]>
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. |
3ec34e55 | 23 | * Copyright (c) 2018, Joyent, Inc. |
77f6826b GA |
24 | * Copyright (c) 2011, 2019, Delphix. All rights reserved. |
25 | * Copyright (c) 2014, Saso Kiselkov. All rights reserved. | |
26 | * Copyright (c) 2017, Nexenta Systems, Inc. All rights reserved. | |
e3570464 | 27 | * Copyright (c) 2019, loli10K <ezomori.nozomu@gmail.com>. All rights reserved. |
77f6826b | 28 | * Copyright (c) 2020, George Amanakis. All rights reserved. |
34dc7c2f BB |
29 | */ |
30 | ||
34dc7c2f BB |
31 | /* |
32 | * DVA-based Adjustable Replacement Cache | |
33 | * | |
34 | * While much of the theory of operation used here is | |
35 | * based on the self-tuning, low overhead replacement cache | |
36 | * presented by Megiddo and Modha at FAST 2003, there are some | |
37 | * significant differences: | |
38 | * | |
39 | * 1. The Megiddo and Modha model assumes any page is evictable. | |
40 | * Pages in its cache cannot be "locked" into memory. This makes | |
41 | * the eviction algorithm simple: evict the last page in the list. | |
42 | * This also make the performance characteristics easy to reason | |
43 | * about. Our cache is not so simple. At any given moment, some | |
44 | * subset of the blocks in the cache are un-evictable because we | |
45 | * have handed out a reference to them. Blocks are only evictable | |
46 | * when there are no external references active. This makes | |
47 | * eviction far more problematic: we choose to evict the evictable | |
48 | * blocks that are the "lowest" in the list. | |
49 | * | |
50 | * There are times when it is not possible to evict the requested | |
51 | * space. In these circumstances we are unable to adjust the cache | |
52 | * size. To prevent the cache growing unbounded at these times we | |
53 | * implement a "cache throttle" that slows the flow of new data | |
54 | * into the cache until we can make space available. | |
55 | * | |
56 | * 2. The Megiddo and Modha model assumes a fixed cache size. | |
57 | * Pages are evicted when the cache is full and there is a cache | |
58 | * miss. Our model has a variable sized cache. It grows with | |
59 | * high use, but also tries to react to memory pressure from the | |
60 | * operating system: decreasing its size when system memory is | |
61 | * tight. | |
62 | * | |
63 | * 3. The Megiddo and Modha model assumes a fixed page size. All | |
d3cc8b15 | 64 | * elements of the cache are therefore exactly the same size. So |
34dc7c2f BB |
65 | * when adjusting the cache size following a cache miss, its simply |
66 | * a matter of choosing a single page to evict. In our model, we | |
e1cfd73f | 67 | * have variable sized cache blocks (ranging from 512 bytes to |
d3cc8b15 | 68 | * 128K bytes). We therefore choose a set of blocks to evict to make |
34dc7c2f BB |
69 | * space for a cache miss that approximates as closely as possible |
70 | * the space used by the new block. | |
71 | * | |
72 | * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" | |
73 | * by N. Megiddo & D. Modha, FAST 2003 | |
74 | */ | |
75 | ||
76 | /* | |
77 | * The locking model: | |
78 | * | |
79 | * A new reference to a cache buffer can be obtained in two | |
80 | * ways: 1) via a hash table lookup using the DVA as a key, | |
81 | * or 2) via one of the ARC lists. The arc_read() interface | |
2aa34383 | 82 | * uses method 1, while the internal ARC algorithms for |
d3cc8b15 | 83 | * adjusting the cache use method 2. We therefore provide two |
34dc7c2f | 84 | * types of locks: 1) the hash table lock array, and 2) the |
2aa34383 | 85 | * ARC list locks. |
34dc7c2f | 86 | * |
5c839890 BC |
87 | * Buffers do not have their own mutexes, rather they rely on the |
88 | * hash table mutexes for the bulk of their protection (i.e. most | |
89 | * fields in the arc_buf_hdr_t are protected by these mutexes). | |
34dc7c2f BB |
90 | * |
91 | * buf_hash_find() returns the appropriate mutex (held) when it | |
92 | * locates the requested buffer in the hash table. It returns | |
93 | * NULL for the mutex if the buffer was not in the table. | |
94 | * | |
95 | * buf_hash_remove() expects the appropriate hash mutex to be | |
96 | * already held before it is invoked. | |
97 | * | |
2aa34383 | 98 | * Each ARC state also has a mutex which is used to protect the |
34dc7c2f | 99 | * buffer list associated with the state. When attempting to |
2aa34383 | 100 | * obtain a hash table lock while holding an ARC list lock you |
34dc7c2f BB |
101 | * must use: mutex_tryenter() to avoid deadlock. Also note that |
102 | * the active state mutex must be held before the ghost state mutex. | |
103 | * | |
ab26409d BB |
104 | * It as also possible to register a callback which is run when the |
105 | * arc_meta_limit is reached and no buffers can be safely evicted. In | |
106 | * this case the arc user should drop a reference on some arc buffers so | |
107 | * they can be reclaimed and the arc_meta_limit honored. For example, | |
108 | * when using the ZPL each dentry holds a references on a znode. These | |
109 | * dentries must be pruned before the arc buffer holding the znode can | |
110 | * be safely evicted. | |
111 | * | |
34dc7c2f BB |
112 | * Note that the majority of the performance stats are manipulated |
113 | * with atomic operations. | |
114 | * | |
b9541d6b | 115 | * The L2ARC uses the l2ad_mtx on each vdev for the following: |
34dc7c2f BB |
116 | * |
117 | * - L2ARC buflist creation | |
118 | * - L2ARC buflist eviction | |
119 | * - L2ARC write completion, which walks L2ARC buflists | |
120 | * - ARC header destruction, as it removes from L2ARC buflists | |
121 | * - ARC header release, as it removes from L2ARC buflists | |
122 | */ | |
123 | ||
d3c2ae1c GW |
124 | /* |
125 | * ARC operation: | |
126 | * | |
127 | * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure. | |
128 | * This structure can point either to a block that is still in the cache or to | |
129 | * one that is only accessible in an L2 ARC device, or it can provide | |
130 | * information about a block that was recently evicted. If a block is | |
131 | * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough | |
132 | * information to retrieve it from the L2ARC device. This information is | |
133 | * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block | |
134 | * that is in this state cannot access the data directly. | |
135 | * | |
136 | * Blocks that are actively being referenced or have not been evicted | |
137 | * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within | |
138 | * the arc_buf_hdr_t that will point to the data block in memory. A block can | |
139 | * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC | |
2aa34383 | 140 | * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and |
a6255b7f | 141 | * also in the arc_buf_hdr_t's private physical data block pointer (b_pabd). |
2aa34383 DK |
142 | * |
143 | * The L1ARC's data pointer may or may not be uncompressed. The ARC has the | |
a6255b7f DQ |
144 | * ability to store the physical data (b_pabd) associated with the DVA of the |
145 | * arc_buf_hdr_t. Since the b_pabd is a copy of the on-disk physical block, | |
2aa34383 DK |
146 | * it will match its on-disk compression characteristics. This behavior can be |
147 | * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the | |
a6255b7f | 148 | * compressed ARC functionality is disabled, the b_pabd will point to an |
2aa34383 DK |
149 | * uncompressed version of the on-disk data. |
150 | * | |
151 | * Data in the L1ARC is not accessed by consumers of the ARC directly. Each | |
152 | * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it. | |
153 | * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC | |
154 | * consumer. The ARC will provide references to this data and will keep it | |
155 | * cached until it is no longer in use. The ARC caches only the L1ARC's physical | |
156 | * data block and will evict any arc_buf_t that is no longer referenced. The | |
157 | * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the | |
d3c2ae1c GW |
158 | * "overhead_size" kstat. |
159 | * | |
2aa34383 DK |
160 | * Depending on the consumer, an arc_buf_t can be requested in uncompressed or |
161 | * compressed form. The typical case is that consumers will want uncompressed | |
162 | * data, and when that happens a new data buffer is allocated where the data is | |
163 | * decompressed for them to use. Currently the only consumer who wants | |
164 | * compressed arc_buf_t's is "zfs send", when it streams data exactly as it | |
165 | * exists on disk. When this happens, the arc_buf_t's data buffer is shared | |
166 | * with the arc_buf_hdr_t. | |
d3c2ae1c | 167 | * |
2aa34383 DK |
168 | * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The |
169 | * first one is owned by a compressed send consumer (and therefore references | |
170 | * the same compressed data buffer as the arc_buf_hdr_t) and the second could be | |
171 | * used by any other consumer (and has its own uncompressed copy of the data | |
172 | * buffer). | |
d3c2ae1c | 173 | * |
2aa34383 DK |
174 | * arc_buf_hdr_t |
175 | * +-----------+ | |
176 | * | fields | | |
177 | * | common to | | |
178 | * | L1- and | | |
179 | * | L2ARC | | |
180 | * +-----------+ | |
181 | * | l2arc_buf_hdr_t | |
182 | * | | | |
183 | * +-----------+ | |
184 | * | l1arc_buf_hdr_t | |
185 | * | | arc_buf_t | |
186 | * | b_buf +------------>+-----------+ arc_buf_t | |
a6255b7f | 187 | * | b_pabd +-+ |b_next +---->+-----------+ |
2aa34383 DK |
188 | * +-----------+ | |-----------| |b_next +-->NULL |
189 | * | |b_comp = T | +-----------+ | |
190 | * | |b_data +-+ |b_comp = F | | |
191 | * | +-----------+ | |b_data +-+ | |
192 | * +->+------+ | +-----------+ | | |
193 | * compressed | | | | | |
194 | * data | |<--------------+ | uncompressed | |
195 | * +------+ compressed, | data | |
196 | * shared +-->+------+ | |
197 | * data | | | |
198 | * | | | |
199 | * +------+ | |
d3c2ae1c GW |
200 | * |
201 | * When a consumer reads a block, the ARC must first look to see if the | |
2aa34383 DK |
202 | * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new |
203 | * arc_buf_t and either copies uncompressed data into a new data buffer from an | |
a6255b7f DQ |
204 | * existing uncompressed arc_buf_t, decompresses the hdr's b_pabd buffer into a |
205 | * new data buffer, or shares the hdr's b_pabd buffer, depending on whether the | |
2aa34383 DK |
206 | * hdr is compressed and the desired compression characteristics of the |
207 | * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the | |
208 | * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be | |
209 | * the last buffer in the hdr's b_buf list, however a shared compressed buf can | |
210 | * be anywhere in the hdr's list. | |
d3c2ae1c GW |
211 | * |
212 | * The diagram below shows an example of an uncompressed ARC hdr that is | |
2aa34383 DK |
213 | * sharing its data with an arc_buf_t (note that the shared uncompressed buf is |
214 | * the last element in the buf list): | |
d3c2ae1c GW |
215 | * |
216 | * arc_buf_hdr_t | |
217 | * +-----------+ | |
218 | * | | | |
219 | * | | | |
220 | * | | | |
221 | * +-----------+ | |
222 | * l2arc_buf_hdr_t| | | |
223 | * | | | |
224 | * +-----------+ | |
225 | * l1arc_buf_hdr_t| | | |
226 | * | | arc_buf_t (shared) | |
227 | * | b_buf +------------>+---------+ arc_buf_t | |
228 | * | | |b_next +---->+---------+ | |
a6255b7f | 229 | * | b_pabd +-+ |---------| |b_next +-->NULL |
d3c2ae1c GW |
230 | * +-----------+ | | | +---------+ |
231 | * | |b_data +-+ | | | |
232 | * | +---------+ | |b_data +-+ | |
233 | * +->+------+ | +---------+ | | |
234 | * | | | | | |
235 | * uncompressed | | | | | |
236 | * data +------+ | | | |
237 | * ^ +->+------+ | | |
238 | * | uncompressed | | | | |
239 | * | data | | | | |
240 | * | +------+ | | |
241 | * +---------------------------------+ | |
242 | * | |
a6255b7f | 243 | * Writing to the ARC requires that the ARC first discard the hdr's b_pabd |
d3c2ae1c | 244 | * since the physical block is about to be rewritten. The new data contents |
2aa34383 DK |
245 | * will be contained in the arc_buf_t. As the I/O pipeline performs the write, |
246 | * it may compress the data before writing it to disk. The ARC will be called | |
247 | * with the transformed data and will bcopy the transformed on-disk block into | |
a6255b7f | 248 | * a newly allocated b_pabd. Writes are always done into buffers which have |
2aa34383 DK |
249 | * either been loaned (and hence are new and don't have other readers) or |
250 | * buffers which have been released (and hence have their own hdr, if there | |
251 | * were originally other readers of the buf's original hdr). This ensures that | |
252 | * the ARC only needs to update a single buf and its hdr after a write occurs. | |
d3c2ae1c | 253 | * |
a6255b7f DQ |
254 | * When the L2ARC is in use, it will also take advantage of the b_pabd. The |
255 | * L2ARC will always write the contents of b_pabd to the L2ARC. This means | |
2aa34383 | 256 | * that when compressed ARC is enabled that the L2ARC blocks are identical |
d3c2ae1c GW |
257 | * to the on-disk block in the main data pool. This provides a significant |
258 | * advantage since the ARC can leverage the bp's checksum when reading from the | |
259 | * L2ARC to determine if the contents are valid. However, if the compressed | |
2aa34383 | 260 | * ARC is disabled, then the L2ARC's block must be transformed to look |
d3c2ae1c GW |
