free_on_write_mtx); | |
7935 | } | |
7936 | ||
34dc7c2f BB |
7937 | /* |
7938 | * A write to a cache device has completed. Update all headers to allow | |
7939 | * reads from these buffers to begin. | |
7940 | */ | |
7941 | static void | |
7942 | l2arc_write_done(zio_t *zio) | |
7943 | { | |
7944 | l2arc_write_callback_t *cb; | |
7945 | l2arc_dev_t *dev; | |
7946 | list_t *buflist; | |
2a432414 | 7947 | arc_buf_hdr_t *head, *hdr, *hdr_prev; |
34dc7c2f | 7948 | kmutex_t *hash_lock; |
3bec585e | 7949 | int64_t bytes_dropped = 0; |
34dc7c2f BB |
7950 | |
7951 | cb = zio->io_private; | |
d3c2ae1c | 7952 | ASSERT3P(cb, !=, NULL); |
34dc7c2f | 7953 | dev = cb->l2wcb_dev; |
d3c2ae1c | 7954 | ASSERT3P(dev, !=, NULL); |
34dc7c2f | 7955 | head = cb->l2wcb_head; |
d3c2ae1c | 7956 | ASSERT3P(head, !=, NULL); |
b9541d6b | 7957 | buflist = &dev->l2ad_buflist; |
d3c2ae1c | 7958 | ASSERT3P(buflist, !=, NULL); |
34dc7c2f BB |
7959 | DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, |
7960 | l2arc_write_callback_t *, cb); | |
7961 | ||
7962 | if (zio->io_error != 0) | |
7963 | ARCSTAT_BUMP(arcstat_l2_writes_error); | |
7964 | ||
34dc7c2f BB |
7965 | /* |
7966 | * All writes completed, or an error was hit. | |
7967 | */ | |
ca0bf58d PS |
7968 | top: |
7969 | mutex_enter(&dev->l2ad_mtx); | |
2a432414 GW |
7970 | for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { |
7971 | hdr_prev = list_prev(buflist, hdr); | |
34dc7c2f | 7972 | |
2a432414 | 7973 | hash_lock = HDR_LOCK(hdr); |
ca0bf58d PS |
7974 | |
7975 | /* | |
7976 | * We cannot use mutex_enter or else we can deadlock | |
7977 | * with l2arc_write_buffers (due to swapping the order | |
7978 | * the hash lock and l2ad_mtx are taken). | |
7979 | */ | |
34dc7c2f BB |
7980 | if (!mutex_tryenter(hash_lock)) { |
7981 | /* | |
ca0bf58d PS |
7982 | * Missed the hash lock. We must retry so we |
7983 | * don't leave the ARC_FLAG_L2_WRITING bit set. | |
34dc7c2f | 7984 | */ |
ca0bf58d PS |
7985 | ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); |
7986 | ||
7987 | /* | |
7988 | * We don't want to rescan the headers we've | |
7989 | * already marked as having been written out, so | |
7990 | * we reinsert the head node so we can pick up | |
7991 | * where we left off. | |
7992 | */ | |
7993 | list_remove(buflist, head); | |
7994 | list_insert_after(buflist, hdr, head); | |
7995 | ||
7996 | mutex_exit(&dev->l2ad_mtx); | |
7997 | ||
7998 | /* | |
7999 | * We wait for the hash lock to become available | |
8000 | * to try and prevent busy waiting, and increase | |
8001 | * the chance we'll be able to acquire the lock | |
8002 | * the next time around. | |
8003 | */ | |
8004 | mutex_enter(hash_lock); | |
8005 | mutex_exit(hash_lock); | |
8006 | goto top; | |
34dc7c2f BB |
8007 | } |
8008 | ||
b9541d6b | 8009 | /* |
ca0bf58d PS |
8010 | * We could not have been moved into the arc_l2c_only |
8011 | * state while in-flight due to our ARC_FLAG_L2_WRITING | |
8012 | * bit being set. Let's just ensure that's being enforced. | |
8013 | */ | |
8014 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
8015 | ||
8a09d5fd BB |
8016 | /* |
8017 | * Skipped - drop L2ARC entry and mark the header as no | |
8018 | * longer L2 eligibile. | |
8019 | */ | |
d3c2ae1c | 8020 | if (zio->io_error != 0) { |
34dc7c2f | 8021 | /* |
b128c09f | 8022 | * Error - drop L2ARC entry. |
34dc7c2f | 8023 | */ |
2a432414 | 8024 | list_remove(buflist, hdr); |
d3c2ae1c | 8025 | arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); |
b9541d6b | 8026 | |
01850391 AG |
8027 | ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr)); |
8028 | ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); | |
d962d5da | 8029 | |
d3c2ae1c | 8030 | bytes_dropped += arc_hdr_size(hdr); |
d962d5da | 8031 | (void) refcount_remove_many(&dev->l2ad_alloc, |
d3c2ae1c | 8032 | arc_hdr_size(hdr), hdr); |
34dc7c2f BB |
8033 | } |
8034 | ||
8035 | /* | |
ca0bf58d PS |
8036 | * Allow ARC to begin reads and ghost list evictions to |
8037 | * this L2ARC entry. | |
34dc7c2f | 8038 | */ |
d3c2ae1c | 8039 | arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); |
34dc7c2f BB |
8040 | |
8041 | mutex_exit(hash_lock); | |
8042 | } | |
8043 | ||
8044 | atomic_inc_64(&l2arc_writes_done); | |
8045 | list_remove(buflist, head); | |
b9541d6b CW |
8046 | ASSERT(!HDR_HAS_L1HDR(head)); |
8047 | kmem_cache_free(hdr_l2only_cache, head); | |
8048 | mutex_exit(&dev->l2ad_mtx); | |
34dc7c2f | 8049 | |
3bec585e SK |
8050 | vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); |
8051 | ||
b128c09f | 8052 | l2arc_do_free_on_write(); |
34dc7c2f BB |
8053 | |
8054 | kmem_free(cb, sizeof (l2arc_write_callback_t)); | |
8055 | } | |
8056 | ||
b5256303 TC |
8057 | static int |
8058 | l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb) | |
8059 | { | |
8060 | int ret; | |
8061 | spa_t *spa = zio->io_spa; | |
8062 | arc_buf_hdr_t *hdr = cb->l2rcb_hdr; | |
8063 | blkptr_t *bp = zio->io_bp; | |
b5256303 TC |
8064 | uint8_t salt[ZIO_DATA_SALT_LEN]; |
8065 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
8066 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
8067 | boolean_t no_crypt = B_FALSE; | |
8068 | ||
8069 | /* | |
8070 | * ZIL data is never be written to the L2ARC, so we don't need | |
8071 | * special handling for its unique MAC storage. | |
8072 | */ | |
8073 | ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG); | |
8074 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); | |
440a3eb9 | 8075 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); |
b5256303 | 8076 | |
440a3eb9 TC |
8077 | /* |
8078 | * If the data was encrypted, decrypt it now. Note that | |
8079 | * we must check the bp here and not the hdr, since the | |
8080 | * hdr does not have its encryption parameters updated | |
8081 | * until arc_read_done(). | |
8082 | */ | |
8083 | if (BP_IS_ENCRYPTED(bp)) { | |
be9a5c35 | 8084 | abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); |
b5256303 TC |
8085 | |
8086 | zio_crypt_decode_params_bp(bp, salt, iv); | |
8087 | zio_crypt_decode_mac_bp(bp, mac); | |
8088 | ||
be9a5c35 TC |
8089 | ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb, |
8090 | BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), | |