261 | * like the physical block in the main data pool before comparing the |
262 | * checksum and determining its validity. | |
b5256303 TC |
263 | * |
264 | * The L1ARC has a slightly different system for storing encrypted data. | |
265 | * Raw (encrypted + possibly compressed) data has a few subtle differences from | |
266 | * data that is just compressed. The biggest difference is that it is not | |
e1cfd73f | 267 | * possible to decrypt encrypted data (or vice-versa) if the keys aren't loaded. |
b5256303 TC |
268 | * The other difference is that encryption cannot be treated as a suggestion. |
269 | * If a caller would prefer compressed data, but they actually wind up with | |
270 | * uncompressed data the worst thing that could happen is there might be a | |
271 | * performance hit. If the caller requests encrypted data, however, we must be | |
272 | * sure they actually get it or else secret information could be leaked. Raw | |
273 | * data is stored in hdr->b_crypt_hdr.b_rabd. An encrypted header, therefore, | |
274 | * may have both an encrypted version and a decrypted version of its data at | |
275 | * once. When a caller needs a raw arc_buf_t, it is allocated and the data is | |
276 | * copied out of this header. To avoid complications with b_pabd, raw buffers | |
277 | * cannot be shared. | |
d3c2ae1c GW |
278 | */ |
279 | ||
34dc7c2f BB |
280 | #include <sys/spa.h> |
281 | #include <sys/zio.h> | |
d3c2ae1c | 282 | #include <sys/spa_impl.h> |
3a17a7a9 | 283 | #include <sys/zio_compress.h> |
d3c2ae1c | 284 | #include <sys/zio_checksum.h> |
34dc7c2f BB |
285 | #include <sys/zfs_context.h> |
286 | #include <sys/arc.h> | |
36da08ef | 287 | #include <sys/refcount.h> |
b128c09f | 288 | #include <sys/vdev.h> |
9babb374 | 289 | #include <sys/vdev_impl.h> |
e8b96c60 | 290 | #include <sys/dsl_pool.h> |
a6255b7f | 291 | #include <sys/zio_checksum.h> |
ca0bf58d | 292 | #include <sys/multilist.h> |
a6255b7f | 293 | #include <sys/abd.h> |
b5256303 TC |
294 | #include <sys/zil.h> |
295 | #include <sys/fm/fs/zfs.h> | |
34dc7c2f BB |
296 | #include <sys/callb.h> |
297 | #include <sys/kstat.h> | |
3ec34e55 | 298 | #include <sys/zthr.h> |
428870ff | 299 | #include <zfs_fletcher.h> |
59ec819a | 300 | #include <sys/arc_impl.h> |
e5d1c27e | 301 | #include <sys/trace_zfs.h> |
37fb3e43 | 302 | #include <sys/aggsum.h> |
3f387973 | 303 | #include <cityhash.h> |
34dc7c2f | 304 | |
498877ba MA |
305 | #ifndef _KERNEL |
306 | /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ | |
307 | boolean_t arc_watch = B_FALSE; | |
308 | #endif | |
309 | ||
3ec34e55 BL |
310 | /* |
311 | * This thread's job is to keep enough free memory in the system, by | |
312 | * calling arc_kmem_reap_soon() plus arc_reduce_target_size(), which improves | |
313 | * arc_available_memory(). | |
314 | */ | |
315 | static zthr_t *arc_reap_zthr; | |
316 | ||
317 | /* | |
318 | * This thread's job is to keep arc_size under arc_c, by calling | |
319 | * arc_adjust(), which improves arc_is_overflowing(). | |
320 | */ | |
c9c9c1e2 | 321 | zthr_t *arc_adjust_zthr; |
3ec34e55 | 322 | |
c9c9c1e2 MM |
323 | kmutex_t arc_adjust_lock; |
324 | kcondvar_t arc_adjust_waiters_cv; | |
325 | boolean_t arc_adjust_needed = B_FALSE; | |
ca0bf58d | 326 | |
e8b96c60 | 327 | /* |
ca0bf58d PS |
328 | * The number of headers to evict in arc_evict_state_impl() before |
329 | * dropping the sublist lock and evicting from another sublist. A lower | |
330 | * value means we're more likely to evict the "correct" header (i.e. the | |
331 | * oldest header in the arc state), but comes with higher overhead | |
332 | * (i.e. more invocations of arc_evict_state_impl()). | |
333 | */ | |
334 | int zfs_arc_evict_batch_limit = 10; | |
335 | ||
34dc7c2f | 336 | /* number of seconds before growing cache again */ |
c9c9c1e2 | 337 | int arc_grow_retry = 5; |
3ec34e55 BL |
338 | |
339 | /* | |
340 | * Minimum time between calls to arc_kmem_reap_soon(). | |
341 | */ | |
342 | int arc_kmem_cache_reap_retry_ms = 1000; | |
34dc7c2f | 343 | |
a6255b7f | 344 | /* shift of arc_c for calculating overflow limit in arc_get_data_impl */ |
3ec34e55 | 345 | int zfs_arc_overflow_shift = 8; |
62422785 | 346 | |
728d6ae9 | 347 | /* shift of arc_c for calculating both min and max arc_p */ |
3ec34e55 | 348 | int arc_p_min_shift = 4; |
728d6ae9 | 349 | |
d164b209 | 350 | /* log2(fraction of arc to reclaim) */ |
c9c9c1e2 | 351 | int arc_shrink_shift = 7; |
d164b209 | 352 | |
03b60eee DB |
353 | /* percent of pagecache to reclaim arc to */ |
354 | #ifdef _KERNEL | |
c9c9c1e2 | 355 | uint_t zfs_arc_pc_percent = 0; |
03b60eee DB |
356 | #endif |
357 | ||
34dc7c2f | 358 | /* |
ca67b33a MA |
359 | * log2(fraction of ARC which must be free to allow growing). |
360 | * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, | |
361 | * when reading a new block into the ARC, we will evict an equal-sized block | |
362 | * from the ARC. | |
363 | * | |
364 | * This must be less than arc_shrink_shift, so that when we shrink the ARC, | |
365 | * we will still not allow it to grow. | |
34dc7c2f | 366 | */ |
ca67b33a | 367 | int arc_no_grow_shift = 5; |
bce45ec9 | 368 | |
49ddb315 | 369 | |
ca0bf58d PS |
370 | /* |
371 | * minimum lifespan of a prefetch block in clock ticks | |
372 | * (initialized in arc_init()) | |
373 | */ | |
d4a72f23 TC |
374 | static int arc_min_prefetch_ms; |
375 | static int arc_min_prescient_prefetch_ms; | |
ca0bf58d | 376 | |
e8b96c60 MA |
377 | /* |
378 | * If this percent of memory is free, don't throttle. | |
379 | */ | |
380 | int arc_lotsfree_percent = 10; | |
381 | ||
3ec34e55 BL |
382 | /* |
383 | * hdr_recl() uses this to determine if the arc is up and running. | |
384 | */ | |
385 | static boolean_t arc_initialized; | |
34dc7c2f | 386 | |
b128c09f BB |
387 | /* |
388 | * The arc has filled available memory and has now warmed up. | |
389 | */ | |
c9c9c1e2 | 390 | boolean_t arc_warm; |
b128c09f | 391 | |
d3c2ae1c GW |
392 | /* |
393 | * log2 fraction of the zio arena to keep free. | |
394 | */ | |
395 | int arc_zio_arena_free_shift = 2; | |
396 | ||
34dc7c2f BB |
397 | /* |
398 | * These tunables are for performance analysis. | |
399 | */ | |
c28b2279 BB |
400 | unsigned long zfs_arc_max = 0; |
401 | unsigned long zfs_arc_min = 0; | |
402 | unsigned long zfs_arc_meta_limit = 0; | |
ca0bf58d | 403 | unsigned long zfs_arc_meta_min = 0; |
25458cbe TC |
404 | unsigned long zfs_arc_dnode_limit = 0; |
405 | unsigned long zfs_arc_dnode_reduce_percent = 10; | |
ca67b33a MA |
406 | int zfs_arc_grow_retry = 0; |
407 | int zfs_arc_shrink_shift = 0; | |
728d6ae9 | 408 | int zfs_arc_p_min_shift = 0; |
ca67b33a | 409 | int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ |
34dc7c2f | 410 | |
dae3e9ea DB |
411 | /* |
412 | * ARC dirty data constraints for arc_tempreserve_space() throttle. | |
413 | */ | |
414 | unsigned long zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */ | |
415 | unsigned long zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */ | |
416 | unsigned long zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */ | |
417 | ||
418 | /* | |
419 | * Enable or disable compressed arc buffers. | |
420 | */ | |
d3c2ae1c GW |
421 | int zfs_compressed_arc_enabled = B_TRUE; |
422 | ||
9907cc1c G |
423 | /* |
424 | * ARC will evict meta buffers that exceed arc_meta_limit. This | |
425 | * tunable make arc_meta_limit adjustable for different workloads. | |
426 | */ | |
427 | unsigned long zfs_arc_meta_limit_percent = 75; | |
428 | ||
429 | /* | |
430 | * Percentage that can be consumed by dnodes of ARC meta buffers. | |
431 | */ | |
432 | unsigned long zfs_arc_dnode_limit_percent = 10; | |
433 | ||
bc888666 | 434 | /* |
ca67b33a | 435 | * These tunables are Linux specific |
bc888666 | 436 | */ |
11f552fa | 437 | unsigned long zfs_arc_sys_free = 0; |
d4a72f23 TC |
438 | int zfs_arc_min_prefetch_ms = 0; |
439 | int zfs_arc_min_prescient_prefetch_ms = 0; | |
ca67b33a MA |
440 | int zfs_arc_p_dampener_disable = 1; |
441 | int zfs_arc_meta_prune = 10000; | |
442 | int zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED; | |
443 | int zfs_arc_meta_adjust_restarts = 4096; | |
7e8bddd0 | 444 | int zfs_arc_lotsfree_percent = 10; |
bc888666 | 445 | |
34dc7c2f | 446 | /* The 6 states: */ |
13a4027a MM |
447 | arc_state_t ARC_anon; |
448 | arc_state_t ARC_mru; | |
449 | arc_state_t ARC_mru_ghost; | |
450 | arc_state_t ARC_mfu; | |
451 | arc_state_t ARC_mfu_ghost; | |
452 | arc_state_t ARC_l2c_only; | |
34dc7c2f | 453 | |
c9c9c1e2 | 454 | arc_stats_t arc_stats = { |
34dc7c2f BB |
455 | { "hits", KSTAT_DATA_UINT64 }, |
456 | { "misses", KSTAT_DATA_UINT64 }, | |
457 | { "demand_data_hits", KSTAT_DATA_UINT64 }, | |
458 | { "demand_data_misses", KSTAT_DATA_UINT64 }, | |
459 | { "demand_metadata_hits", KSTAT_DATA_UINT64 }, | |
460 | { "demand_metadata_misses", KSTAT_DATA_UINT64 }, | |
461 | { "prefetch_data_hits", KSTAT_DATA_UINT64 }, | |
462 | { "prefetch_data_misses", KSTAT_DATA_UINT64 }, | |
463 | { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, | |
464 | { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, | |
465 | { "mru_hits", KSTAT_DATA_UINT64 }, | |
466 | { "mru_ghost_hits", KSTAT_DATA_UINT64 }, | |
467 | { "mfu_hits", KSTAT_DATA_UINT64 }, | |
468 | { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, | |
469 | { "deleted", KSTAT_DATA_UINT64 }, | |
34dc7c2f | 470 | { "mutex_miss", KSTAT_DATA_UINT64 }, |
0873bb63 | 471 | { "access_skip", KSTAT_DATA_UINT64 }, |
34dc7c2f | 472 | { "evict_skip", KSTAT_DATA_UINT64 }, |
ca0bf58d | 473 | { "evict_not_enough", KSTAT_DATA_UINT64 }, |
428870ff BB |
474 | { "evict_l2_cached", KSTAT_DATA_UINT64 }, |
475 | { "evict_l2_eligible", KSTAT_DATA_UINT64 }, | |
476 | { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, | |
ca0bf58d | 477 | { "evict_l2_skip", KSTAT_DATA_UINT64 }, |
34dc7c2f BB |
478 | { "hash_elements", KSTAT_DATA_UINT64 }, |
479 | { "hash_elements_max", KSTAT_DATA_UINT64 }, | |
480 | { "hash_collisions", KSTAT_DATA_UINT64 }, | |
481 | { "hash_chains", KSTAT_DATA_UINT64 }, | |
482 | { "hash_chain_max", KSTAT_DATA_UINT64 }, | |
483 | { "p", KSTAT_DATA_UINT64 }, | |
484 | { "c", KSTAT_DATA_UINT64 }, | |
485 | { "c_min", KSTAT_DATA_UINT64 }, | |
486 | { "c_max", KSTAT_DATA_UINT64 }, | |
487 | { "size", KSTAT_DATA_UINT64 }, | |
d3c2ae1c GW |
488 | { "compressed_size", KSTAT_DATA_UINT64 }, |
489 | { "uncompressed_size", KSTAT_DATA_UINT64 }, | |
490 | { "overhead_size", KSTAT_DATA_UINT64 }, | |
34dc7c2f | 491 | { "hdr_size", KSTAT_DATA_UINT64 }, |
d164b209 | 492 | { "data_size", KSTAT_DATA_UINT64 }, |
500445c0 | 493 | { "metadata_size", KSTAT_DATA_UINT64 }, |
25458cbe TC |
494 | { "dbuf_size", KSTAT_DATA_UINT64 }, |
495 | { "dnode_size", KSTAT_DATA_UINT64 }, | |
496 | { "bonus_size", KSTAT_DATA_UINT64 }, | |
13be560d | 497 | { "anon_size", KSTAT_DATA_UINT64 }, |
500445c0 PS |
498 | { "anon_evictable_data", KSTAT_DATA_UINT64 }, |
499 | { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, | |
13be560d | 500 | { "mru_size", KSTAT_DATA_UINT64 }, |
500445c0 PS |
501 | { "mru_evictable_data", KSTAT_DATA_UINT64 }, |
502 | { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, | |
13be560d | 503 | { "mru_ghost_size", KSTAT_DATA_UINT64 }, |
500445c0 PS |
504 | { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, |
505 | { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, | |
13be560d | 506 | { "mfu_size", KSTAT_DATA_UINT64 }, |
500445c0 PS |
507 | { "mfu_evictable_data", KSTAT_DATA_UINT64 }, |
508 | { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, | |
13be560d | 509 | { "mfu_ghost_size", KSTAT_DATA_UINT64 }, |
500445c0 PS |
510 | { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, |
511 | { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, | |
34dc7c2f BB |
512 | { "l2_hits", KSTAT_DATA_UINT64 }, |
513 | { "l2_misses", KSTAT_DATA_UINT64 }, | |
514 | { "l2_feeds", KSTAT_DATA_UINT64 }, | |
515 | { "l2_rw_clash", KSTAT_DATA_UINT64 }, | |
d164b209 BB |
516 | { "l2_read_bytes", KSTAT_DATA_UINT64 }, |
517 | { "l2_write_bytes", KSTAT_DATA_UINT64 }, | |
34dc7c2f BB |
518 | { "l2_writes_sent", KSTAT_DATA_UINT64 }, |
519 | { "l2_writes_done", KSTAT_DATA_UINT64 }, | |
520 | { "l2_writes_error", KSTAT_DATA_UINT64 }, | |
ca0bf58d | 521 | { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, |
34dc7c2f BB |
522 | { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, |
523 | { "l2_evict_reading", KSTAT_DATA_UINT64 }, | |
b9541d6b | 524 | { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, |
34dc7c2f BB |
525 | { "l2_free_on_write", KSTAT_DATA_UINT64 }, |
526 | { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, | |
527 | { "l2_cksum_bad", KSTAT_DATA_UINT64 }, | |
528 | { "l2_io_error", KSTAT_DATA_UINT64 }, | |
529 | { "l2_size", KSTAT_DATA_UINT64 }, | |
3a17a7a9 | 530 | { "l2_asize", KSTAT_DATA_UINT64 }, |
34dc7c2f | 531 | { "l2_hdr_size", KSTAT_DATA_UINT64 }, |
77f6826b | 532 | { "l2_log_blk_writes", KSTAT_DATA_UINT64 }, |
657fd33b GA |
533 | { "l2_log_blk_avg_asize", KSTAT_DATA_UINT64 }, |
534 | { "l2_log_blk_asize", KSTAT_DATA_UINT64 }, | |
535 | { "l2_log_blk_count", KSTAT_DATA_UINT64 }, | |
77f6826b GA |
536 | { "l2_data_to_meta_ratio", KSTAT_DATA_UINT64 }, |
537 | { "l2_rebuild_success", KSTAT_DATA_UINT64 }, | |
538 | { "l2_rebuild_unsupported", KSTAT_DATA_UINT64 }, | |
539 | { "l2_rebuild_io_errors", KSTAT_DATA_UINT64 }, | |
540 | { "l2_rebuild_dh_errors", KSTAT_DATA_UINT64 }, | |
541 | { "l2_rebuild_cksum_lb_errors", KSTAT_DATA_UINT64 }, | |
542 | { "l2_rebuild_lowmem", KSTAT_DATA_UINT64 }, | |
543 | { "l2_rebuild_size", KSTAT_DATA_UINT64 }, | |
657fd33b | 544 | { "l2_rebuild_asize", KSTAT_DATA_UINT64 }, |
77f6826b GA |
545 | { "l2_rebuild_bufs", KSTAT_DATA_UINT64 }, |
546 | { "l2_rebuild_bufs_precached", KSTAT_DATA_UINT64 }, | |
77f6826b | 547 | { "l2_rebuild_log_blks", KSTAT_DATA_UINT64 }, |
1834f2d8 | 548 | { "memory_throttle_count", KSTAT_DATA_UINT64 }, |
7cb67b45 BB |
549 | { "memory_direct_count", KSTAT_DATA_UINT64 }, |
550 | { "memory_indirect_count", KSTAT_DATA_UINT64 }, | |
70f02287 BB |
551 | { "memory_all_bytes", KSTAT_DATA_UINT64 }, |
552 | { "memory_free_bytes", KSTAT_DATA_UINT64 }, | |
553 | { "memory_available_bytes", KSTAT_DATA_INT64 }, | |
1834f2d8 BB |
554 | { "arc_no_grow", KSTAT_DATA_UINT64 }, |
555 | { "arc_tempreserve", KSTAT_DATA_UINT64 }, | |
556 | { "arc_loaned_bytes", KSTAT_DATA_UINT64 }, | |
ab26409d | 557 | { "arc_prune", KSTAT_DATA_UINT64 }, |
1834f2d8 BB |
558 | { "arc_meta_used", KSTAT_DATA_UINT64 }, |
559 | { "arc_meta_limit", KSTAT_DATA_UINT64 }, | |
25458cbe | 560 | { "arc_dnode_limit", KSTAT_DATA_UINT64 }, |
1834f2d8 | 561 | { "arc_meta_max", KSTAT_DATA_UINT64 }, |
11f552fa | 562 | { "arc_meta_min", KSTAT_DATA_UINT64 }, |
a8b2e306 | 563 | { "async_upgrade_sync", KSTAT_DATA_UINT64 }, |
7f60329a | 564 | { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, |
d4a72f23 | 565 | { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 }, |
11f552fa | 566 | { "arc_need_free", KSTAT_DATA_UINT64 }, |
b5256303 | 567 | { "arc_sys_free", KSTAT_DATA_UINT64 }, |
1dc32a67 MA |
568 | { "arc_raw_size", KSTAT_DATA_UINT64 }, |
569 | { "cached_only_in_progress", KSTAT_DATA_UINT64 }, | |
34dc7c2f BB |
570 | }; |
571 | ||
34dc7c2f BB |
572 | #define ARCSTAT_MAX(stat, val) { \ |
573 | uint64_t m; \ | |
574 | while ((val) > (m = arc_stats.stat.value.ui64) && \ | |
575 | (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ | |
576 | continue; \ | |
577 | } | |
578 | ||
579 | #define ARCSTAT_MAXSTAT(stat) \ | |
580 | ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) | |
581 | ||
582 | /* | |
583 | * We define a macro to allow ARC hits/misses to be easily broken down by | |
584 | * two separate conditions, giving a total of four different subtypes for | |
585 | * each of hits and misses (so eight statistics total). | |
586 | */ | |
587 | #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ | |
588 | if (cond1) { \ | |
589 | if (cond2) { \ | |
590 | ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ | |
591 | } else { \ | |
592 | ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ | |
593 | } \ | |
594 | } else { \ | |
595 | if (cond2) { \ | |
596 | ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ | |
597 | } else { \ | |
598 | ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ | |
599 | } \ | |
600 | } | |
601 | ||
77f6826b GA |
602 | /* |
603 | * This macro allows us to use kstats as floating averages. Each time we | |
604 | * update this kstat, we first factor it and the update value by | |
605 | * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall | |
606 | * average. This macro assumes that integer loads and stores are atomic, but | |
607 | * is not safe for multiple writers updating the kstat in parallel (only the | |
608 | * last writer's update will remain). | |
609 | */ | |
610 | #define ARCSTAT_F_AVG_FACTOR 3 | |
611 | #define ARCSTAT_F_AVG(stat, value) \ | |
612 | do { \ | |
613 | uint64_t x = ARCSTAT(stat); \ | |
614 | x = x - x / ARCSTAT_F_AVG_FACTOR + \ | |
615 | (value) / ARCSTAT_F_AVG_FACTOR; \ | |
616 | ARCSTAT(stat) = x; \ | |
617 | _NOTE(CONSTCOND) \ | |
618 | } while (0) | |
619 | ||
34dc7c2f | 620 | kstat_t *arc_ksp; |
428870ff | 621 | static arc_state_t *arc_anon; |
34dc7c2f | 622 | static arc_state_t *arc_mru_ghost; |
34dc7c2f BB |
623 | static arc_state_t *arc_mfu_ghost; |
624 | static arc_state_t *arc_l2c_only; | |
625 | ||
c9c9c1e2 MM |
626 | arc_state_t *arc_mru; |
627 | arc_state_t *arc_mfu; | |
628 | ||
34dc7c2f BB |
629 | /* |
630 | * There are several ARC variables that are critical to export as kstats -- | |
631 | * but we don't want to have to grovel around in the kstat whenever we wish to | |
632 | * manipulate them. For these variables, we therefore define them to be in | |
633 | * terms of the statistic variable. This assures that we are not introducing | |
634 | * the possibility of inconsistency by having shadow copies of the variables, | |
635 | * while still allowing the code to be readable. | |
636 | */ | |
1834f2d8 BB |
637 | #define arc_tempreserve ARCSTAT(arcstat_tempreserve) |
638 | #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes) | |
23c0a133 | 639 | #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ |
03fdcb9a MM |
640 | /* max size for dnodes */ |
641 | #define arc_dnode_size_limit ARCSTAT(arcstat_dnode_limit) | |
ca0bf58d | 642 | #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ |
23c0a133 | 643 | #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ |
34dc7c2f | 644 | |
b5256303 TC |
645 | /* size of all b_rabd's in entire arc */ |
646 | #define arc_raw_size ARCSTAT(arcstat_raw_size) | |
d3c2ae1c GW |
647 | /* compressed size of entire arc */ |
648 | #define arc_compressed_size ARCSTAT(arcstat_compressed_size) | |
649 | /* uncompressed size of entire arc */ | |
650 | #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) | |
651 | /* number of bytes in the arc from arc_buf_t's */ | |
652 | #define arc_overhead_size ARCSTAT(arcstat_overhead_size) | |
3a17a7a9 | 653 | |
37fb3e43 PD |
654 | /* |
655 | * There are also some ARC variables that we want to export, but that are | |
656 | * updated so often that having the canonical representation be the statistic | |
657 | * variable causes a performance bottleneck. We want to use aggsum_t's for these | |
658 | * instead, but still be able to export the kstat in the same way as before. | |
659 | * The solution is to always use the aggsum version, except in the kstat update | |
660 | * callback. | |
661 | */ | |
662 | aggsum_t arc_size; | |
663 | aggsum_t arc_meta_used; | |
664 | aggsum_t astat_data_size; | |
665 | aggsum_t astat_metadata_size; | |
666 | aggsum_t astat_dbuf_size; | |
667 | aggsum_t astat_dnode_size; | |
668 | aggsum_t astat_bonus_size; | |
669 | aggsum_t astat_hdr_size; | |
670 | aggsum_t astat_l2_hdr_size; | |
671 | ||
c9c9c1e2 MM |
672 | hrtime_t arc_growtime; |
673 | list_t arc_prune_list; | |
674 | kmutex_t arc_prune_mtx; | |
675 | taskq_t *arc_prune_taskq; | |
428870ff | 676 | |
34dc7c2f BB |
677 | #define GHOST_STATE(state) \ |
678 | ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ | |
679 | (state) == arc_l2c_only) | |
680 | ||
2a432414 GW |
681 | #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) |
682 | #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) | |
683 | #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) | |
684 | #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) | |
d4a72f23 TC |
685 | #define HDR_PRESCIENT_PREFETCH(hdr) \ |
686 | ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) | |
d3c2ae1c GW |
687 | #define HDR_COMPRESSION_ENABLED(hdr) \ |
688 | ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) | |
b9541d6b | 689 | |
2a432414 GW |
690 | #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) |
691 | #define HDR_L2_READING(hdr) \ | |
d3c2ae1c GW |
692 | (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ |
693 | ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) | |
2a432414 GW |
694 | #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) |
695 | #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) | |
696 | #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) | |
b5256303 TC |
697 | #define HDR_PROTECTED(hdr) ((hdr)->b_flags & ARC_FLAG_PROTECTED) |
698 | #define HDR_NOAUTH(hdr) ((hdr)->b_flags & ARC_FLAG_NOAUTH) | |
d3c2ae1c | 699 | #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) |
34dc7c2f | 700 | |
b9541d6b | 701 | #define HDR_ISTYPE_METADATA(hdr) \ |
d3c2ae1c | 702 | ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) |
b9541d6b CW |
703 | #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) |
704 | ||
705 | #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) | |
706 | #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) | |
b5256303 TC |
707 | #define HDR_HAS_RABD(hdr) \ |
708 | (HDR_HAS_L1HDR(hdr) && HDR_PROTECTED(hdr) && \ | |
709 | (hdr)->b_crypt_hdr.b_rabd != NULL) | |
710 | #define HDR_ENCRYPTED(hdr) \ | |
711 | (HDR_PROTECTED(hdr) && DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot)) | |
712 | #define HDR_AUTHENTICATED(hdr) \ | |
713 | (HDR_PROTECTED(hdr) && !DMU_OT_IS_ENCRYPTED((hdr)->b_crypt_hdr.b_ot)) | |
b9541d6b | 714 | |
d3c2ae1c GW |
715 | /* For storing compression mode in b_flags */ |
716 | #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) | |
717 | ||
718 | #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ | |
719 | HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) | |
720 | #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ | |
721 | HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); | |
722 | ||
723 | #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) | |
524b4217 DK |
724 | #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) |
725 | #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) | |
b5256303 | 726 | #define ARC_BUF_ENCRYPTED(buf) ((buf)->b_flags & ARC_BUF_FLAG_ENCRYPTED) |
d3c2ae1c | 727 | |
34dc7c2f BB |
728 | /* |
729 | * Other sizes | |
730 | */ | |
731 | ||
b5256303 TC |
732 | #define HDR_FULL_CRYPT_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) |
733 | #define HDR_FULL_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_crypt_hdr)) | |
b9541d6b | 734 | #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) |
34dc7c2f BB |
735 | |
736 | /* | |
737 | * Hash table routines | |
738 | */ | |
739 | ||
00b46022 BB |
740 | #define HT_LOCK_ALIGN 64 |
741 | #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN))) | |
34dc7c2f BB |
742 | |
743 | struct ht_lock { | |
744 | kmutex_t ht_lock; | |
745 | #ifdef _KERNEL | |
00b46022 | 746 | unsigned char pad[HT_LOCK_PAD]; |
34dc7c2f BB |
747 | #endif |
748 | }; | |
749 | ||
b31d8ea7 | 750 | #define BUF_LOCKS 8192 |
34dc7c2f BB |
751 | typedef struct buf_hash_table { |
752 | uint64_t ht_mask; | |
753 | arc_buf_hdr_t **ht_table; | |
754 | struct ht_lock ht_locks[BUF_LOCKS]; | |
755 | } buf_hash_table_t; | |
756 | ||
757 | static buf_hash_table_t buf_hash_table; | |
758 | ||
759 | #define BUF_HASH_INDEX(spa, dva, birth) \ | |
760 | (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) | |
761 | #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) | |
762 | #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) | |
428870ff BB |
763 | #define HDR_LOCK(hdr) \ |
764 | (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) | |
34dc7c2f BB |
765 | |
766 | uint64_t zfs_crc64_table[256]; | |
767 | ||
768 | /* | |
769 | * Level 2 ARC | |
770 | */ | |
771 | ||
772 | #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ | |
3a17a7a9 | 773 | #define L2ARC_HEADROOM 2 /* num of writes */ |
8a09d5fd | 774 | |
3a17a7a9 SK |
775 | /* |
776 | * If we discover during ARC scan any buffers to be compressed, we boost | |
777 | * our headroom for the next scanning cycle by this percentage multiple. | |
778 | */ | |
779 | #define L2ARC_HEADROOM_BOOST 200 | |
d164b209 BB |
780 | #define L2ARC_FEED_SECS 1 /* caching interval secs */ |
781 | #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ | |
34dc7c2f | 782 | |
4aafab91 G |
783 | /* |
784 | * We can feed L2ARC from two states of ARC buffers, mru and mfu, | |
785 | * and each of the state has two types: data and metadata. | |
786 | */ | |
787 | #define L2ARC_FEED_TYPES 4 | |
788 | ||
34dc7c2f BB |
789 | #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) |
790 | #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) | |
791 | ||
d3cc8b15 | 792 | /* L2ARC Performance Tunables */ |
abd8610c BB |
793 | unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */ |
794 | unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */ | |
795 | unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */ | |
3a17a7a9 | 796 | unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; |
abd8610c BB |
797 | unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ |
798 | unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */ | |
799 | int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ | |
800 | int l2arc_feed_again = B_TRUE; /* turbo warmup */ | |
c93504f0 | 801 | int l2arc_norw = B_FALSE; /* no reads during writes */ |
34dc7c2f BB |
802 | |
803 | /* | |
804 | * L2ARC Internals | |
805 | */ | |
34dc7c2f BB |
806 | static list_t L2ARC_dev_list; /* device list */ |
807 | static list_t *l2arc_dev_list; /* device list pointer */ | |
808 | static kmutex_t l2arc_dev_mtx; /* device list mutex */ | |
809 | static l2arc_dev_t *l2arc_dev_last; /* last device used */ | |
34dc7c2f BB |
810 | static list_t L2ARC_free_on_write; /* free after write buf list */ |
811 | static list_t *l2arc_free_on_write; /* free after write list ptr */ | |
812 | static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ | |
813 | static uint64_t l2arc_ndev; /* number of devices */ | |
814 | ||
815 | typedef struct l2arc_read_callback { | |
2aa34383 | 816 | arc_buf_hdr_t *l2rcb_hdr; /* read header */ |
3a17a7a9 | 817 | blkptr_t l2rcb_bp; /* original blkptr */ |
5dbd68a3 | 818 | zbookmark_phys_t l2rcb_zb; /* original bookmark */ |