8091 | salt, iv, mac, HDR_GET_PSIZE(hdr), eabd, | |
8092 | hdr->b_l1hdr.b_pabd, &no_crypt); | |
b5256303 TC |
8093 | if (ret != 0) { |
8094 | arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr); | |
b5256303 TC |
8095 | goto error; |
8096 | } | |
8097 | ||
b5256303 TC |
8098 | /* |
8099 | * If we actually performed decryption, replace b_pabd | |
8100 | * with the decrypted data. Otherwise we can just throw | |
8101 | * our decryption buffer away. | |
8102 | */ | |
8103 | if (!no_crypt) { | |
8104 | arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, | |
8105 | arc_hdr_size(hdr), hdr); | |
8106 | hdr->b_l1hdr.b_pabd = eabd; | |
8107 | zio->io_abd = eabd; | |
8108 | } else { | |
8109 | arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr); | |
8110 | } | |
8111 | } | |
8112 | ||
8113 | /* | |
8114 | * If the L2ARC block was compressed, but ARC compression | |
8115 | * is disabled we decompress the data into a new buffer and | |
8116 | * replace the existing data. | |
8117 | */ | |
8118 | if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && | |
8119 | !HDR_COMPRESSION_ENABLED(hdr)) { | |
8120 | abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); | |
8121 | void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr)); | |
8122 | ||
8123 | ret = zio_decompress_data(HDR_GET_COMPRESS(hdr), | |
8124 | hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr), | |
8125 | HDR_GET_LSIZE(hdr)); | |
8126 | if (ret != 0) { | |
8127 | abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr)); | |
8128 | arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr); | |
8129 | goto error; | |
8130 | } | |
8131 | ||
8132 | abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr)); | |
8133 | arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, | |
8134 | arc_hdr_size(hdr), hdr); | |
8135 | hdr->b_l1hdr.b_pabd = cabd; | |
8136 | zio->io_abd = cabd; | |
8137 | zio->io_size = HDR_GET_LSIZE(hdr); | |
8138 | } | |
8139 | ||
8140 | return (0); | |
8141 | ||
8142 | error: | |
8143 | return (ret); | |
8144 | } | |
8145 | ||
8146 | ||
34dc7c2f BB |
8147 | /* |
8148 | * A read to a cache device completed. Validate buffer contents before | |
8149 | * handing over to the regular ARC routines. | |
8150 | */ | |
8151 | static void | |
8152 | l2arc_read_done(zio_t *zio) | |
8153 | { | |
b5256303 | 8154 | int tfm_error = 0; |
34dc7c2f BB |
8155 | l2arc_read_callback_t *cb; |
8156 | arc_buf_hdr_t *hdr; | |
34dc7c2f | 8157 | kmutex_t *hash_lock; |
b5256303 | 8158 | boolean_t valid_cksum, using_rdata; |
b128c09f | 8159 | |
d3c2ae1c | 8160 | ASSERT3P(zio->io_vd, !=, NULL); |
b128c09f BB |
8161 | ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); |
8162 | ||
8163 | spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); | |
34dc7c2f BB |
8164 | |
8165 | cb = zio->io_private; | |
d3c2ae1c GW |
8166 | ASSERT3P(cb, !=, NULL); |
8167 | hdr = cb->l2rcb_hdr; | |
8168 | ASSERT3P(hdr, !=, NULL); | |
34dc7c2f | 8169 | |
d3c2ae1c | 8170 | hash_lock = HDR_LOCK(hdr); |
34dc7c2f | 8171 | mutex_enter(hash_lock); |
428870ff | 8172 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); |
34dc7c2f | 8173 | |
82710e99 GDN |
8174 | /* |
8175 | * If the data was read into a temporary buffer, | |
8176 | * move it and free the buffer. | |
8177 | */ | |
8178 | if (cb->l2rcb_abd != NULL) { | |
8179 | ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); | |
8180 | if (zio->io_error == 0) { | |
8181 | abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd, | |
8182 | arc_hdr_size(hdr)); | |
8183 | } | |
8184 | ||
8185 | /* | |
8186 | * The following must be done regardless of whether | |
8187 | * there was an error: | |
8188 | * - free the temporary buffer | |
8189 | * - point zio to the real ARC buffer | |
8190 | * - set zio size accordingly | |
8191 | * These are required because zio is either re-used for | |
8192 | * an I/O of the block in the case of the error | |
8193 | * or the zio is passed to arc_read_done() and it | |
8194 | * needs real data. | |
8195 | */ | |
8196 | abd_free(cb->l2rcb_abd); | |
8197 | zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); | |
440a3eb9 TC |
8198 | |
8199 | if (BP_IS_ENCRYPTED(&cb->l2rcb_bp) && | |
8200 | (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT)) { | |
8201 | ASSERT(HDR_HAS_RABD(hdr)); | |
8202 | zio->io_abd = zio->io_orig_abd = | |
8203 | hdr->b_crypt_hdr.b_rabd; | |
8204 | } else { | |
8205 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); | |
8206 | zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd; | |
8207 | } | |
82710e99 GDN |
8208 | } |
8209 | ||
a6255b7f | 8210 | ASSERT3P(zio->io_abd, !=, NULL); |
3a17a7a9 | 8211 | |
34dc7c2f BB |
8212 | /* |
8213 | * Check this survived the L2ARC journey. | |
8214 | */ | |
b5256303 TC |
8215 | ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd || |
8216 | (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd)); | |
d3c2ae1c GW |
8217 | zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ |
8218 | zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ | |
8219 | ||
8220 | valid_cksum = arc_cksum_is_equal(hdr, zio); | |
b5256303 TC |
8221 | using_rdata = (HDR_HAS_RABD(hdr) && |
8222 | zio->io_abd == hdr->b_crypt_hdr.b_rabd); | |
8223 | ||
8224 | /* | |
8225 | * b_rabd will always match the data as it exists on disk if it is | |
8226 | * being used. Therefore if we are reading into b_rabd we do not | |
8227 | * attempt to untransform the data. | |
8228 | */ | |
8229 | if (valid_cksum && !using_rdata) | |
8230 | tfm_error = l2arc_untransform(zio, cb); | |
8231 | ||
8232 | if (valid_cksum && tfm_error == 0 && zio->io_error == 0 && | |
8233 | !HDR_L2_EVICTED(hdr)) { | |
34dc7c2f | 8234 | mutex_exit(hash_lock); |
d3c2ae1c | 8235 | zio->io_private = hdr; |
34dc7c2f BB |
8236 | arc_read_done(zio); |
8237 | } else { | |
8238 | mutex_exit(hash_lock); | |
8239 | /* | |
8240 | * Buffer didn't survive caching. Increment stats and | |
8241 | * reissue to the original storage device. | |
8242 | */ | |
b128c09f | 8243 | if (zio->io_error != 0) { |
34dc7c2f | 8244 | ARCSTAT_BUMP(arcstat_l2_io_error); |
b128c09f | 8245 | } else { |