3a17a7a9 | 819 | int l2rcb_flags; /* original flags */ |
82710e99 | 820 | abd_t *l2rcb_abd; /* temporary buffer */ |
34dc7c2f BB |
821 | } l2arc_read_callback_t; |
822 | ||
34dc7c2f BB |
823 | typedef struct l2arc_data_free { |
824 | /* protected by l2arc_free_on_write_mtx */ | |
a6255b7f | 825 | abd_t *l2df_abd; |
34dc7c2f | 826 | size_t l2df_size; |
d3c2ae1c | 827 | arc_buf_contents_t l2df_type; |
34dc7c2f BB |
828 | list_node_t l2df_list_node; |
829 | } l2arc_data_free_t; | |
830 | ||
b5256303 TC |
831 | typedef enum arc_fill_flags { |
832 | ARC_FILL_LOCKED = 1 << 0, /* hdr lock is held */ | |
833 | ARC_FILL_COMPRESSED = 1 << 1, /* fill with compressed data */ | |
834 | ARC_FILL_ENCRYPTED = 1 << 2, /* fill with encrypted data */ | |
835 | ARC_FILL_NOAUTH = 1 << 3, /* don't attempt to authenticate */ | |
836 | ARC_FILL_IN_PLACE = 1 << 4 /* fill in place (special case) */ | |
837 | } arc_fill_flags_t; | |
838 | ||
34dc7c2f BB |
839 | static kmutex_t l2arc_feed_thr_lock; |
840 | static kcondvar_t l2arc_feed_thr_cv; | |
841 | static uint8_t l2arc_thread_exit; | |
842 | ||
77f6826b GA |
843 | static kmutex_t l2arc_rebuild_thr_lock; |
844 | static kcondvar_t l2arc_rebuild_thr_cv; | |
845 | ||
a6255b7f | 846 | static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *); |
d3c2ae1c | 847 | static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); |
a6255b7f DQ |
848 | static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *); |
849 | static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *); | |
d3c2ae1c | 850 | static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); |
a6255b7f | 851 | static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag); |
b5256303 TC |
852 | static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t); |
853 | static void arc_hdr_alloc_abd(arc_buf_hdr_t *, boolean_t); | |
2a432414 | 854 | static void arc_access(arc_buf_hdr_t *, kmutex_t *); |
ca0bf58d | 855 | static boolean_t arc_is_overflowing(void); |
2a432414 | 856 | static void arc_buf_watch(arc_buf_t *); |
77f6826b | 857 | static l2arc_dev_t *l2arc_vdev_get(vdev_t *vd); |
2a432414 | 858 | |
b9541d6b CW |
859 | static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); |
860 | static uint32_t arc_bufc_to_flags(arc_buf_contents_t); | |
d3c2ae1c GW |
861 | static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); |
862 | static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); | |
b9541d6b | 863 | |
2a432414 GW |
864 | static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); |
865 | static void l2arc_read_done(zio_t *); | |
34dc7c2f | 866 | |
77f6826b GA |
867 | /* |
868 | * Performance tuning of L2ARC persistence: | |
869 | * | |
870 | * l2arc_rebuild_enabled : A ZFS module parameter that controls whether adding | |
871 | * an L2ARC device (either at pool import or later) will attempt | |
872 | * to rebuild L2ARC buffer contents. | |
873 | * l2arc_rebuild_blocks_min_l2size : A ZFS module parameter that controls | |
874 | * whether log blocks are written to the L2ARC device. If the L2ARC | |
875 | * device is less than 1GB, the amount of data l2arc_evict() | |
876 | * evicts is significant compared to the amount of restored L2ARC | |
877 | * data. In this case do not write log blocks in L2ARC in order | |
878 | * not to waste space. | |
879 | */ | |
880 | int l2arc_rebuild_enabled = B_TRUE; | |
881 | unsigned long l2arc_rebuild_blocks_min_l2size = 1024 * 1024 * 1024; | |
882 | ||
883 | /* L2ARC persistence rebuild control routines. */ | |
884 | void l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen); | |
885 | static void l2arc_dev_rebuild_start(l2arc_dev_t *dev); | |
886 | static int l2arc_rebuild(l2arc_dev_t *dev); | |
887 | ||
888 | /* L2ARC persistence read I/O routines. */ | |
889 | static int l2arc_dev_hdr_read(l2arc_dev_t *dev); | |
890 | static int l2arc_log_blk_read(l2arc_dev_t *dev, | |
891 | const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp, | |
892 | l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb, | |
893 | zio_t *this_io, zio_t **next_io); | |
894 | static zio_t *l2arc_log_blk_fetch(vdev_t *vd, | |
895 | const l2arc_log_blkptr_t *lp, l2arc_log_blk_phys_t *lb); | |
896 | static void l2arc_log_blk_fetch_abort(zio_t *zio); | |
897 | ||
898 | /* L2ARC persistence block restoration routines. */ | |
899 | static void l2arc_log_blk_restore(l2arc_dev_t *dev, | |
657fd33b | 900 | const l2arc_log_blk_phys_t *lb, uint64_t lb_asize, uint64_t lb_daddr); |
77f6826b GA |
901 | static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, |
902 | l2arc_dev_t *dev); | |
903 | ||
904 | /* L2ARC persistence write I/O routines. */ | |
905 | static void l2arc_dev_hdr_update(l2arc_dev_t *dev); | |
906 | static void l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, | |
907 | l2arc_write_callback_t *cb); | |
908 | ||
909 | /* L2ARC persistence auxilliary routines. */ | |
910 | boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev, | |
911 | const l2arc_log_blkptr_t *lbp); | |
912 | static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev, | |
913 | const arc_buf_hdr_t *ab); | |
914 | boolean_t l2arc_range_check_overlap(uint64_t bottom, | |
915 | uint64_t top, uint64_t check); | |
916 | static void l2arc_blk_fetch_done(zio_t *zio); | |
917 | static inline uint64_t | |
918 | l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev); | |
37fb3e43 PD |
919 | |
920 | /* | |
921 | * We use Cityhash for this. It's fast, and has good hash properties without | |
922 | * requiring any large static buffers. | |
923 | */ | |
34dc7c2f | 924 | static uint64_t |
d164b209 | 925 | buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) |
34dc7c2f | 926 | { |
37fb3e43 | 927 | return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth)); |
34dc7c2f BB |
928 | } |
929 | ||
d3c2ae1c GW |
930 | #define HDR_EMPTY(hdr) \ |
931 | ((hdr)->b_dva.dva_word[0] == 0 && \ | |
932 | (hdr)->b_dva.dva_word[1] == 0) | |
34dc7c2f | 933 | |
ca6c7a94 BB |
934 | #define HDR_EMPTY_OR_LOCKED(hdr) \ |
935 | (HDR_EMPTY(hdr) || MUTEX_HELD(HDR_LOCK(hdr))) | |
936 | ||
d3c2ae1c GW |
937 | #define HDR_EQUAL(spa, dva, birth, hdr) \ |
938 | ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ | |
939 | ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ | |
940 | ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) | |
34dc7c2f | 941 | |
428870ff BB |
942 | static void |
943 | buf_discard_identity(arc_buf_hdr_t *hdr) | |
944 | { | |
945 | hdr->b_dva.dva_word[0] = 0; | |
946 | hdr->b_dva.dva_word[1] = 0; | |
947 | hdr->b_birth = 0; | |
428870ff BB |
948 | } |
949 | ||
34dc7c2f | 950 | static arc_buf_hdr_t * |
9b67f605 | 951 | buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) |
34dc7c2f | 952 | { |
9b67f605 MA |
953 | const dva_t *dva = BP_IDENTITY(bp); |
954 | uint64_t birth = BP_PHYSICAL_BIRTH(bp); | |
34dc7c2f BB |
955 | uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); |
956 | kmutex_t *hash_lock = BUF_HASH_LOCK(idx); | |
2a432414 | 957 | arc_buf_hdr_t *hdr; |
34dc7c2f BB |
958 | |
959 | mutex_enter(hash_lock); | |
2a432414 GW |
960 | for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; |
961 | hdr = hdr->b_hash_next) { | |
d3c2ae1c | 962 | if (HDR_EQUAL(spa, dva, birth, hdr)) { |
34dc7c2f | 963 | *lockp = hash_lock; |
2a432414 | 964 | return (hdr); |
34dc7c2f BB |
965 | } |
966 | } | |
967 | mutex_exit(hash_lock); | |
968 | *lockp = NULL; | |
969 | return (NULL); | |
970 | } | |
971 | ||
972 | /* | |
973 | * Insert an entry into the hash table. If there is already an element | |
974 | * equal to elem in the hash table, then the already existing element | |
975 | * will be returned and the new element will not be inserted. | |
976 | * Otherwise returns NULL. | |
b9541d6b | 977 | * If lockp == NULL, the caller is assumed to already hold the hash lock. |
34dc7c2f BB |
978 | */ |
979 | static arc_buf_hdr_t * | |
2a432414 | 980 | buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) |
34dc7c2f | 981 | { |
2a432414 | 982 | uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); |
34dc7c2f | 983 | kmutex_t *hash_lock = BUF_HASH_LOCK(idx); |
2a432414 | 984 | arc_buf_hdr_t *fhdr; |
34dc7c2f BB |
985 | uint32_t i; |
986 | ||
2a432414 GW |
987 | ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); |
988 | ASSERT(hdr->b_birth != 0); | |
989 | ASSERT(!HDR_IN_HASH_TABLE(hdr)); | |
b9541d6b CW |
990 | |
991 | if (lockp != NULL) { | |
992 | *lockp = hash_lock; | |
993 | mutex_enter(hash_lock); | |
994 | } else { | |
995 | ASSERT(MUTEX_HELD(hash_lock)); | |
996 | } | |
997 | ||
2a432414 GW |
998 | for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; |
999 | fhdr = fhdr->b_hash_next, i++) { | |
d3c2ae1c | 1000 | if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) |
2a432414 | 1001 | return (fhdr); |
34dc7c2f BB |
1002 | } |
1003 | ||
2a432414 GW |
1004 | hdr->b_hash_next = buf_hash_table.ht_table[idx]; |
1005 | buf_hash_table.ht_table[idx] = hdr; | |
d3c2ae1c | 1006 | arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); |
34dc7c2f BB |
1007 | |
1008 | /* collect some hash table performance data */ | |
1009 | if (i > 0) { | |
1010 | ARCSTAT_BUMP(arcstat_hash_collisions); | |
1011 | if (i == 1) | |
1012 | ARCSTAT_BUMP(arcstat_hash_chains); | |
1013 | ||
1014 | ARCSTAT_MAX(arcstat_hash_chain_max, i); | |
1015 | } | |
1016 | ||
1017 | ARCSTAT_BUMP(arcstat_hash_elements); | |
1018 | ARCSTAT_MAXSTAT(arcstat_hash_elements); | |
1019 | ||
1020 | return (NULL); | |
1021 | } | |
1022 | ||
1023 | static void | |
2a432414 | 1024 | buf_hash_remove(arc_buf_hdr_t *hdr) |
34dc7c2f | 1025 | { |
2a432414 GW |
1026 | arc_buf_hdr_t *fhdr, **hdrp; |
1027 | uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); | |
34dc7c2f BB |
1028 | |
1029 | ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); | |
2a432414 | 1030 | ASSERT(HDR_IN_HASH_TABLE(hdr)); |
34dc7c2f | 1031 | |
2a432414 GW |
1032 | hdrp = &buf_hash_table.ht_table[idx]; |
1033 | while ((fhdr = *hdrp) != hdr) { | |
d3c2ae1c | 1034 | ASSERT3P(fhdr, !=, NULL); |
2a432414 | 1035 | hdrp = &fhdr->b_hash_next; |
34dc7c2f | 1036 | } |
2a432414 GW |
1037 | *hdrp = hdr->b_hash_next; |
1038 | hdr->b_hash_next = NULL; | |
d3c2ae1c | 1039 | arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); |
34dc7c2f BB |
1040 | |
1041 | /* collect some hash table performance data */ | |
1042 | ARCSTAT_BUMPDOWN(arcstat_hash_elements); | |
1043 | ||
1044 | if (buf_hash_table.ht_table[idx] && | |
1045 | buf_hash_table.ht_table[idx]->b_hash_next == NULL) | |
1046 | ARCSTAT_BUMPDOWN(arcstat_hash_chains); | |
1047 | } | |
1048 | ||
1049 | /* | |
1050 | * Global data structures and functions for the buf kmem cache. | |
1051 | */ | |
b5256303 | 1052 | |
b9541d6b | 1053 | static kmem_cache_t *hdr_full_cache; |
b5256303 | 1054 | static kmem_cache_t *hdr_full_crypt_cache; |
b9541d6b | 1055 | static kmem_cache_t *hdr_l2only_cache; |
34dc7c2f BB |
1056 | static kmem_cache_t *buf_cache; |
1057 | ||
1058 | static void | |
1059 | buf_fini(void) | |
1060 | { | |
1061 | int i; | |
1062 | ||
93ce2b4c | 1063 | #if defined(_KERNEL) |
d1d7e268 MK |
1064 | /* |
1065 | * Large allocations which do not require contiguous pages | |
1066 | * should be using vmem_free() in the linux kernel\ | |
1067 | */ | |
00b46022 BB |
1068 | vmem_free(buf_hash_table.ht_table, |
1069 | (buf_hash_table.ht_mask + 1) * sizeof (void *)); | |
1070 | #else | |
34dc7c2f BB |
1071 | kmem_free(buf_hash_table.ht_table, |
1072 | (buf_hash_table.ht_mask + 1) * sizeof (void *)); | |
00b46022 | 1073 | #endif |
34dc7c2f BB |
1074 | for (i = 0; i < BUF_LOCKS; i++) |
1075 | mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); | |
b9541d6b | 1076 | kmem_cache_destroy(hdr_full_cache); |
b5256303 | 1077 | kmem_cache_destroy(hdr_full_crypt_cache); |
b9541d6b | 1078 | kmem_cache_destroy(hdr_l2only_cache); |
34dc7c2f BB |
1079 | kmem_cache_destroy(buf_cache); |
1080 | } | |
1081 | ||
1082 | /* | |
1083 | * Constructor callback - called when the cache is empty | |
1084 | * and a new buf is requested. | |
1085 | */ | |
1086 | /* ARGSUSED */ | |
1087 | static int | |
b9541d6b CW |
1088 | hdr_full_cons(void *vbuf, void *unused, int kmflag) |
1089 | { | |
1090 | arc_buf_hdr_t *hdr = vbuf; | |
1091 | ||
1092 | bzero(hdr, HDR_FULL_SIZE); | |
ae76f45c | 1093 | hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; |
b9541d6b | 1094 | cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); |
424fd7c3 | 1095 | zfs_refcount_create(&hdr->b_l1hdr.b_refcnt); |
b9541d6b CW |
1096 | mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); |
1097 | list_link_init(&hdr->b_l1hdr.b_arc_node); | |
1098 | list_link_init(&hdr->b_l2hdr.b_l2node); | |
ca0bf58d | 1099 | multilist_link_init(&hdr->b_l1hdr.b_arc_node); |
b9541d6b CW |
1100 | arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); |
1101 | ||
1102 | return (0); | |
1103 | } | |
1104 | ||
b5256303 TC |
1105 | /* ARGSUSED */ |
1106 | static int | |
1107 | hdr_full_crypt_cons(void *vbuf, void *unused, int kmflag) | |
1108 | { | |
1109 | arc_buf_hdr_t *hdr = vbuf; | |
1110 | ||
1111 | hdr_full_cons(vbuf, unused, kmflag); | |
1112 | bzero(&hdr->b_crypt_hdr, sizeof (hdr->b_crypt_hdr)); | |
1113 | arc_space_consume(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS); | |
1114 | ||
1115 | return (0); | |
1116 | } | |
1117 | ||
b9541d6b CW |
1118 | /* ARGSUSED */ |
1119 | static int | |
1120 | hdr_l2only_cons(void *vbuf, void *unused, int kmflag) | |
34dc7c2f | 1121 | { |
2a432414 GW |
1122 | arc_buf_hdr_t *hdr = vbuf; |