2e528b49 | 8246 | zio->io_error = SET_ERROR(EIO); |
b128c09f | 8247 | } |
b5256303 | 8248 | if (!valid_cksum || tfm_error != 0) |
34dc7c2f BB |
8249 | ARCSTAT_BUMP(arcstat_l2_cksum_bad); |
8250 | ||
34dc7c2f | 8251 | /* |
b128c09f BB |
8252 | * If there's no waiter, issue an async i/o to the primary |
8253 | * storage now. If there *is* a waiter, the caller must | |
8254 | * issue the i/o in a context where it's OK to block. | |
34dc7c2f | 8255 | */ |
d164b209 BB |
8256 | if (zio->io_waiter == NULL) { |
8257 | zio_t *pio = zio_unique_parent(zio); | |
b5256303 TC |
8258 | void *abd = (using_rdata) ? |
8259 | hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd; | |
d164b209 BB |
8260 | |
8261 | ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); | |
8262 | ||
d3c2ae1c | 8263 | zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, |
b5256303 | 8264 | abd, zio->io_size, arc_read_done, |
d3c2ae1c GW |
8265 | hdr, zio->io_priority, cb->l2rcb_flags, |
8266 | &cb->l2rcb_zb)); | |
d164b209 | 8267 | } |
34dc7c2f BB |
8268 | } |
8269 | ||
8270 | kmem_free(cb, sizeof (l2arc_read_callback_t)); | |
8271 | } | |
8272 | ||
8273 | /* | |
8274 | * This is the list priority from which the L2ARC will search for pages to | |
8275 | * cache. This is used within loops (0..3) to cycle through lists in the | |
8276 | * desired order. This order can have a significant effect on cache | |
8277 | * performance. | |
8278 | * | |
8279 | * Currently the metadata lists are hit first, MFU then MRU, followed by | |
8280 | * the data lists. This function returns a locked list, and also returns | |
8281 | * the lock pointer. | |
8282 | */ | |
ca0bf58d PS |
8283 | static multilist_sublist_t * |
8284 | l2arc_sublist_lock(int list_num) | |
34dc7c2f | 8285 | { |
ca0bf58d PS |
8286 | multilist_t *ml = NULL; |
8287 | unsigned int idx; | |
34dc7c2f | 8288 | |
4aafab91 | 8289 | ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES); |
34dc7c2f BB |
8290 | |
8291 | switch (list_num) { | |
8292 | case 0: | |
64fc7762 | 8293 | ml = arc_mfu->arcs_list[ARC_BUFC_METADATA]; |
34dc7c2f BB |
8294 | break; |
8295 | case 1: | |
64fc7762 | 8296 | ml = arc_mru->arcs_list[ARC_BUFC_METADATA]; |
34dc7c2f BB |
8297 | break; |
8298 | case 2: | |
64fc7762 | 8299 | ml = arc_mfu->arcs_list[ARC_BUFC_DATA]; |
34dc7c2f BB |
8300 | break; |
8301 | case 3: | |
64fc7762 | 8302 | ml = arc_mru->arcs_list[ARC_BUFC_DATA]; |
34dc7c2f | 8303 | break; |
4aafab91 G |
8304 | default: |
8305 | return (NULL); | |
34dc7c2f BB |
8306 | } |
8307 | ||
ca0bf58d PS |
8308 | /* |
8309 | * Return a randomly-selected sublist. This is acceptable | |
8310 | * because the caller feeds only a little bit of data for each | |
8311 | * call (8MB). Subsequent calls will result in different | |
8312 | * sublists being selected. | |
8313 | */ | |
8314 | idx = multilist_get_random_index(ml); | |
8315 | return (multilist_sublist_lock(ml, idx)); | |
34dc7c2f BB |
8316 | } |
8317 | ||
8318 | /* | |
8319 | * Evict buffers from the device write hand to the distance specified in | |
8320 | * bytes. This distance may span populated buffers, it may span nothing. | |
8321 | * This is clearing a region on the L2ARC device ready for writing. | |
8322 | * If the 'all' boolean is set, every buffer is evicted. | |
8323 | */ | |
8324 | static void | |
8325 | l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) | |
8326 | { | |
8327 | list_t *buflist; | |
2a432414 | 8328 | arc_buf_hdr_t *hdr, *hdr_prev; |
34dc7c2f BB |
8329 | kmutex_t *hash_lock; |
8330 | uint64_t taddr; | |
8331 | ||
b9541d6b | 8332 | buflist = &dev->l2ad_buflist; |
34dc7c2f BB |
8333 | |
8334 | if (!all && dev->l2ad_first) { | |
8335 | /* | |
8336 | * This is the first sweep through the device. There is | |
8337 | * nothing to evict. | |
8338 | */ | |
8339 | return; | |
8340 | } | |
8341 | ||
b128c09f | 8342 | if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { |
34dc7c2f BB |
8343 | /* |
8344 | * When nearing the end of the device, evict to the end | |
8345 | * before the device write hand jumps to the start. | |
8346 | */ | |
8347 | taddr = dev->l2ad_end; | |
8348 | } else { | |
8349 | taddr = dev->l2ad_hand + distance; | |
8350 | } | |
8351 | DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, | |
8352 | uint64_t, taddr, boolean_t, all); | |
8353 | ||
8354 | top: | |
b9541d6b | 8355 | mutex_enter(&dev->l2ad_mtx); |
2a432414 GW |
8356 | for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { |
8357 | hdr_prev = list_prev(buflist, hdr); | |
34dc7c2f | 8358 | |
2a432414 | 8359 | hash_lock = HDR_LOCK(hdr); |
ca0bf58d PS |
8360 | |
8361 | /* | |
8362 | * We cannot use mutex_enter or else we can deadlock | |
8363 | * with l2arc_write_buffers (due to swapping the order | |
8364 | * the hash lock and l2ad_mtx are taken). | |
8365 | */ | |
34dc7c2f BB |
8366 | if (!mutex_tryenter(hash_lock)) { |
8367 | /* | |
8368 | * Missed the hash lock. Retry. | |
8369 | */ | |
8370 | ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); | |
b9541d6b | 8371 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f BB |
8372 | mutex_enter(hash_lock); |
8373 | mutex_exit(hash_lock); | |
8374 | goto top; | |
8375 | } | |
8376 | ||
f06f53fa AG |
8377 | /* |
8378 | * A header can't be on this list if it doesn't have L2 header. | |
8379 | */ | |
8380 | ASSERT(HDR_HAS_L2HDR(hdr)); | |
34dc7c2f | 8381 | |
f06f53fa AG |
8382 | /* Ensure this header has finished being written. */ |
8383 | ASSERT(!HDR_L2_WRITING(hdr)); | |
8384 | ASSERT(!HDR_L2_WRITE_HEAD(hdr)); | |
8385 | ||
8386 | if (!all && (hdr->b_l2hdr.b_daddr >= taddr || | |
b9541d6b | 8387 | hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { |
34dc7c2f BB |
8388 | /* |
8389 | * We've evicted to the target address, | |
8390 | * or the end of the device. | |
8391 | */ | |
8392 | mutex_exit(hash_lock); | |
8393 | break; | |
8394 | } | |
8395 | ||
b9541d6b | 8396 | if (!HDR_HAS_L1HDR(hdr)) { |
2a432414 | 8397 | ASSERT(!HDR_L2_READING(hdr)); |
34dc7c2f BB |
8398 | /* |
8399 | * This doesn't exist in the ARC. Destroy. | |
8400 | * arc_hdr_destroy() will call list_remove() | |
01850391 | 8401 | * and decrement arcstat_l2_lsize. |
34dc7c2f | 8402 | */ |
2a432414 GW |
8403 | arc_change_state(arc_anon, hdr, hash_lock); |
8404 | arc_hdr_destroy(hdr); | |
34dc7c2f | 8405 | } else { |
b9541d6b CW |
8406 | ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); |
8407 | ARCSTAT_BUMP(arcstat_l2_evict_l1cached); | |
b128c09f BB |
8408 | /* |
8409 | * Invalidate issued or about to be issued | |
8410 | * reads, since we may be about to write | |
8411 | * over this location. | |
8412 | */ | |
2a432414 | 8413 | if (HDR_L2_READING(hdr)) { |
b128c09f | 8414 | ARCSTAT_BUMP(arcstat_l2_evict_reading); |
d3c2ae1c | 8415 | arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); |
b128c09f BB |
8416 | } |
8417 | ||
d962d5da | 8418 | arc_hdr_l2hdr_destroy(hdr); |
34dc7c2f BB |
8419 | } |
8420 | mutex_exit(hash_lock); | |
8421 | } | |
b9541d6b | 8422 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f BB |
8423 | } |
8424 | ||
b5256303 TC |
8425 | /* |
8426 | * Handle any abd transforms that might be required for writing to the L2ARC. | |
8427 | * If successful, this function will always return an abd with the data | |
8428 | * transformed as it is on disk in a new abd of asize bytes. | |
8429 | */ | |
8430 | static int | |
8431 | l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize, | |
8432 | abd_t **abd_out) | |
8433 | { | |
8434 | int ret; | |
8435 | void *tmp = NULL; | |
8436 | abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd; | |
8437 | enum zio_compress compress = HDR_GET_COMPRESS(hdr); | |
8438 | uint64_t psize = HDR_GET_PSIZE(hdr); | |
8439 | uint64_t size = arc_hdr_size(hdr); | |
8440 | boolean_t ismd = HDR_ISTYPE_METADATA(hdr); | |
8441 | boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS); | |
8442 | dsl_crypto_key_t *dck = NULL; | |
8443 | uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 }; | |
4807c0ba | 8444 | boolean_t no_crypt = B_FALSE; |
b5256303 TC |
8445 | |
8446 | ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && | |
8447 | !HDR_COMPRESSION_ENABLED(hdr)) || | |
8448 | HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize); | |
8449 | ASSERT3U(psize, <=, asize); | |
8450 | ||
8451 | /* | |
8452 | * If this data simply needs its own buffer, we simply allocate it | |
8453 | * and copy the data. This may be done to elimiate a depedency on a | |
8454 | * shared buffer or to reallocate the buffer to match asize. | |
8455 | */ | |
4807c0ba | 8456 | if (HDR_HAS_RABD(hdr) && asize != psize) { |
10adee27 | 8457 | ASSERT3U(asize, >=, psize); |
4807c0ba | 8458 | to_write = abd_alloc_for_io(asize, ismd); |
10adee27 TC |
8459 | abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize); |
8460 | if (psize != asize) | |
8461 | abd_zero_off(to_write, psize, asize - psize); | |
4807c0ba TC |
8462 | goto out; |
8463 | } | |
8464 | ||
b5256303 TC |
8465 | if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) && |
8466 | !HDR_ENCRYPTED(hdr)) { | |
8467 | ASSERT3U(size, ==, psize); | |
8468 | to_write = abd_alloc_for_io(asize, ismd); | |
8469 | abd_copy(to_write, hdr->b_l1hdr.b_pabd, size); | |
8470 | if (size != asize) | |
8471 | abd_zero_off(to_write, size, asize - size); | |
8472 | goto out; | |
8473 | } | |
8474 | ||
8475 | if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) { | |
8476 | cabd = abd_alloc_for_io(asize, ismd); | |
8477 | tmp = abd_borrow_buf(cabd, asize); | |
8478 | ||
8479 | psize = zio_compress_data(compress, to_write, tmp, size); | |
8480 | ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr)); | |
8481 | if (psize < asize) | |
8482 | bzero((char *)tmp + psize, asize - psize); | |
8483 | psize = HDR_GET_PSIZE(hdr); | |
8484 | abd_return_buf_copy(cabd, tmp, asize); | |
8485 | to_write = cabd; | |
8486 | } | |
8487 | ||
8488 | if (HDR_ENCRYPTED(hdr)) { | |
8489 | eabd = abd_alloc_for_io(asize, ismd); | |
8490 | ||
8491 | /* | |
8492 | * If the dataset was disowned before the buffer | |
8493 | * made it to this point, the key to re-encrypt | |
8494 | * it won't be available. In this case we simply | |
8495 | * won't write the buffer to the L2ARC. | |
8496 | */ | |
8497 | ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj, | |
8498 | FTAG, &dck); | |
8499 | if (ret != 0) | |
8500 | goto error; | |
8501 | ||
8502 | ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key, | |
be9a5c35 TC |
8503 | hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt, |
8504 | hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd, | |
8505 | &no_crypt); | |
b5256303 TC |
8506 | if (ret != 0) |
8507 | goto error; | |
8508 | ||
4807c0ba TC |
8509 | if (no_crypt) |
8510 | abd_copy(eabd, to_write, psize); | |
b5256303 TC |
8511 | |
8512 | if (psize != asize) | |
8513 | abd_zero_off(eabd, psize, asize - psize); | |
8514 | ||
8515 | /* assert that the MAC we got here matches the one we saved */ | |
8516 | ASSERT0(bcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN)); | |
8517 | spa_keystore_dsl_key_rele(spa, dck, FTAG); | |
8518 | ||
8519 | if (to_write == cabd) | |
8520 | abd_free(cabd); | |
8521 | ||
8522 | to_write = eabd; | |
8523 | } | |
8524 | ||
8525 | out: | |
8526 | ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd); | |
8527 | *abd_out = to_write; | |
8528 | return (0); | |
8529 | ||
8530 | error: | |
8531 | if (dck != NULL) | |
8532 | spa_keystore_dsl_key_rele(spa, dck, FTAG); | |
8533 | if (cabd != NULL) | |
8534 | abd_free(cabd); | |
8535 | if (eabd != NULL) | |
8536 | abd_free(eabd); | |
8537 | ||
8538 | *abd_out = NULL; | |
8539 | return (ret); | |
8540 | } | |
8541 | ||
34dc7c2f BB |
8542 | /* |
8543 | * Find and write ARC buffers to the L2ARC device. | |
8544 | * | |
2a432414 | 8545 | * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid |
34dc7c2f | 8546 | * for reading until they have completed writing. |
3a17a7a9 SK |
8547 | * The headroom_boost is an in-out parameter used to maintain headroom boost |
8548 | * state between calls to this function. | |
8549 | * | |
8550 | * Returns the number of bytes actually written (which may be smaller than | |
8551 | * the delta by which the device hand has changed due to alignment). | |
34dc7c2f | 8552 | */ |
d164b209 | 8553 | static uint64_t |
d3c2ae1c | 8554 | l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) |
34dc7c2f | 8555 | { |