1123 | ||
b9541d6b CW |
1124 | bzero(hdr, HDR_L2ONLY_SIZE); |
1125 | arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); | |
34dc7c2f | 1126 | |
34dc7c2f BB |
1127 | return (0); |
1128 | } | |
1129 | ||
b128c09f BB |
1130 | /* ARGSUSED */ |
1131 | static int | |
1132 | buf_cons(void *vbuf, void *unused, int kmflag) | |
1133 | { | |
1134 | arc_buf_t *buf = vbuf; | |
1135 | ||
1136 | bzero(buf, sizeof (arc_buf_t)); | |
428870ff | 1137 | mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); |
d164b209 BB |
1138 | arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); |
1139 | ||
b128c09f BB |
1140 | return (0); |
1141 | } | |
1142 | ||
34dc7c2f BB |
1143 | /* |
1144 | * Destructor callback - called when a cached buf is | |
1145 | * no longer required. | |
1146 | */ | |
1147 | /* ARGSUSED */ | |
1148 | static void | |
b9541d6b | 1149 | hdr_full_dest(void *vbuf, void *unused) |
34dc7c2f | 1150 | { |
2a432414 | 1151 | arc_buf_hdr_t *hdr = vbuf; |
34dc7c2f | 1152 | |
d3c2ae1c | 1153 | ASSERT(HDR_EMPTY(hdr)); |
b9541d6b | 1154 | cv_destroy(&hdr->b_l1hdr.b_cv); |
424fd7c3 | 1155 | zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt); |
b9541d6b | 1156 | mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); |
ca0bf58d | 1157 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); |
b9541d6b CW |
1158 | arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); |
1159 | } | |
1160 | ||
b5256303 TC |
1161 | /* ARGSUSED */ |
1162 | static void | |
1163 | hdr_full_crypt_dest(void *vbuf, void *unused) | |
1164 | { | |
1165 | arc_buf_hdr_t *hdr = vbuf; | |
1166 | ||
1167 | hdr_full_dest(vbuf, unused); | |
1168 | arc_space_return(sizeof (hdr->b_crypt_hdr), ARC_SPACE_HDRS); | |
1169 | } | |
1170 | ||
b9541d6b CW |
1171 | /* ARGSUSED */ |
1172 | static void | |
1173 | hdr_l2only_dest(void *vbuf, void *unused) | |
1174 | { | |
2a8ba608 | 1175 | arc_buf_hdr_t *hdr __maybe_unused = vbuf; |
b9541d6b | 1176 | |
d3c2ae1c | 1177 | ASSERT(HDR_EMPTY(hdr)); |
b9541d6b | 1178 | arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); |
34dc7c2f BB |
1179 | } |
1180 | ||
b128c09f BB |
1181 | /* ARGSUSED */ |
1182 | static void | |
1183 | buf_dest(void *vbuf, void *unused) | |
1184 | { | |
1185 | arc_buf_t *buf = vbuf; | |
1186 | ||
428870ff | 1187 | mutex_destroy(&buf->b_evict_lock); |
d164b209 | 1188 | arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); |
b128c09f BB |
1189 | } |
1190 | ||
8c8af9d8 BB |
1191 | /* |
1192 | * Reclaim callback -- invoked when memory is low. | |
1193 | */ | |
1194 | /* ARGSUSED */ | |
1195 | static void | |
1196 | hdr_recl(void *unused) | |
1197 | { | |
1198 | dprintf("hdr_recl called\n"); | |
1199 | /* | |
1200 | * umem calls the reclaim func when we destroy the buf cache, | |
1201 | * which is after we do arc_fini(). | |
1202 | */ | |
3ec34e55 BL |
1203 | if (arc_initialized) |
1204 | zthr_wakeup(arc_reap_zthr); | |
8c8af9d8 BB |
1205 | } |
1206 | ||
34dc7c2f BB |
1207 | static void |
1208 | buf_init(void) | |
1209 | { | |
2db28197 | 1210 | uint64_t *ct = NULL; |
34dc7c2f BB |
1211 | uint64_t hsize = 1ULL << 12; |
1212 | int i, j; | |
1213 | ||
1214 | /* | |
1215 | * The hash table is big enough to fill all of physical memory | |
49ddb315 MA |
1216 | * with an average block size of zfs_arc_average_blocksize (default 8K). |
1217 | * By default, the table will take up | |
1218 | * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). | |
34dc7c2f | 1219 | */ |
9edb3695 | 1220 | while (hsize * zfs_arc_average_blocksize < arc_all_memory()) |
34dc7c2f BB |
1221 | hsize <<= 1; |
1222 | retry: | |
1223 | buf_hash_table.ht_mask = hsize - 1; | |
93ce2b4c | 1224 | #if defined(_KERNEL) |
d1d7e268 MK |
1225 | /* |
1226 | * Large allocations which do not require contiguous pages | |
1227 | * should be using vmem_alloc() in the linux kernel | |
1228 | */ | |
00b46022 BB |
1229 | buf_hash_table.ht_table = |
1230 | vmem_zalloc(hsize * sizeof (void*), KM_SLEEP); | |
1231 | #else | |
34dc7c2f BB |
1232 | buf_hash_table.ht_table = |
1233 | kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); | |
00b46022 | 1234 | #endif |
34dc7c2f BB |
1235 | if (buf_hash_table.ht_table == NULL) { |
1236 | ASSERT(hsize > (1ULL << 8)); | |
1237 | hsize >>= 1; | |
1238 | goto retry; | |
1239 | } | |
1240 | ||
b9541d6b | 1241 | hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, |
8c8af9d8 | 1242 | 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); |
b5256303 TC |
1243 | hdr_full_crypt_cache = kmem_cache_create("arc_buf_hdr_t_full_crypt", |
1244 | HDR_FULL_CRYPT_SIZE, 0, hdr_full_crypt_cons, hdr_full_crypt_dest, | |
1245 | hdr_recl, NULL, NULL, 0); | |
b9541d6b | 1246 | hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", |
8c8af9d8 | 1247 | HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, |
b9541d6b | 1248 | NULL, NULL, 0); |
34dc7c2f | 1249 | buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), |
b128c09f | 1250 | 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); |
34dc7c2f BB |
1251 | |
1252 | for (i = 0; i < 256; i++) | |
1253 | for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) | |
1254 | *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); | |
1255 | ||
1256 | for (i = 0; i < BUF_LOCKS; i++) { | |
1257 | mutex_init(&buf_hash_table.ht_locks[i].ht_lock, | |
40d06e3c | 1258 | NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f BB |
1259 | } |
1260 | } | |
1261 | ||
d3c2ae1c | 1262 | #define ARC_MINTIME (hz>>4) /* 62 ms */ |
ca0bf58d | 1263 | |
2aa34383 DK |
1264 | /* |
1265 | * This is the size that the buf occupies in memory. If the buf is compressed, | |
1266 | * it will correspond to the compressed size. You should use this method of | |
1267 | * getting the buf size unless you explicitly need the logical size. | |
1268 | */ | |
1269 | uint64_t | |
1270 | arc_buf_size(arc_buf_t *buf) | |
1271 | { | |
1272 | return (ARC_BUF_COMPRESSED(buf) ? | |
1273 | HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); | |
1274 | } | |
1275 | ||
1276 | uint64_t | |
1277 | arc_buf_lsize(arc_buf_t *buf) | |
1278 | { | |
1279 | return (HDR_GET_LSIZE(buf->b_hdr)); | |
1280 | } | |
1281 | ||
b5256303 TC |
1282 | /* |
1283 | * This function will return B_TRUE if the buffer is encrypted in memory. | |
1284 | * This buffer can be decrypted by calling arc_untransform(). | |
1285 | */ | |
1286 | boolean_t | |
1287 | arc_is_encrypted(arc_buf_t *buf) | |
1288 | { | |
1289 | return (ARC_BUF_ENCRYPTED(buf) != 0); | |
1290 | } | |
1291 | ||
1292 | /* | |
1293 | * Returns B_TRUE if the buffer represents data that has not had its MAC | |
1294 | * verified yet. | |
1295 | */ | |
1296 | boolean_t | |
1297 | arc_is_unauthenticated(arc_buf_t *buf) | |
1298 | { | |
1299 | return (HDR_NOAUTH(buf->b_hdr) != 0); | |
1300 | } | |
1301 | ||
1302 | void | |
1303 | arc_get_raw_params(arc_buf_t *buf, boolean_t *byteorder, uint8_t *salt, | |
1304 | uint8_t *iv, uint8_t *mac) | |
1305 | { | |
1306 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
1307 | ||
1308 | ASSERT(HDR_PROTECTED(hdr)); | |
1309 | ||
1310 | bcopy(hdr->b_crypt_hdr.b_salt, salt, ZIO_DATA_SALT_LEN); | |
1311 | bcopy(hdr->b_crypt_hdr.b_iv, iv, ZIO_DATA_IV_LEN); | |
1312 | bcopy(hdr->b_crypt_hdr.b_mac, mac, ZIO_DATA_MAC_LEN); | |
1313 | *byteorder = (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ? | |
1314 | ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER; | |
1315 | } | |
1316 | ||
1317 | /* | |
1318 | * Indicates how this buffer is compressed in memory. If it is not compressed | |
1319 | * the value will be ZIO_COMPRESS_OFF. It can be made normally readable with | |
1320 | * arc_untransform() as long as it is also unencrypted. | |
1321 | */ | |
2aa34383 DK |
1322 | enum zio_compress |
1323 | arc_get_compression(arc_buf_t *buf) | |
1324 | { | |
1325 | return (ARC_BUF_COMPRESSED(buf) ? | |
1326 | HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); | |
1327 | } | |
1328 | ||
b5256303 TC |
1329 | /* |
1330 | * Return the compression algorithm used to store this data in the ARC. If ARC | |
1331 | * compression is enabled or this is an encrypted block, this will be the same | |
1332 | * as what's used to store it on-disk. Otherwise, this will be ZIO_COMPRESS_OFF. | |
1333 | */ | |
1334 | static inline enum zio_compress | |
1335 | arc_hdr_get_compress(arc_buf_hdr_t *hdr) | |
1336 | { | |
1337 | return (HDR_COMPRESSION_ENABLED(hdr) ? | |
1338 | HDR_GET_COMPRESS(hdr) : ZIO_COMPRESS_OFF); | |
1339 | } | |
1340 | ||
d3c2ae1c GW |
1341 | static inline boolean_t |
1342 | arc_buf_is_shared(arc_buf_t *buf) | |
1343 | { | |
1344 | boolean_t shared = (buf->b_data != NULL && | |
a6255b7f DQ |
1345 | buf->b_hdr->b_l1hdr.b_pabd != NULL && |
1346 | abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) && | |
1347 | buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd)); | |
d3c2ae1c | 1348 | IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); |
2aa34383 DK |
1349 | IMPLY(shared, ARC_BUF_SHARED(buf)); |
1350 | IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); | |
524b4217 DK |
1351 | |
1352 | /* | |
1353 | * It would be nice to assert arc_can_share() too, but the "hdr isn't | |
1354 | * already being shared" requirement prevents us from doing that. | |
1355 | */ | |
1356 | ||
d3c2ae1c GW |
1357 | return (shared); |
1358 | } | |
ca0bf58d | 1359 | |
a7004725 DK |
1360 | /* |
1361 | * Free the checksum associated with this header. If there is no checksum, this | |
1362 | * is a no-op. | |
1363 | */ | |
d3c2ae1c GW |
1364 | static inline void |
1365 | arc_cksum_free(arc_buf_hdr_t *hdr) | |
1366 | { | |
1367 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
b5256303 | 1368 | |
d3c2ae1c GW |
1369 | mutex_enter(&hdr->b_l1hdr.b_freeze_lock); |
1370 | if (hdr->b_l1hdr.b_freeze_cksum != NULL) { | |
1371 | kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); | |
1372 | hdr->b_l1hdr.b_freeze_cksum = NULL; | |
b9541d6b | 1373 | } |
d3c2ae1c | 1374 | mutex_exit(&hdr->b_l1hdr.b_freeze_lock); |
b9541d6b CW |
1375 | } |
1376 | ||
a7004725 DK |
1377 | /* |
1378 | * Return true iff at least one of the bufs on hdr is not compressed. | |
b5256303 | 1379 | * Encrypted buffers count as compressed. |
a7004725 DK |
1380 | */ |
1381 | static boolean_t | |
1382 | arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) | |
1383 | { | |
ca6c7a94 | 1384 | ASSERT(hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY_OR_LOCKED(hdr)); |
149ce888 | 1385 | |
a7004725 DK |
1386 | for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { |
1387 | if (!ARC_BUF_COMPRESSED(b)) { | |
1388 | return (B_TRUE); | |
1389 | } | |
1390 | } | |
1391 | return (B_FALSE); | |
1392 | } | |
1393 | ||
1394 | ||
524b4217 DK |
1395 | /* |
1396 | * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data | |
1397 | * matches the checksum that is stored in the hdr. If there is no checksum, | |
1398 | * or if the buf is compressed, this is a no-op. | |
1399 | */ | |
34dc7c2f BB |
1400 | static void |
1401 | arc_cksum_verify(arc_buf_t *buf) | |
1402 | { | |
d3c2ae1c | 1403 | arc_buf_hdr_t *hdr = buf->b_hdr; |
34dc7c2f BB |
1404 | zio_cksum_t zc; |
1405 | ||
1406 | if (!(zfs_flags & ZFS_DEBUG_MODIFY)) | |
1407 | return; | |
1408 | ||
149ce888 | 1409 | if (ARC_BUF_COMPRESSED(buf)) |
524b4217 | 1410 | return; |
524b4217 | 1411 | |
d3c2ae1c GW |
1412 | ASSERT(HDR_HAS_L1HDR(hdr)); |
1413 | ||
1414 | mutex_enter(&hdr->b_l1hdr.b_freeze_lock); | |
149ce888 | 1415 | |
d3c2ae1c GW |
1416 | if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { |
1417 | mutex_exit(&hdr->b_l1hdr.b_freeze_lock); | |
34dc7c2f BB |
1418 | return; |
1419 | } | |
2aa34383 | 1420 | |
3c67d83a | 1421 | fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); |
d3c2ae1c | 1422 | if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) |
34dc7c2f | 1423 | panic("buffer modified while frozen!"); |
d3c2ae1c | 1424 | mutex_exit(&hdr->b_l1hdr.b_freeze_lock); |
34dc7c2f BB |
1425 | } |
1426 | ||
b5256303 TC |
1427 | /* |
1428 | * This function makes the assumption that data stored in the L2ARC | |
1429 | * will be transformed exactly as it is in the main pool. Because of | |
1430 | * this we can verify the checksum against the reading process's bp. | |
1431 | */ | |
d3c2ae1c GW |
1432 | static boolean_t |
1433 | arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) | |
34dc7c2f | 1434 | { |
d3c2ae1c GW |
1435 | ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); |
1436 | VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); | |
34dc7c2f | 1437 | |
d3c2ae1c GW |
1438 | /* |
1439 | * Block pointers always store the checksum for the logical data. | |
1440 | * If the block pointer has the gang bit set, then the checksum | |
1441 | * it represents is for the reconstituted data and not for an | |
1442 | * individual gang member. The zio pipeline, however, must be able to | |
1443 | * determine the checksum of each of the gang constituents so it | |
1444 | * treats the checksum comparison differently than what we need | |
1445 | * for l2arc blocks. This prevents us from using the | |
1446 | * zio_checksum_error() interface directly. Instead we must call the | |
1447 | * zio_checksum_error_impl() so that we can ensure the checksum is | |
1448 | * generated using the correct checksum algorithm and accounts for the | |
1449 | * logical I/O size and not just a gang fragment. | |
1450 | */ | |
b5256303 | 1451 | return (zio_checksum_error_impl(zio->io_spa, zio->io_bp, |
a6255b7f | 1452 | BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size, |