2a432414 | 8556 | arc_buf_hdr_t *hdr, *hdr_prev, *head; |
01850391 | 8557 | uint64_t write_asize, write_psize, write_lsize, headroom; |
3a17a7a9 | 8558 | boolean_t full; |
34dc7c2f BB |
8559 | l2arc_write_callback_t *cb; |
8560 | zio_t *pio, *wzio; | |
3541dc6d | 8561 | uint64_t guid = spa_load_guid(spa); |
34dc7c2f | 8562 | |
d3c2ae1c | 8563 | ASSERT3P(dev->l2ad_vdev, !=, NULL); |
3a17a7a9 | 8564 | |
34dc7c2f | 8565 | pio = NULL; |
01850391 | 8566 | write_lsize = write_asize = write_psize = 0; |
34dc7c2f | 8567 | full = B_FALSE; |
b9541d6b | 8568 | head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); |
d3c2ae1c | 8569 | arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); |
3a17a7a9 | 8570 | |
34dc7c2f BB |
8571 | /* |
8572 | * Copy buffers for L2ARC writing. | |
8573 | */ | |
1c27024e | 8574 | for (int try = 0; try < L2ARC_FEED_TYPES; try++) { |
ca0bf58d | 8575 | multilist_sublist_t *mls = l2arc_sublist_lock(try); |
3a17a7a9 SK |
8576 | uint64_t passed_sz = 0; |
8577 | ||
4aafab91 G |
8578 | VERIFY3P(mls, !=, NULL); |
8579 | ||
b128c09f BB |
8580 | /* |
8581 | * L2ARC fast warmup. | |
8582 | * | |
8583 | * Until the ARC is warm and starts to evict, read from the | |
8584 | * head of the ARC lists rather than the tail. | |
8585 | */ | |
b128c09f | 8586 | if (arc_warm == B_FALSE) |
ca0bf58d | 8587 | hdr = multilist_sublist_head(mls); |
b128c09f | 8588 | else |
ca0bf58d | 8589 | hdr = multilist_sublist_tail(mls); |
b128c09f | 8590 | |
3a17a7a9 | 8591 | headroom = target_sz * l2arc_headroom; |
d3c2ae1c | 8592 | if (zfs_compressed_arc_enabled) |
3a17a7a9 SK |
8593 | headroom = (headroom * l2arc_headroom_boost) / 100; |
8594 | ||
2a432414 | 8595 | for (; hdr; hdr = hdr_prev) { |
3a17a7a9 | 8596 | kmutex_t *hash_lock; |
b5256303 | 8597 | abd_t *to_write = NULL; |
3a17a7a9 | 8598 | |
b128c09f | 8599 | if (arc_warm == B_FALSE) |
ca0bf58d | 8600 | hdr_prev = multilist_sublist_next(mls, hdr); |
b128c09f | 8601 | else |
ca0bf58d | 8602 | hdr_prev = multilist_sublist_prev(mls, hdr); |
34dc7c2f | 8603 | |
2a432414 | 8604 | hash_lock = HDR_LOCK(hdr); |
3a17a7a9 | 8605 | if (!mutex_tryenter(hash_lock)) { |
34dc7c2f BB |
8606 | /* |
8607 | * Skip this buffer rather than waiting. | |
8608 | */ | |
8609 | continue; | |
8610 | } | |
8611 | ||
d3c2ae1c | 8612 | passed_sz += HDR_GET_LSIZE(hdr); |
34dc7c2f BB |
8613 | if (passed_sz > headroom) { |
8614 | /* | |
8615 | * Searched too far. | |
8616 | */ | |
8617 | mutex_exit(hash_lock); | |
8618 | break; | |
8619 | } | |
8620 | ||
2a432414 | 8621 | if (!l2arc_write_eligible(guid, hdr)) { |
34dc7c2f BB |
8622 | mutex_exit(hash_lock); |
8623 | continue; | |
8624 | } | |
8625 | ||
01850391 AG |
8626 | /* |
8627 | * We rely on the L1 portion of the header below, so | |
8628 | * it's invalid for this header to have been evicted out | |
8629 | * of the ghost cache, prior to being written out. The | |
8630 | * ARC_FLAG_L2_WRITING bit ensures this won't happen. | |
8631 | */ | |
8632 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
8633 | ||
8634 | ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); | |
01850391 | 8635 | ASSERT3U(arc_hdr_size(hdr), >, 0); |
b5256303 TC |
8636 | ASSERT(hdr->b_l1hdr.b_pabd != NULL || |
8637 | HDR_HAS_RABD(hdr)); | |
8638 | uint64_t psize = HDR_GET_PSIZE(hdr); | |
01850391 AG |
8639 | uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, |
8640 | psize); | |
8641 | ||
8642 | if ((write_asize + asize) > target_sz) { | |
34dc7c2f BB |
8643 | full = B_TRUE; |
8644 | mutex_exit(hash_lock); | |
8645 | break; | |
8646 | } | |
8647 | ||
b5256303 TC |
8648 | /* |
8649 | * We rely on the L1 portion of the header below, so | |
8650 | * it's invalid for this header to have been evicted out | |
8651 | * of the ghost cache, prior to being written out. The | |
8652 | * ARC_FLAG_L2_WRITING bit ensures this won't happen. | |
8653 | */ | |
8654 | arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING); | |
8655 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
8656 | ||
8657 | ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); | |
8658 | ASSERT(hdr->b_l1hdr.b_pabd != NULL || | |
8659 | HDR_HAS_RABD(hdr)); | |
8660 | ASSERT3U(arc_hdr_size(hdr), >, 0); | |
8661 | ||
8662 | /* | |
8663 | * If this header has b_rabd, we can use this since it | |
8664 | * must always match the data exactly as it exists on | |
8665 | * disk. Otherwise, the L2ARC can normally use the | |
8666 | * hdr's data, but if we're sharing data between the | |
8667 | * hdr and one of its bufs, L2ARC needs its own copy of | |
8668 | * the data so that the ZIO below can't race with the | |
8669 | * buf consumer. To ensure that this copy will be | |
8670 | * available for the lifetime of the ZIO and be cleaned | |
8671 | * up afterwards, we add it to the l2arc_free_on_write | |
8672 | * queue. If we need to apply any transforms to the | |
8673 | * data (compression, encryption) we will also need the | |
8674 | * extra buffer. | |
8675 | */ | |
8676 | if (HDR_HAS_RABD(hdr) && psize == asize) { | |
8677 | to_write = hdr->b_crypt_hdr.b_rabd; | |
8678 | } else if ((HDR_COMPRESSION_ENABLED(hdr) || | |
8679 | HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) && | |
8680 | !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) && | |
8681 | psize == asize) { | |
8682 | to_write = hdr->b_l1hdr.b_pabd; | |
8683 | } else { | |
8684 | int ret; | |
8685 | arc_buf_contents_t type = arc_buf_type(hdr); | |
8686 | ||
8687 | ret = l2arc_apply_transforms(spa, hdr, asize, | |
8688 | &to_write); | |
8689 | if (ret != 0) { | |
8690 | arc_hdr_clear_flags(hdr, | |
8691 | ARC_FLAG_L2_WRITING); | |
8692 | mutex_exit(hash_lock); | |
8693 | continue; | |
8694 | } | |
8695 | ||
8696 | l2arc_free_abd_on_write(to_write, asize, type); | |
8697 | } | |
8698 | ||
34dc7c2f BB |
8699 | if (pio == NULL) { |
8700 | /* | |
8701 | * Insert a dummy header on the buflist so | |
8702 | * l2arc_write_done() can find where the | |
8703 | * write buffers begin without searching. | |
8704 | */ | |
ca0bf58d | 8705 | mutex_enter(&dev->l2ad_mtx); |
b9541d6b | 8706 | list_insert_head(&dev->l2ad_buflist, head); |
ca0bf58d | 8707 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f | 8708 | |
96c080cb BB |
8709 | cb = kmem_alloc( |
8710 | sizeof (l2arc_write_callback_t), KM_SLEEP); | |
34dc7c2f BB |
8711 | cb->l2wcb_dev = dev; |
8712 | cb->l2wcb_head = head; | |
8713 | pio = zio_root(spa, l2arc_write_done, cb, | |
8714 | ZIO_FLAG_CANFAIL); | |
8715 | } | |
8716 | ||
b9541d6b | 8717 | hdr->b_l2hdr.b_dev = dev; |
b9541d6b | 8718 | hdr->b_l2hdr.b_hits = 0; |
3a17a7a9 | 8719 | |
d3c2ae1c | 8720 | hdr->b_l2hdr.b_daddr = dev->l2ad_hand; |
b5256303 | 8721 | arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR); |
3a17a7a9 | 8722 | |
ca0bf58d | 8723 | mutex_enter(&dev->l2ad_mtx); |
b9541d6b | 8724 | list_insert_head(&dev->l2ad_buflist, hdr); |
ca0bf58d | 8725 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f | 8726 | |
b5256303 TC |
8727 | (void) refcount_add_many(&dev->l2ad_alloc, |
8728 | arc_hdr_size(hdr), hdr); | |
3a17a7a9 | 8729 | |
34dc7c2f | 8730 | wzio = zio_write_phys(pio, dev->l2ad_vdev, |
82710e99 | 8731 | hdr->b_l2hdr.b_daddr, asize, to_write, |
d3c2ae1c GW |
8732 | ZIO_CHECKSUM_OFF, NULL, hdr, |
8733 | ZIO_PRIORITY_ASYNC_WRITE, | |
34dc7c2f BB |
8734 | ZIO_FLAG_CANFAIL, B_FALSE); |
8735 | ||
01850391 | 8736 | write_lsize += HDR_GET_LSIZE(hdr); |
34dc7c2f BB |
8737 | DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, |
8738 | zio_t *, wzio); | |
d962d5da | 8739 | |
01850391 AG |
8740 | write_psize += psize; |
8741 | write_asize += asize; | |
d3c2ae1c GW |
8742 | dev->l2ad_hand += asize; |
8743 | ||
8744 | mutex_exit(hash_lock); | |
8745 | ||
8746 | (void) zio_nowait(wzio); | |
34dc7c2f | 8747 | } |
d3c2ae1c GW |
8748 | |
8749 | multilist_sublist_unlock(mls); | |
8750 | ||
8751 | if (full == B_TRUE) | |
8752 | break; | |
34dc7c2f | 8753 | } |
34dc7c2f | 8754 | |
d3c2ae1c GW |
8755 | /* No buffers selected for writing? */ |
8756 | if (pio == NULL) { | |
01850391 | 8757 | ASSERT0(write_lsize); |
d3c2ae1c GW |
8758 | ASSERT(!HDR_HAS_L1HDR(head)); |
8759 | kmem_cache_free(hdr_l2only_cache, head); | |
8760 | return (0); | |
8761 | } | |
34dc7c2f | 8762 | |
3a17a7a9 | 8763 | ASSERT3U(write_asize, <=, target_sz); |
34dc7c2f | 8764 | ARCSTAT_BUMP(arcstat_l2_writes_sent); |
01850391 AG |
8765 | ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize); |
8766 | ARCSTAT_INCR(arcstat_l2_lsize, write_lsize); | |
8767 | ARCSTAT_INCR(arcstat_l2_psize, write_psize); | |
8768 | vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); | |
34dc7c2f BB |
8769 | |
8770 | /* | |
8771 | * Bump device hand to the device start if it is approaching the end. | |
8772 | * l2arc_evict() will already have evicted ahead for this case. | |
8773 | */ | |
b128c09f | 8774 | if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { |
34dc7c2f | 8775 | dev->l2ad_hand = dev->l2ad_start; |
34dc7c2f BB |
8776 | dev->l2ad_first = B_FALSE; |
8777 | } | |
8778 | ||
d164b209 | 8779 | dev->l2ad_writing = B_TRUE; |
34dc7c2f | 8780 | (void) zio_wait(pio); |
d164b209 BB |
8781 | dev->l2ad_writing = B_FALSE; |
8782 | ||
3a17a7a9 SK |
8783 | return (write_asize); |
8784 | } | |
8785 | ||
34dc7c2f BB |
8786 | /* |
8787 | * This thread feeds the L2ARC at regular intervals. This is the beating | |
8788 | * heart of the L2ARC. | |
8789 | */ | |
867959b5 | 8790 | /* ARGSUSED */ |
34dc7c2f | 8791 | static void |
c25b8f99 | 8792 | l2arc_feed_thread(void *unused) |
34dc7c2f BB |
8793 | { |
8794 | callb_cpr_t cpr; | |
8795 | l2arc_dev_t *dev; | |
8796 | spa_t *spa; | |
d164b209 | 8797 | uint64_t size, wrote; |
428870ff | 8798 | clock_t begin, next = ddi_get_lbolt(); |
40d06e3c | 8799 | fstrans_cookie_t cookie; |
34dc7c2f BB |
8800 | |
8801 | CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); | |
8802 | ||
8803 | mutex_enter(&l2arc_feed_thr_lock); | |
8804 | ||
40d06e3c | 8805 | cookie = spl_fstrans_mark(); |
34dc7c2f | 8806 | while (l2arc_thread_exit == 0) { |
34dc7c2f | 8807 | CALLB_CPR_SAFE_BEGIN(&cpr); |
b64ccd6c | 8808 | (void) cv_timedwait_sig(&l2arc_feed_thr_cv, |
5b63b3eb | 8809 | &l2arc_feed_thr_lock, next); |
34dc7c2f | 8810 | CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); |
428870ff | 8811 | next = ddi_get_lbolt() + hz; |
34dc7c2f BB |
8812 | |
8813 | /* | |
b128c09f | 8814 | * Quick check for L2ARC devices. |
34dc7c2f BB |
8815 | */ |
8816 | mutex_enter(&l2arc_dev_mtx); | |
8817 | if (l2arc_ndev == 0) { | |
8818 | mutex_exit(&l2arc_dev_mtx); | |
8819 | continue; | |
8820 | } | |
b128c09f | 8821 | mutex_exit(&l2arc_dev_mtx); |
428870ff | 8822 | begin = ddi_get_lbolt(); |
34dc7c2f BB |
8823 | |
8824 | /* | |
b128c09f BB |
8825 | * This selects the next l2arc device to write to, and in |
8826 | * doing so the next spa to feed from: dev->l2ad_spa. This | |
8827 | * will return NULL if there are now no l2arc devices or if | |
8828 | * they are all faulted. | |
8829 | * | |
8830 | * If a device is returned, its spa's config lock is also | |
8831 | * held to prevent device removal. l2arc_dev_get_next() | |
8832 | * will grab and release l2arc_dev_mtx. | |
34dc7c2f | 8833 | */ |
b128c09f | 8834 | if ((dev = l2arc_dev_get_next()) == NULL) |
34dc7c2f | 8835 | continue; |
b128c09f BB |
8836 | |
8837 | spa = dev->l2ad_spa; | |
d3c2ae1c | 8838 | ASSERT3P(spa, !=, NULL); |
34dc7c2f | 8839 | |
572e2857 BB |
8840 | /* |
8841 | * If the pool is read-only then force the feed thread to | |
8842 | * sleep a little longer. | |
8843 | */ | |
8844 | if (!spa_writeable(spa)) { | |
8845 | next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; | |
8846 | spa_config_exit(spa, SCL_L2ARC, dev); | |
8847 | continue; | |
8848 | } | |
8849 | ||
34dc7c2f | 8850 | /* |
b128c09f | 8851 | * Avoid contributing to memory pressure. |
34dc7c2f | 8852 | */ |
ca67b33a | 8853 | if (arc_reclaim_needed()) { |
b128c09f BB |
8854 | ARCSTAT_BUMP(arcstat_l2_abort_lowmem); |
8855 | spa_config_exit(spa, SCL_L2ARC, dev); | |
34dc7c2f BB |
8856 | continue; |
8857 | } | |
b128c09f | 8858 | |
34dc7c2f BB |
8859 | ARCSTAT_BUMP(arcstat_l2_feeds); |
8860 | ||
3a17a7a9 | 8861 | size = l2arc_write_size(); |
b128c09f | 8862 | |
34dc7c2f BB |
8863 | /* |
8864 | * Evict L2ARC buffers that will be overwritten. | |
8865 | */ | |
b128c09f | 8866 | l2arc_evict(dev, size, B_FALSE); |
34dc7c2f BB |
8867 | |
8868 | /* | |