d3c2ae1c | 1453 | zio->io_offset, NULL) == 0); |
34dc7c2f BB |
1454 | } |
1455 | ||
524b4217 DK |
1456 | /* |
1457 | * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a | |
1458 | * checksum and attaches it to the buf's hdr so that we can ensure that the buf | |
1459 | * isn't modified later on. If buf is compressed or there is already a checksum | |
1460 | * on the hdr, this is a no-op (we only checksum uncompressed bufs). | |
1461 | */ | |
34dc7c2f | 1462 | static void |
d3c2ae1c | 1463 | arc_cksum_compute(arc_buf_t *buf) |
34dc7c2f | 1464 | { |
d3c2ae1c GW |
1465 | arc_buf_hdr_t *hdr = buf->b_hdr; |
1466 | ||
1467 | if (!(zfs_flags & ZFS_DEBUG_MODIFY)) | |
34dc7c2f BB |
1468 | return; |
1469 | ||
d3c2ae1c | 1470 | ASSERT(HDR_HAS_L1HDR(hdr)); |
2aa34383 | 1471 | |
b9541d6b | 1472 | mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); |
149ce888 | 1473 | if (hdr->b_l1hdr.b_freeze_cksum != NULL || ARC_BUF_COMPRESSED(buf)) { |
d3c2ae1c | 1474 | mutex_exit(&hdr->b_l1hdr.b_freeze_lock); |
34dc7c2f BB |
1475 | return; |
1476 | } | |
2aa34383 | 1477 | |
b5256303 | 1478 | ASSERT(!ARC_BUF_ENCRYPTED(buf)); |
2aa34383 | 1479 | ASSERT(!ARC_BUF_COMPRESSED(buf)); |
d3c2ae1c GW |
1480 | hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), |
1481 | KM_SLEEP); | |
3c67d83a | 1482 | fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, |
d3c2ae1c GW |
1483 | hdr->b_l1hdr.b_freeze_cksum); |
1484 | mutex_exit(&hdr->b_l1hdr.b_freeze_lock); | |
498877ba MA |
1485 | arc_buf_watch(buf); |
1486 | } | |
1487 | ||
1488 | #ifndef _KERNEL | |
1489 | void | |
1490 | arc_buf_sigsegv(int sig, siginfo_t *si, void *unused) | |
1491 | { | |
02730c33 | 1492 | panic("Got SIGSEGV at address: 0x%lx\n", (long)si->si_addr); |
498877ba MA |
1493 | } |
1494 | #endif | |
1495 | ||
1496 | /* ARGSUSED */ | |
1497 | static void | |
1498 | arc_buf_unwatch(arc_buf_t *buf) | |
1499 | { | |
1500 | #ifndef _KERNEL | |
1501 | if (arc_watch) { | |
a7004725 | 1502 | ASSERT0(mprotect(buf->b_data, arc_buf_size(buf), |
498877ba MA |
1503 | PROT_READ | PROT_WRITE)); |
1504 | } | |
1505 | #endif | |
1506 | } | |
1507 | ||
1508 | /* ARGSUSED */ | |
1509 | static void | |
1510 | arc_buf_watch(arc_buf_t *buf) | |
1511 | { | |
1512 | #ifndef _KERNEL | |
1513 | if (arc_watch) | |
2aa34383 | 1514 | ASSERT0(mprotect(buf->b_data, arc_buf_size(buf), |
d3c2ae1c | 1515 | PROT_READ)); |
498877ba | 1516 | #endif |
34dc7c2f BB |
1517 | } |
1518 | ||
b9541d6b CW |
1519 | static arc_buf_contents_t |
1520 | arc_buf_type(arc_buf_hdr_t *hdr) | |
1521 | { | |
d3c2ae1c | 1522 | arc_buf_contents_t type; |
b9541d6b | 1523 | if (HDR_ISTYPE_METADATA(hdr)) { |
d3c2ae1c | 1524 | type = ARC_BUFC_METADATA; |
b9541d6b | 1525 | } else { |
d3c2ae1c | 1526 | type = ARC_BUFC_DATA; |
b9541d6b | 1527 | } |
d3c2ae1c GW |
1528 | VERIFY3U(hdr->b_type, ==, type); |
1529 | return (type); | |
b9541d6b CW |
1530 | } |
1531 | ||
2aa34383 DK |
1532 | boolean_t |
1533 | arc_is_metadata(arc_buf_t *buf) | |
1534 | { | |
1535 | return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); | |
1536 | } | |
1537 | ||
b9541d6b CW |
1538 | static uint32_t |
1539 | arc_bufc_to_flags(arc_buf_contents_t type) | |
1540 | { | |
1541 | switch (type) { | |
1542 | case ARC_BUFC_DATA: | |
1543 | /* metadata field is 0 if buffer contains normal data */ | |
1544 | return (0); | |
1545 | case ARC_BUFC_METADATA: | |
1546 | return (ARC_FLAG_BUFC_METADATA); | |
1547 | default: | |
1548 | break; | |
1549 | } | |
1550 | panic("undefined ARC buffer type!"); | |
1551 | return ((uint32_t)-1); | |
1552 | } | |
1553 | ||
34dc7c2f BB |
1554 | void |
1555 | arc_buf_thaw(arc_buf_t *buf) | |
1556 | { | |
d3c2ae1c GW |
1557 | arc_buf_hdr_t *hdr = buf->b_hdr; |
1558 | ||
2aa34383 DK |
1559 | ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); |
1560 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
1561 | ||
524b4217 | 1562 | arc_cksum_verify(buf); |
34dc7c2f | 1563 | |
2aa34383 | 1564 | /* |
149ce888 | 1565 | * Compressed buffers do not manipulate the b_freeze_cksum. |
2aa34383 | 1566 | */ |
149ce888 | 1567 | if (ARC_BUF_COMPRESSED(buf)) |
2aa34383 | 1568 | return; |
2aa34383 | 1569 | |
d3c2ae1c GW |
1570 | ASSERT(HDR_HAS_L1HDR(hdr)); |
1571 | arc_cksum_free(hdr); | |
498877ba | 1572 | arc_buf_unwatch(buf); |
34dc7c2f BB |
1573 | } |
1574 | ||
1575 | void | |
1576 | arc_buf_freeze(arc_buf_t *buf) | |
1577 | { | |
1578 | if (!(zfs_flags & ZFS_DEBUG_MODIFY)) | |
1579 | return; | |
1580 | ||
149ce888 | 1581 | if (ARC_BUF_COMPRESSED(buf)) |
2aa34383 | 1582 | return; |
428870ff | 1583 | |
149ce888 | 1584 | ASSERT(HDR_HAS_L1HDR(buf->b_hdr)); |
d3c2ae1c | 1585 | arc_cksum_compute(buf); |
34dc7c2f BB |
1586 | } |
1587 | ||
d3c2ae1c GW |
1588 | /* |
1589 | * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, | |
1590 | * the following functions should be used to ensure that the flags are | |
1591 | * updated in a thread-safe way. When manipulating the flags either | |
1592 | * the hash_lock must be held or the hdr must be undiscoverable. This | |
1593 | * ensures that we're not racing with any other threads when updating | |
1594 | * the flags. | |
1595 | */ | |
1596 | static inline void | |
1597 | arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) | |
1598 | { | |
ca6c7a94 | 1599 | ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); |
d3c2ae1c GW |
1600 | hdr->b_flags |= flags; |
1601 | } | |
1602 | ||
1603 | static inline void | |
1604 | arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) | |
1605 | { | |
ca6c7a94 | 1606 | ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); |
d3c2ae1c GW |
1607 | hdr->b_flags &= ~flags; |
1608 | } | |
1609 | ||
1610 | /* | |
1611 | * Setting the compression bits in the arc_buf_hdr_t's b_flags is | |
1612 | * done in a special way since we have to clear and set bits | |
1613 | * at the same time. Consumers that wish to set the compression bits | |
1614 | * must use this function to ensure that the flags are updated in | |
1615 | * thread-safe manner. | |
1616 | */ | |
1617 | static void | |
1618 | arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) | |
1619 | { | |
ca6c7a94 | 1620 | ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); |
d3c2ae1c GW |
1621 | |
1622 | /* | |
1623 | * Holes and embedded blocks will always have a psize = 0 so | |
1624 | * we ignore the compression of the blkptr and set the | |
d3c2ae1c GW |
1625 | * want to uncompress them. Mark them as uncompressed. |
1626 | */ | |
1627 | if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { | |
1628 | arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); | |
d3c2ae1c | 1629 | ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); |
d3c2ae1c GW |
1630 | } else { |
1631 | arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); | |
d3c2ae1c GW |
1632 | ASSERT(HDR_COMPRESSION_ENABLED(hdr)); |
1633 | } | |
b5256303 TC |
1634 | |
1635 | HDR_SET_COMPRESS(hdr, cmp); | |
1636 | ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); | |
d3c2ae1c GW |
1637 | } |
1638 | ||
524b4217 DK |
1639 | /* |
1640 | * Looks for another buf on the same hdr which has the data decompressed, copies | |
1641 | * from it, and returns true. If no such buf exists, returns false. | |
1642 | */ | |
1643 | static boolean_t | |
1644 | arc_buf_try_copy_decompressed_data(arc_buf_t *buf) | |
1645 | { | |
1646 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
524b4217 DK |
1647 | boolean_t copied = B_FALSE; |
1648 | ||
1649 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
1650 | ASSERT3P(buf->b_data, !=, NULL); | |
1651 | ASSERT(!ARC_BUF_COMPRESSED(buf)); | |
1652 | ||
a7004725 | 1653 | for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; |
524b4217 DK |
1654 | from = from->b_next) { |
1655 | /* can't use our own data buffer */ | |
1656 | if (from == buf) { | |
1657 | continue; | |
1658 | } | |
1659 | ||
1660 | if (!ARC_BUF_COMPRESSED(from)) { | |
1661 | bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); | |
1662 | copied = B_TRUE; | |
1663 | break; | |
1664 | } | |
1665 | } | |
1666 | ||
1667 | /* | |
1668 | * There were no decompressed bufs, so there should not be a | |
1669 | * checksum on the hdr either. | |
1670 | */ | |
46db9d61 BB |
1671 | if (zfs_flags & ZFS_DEBUG_MODIFY) |
1672 | EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); | |
524b4217 DK |
1673 | |
1674 | return (copied); | |
1675 | } | |
1676 | ||
77f6826b GA |
1677 | /* |
1678 | * Allocates an ARC buf header that's in an evicted & L2-cached state. | |
1679 | * This is used during l2arc reconstruction to make empty ARC buffers | |
1680 | * which circumvent the regular disk->arc->l2arc path and instead come | |
1681 | * into being in the reverse order, i.e. l2arc->arc. | |
1682 | */ | |
1683 | arc_buf_hdr_t * | |
1684 | arc_buf_alloc_l2only(size_t size, arc_buf_contents_t type, l2arc_dev_t *dev, | |
1685 | dva_t dva, uint64_t daddr, int32_t psize, uint64_t birth, | |
1686 | enum zio_compress compress, boolean_t protected, boolean_t prefetch) | |
1687 | { | |
1688 | arc_buf_hdr_t *hdr; | |
1689 | ||
1690 | ASSERT(size != 0); | |
1691 | hdr = kmem_cache_alloc(hdr_l2only_cache, KM_SLEEP); | |
1692 | hdr->b_birth = birth; | |
1693 | hdr->b_type = type; | |
1694 | hdr->b_flags = 0; | |
1695 | arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR); | |
1696 | HDR_SET_LSIZE(hdr, size); | |
1697 | HDR_SET_PSIZE(hdr, psize); | |
1698 | arc_hdr_set_compress(hdr, compress); | |
1699 | if (protected) | |
1700 | arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED); | |
1701 | if (prefetch) | |
1702 | arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); | |
1703 | hdr->b_spa = spa_load_guid(dev->l2ad_vdev->vdev_spa); | |
1704 | ||
1705 | hdr->b_dva = dva; | |
1706 | ||
1707 | hdr->b_l2hdr.b_dev = dev; | |
1708 | hdr->b_l2hdr.b_daddr = daddr; | |
1709 | ||
1710 | return (hdr); | |
1711 | } | |
1712 | ||
b5256303 TC |
1713 | /* |
1714 | * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t. | |
1715 | */ | |
1716 | static uint64_t | |
1717 | arc_hdr_size(arc_buf_hdr_t *hdr) | |
1718 | { | |
1719 | uint64_t size; | |
1720 | ||
1721 | if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF && | |
1722 | HDR_GET_PSIZE(hdr) > 0) { | |
1723 | size = HDR_GET_PSIZE(hdr); | |
1724 | } else { | |
1725 | ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); | |
1726 | size = HDR_GET_LSIZE(hdr); | |
1727 | } | |
1728 | return (size); | |
1729 | } | |
1730 | ||
1731 | static int | |
1732 | arc_hdr_authenticate(arc_buf_hdr_t *hdr, spa_t *spa, uint64_t dsobj) | |
1733 | { | |
1734 | int ret; | |
1735 | uint64_t csize; | |
1736 | uint64_t lsize = HDR_GET_LSIZE(hdr); | |
1737 | uint64_t psize = HDR_GET_PSIZE(hdr); | |
1738 | void *tmpbuf = NULL; | |
1739 | abd_t *abd = hdr->b_l1hdr.b_pabd; | |
1740 | ||
ca6c7a94 | 1741 | ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); |
b5256303 TC |
1742 | ASSERT(HDR_AUTHENTICATED(hdr)); |
1743 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); | |
1744 | ||
1745 | /* | |
1746 | * The MAC is calculated on the compressed data that is stored on disk. | |
1747 | * However, if compressed arc is disabled we will only have the | |
1748 | * decompressed data available to us now. Compress it into a temporary | |
1749 | * abd so we can verify the MAC. The performance overhead of this will | |
1750 | * be relatively low, since most objects in an encrypted objset will | |
1751 | * be encrypted (instead of authenticated) anyway. | |
1752 | */ | |
1753 | if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && | |
1754 | !HDR_COMPRESSION_ENABLED(hdr)) { | |
1755 | tmpbuf = zio_buf_alloc(lsize); | |
1756 | abd = abd_get_from_buf(tmpbuf, lsize); | |
1757 | abd_take_ownership_of_buf(abd, B_TRUE); | |
1758 | ||
1759 | csize = zio_compress_data(HDR_GET_COMPRESS(hdr), | |
1760 | hdr->b_l1hdr.b_pabd, tmpbuf, lsize); | |
1761 | ASSERT3U(csize, <=, psize); | |
1762 | abd_zero_off(abd, csize, psize - csize); | |
1763 | } | |
1764 | ||
1765 | /* | |
1766 | * Authentication is best effort. We authenticate whenever the key is | |
1767 | * available. If we succeed we clear ARC_FLAG_NOAUTH. | |
1768 | */ | |
1769 | if (hdr->b_crypt_hdr.b_ot == DMU_OT_OBJSET) { | |
1770 | ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); | |
1771 | ASSERT3U(lsize, ==, psize); | |
1772 | ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, dsobj, abd, | |
1773 | psize, hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS); | |
1774 | } else { | |
1775 | ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, abd, psize, | |
1776 | hdr->b_crypt_hdr.b_mac); | |
1777 | } | |
1778 | ||
1779 | if (ret == 0) | |
1780 | arc_hdr_clear_flags(hdr, ARC_FLAG_NOAUTH); | |
1781 | else if (ret != ENOENT) | |
1782 | goto error; | |
1783 | ||
1784 | if (tmpbuf != NULL) | |
1785 | abd_free(abd); | |
1786 | ||
1787 | return (0); | |
1788 | ||
1789 | error: | |
1790 | if (tmpbuf != NULL) | |
1791 | abd_free(abd); | |
1792 | ||
1793 | return (ret); | |
1794 | } | |
1795 | ||
1796 | /* | |
1797 | * This function will take a header that only has raw encrypted data in | |
1798 | * b_crypt_hdr.b_rabd and decrypt it into a new buffer which is stored in | |
1799 | * b_l1hdr.b_pabd. If designated in the header flags, this function will | |
1800 | * also decompress the data. | |
1801 | */ | |
1802 | static int | |
be9a5c35 | 1803 | arc_hdr_decrypt(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb) |
b5256303 TC |
1804 | { |
1805 | int ret; | |
b5256303 TC |
1806 | abd_t *cabd = NULL; |