8869 | * Write ARC buffers. | |
8870 | */ | |
d3c2ae1c | 8871 | wrote = l2arc_write_buffers(spa, dev, size); |
d164b209 BB |
8872 | |
8873 | /* | |
8874 | * Calculate interval between writes. | |
8875 | */ | |
8876 | next = l2arc_write_interval(begin, size, wrote); | |
b128c09f | 8877 | spa_config_exit(spa, SCL_L2ARC, dev); |
34dc7c2f | 8878 | } |
40d06e3c | 8879 | spl_fstrans_unmark(cookie); |
34dc7c2f BB |
8880 | |
8881 | l2arc_thread_exit = 0; | |
8882 | cv_broadcast(&l2arc_feed_thr_cv); | |
8883 | CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ | |
8884 | thread_exit(); | |
8885 | } | |
8886 | ||
b128c09f BB |
8887 | boolean_t |
8888 | l2arc_vdev_present(vdev_t *vd) | |
8889 | { | |
8890 | l2arc_dev_t *dev; | |
8891 | ||
8892 | mutex_enter(&l2arc_dev_mtx); | |
8893 | for (dev = list_head(l2arc_dev_list); dev != NULL; | |
8894 | dev = list_next(l2arc_dev_list, dev)) { | |
8895 | if (dev->l2ad_vdev == vd) | |
8896 | break; | |
8897 | } | |
8898 | mutex_exit(&l2arc_dev_mtx); | |
8899 | ||
8900 | return (dev != NULL); | |
8901 | } | |
8902 | ||
34dc7c2f BB |
8903 | /* |
8904 | * Add a vdev for use by the L2ARC. By this point the spa has already | |
8905 | * validated the vdev and opened it. | |
8906 | */ | |
8907 | void | |
9babb374 | 8908 | l2arc_add_vdev(spa_t *spa, vdev_t *vd) |
34dc7c2f BB |
8909 | { |
8910 | l2arc_dev_t *adddev; | |
8911 | ||
b128c09f BB |
8912 | ASSERT(!l2arc_vdev_present(vd)); |
8913 | ||
34dc7c2f BB |
8914 | /* |
8915 | * Create a new l2arc device entry. | |
8916 | */ | |
8917 | adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); | |
8918 | adddev->l2ad_spa = spa; | |
8919 | adddev->l2ad_vdev = vd; | |
9babb374 BB |
8920 | adddev->l2ad_start = VDEV_LABEL_START_SIZE; |
8921 | adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); | |
34dc7c2f | 8922 | adddev->l2ad_hand = adddev->l2ad_start; |
34dc7c2f | 8923 | adddev->l2ad_first = B_TRUE; |
d164b209 | 8924 | adddev->l2ad_writing = B_FALSE; |
98f72a53 | 8925 | list_link_init(&adddev->l2ad_node); |
34dc7c2f | 8926 | |
b9541d6b | 8927 | mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f BB |
8928 | /* |
8929 | * This is a list of all ARC buffers that are still valid on the | |
8930 | * device. | |
8931 | */ | |
b9541d6b CW |
8932 | list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), |
8933 | offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); | |
34dc7c2f | 8934 | |
428870ff | 8935 | vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); |
d962d5da | 8936 | refcount_create(&adddev->l2ad_alloc); |
34dc7c2f BB |
8937 | |
8938 | /* | |
8939 | * Add device to global list | |
8940 | */ | |
8941 | mutex_enter(&l2arc_dev_mtx); | |
8942 | list_insert_head(l2arc_dev_list, adddev); | |
8943 | atomic_inc_64(&l2arc_ndev); | |
8944 | mutex_exit(&l2arc_dev_mtx); | |
8945 | } | |
8946 | ||
8947 | /* | |
8948 | * Remove a vdev from the L2ARC. | |
8949 | */ | |
8950 | void | |
8951 | l2arc_remove_vdev(vdev_t *vd) | |
8952 | { | |
8953 | l2arc_dev_t *dev, *nextdev, *remdev = NULL; | |
8954 | ||
34dc7c2f BB |
8955 | /* |
8956 | * Find the device by vdev | |
8957 | */ | |
8958 | mutex_enter(&l2arc_dev_mtx); | |
8959 | for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { | |
8960 | nextdev = list_next(l2arc_dev_list, dev); | |
8961 | if (vd == dev->l2ad_vdev) { | |
8962 | remdev = dev; | |
8963 | break; | |
8964 | } | |
8965 | } | |
d3c2ae1c | 8966 | ASSERT3P(remdev, !=, NULL); |
34dc7c2f BB |
8967 | |
8968 | /* | |
8969 | * Remove device from global list | |
8970 | */ | |
8971 | list_remove(l2arc_dev_list, remdev); | |
8972 | l2arc_dev_last = NULL; /* may have been invalidated */ | |
b128c09f BB |
8973 | atomic_dec_64(&l2arc_ndev); |
8974 | mutex_exit(&l2arc_dev_mtx); | |
34dc7c2f BB |
8975 | |
8976 | /* | |
8977 | * Clear all buflists and ARC references. L2ARC device flush. | |
8978 | */ | |
8979 | l2arc_evict(remdev, 0, B_TRUE); | |
b9541d6b CW |
8980 | list_destroy(&remdev->l2ad_buflist); |
8981 | mutex_destroy(&remdev->l2ad_mtx); | |
d962d5da | 8982 | refcount_destroy(&remdev->l2ad_alloc); |
34dc7c2f | 8983 | kmem_free(remdev, sizeof (l2arc_dev_t)); |
34dc7c2f BB |
8984 | } |
8985 | ||
8986 | void | |
b128c09f | 8987 | l2arc_init(void) |
34dc7c2f BB |
8988 | { |
8989 | l2arc_thread_exit = 0; | |
8990 | l2arc_ndev = 0; | |
8991 | l2arc_writes_sent = 0; | |
8992 | l2arc_writes_done = 0; | |
8993 | ||
8994 | mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); | |
8995 | cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); | |
8996 | mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); | |
34dc7c2f BB |
8997 | mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); |
8998 | ||
8999 | l2arc_dev_list = &L2ARC_dev_list; | |
9000 | l2arc_free_on_write = &L2ARC_free_on_write; | |
9001 | list_create(l2arc_dev_list, sizeof (l2arc_dev_t), | |
9002 | offsetof(l2arc_dev_t, l2ad_node)); | |
9003 | list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), | |
9004 | offsetof(l2arc_data_free_t, l2df_list_node)); | |
34dc7c2f BB |
9005 | } |
9006 | ||
9007 | void | |
b128c09f | 9008 | l2arc_fini(void) |
34dc7c2f | 9009 | { |
b128c09f BB |
9010 | /* |
9011 | * This is called from dmu_fini(), which is called from spa_fini(); | |
9012 | * Because of this, we can assume that all l2arc devices have | |
9013 | * already been removed when the pools themselves were removed. | |
9014 | */ | |
9015 | ||
9016 | l2arc_do_free_on_write(); | |
34dc7c2f BB |
9017 | |
9018 | mutex_destroy(&l2arc_feed_thr_lock); | |
9019 | cv_destroy(&l2arc_feed_thr_cv); | |
9020 | mutex_destroy(&l2arc_dev_mtx); | |
34dc7c2f BB |
9021 | mutex_destroy(&l2arc_free_on_write_mtx); |
9022 | ||
9023 | list_destroy(l2arc_dev_list); | |
9024 | list_destroy(l2arc_free_on_write); | |
9025 | } | |
b128c09f BB |
9026 | |
9027 | void | |
9028 | l2arc_start(void) | |
9029 | { | |
fb5f0bc8 | 9030 | if (!(spa_mode_global & FWRITE)) |
b128c09f BB |
9031 | return; |
9032 | ||
9033 | (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, | |
1229323d | 9034 | TS_RUN, defclsyspri); |
b128c09f BB |
9035 | } |
9036 | ||
9037 | void | |
9038 | l2arc_stop(void) | |
9039 | { | |
fb5f0bc8 | 9040 | if (!(spa_mode_global & FWRITE)) |