1807 | void *tmp = NULL; | |
1808 | boolean_t no_crypt = B_FALSE; | |
1809 | boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS); | |
1810 | ||
ca6c7a94 | 1811 | ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); |
b5256303 TC |
1812 | ASSERT(HDR_ENCRYPTED(hdr)); |
1813 | ||
1814 | arc_hdr_alloc_abd(hdr, B_FALSE); | |
1815 | ||
be9a5c35 TC |
1816 | ret = spa_do_crypt_abd(B_FALSE, spa, zb, hdr->b_crypt_hdr.b_ot, |
1817 | B_FALSE, bswap, hdr->b_crypt_hdr.b_salt, hdr->b_crypt_hdr.b_iv, | |
1818 | hdr->b_crypt_hdr.b_mac, HDR_GET_PSIZE(hdr), hdr->b_l1hdr.b_pabd, | |
b5256303 TC |
1819 | hdr->b_crypt_hdr.b_rabd, &no_crypt); |
1820 | if (ret != 0) | |
1821 | goto error; | |
1822 | ||
1823 | if (no_crypt) { | |
1824 | abd_copy(hdr->b_l1hdr.b_pabd, hdr->b_crypt_hdr.b_rabd, | |
1825 | HDR_GET_PSIZE(hdr)); | |
1826 | } | |
1827 | ||
1828 | /* | |
1829 | * If this header has disabled arc compression but the b_pabd is | |
1830 | * compressed after decrypting it, we need to decompress the newly | |
1831 | * decrypted data. | |
1832 | */ | |
1833 | if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && | |
1834 | !HDR_COMPRESSION_ENABLED(hdr)) { | |
1835 | /* | |
1836 | * We want to make sure that we are correctly honoring the | |
1837 | * zfs_abd_scatter_enabled setting, so we allocate an abd here | |
1838 | * and then loan a buffer from it, rather than allocating a | |
1839 | * linear buffer and wrapping it in an abd later. | |
1840 | */ | |
1841 | cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); | |
1842 | tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr)); | |
1843 | ||
1844 | ret = zio_decompress_data(HDR_GET_COMPRESS(hdr), | |
1845 | hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr), | |
1846 | HDR_GET_LSIZE(hdr)); | |
1847 | if (ret != 0) { | |
1848 | abd_return_buf(cabd, tmp, arc_hdr_size(hdr)); | |
1849 | goto error; | |
1850 | } | |
1851 | ||
1852 | abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr)); | |
1853 | arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, | |
1854 | arc_hdr_size(hdr), hdr); | |
1855 | hdr->b_l1hdr.b_pabd = cabd; | |
1856 | } | |
1857 | ||
b5256303 TC |
1858 | return (0); |
1859 | ||
1860 | error: | |
1861 | arc_hdr_free_abd(hdr, B_FALSE); | |
b5256303 TC |
1862 | if (cabd != NULL) |
1863 | arc_free_data_buf(hdr, cabd, arc_hdr_size(hdr), hdr); | |
1864 | ||
1865 | return (ret); | |
1866 | } | |
1867 | ||
1868 | /* | |
1869 | * This function is called during arc_buf_fill() to prepare the header's | |
1870 | * abd plaintext pointer for use. This involves authenticated protected | |
1871 | * data and decrypting encrypted data into the plaintext abd. | |
1872 | */ | |
1873 | static int | |
1874 | arc_fill_hdr_crypt(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, spa_t *spa, | |
be9a5c35 | 1875 | const zbookmark_phys_t *zb, boolean_t noauth) |
b5256303 TC |
1876 | { |
1877 | int ret; | |
1878 | ||
1879 | ASSERT(HDR_PROTECTED(hdr)); | |
1880 | ||
1881 | if (hash_lock != NULL) | |
1882 | mutex_enter(hash_lock); | |
1883 | ||
1884 | if (HDR_NOAUTH(hdr) && !noauth) { | |
1885 | /* | |
1886 | * The caller requested authenticated data but our data has | |
1887 | * not been authenticated yet. Verify the MAC now if we can. | |
1888 | */ | |
be9a5c35 | 1889 | ret = arc_hdr_authenticate(hdr, spa, zb->zb_objset); |
b5256303 TC |
1890 | if (ret != 0) |
1891 | goto error; | |
1892 | } else if (HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd == NULL) { | |
1893 | /* | |
1894 | * If we only have the encrypted version of the data, but the | |
1895 | * unencrypted version was requested we take this opportunity | |
1896 | * to store the decrypted version in the header for future use. | |
1897 | */ | |
be9a5c35 | 1898 | ret = arc_hdr_decrypt(hdr, spa, zb); |
b5256303 TC |
1899 | if (ret != 0) |
1900 | goto error; | |
1901 | } | |
1902 | ||
1903 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); | |
1904 | ||
1905 | if (hash_lock != NULL) | |
1906 | mutex_exit(hash_lock); | |
1907 | ||
1908 | return (0); | |
1909 | ||
1910 | error: | |
1911 | if (hash_lock != NULL) | |
1912 | mutex_exit(hash_lock); | |
1913 | ||
1914 | return (ret); | |
1915 | } | |
1916 | ||
1917 | /* | |
1918 | * This function is used by the dbuf code to decrypt bonus buffers in place. | |
1919 | * The dbuf code itself doesn't have any locking for decrypting a shared dnode | |
1920 | * block, so we use the hash lock here to protect against concurrent calls to | |
1921 | * arc_buf_fill(). | |
1922 | */ | |
1923 | static void | |
1924 | arc_buf_untransform_in_place(arc_buf_t *buf, kmutex_t *hash_lock) | |
1925 | { | |
1926 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
1927 | ||
1928 | ASSERT(HDR_ENCRYPTED(hdr)); | |
1929 | ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE); | |
ca6c7a94 | 1930 | ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); |
b5256303 TC |
1931 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); |
1932 | ||
1933 | zio_crypt_copy_dnode_bonus(hdr->b_l1hdr.b_pabd, buf->b_data, | |
1934 | arc_buf_size(buf)); | |
1935 | buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED; | |
1936 | buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; | |
1937 | hdr->b_crypt_hdr.b_ebufcnt -= 1; | |
1938 | } | |
1939 | ||
524b4217 DK |
1940 | /* |
1941 | * Given a buf that has a data buffer attached to it, this function will | |
1942 | * efficiently fill the buf with data of the specified compression setting from | |
1943 | * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr | |
1944 | * are already sharing a data buf, no copy is performed. | |
1945 | * | |
1946 | * If the buf is marked as compressed but uncompressed data was requested, this | |
1947 | * will allocate a new data buffer for the buf, remove that flag, and fill the | |
1948 | * buf with uncompressed data. You can't request a compressed buf on a hdr with | |
1949 | * uncompressed data, and (since we haven't added support for it yet) if you | |
1950 | * want compressed data your buf must already be marked as compressed and have | |
1951 | * the correct-sized data buffer. | |
1952 | */ | |
1953 | static int | |
be9a5c35 TC |
1954 | arc_buf_fill(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb, |
1955 | arc_fill_flags_t flags) | |
d3c2ae1c | 1956 | { |
b5256303 | 1957 | int error = 0; |
d3c2ae1c | 1958 | arc_buf_hdr_t *hdr = buf->b_hdr; |
b5256303 TC |
1959 | boolean_t hdr_compressed = |
1960 | (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF); | |
1961 | boolean_t compressed = (flags & ARC_FILL_COMPRESSED) != 0; | |
1962 | boolean_t encrypted = (flags & ARC_FILL_ENCRYPTED) != 0; | |
d3c2ae1c | 1963 | dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; |
b5256303 | 1964 | kmutex_t *hash_lock = (flags & ARC_FILL_LOCKED) ? NULL : HDR_LOCK(hdr); |
d3c2ae1c | 1965 | |
524b4217 | 1966 | ASSERT3P(buf->b_data, !=, NULL); |
b5256303 | 1967 | IMPLY(compressed, hdr_compressed || ARC_BUF_ENCRYPTED(buf)); |
524b4217 | 1968 | IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); |
b5256303 TC |
1969 | IMPLY(encrypted, HDR_ENCRYPTED(hdr)); |
1970 | IMPLY(encrypted, ARC_BUF_ENCRYPTED(buf)); | |
1971 | IMPLY(encrypted, ARC_BUF_COMPRESSED(buf)); | |
1972 | IMPLY(encrypted, !ARC_BUF_SHARED(buf)); | |
1973 | ||
1974 | /* | |
1975 | * If the caller wanted encrypted data we just need to copy it from | |
1976 | * b_rabd and potentially byteswap it. We won't be able to do any | |
1977 | * further transforms on it. | |
1978 | */ | |
1979 | if (encrypted) { | |
1980 | ASSERT(HDR_HAS_RABD(hdr)); | |
1981 | abd_copy_to_buf(buf->b_data, hdr->b_crypt_hdr.b_rabd, | |
1982 | HDR_GET_PSIZE(hdr)); | |
1983 | goto byteswap; | |
1984 | } | |
1985 | ||
1986 | /* | |
e1cfd73f | 1987 | * Adjust encrypted and authenticated headers to accommodate |
69830602 TC |
1988 | * the request if needed. Dnode blocks (ARC_FILL_IN_PLACE) are |
1989 | * allowed to fail decryption due to keys not being loaded | |
1990 | * without being marked as an IO error. | |
b5256303 TC |
1991 | */ |
1992 | if (HDR_PROTECTED(hdr)) { | |
1993 | error = arc_fill_hdr_crypt(hdr, hash_lock, spa, | |
be9a5c35 | 1994 | zb, !!(flags & ARC_FILL_NOAUTH)); |
69830602 TC |
1995 | if (error == EACCES && (flags & ARC_FILL_IN_PLACE) != 0) { |
1996 | return (error); | |
1997 | } else if (error != 0) { | |
e7504d7a TC |
1998 | if (hash_lock != NULL) |
1999 | mutex_enter(hash_lock); | |
2c24b5b1 | 2000 | arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); |
e7504d7a TC |
2001 | if (hash_lock != NULL) |
2002 | mutex_exit(hash_lock); | |
b5256303 | 2003 | return (error); |
2c24b5b1 | 2004 | } |
b5256303 TC |
2005 | } |
2006 | ||
2007 | /* | |
2008 | * There is a special case here for dnode blocks which are | |
2009 | * decrypting their bonus buffers. These blocks may request to | |
2010 | * be decrypted in-place. This is necessary because there may | |
2011 | * be many dnodes pointing into this buffer and there is | |
2012 | * currently no method to synchronize replacing the backing | |
2013 | * b_data buffer and updating all of the pointers. Here we use | |
2014 | * the hash lock to ensure there are no races. If the need | |
2015 | * arises for other types to be decrypted in-place, they must | |
2016 | * add handling here as well. | |
2017 | */ | |
2018 | if ((flags & ARC_FILL_IN_PLACE) != 0) { | |
2019 | ASSERT(!hdr_compressed); | |
2020 | ASSERT(!compressed); | |
2021 | ASSERT(!encrypted); | |
2022 | ||
2023 | if (HDR_ENCRYPTED(hdr) && ARC_BUF_ENCRYPTED(buf)) { | |
2024 | ASSERT3U(hdr->b_crypt_hdr.b_ot, ==, DMU_OT_DNODE); | |
2025 | ||
2026 | if (hash_lock != NULL) | |
2027 | mutex_enter(hash_lock); | |
2028 | arc_buf_untransform_in_place(buf, hash_lock); | |
2029 | if (hash_lock != NULL) | |
2030 | mutex_exit(hash_lock); | |
2031 | ||
2032 | /* Compute the hdr's checksum if necessary */ | |
2033 | arc_cksum_compute(buf); | |
2034 | } | |
2035 | ||
2036 | return (0); | |
2037 | } | |
524b4217 DK |
2038 | |
2039 | if (hdr_compressed == compressed) { | |
2aa34383 | 2040 | if (!arc_buf_is_shared(buf)) { |
a6255b7f | 2041 | abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd, |
524b4217 | 2042 | arc_buf_size(buf)); |
2aa34383 | 2043 | } |
d3c2ae1c | 2044 | } else { |
524b4217 DK |
2045 | ASSERT(hdr_compressed); |
2046 | ASSERT(!compressed); | |
d3c2ae1c | 2047 | ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); |
2aa34383 DK |
2048 | |
2049 | /* | |
524b4217 DK |
2050 | * If the buf is sharing its data with the hdr, unlink it and |
2051 | * allocate a new data buffer for the buf. | |
2aa34383 | 2052 | */ |
524b4217 DK |
2053 | if (arc_buf_is_shared(buf)) { |
2054 | ASSERT(ARC_BUF_COMPRESSED(buf)); | |
2055 | ||
e1cfd73f | 2056 | /* We need to give the buf its own b_data */ |
524b4217 | 2057 | buf->b_flags &= ~ARC_BUF_FLAG_SHARED; |
2aa34383 DK |
2058 | buf->b_data = |
2059 | arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); | |
2060 | arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); | |
2061 | ||
524b4217 | 2062 | /* Previously overhead was 0; just add new overhead */ |
2aa34383 | 2063 | ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); |
524b4217 DK |
2064 | } else if (ARC_BUF_COMPRESSED(buf)) { |
2065 | /* We need to reallocate the buf's b_data */ | |
2066 | arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), | |
2067 | buf); | |
2068 | buf->b_data = | |
2069 | arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); | |
2070 | ||
2071 | /* We increased the size of b_data; update overhead */ | |
2072 | ARCSTAT_INCR(arcstat_overhead_size, | |
2073 | HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); | |
2aa34383 DK |
2074 | } |
2075 | ||
524b4217 DK |
2076 | /* |
2077 | * Regardless of the buf's previous compression settings, it | |
2078 | * should not be compressed at the end of this function. | |
2079 | */ | |
2080 | buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; | |
2081 | ||
2082 | /* | |
2083 | * Try copying the data from another buf which already has a | |
2084 | * decompressed version. If that's not possible, it's time to | |
2085 | * bite the bullet and decompress the data from the hdr. | |
2086 | */ | |
2087 | if (arc_buf_try_copy_decompressed_data(buf)) { | |
2088 | /* Skip byteswapping and checksumming (already done) */ | |
524b4217 DK |
2089 | return (0); |
2090 | } else { | |
b5256303 | 2091 | error = zio_decompress_data(HDR_GET_COMPRESS(hdr), |
a6255b7f | 2092 | hdr->b_l1hdr.b_pabd, buf->b_data, |
524b4217 DK |
2093 | HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); |
2094 | ||
2095 | /* | |
2096 | * Absent hardware errors or software bugs, this should | |
2097 | * be impossible, but log it anyway so we can debug it. | |
2098 | */ | |
2099 | if (error != 0) { | |
2100 | zfs_dbgmsg( | |
a887d653 | 2101 | "hdr %px, compress %d, psize %d, lsize %d", |
b5256303 | 2102 | hdr, arc_hdr_get_compress(hdr), |
524b4217 | 2103 | HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); |
e7504d7a TC |
2104 | if (hash_lock != NULL) |
2105 | mutex_enter(hash_lock); | |
2c24b5b1 | 2106 | arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); |
e7504d7a TC |
2107 | if (hash_lock != NULL) |
2108 | mutex_exit(hash_lock); | |
524b4217 DK |
2109 | return (SET_ERROR(EIO)); |
2110 | } | |
d3c2ae1c GW |
2111 | } |
2112 | } | |
524b4217 | 2113 | |
b5256303 | 2114 | byteswap: |
524b4217 | 2115 | /* Byteswap the buf's data if necessary */ |
d3c2ae1c GW |
2116 | if (bswap != DMU_BSWAP_NUMFUNCS) { |
2117 | ASSERT(!HDR_SHARED_DATA(hdr)); | |