b128c09f BB |
9041 | return; |
9042 | ||
9043 | mutex_enter(&l2arc_feed_thr_lock); | |
9044 | cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ | |
9045 | l2arc_thread_exit = 1; | |
9046 | while (l2arc_thread_exit != 0) | |
9047 | cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); | |
9048 | mutex_exit(&l2arc_feed_thr_lock); | |
9049 | } | |
c28b2279 | 9050 | |
93ce2b4c | 9051 | #if defined(_KERNEL) |
0f699108 AZ |
9052 | EXPORT_SYMBOL(arc_buf_size); |
9053 | EXPORT_SYMBOL(arc_write); | |
c28b2279 | 9054 | EXPORT_SYMBOL(arc_read); |
e0b0ca98 | 9055 | EXPORT_SYMBOL(arc_buf_info); |
c28b2279 | 9056 | EXPORT_SYMBOL(arc_getbuf_func); |
ab26409d BB |
9057 | EXPORT_SYMBOL(arc_add_prune_callback); |
9058 | EXPORT_SYMBOL(arc_remove_prune_callback); | |
c28b2279 | 9059 | |
02730c33 | 9060 | /* BEGIN CSTYLED */ |
bce45ec9 | 9061 | module_param(zfs_arc_min, ulong, 0644); |
c409e464 | 9062 | MODULE_PARM_DESC(zfs_arc_min, "Min arc size"); |
c28b2279 | 9063 | |
bce45ec9 | 9064 | module_param(zfs_arc_max, ulong, 0644); |
c409e464 | 9065 | MODULE_PARM_DESC(zfs_arc_max, "Max arc size"); |
c28b2279 | 9066 | |
bce45ec9 | 9067 | module_param(zfs_arc_meta_limit, ulong, 0644); |
c28b2279 | 9068 | MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size"); |
6a8f9b6b | 9069 | |
9907cc1c G |
9070 | module_param(zfs_arc_meta_limit_percent, ulong, 0644); |
9071 | MODULE_PARM_DESC(zfs_arc_meta_limit_percent, | |
9072 | "Percent of arc size for arc meta limit"); | |
9073 | ||
ca0bf58d PS |
9074 | module_param(zfs_arc_meta_min, ulong, 0644); |
9075 | MODULE_PARM_DESC(zfs_arc_meta_min, "Min arc metadata"); | |
9076 | ||
bce45ec9 | 9077 | module_param(zfs_arc_meta_prune, int, 0644); |
2cbb06b5 | 9078 | MODULE_PARM_DESC(zfs_arc_meta_prune, "Meta objects to scan for prune"); |
c409e464 | 9079 | |
ca67b33a | 9080 | module_param(zfs_arc_meta_adjust_restarts, int, 0644); |
bc888666 BB |
9081 | MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts, |
9082 | "Limit number of restarts in arc_adjust_meta"); | |
9083 | ||
f6046738 BB |
9084 | module_param(zfs_arc_meta_strategy, int, 0644); |
9085 | MODULE_PARM_DESC(zfs_arc_meta_strategy, "Meta reclaim strategy"); | |
9086 | ||
bce45ec9 | 9087 | module_param(zfs_arc_grow_retry, int, 0644); |
c409e464 BB |
9088 | MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size"); |
9089 | ||
62422785 PS |
9090 | module_param(zfs_arc_p_dampener_disable, int, 0644); |
9091 | MODULE_PARM_DESC(zfs_arc_p_dampener_disable, "disable arc_p adapt dampener"); | |
9092 | ||
bce45ec9 | 9093 | module_param(zfs_arc_shrink_shift, int, 0644); |
c409e464 BB |
9094 | MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)"); |
9095 | ||
03b60eee DB |
9096 | module_param(zfs_arc_pc_percent, uint, 0644); |
9097 | MODULE_PARM_DESC(zfs_arc_pc_percent, | |
9098 | "Percent of pagecache to reclaim arc to"); | |
9099 | ||
728d6ae9 BB |
9100 | module_param(zfs_arc_p_min_shift, int, 0644); |
9101 | MODULE_PARM_DESC(zfs_arc_p_min_shift, "arc_c shift to calc min/max arc_p"); | |
9102 | ||
49ddb315 MA |
9103 | module_param(zfs_arc_average_blocksize, int, 0444); |
9104 | MODULE_PARM_DESC(zfs_arc_average_blocksize, "Target average block size"); | |
9105 | ||
d3c2ae1c | 9106 | module_param(zfs_compressed_arc_enabled, int, 0644); |
544596c5 | 9107 | MODULE_PARM_DESC(zfs_compressed_arc_enabled, "Disable compressed arc buffers"); |
d3c2ae1c | 9108 | |
d4a72f23 TC |
9109 | module_param(zfs_arc_min_prefetch_ms, int, 0644); |
9110 | MODULE_PARM_DESC(zfs_arc_min_prefetch_ms, "Min life of prefetch block in ms"); | |
9111 | ||
9112 | module_param(zfs_arc_min_prescient_prefetch_ms, int, 0644); | |
9113 | MODULE_PARM_DESC(zfs_arc_min_prescient_prefetch_ms, | |
9114 | "Min life of prescient prefetched block in ms"); | |
bce45ec9 BB |
9115 | |
9116 | module_param(l2arc_write_max, ulong, 0644); | |
abd8610c BB |
9117 | MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval"); |
9118 | ||
bce45ec9 | 9119 | module_param(l2arc_write_boost, ulong, 0644); |
abd8610c BB |
9120 | MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup"); |
9121 | ||
bce45ec9 | 9122 | module_param(l2arc_headroom, ulong, 0644); |
abd8610c BB |
9123 | MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache"); |
9124 | ||
3a17a7a9 SK |
9125 | module_param(l2arc_headroom_boost, ulong, 0644); |
9126 | MODULE_PARM_DESC(l2arc_headroom_boost, "Compressed l2arc_headroom multiplier"); | |
9127 | ||
bce45ec9 | 9128 | module_param(l2arc_feed_secs, ulong, 0644); |
abd8610c BB |
9129 | MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing"); |
9130 | ||
bce45ec9 | 9131 | module_param(l2arc_feed_min_ms, ulong, 0644); |
abd8610c BB |
9132 | MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds"); |
9133 | ||
bce45ec9 | 9134 | module_param(l2arc_noprefetch, int, 0644); |
abd8610c BB |
9135 | MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers"); |
9136 | ||
bce45ec9 | 9137 | module_param(l2arc_feed_again, int, 0644); |
abd8610c BB |
9138 | MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup"); |
9139 | ||
bce45ec9 | 9140 | module_param(l2arc_norw, int, 0644); |
abd8610c BB |
9141 | MODULE_PARM_DESC(l2arc_norw, "No reads during writes"); |
9142 | ||
7e8bddd0 BB |
9143 | module_param(zfs_arc_lotsfree_percent, int, 0644); |
9144 | MODULE_PARM_DESC(zfs_arc_lotsfree_percent, | |
9145 | "System free memory I/O throttle in bytes"); | |
9146 | ||
11f552fa BB |
9147 | module_param(zfs_arc_sys_free, ulong, 0644); |
9148 | MODULE_PARM_DESC(zfs_arc_sys_free, "System free memory target size in bytes"); | |
9149 | ||
25458cbe TC |
9150 | module_param(zfs_arc_dnode_limit, ulong, 0644); |
9151 | MODULE_PARM_DESC(zfs_arc_dnode_limit, "Minimum bytes of dnodes in arc"); | |
9152 | ||
9907cc1c G |
9153 | module_param(zfs_arc_dnode_limit_percent, ulong, 0644); |
9154 | MODULE_PARM_DESC(zfs_arc_dnode_limit_percent, | |
9155 | "Percent of ARC meta buffers for dnodes"); | |
9156 | ||
25458cbe TC |
9157 | module_param(zfs_arc_dnode_reduce_percent, ulong, 0644); |
9158 | MODULE_PARM_DESC(zfs_arc_dnode_reduce_percent, | |
9159 | "Percentage of excess dnodes to try to unpin"); | |
02730c33 | 9160 | /* END CSTYLED */ |
c28b2279 | 9161 | #endif |