2118 | ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); | |
2119 | dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); | |
2120 | } | |
524b4217 DK |
2121 | |
2122 | /* Compute the hdr's checksum if necessary */ | |
d3c2ae1c | 2123 | arc_cksum_compute(buf); |
524b4217 | 2124 | |
d3c2ae1c GW |
2125 | return (0); |
2126 | } | |
2127 | ||
2128 | /* | |
b5256303 TC |
2129 | * If this function is being called to decrypt an encrypted buffer or verify an |
2130 | * authenticated one, the key must be loaded and a mapping must be made | |
2131 | * available in the keystore via spa_keystore_create_mapping() or one of its | |
2132 | * callers. | |
d3c2ae1c | 2133 | */ |
b5256303 | 2134 | int |
a2c2ed1b TC |
2135 | arc_untransform(arc_buf_t *buf, spa_t *spa, const zbookmark_phys_t *zb, |
2136 | boolean_t in_place) | |
d3c2ae1c | 2137 | { |
a2c2ed1b | 2138 | int ret; |
b5256303 | 2139 | arc_fill_flags_t flags = 0; |
d3c2ae1c | 2140 | |
b5256303 TC |
2141 | if (in_place) |
2142 | flags |= ARC_FILL_IN_PLACE; | |
2143 | ||
be9a5c35 | 2144 | ret = arc_buf_fill(buf, spa, zb, flags); |
a2c2ed1b TC |
2145 | if (ret == ECKSUM) { |
2146 | /* | |
2147 | * Convert authentication and decryption errors to EIO | |
2148 | * (and generate an ereport) before leaving the ARC. | |
2149 | */ | |
2150 | ret = SET_ERROR(EIO); | |
be9a5c35 | 2151 | spa_log_error(spa, zb); |
a2c2ed1b TC |
2152 | zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, |
2153 | spa, NULL, zb, NULL, 0, 0); | |
2154 | } | |
2155 | ||
2156 | return (ret); | |
d3c2ae1c GW |
2157 | } |
2158 | ||
2159 | /* | |
2160 | * Increment the amount of evictable space in the arc_state_t's refcount. | |
2161 | * We account for the space used by the hdr and the arc buf individually | |
2162 | * so that we can add and remove them from the refcount individually. | |
2163 | */ | |
34dc7c2f | 2164 | static void |
d3c2ae1c GW |
2165 | arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) |
2166 | { | |
2167 | arc_buf_contents_t type = arc_buf_type(hdr); | |
d3c2ae1c GW |
2168 | |
2169 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
2170 | ||
2171 | if (GHOST_STATE(state)) { | |
2172 | ASSERT0(hdr->b_l1hdr.b_bufcnt); | |
2173 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
a6255b7f | 2174 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 2175 | ASSERT(!HDR_HAS_RABD(hdr)); |
424fd7c3 | 2176 | (void) zfs_refcount_add_many(&state->arcs_esize[type], |
2aa34383 | 2177 | HDR_GET_LSIZE(hdr), hdr); |
d3c2ae1c GW |
2178 | return; |
2179 | } | |
2180 | ||
2181 | ASSERT(!GHOST_STATE(state)); | |
a6255b7f | 2182 | if (hdr->b_l1hdr.b_pabd != NULL) { |
424fd7c3 | 2183 | (void) zfs_refcount_add_many(&state->arcs_esize[type], |
d3c2ae1c GW |
2184 | arc_hdr_size(hdr), hdr); |
2185 | } | |
b5256303 | 2186 | if (HDR_HAS_RABD(hdr)) { |
424fd7c3 | 2187 | (void) zfs_refcount_add_many(&state->arcs_esize[type], |
b5256303 TC |
2188 | HDR_GET_PSIZE(hdr), hdr); |
2189 | } | |
2190 | ||
1c27024e DB |
2191 | for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; |
2192 | buf = buf->b_next) { | |
2aa34383 | 2193 | if (arc_buf_is_shared(buf)) |
d3c2ae1c | 2194 | continue; |
424fd7c3 | 2195 | (void) zfs_refcount_add_many(&state->arcs_esize[type], |
2aa34383 | 2196 | arc_buf_size(buf), buf); |
d3c2ae1c GW |
2197 | } |
2198 | } | |
2199 | ||
2200 | /* | |
2201 | * Decrement the amount of evictable space in the arc_state_t's refcount. | |
2202 | * We account for the space used by the hdr and the arc buf individually | |
2203 | * so that we can add and remove them from the refcount individually. | |
2204 | */ | |
2205 | static void | |
2aa34383 | 2206 | arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) |
d3c2ae1c GW |
2207 | { |
2208 | arc_buf_contents_t type = arc_buf_type(hdr); | |
d3c2ae1c GW |
2209 | |
2210 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
2211 | ||
2212 | if (GHOST_STATE(state)) { | |
2213 | ASSERT0(hdr->b_l1hdr.b_bufcnt); | |
2214 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
a6255b7f | 2215 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 2216 | ASSERT(!HDR_HAS_RABD(hdr)); |
424fd7c3 | 2217 | (void) zfs_refcount_remove_many(&state->arcs_esize[type], |
2aa34383 | 2218 | HDR_GET_LSIZE(hdr), hdr); |
d3c2ae1c GW |
2219 | return; |
2220 | } | |
2221 | ||
2222 | ASSERT(!GHOST_STATE(state)); | |
a6255b7f | 2223 | if (hdr->b_l1hdr.b_pabd != NULL) { |
424fd7c3 | 2224 | (void) zfs_refcount_remove_many(&state->arcs_esize[type], |
d3c2ae1c GW |
2225 | arc_hdr_size(hdr), hdr); |
2226 | } | |
b5256303 | 2227 | if (HDR_HAS_RABD(hdr)) { |
424fd7c3 | 2228 | (void) zfs_refcount_remove_many(&state->arcs_esize[type], |
b5256303 TC |
2229 | HDR_GET_PSIZE(hdr), hdr); |
2230 | } | |
2231 | ||
1c27024e DB |
2232 | for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; |
2233 | buf = buf->b_next) { | |
2aa34383 | 2234 | if (arc_buf_is_shared(buf)) |
d3c2ae1c | 2235 | continue; |
424fd7c3 | 2236 | (void) zfs_refcount_remove_many(&state->arcs_esize[type], |
2aa34383 | 2237 | arc_buf_size(buf), buf); |
d3c2ae1c GW |
2238 | } |
2239 | } | |
2240 | ||
2241 | /* | |
2242 | * Add a reference to this hdr indicating that someone is actively | |
2243 | * referencing that memory. When the refcount transitions from 0 to 1, | |
2244 | * we remove it from the respective arc_state_t list to indicate that | |
2245 | * it is not evictable. | |
2246 | */ | |
2247 | static void | |
2248 | add_reference(arc_buf_hdr_t *hdr, void *tag) | |
34dc7c2f | 2249 | { |
b9541d6b CW |
2250 | arc_state_t *state; |
2251 | ||
2252 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
ca6c7a94 | 2253 | if (!HDR_EMPTY(hdr) && !MUTEX_HELD(HDR_LOCK(hdr))) { |
d3c2ae1c | 2254 | ASSERT(hdr->b_l1hdr.b_state == arc_anon); |
424fd7c3 | 2255 | ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
d3c2ae1c GW |
2256 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); |
2257 | } | |
34dc7c2f | 2258 | |
b9541d6b CW |
2259 | state = hdr->b_l1hdr.b_state; |
2260 | ||
c13060e4 | 2261 | if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && |
b9541d6b CW |
2262 | (state != arc_anon)) { |
2263 | /* We don't use the L2-only state list. */ | |
2264 | if (state != arc_l2c_only) { | |
64fc7762 | 2265 | multilist_remove(state->arcs_list[arc_buf_type(hdr)], |
d3c2ae1c | 2266 | hdr); |
2aa34383 | 2267 | arc_evictable_space_decrement(hdr, state); |
34dc7c2f | 2268 | } |
b128c09f | 2269 | /* remove the prefetch flag if we get a reference */ |
d3c2ae1c | 2270 | arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); |
34dc7c2f BB |
2271 | } |
2272 | } | |
2273 | ||
d3c2ae1c GW |
2274 | /* |
2275 | * Remove a reference from this hdr. When the reference transitions from | |
2276 | * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's | |
2277 | * list making it eligible for eviction. | |
2278 | */ | |
34dc7c2f | 2279 | static int |
2a432414 | 2280 | remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) |
34dc7c2f BB |
2281 | { |
2282 | int cnt; | |
b9541d6b | 2283 | arc_state_t *state = hdr->b_l1hdr.b_state; |
34dc7c2f | 2284 | |
b9541d6b | 2285 | ASSERT(HDR_HAS_L1HDR(hdr)); |
34dc7c2f BB |
2286 | ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); |
2287 | ASSERT(!GHOST_STATE(state)); | |
2288 | ||
b9541d6b CW |
2289 | /* |
2290 | * arc_l2c_only counts as a ghost state so we don't need to explicitly | |
2291 | * check to prevent usage of the arc_l2c_only list. | |
2292 | */ | |
424fd7c3 | 2293 | if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && |
34dc7c2f | 2294 | (state != arc_anon)) { |
64fc7762 | 2295 | multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr); |
d3c2ae1c GW |
2296 | ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); |
2297 | arc_evictable_space_increment(hdr, state); | |
34dc7c2f BB |
2298 | } |
2299 | return (cnt); | |
2300 | } | |
2301 | ||
e0b0ca98 BB |
2302 | /* |
2303 | * Returns detailed information about a specific arc buffer. When the | |
2304 | * state_index argument is set the function will calculate the arc header | |
2305 | * list position for its arc state. Since this requires a linear traversal | |
2306 | * callers are strongly encourage not to do this. However, it can be helpful | |
2307 | * for targeted analysis so the functionality is provided. | |
2308 | */ | |
2309 | void | |
2310 | arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index) | |
2311 | { | |
2312 | arc_buf_hdr_t *hdr = ab->b_hdr; | |
b9541d6b CW |
2313 | l1arc_buf_hdr_t *l1hdr = NULL; |
2314 | l2arc_buf_hdr_t *l2hdr = NULL; | |
2315 | arc_state_t *state = NULL; | |
2316 | ||
8887c7d7 TC |
2317 | memset(abi, 0, sizeof (arc_buf_info_t)); |
2318 | ||
2319 | if (hdr == NULL) | |
2320 | return; | |
2321 | ||
2322 | abi->abi_flags = hdr->b_flags; | |
2323 | ||
b9541d6b CW |
2324 | if (HDR_HAS_L1HDR(hdr)) { |
2325 | l1hdr = &hdr->b_l1hdr; | |
2326 | state = l1hdr->b_state; | |
2327 | } | |
2328 | if (HDR_HAS_L2HDR(hdr)) | |
2329 | l2hdr = &hdr->b_l2hdr; | |
e0b0ca98 | 2330 | |
b9541d6b | 2331 | if (l1hdr) { |
d3c2ae1c | 2332 | abi->abi_bufcnt = l1hdr->b_bufcnt; |
b9541d6b CW |
2333 | abi->abi_access = l1hdr->b_arc_access; |
2334 | abi->abi_mru_hits = l1hdr->b_mru_hits; | |
2335 | abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits; | |
2336 | abi->abi_mfu_hits = l1hdr->b_mfu_hits; | |
2337 | abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits; | |
424fd7c3 | 2338 | abi->abi_holds = zfs_refcount_count(&l1hdr->b_refcnt); |
b9541d6b CW |
2339 | } |
2340 | ||
2341 | if (l2hdr) { | |
2342 | abi->abi_l2arc_dattr = l2hdr->b_daddr; | |
b9541d6b CW |
2343 | abi->abi_l2arc_hits = l2hdr->b_hits; |
2344 | } | |
2345 | ||
e0b0ca98 | 2346 | abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON; |
b9541d6b | 2347 | abi->abi_state_contents = arc_buf_type(hdr); |
d3c2ae1c | 2348 | abi->abi_size = arc_hdr_size(hdr); |
e0b0ca98 BB |
2349 | } |
2350 | ||
34dc7c2f | 2351 | /* |
ca0bf58d | 2352 | * Move the supplied buffer to the indicated state. The hash lock |
34dc7c2f BB |
2353 | * for the buffer must be held by the caller. |
2354 | */ | |
2355 | static void | |
2a432414 GW |
2356 | arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, |
2357 | kmutex_t *hash_lock) | |
34dc7c2f | 2358 | { |
b9541d6b CW |
2359 | arc_state_t *old_state; |
2360 | int64_t refcnt; | |
d3c2ae1c GW |
2361 | uint32_t bufcnt; |
2362 | boolean_t update_old, update_new; | |
b9541d6b CW |
2363 | arc_buf_contents_t buftype = arc_buf_type(hdr); |
2364 | ||
2365 | /* | |
2366 | * We almost always have an L1 hdr here, since we call arc_hdr_realloc() | |
2367 | * in arc_read() when bringing a buffer out of the L2ARC. However, the | |
2368 | * L1 hdr doesn't always exist when we change state to arc_anon before | |
2369 | * destroying a header, in which case reallocating to add the L1 hdr is | |
2370 | * pointless. | |
2371 | */ | |
2372 | if (HDR_HAS_L1HDR(hdr)) { | |
2373 | old_state = hdr->b_l1hdr.b_state; | |
424fd7c3 | 2374 | refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt); |
d3c2ae1c | 2375 | bufcnt = hdr->b_l1hdr.b_bufcnt; |
b5256303 TC |
2376 | update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL || |
2377 | HDR_HAS_RABD(hdr)); | |
b9541d6b CW |
2378 | } else { |
2379 | old_state = arc_l2c_only; | |
2380 | refcnt = 0; | |
d3c2ae1c GW |
2381 | bufcnt = 0; |
2382 | update_old = B_FALSE; | |
b9541d6b | 2383 | } |
d3c2ae1c | 2384 | update_new = update_old; |
34dc7c2f BB |
2385 | |
2386 | ASSERT(MUTEX_HELD(hash_lock)); | |
e8b96c60 | 2387 | ASSERT3P(new_state, !=, old_state); |
d3c2ae1c GW |
2388 | ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); |
2389 | ASSERT(old_state != arc_anon || bufcnt <= 1); | |
34dc7c2f BB |
2390 | |
2391 | /* | |
2392 | * If this buffer is evictable, transfer it from the | |
2393 | * old state list to the new state list. | |
2394 | */ | |
2395 | if (refcnt == 0) { | |
b9541d6b | 2396 | if (old_state != arc_anon && old_state != arc_l2c_only) { |
b9541d6b | 2397 | ASSERT(HDR_HAS_L1HDR(hdr)); |
64fc7762 | 2398 | multilist_remove(old_state->arcs_list[buftype], hdr); |
34dc7c2f | 2399 | |
d3c2ae1c GW |
2400 | if (GHOST_STATE(old_state)) { |
2401 | ASSERT0(bufcnt); | |
2402 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
2403 | update_old = B_TRUE; | |
34dc7c2f | 2404 | } |
2aa34383 | 2405 | arc_evictable_space_decrement(hdr, old_state); |
34dc7c2f | 2406 | } |
b9541d6b | 2407 | if (new_state != arc_anon && new_state != arc_l2c_only) { |
b9541d6b CW |
2408 | /* |
2409 | * An L1 header always exists here, since if we're | |
2410 | * moving to some L1-cached state (i.e. not l2c_only or | |
2411 | * anonymous), we realloc the header to add an L1hdr | |
2412 | * beforehand. | |
2413 | */ | |
2414 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
64fc7762 | 2415 | multilist_insert(new_state->arcs_list[buftype], hdr); |
34dc7c2f | 2416 | |
34dc7c2f | 2417 | if (GHOST_STATE(new_state)) { |
d3c2ae1c GW |
2418 | ASSERT0(bufcnt); |
2419 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
2420 | update_new = B_TRUE; | |
34dc7c2f | 2421 | } |
d3c2ae1c | 2422 | arc_evictable_space_increment(hdr, new_state); |
34dc7c2f BB |
2423 | } |
2424 | } | |
2425 | ||
d3c2ae1c | 2426 | ASSERT(!HDR_EMPTY(hdr)); |
2a432414 GW |
2427 | if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) |
2428 | buf_hash_remove(hdr); | |
34dc7c2f | 2429 | |
b9541d6b | 2430 | /* adjust state sizes (ignore arc_l2c_only) */ |
36da08ef | 2431 | |
d3c2ae1c | 2432 | if (update_new && new_state != arc_l2c_only) { |
36da08ef PS |
2433 | ASSERT(HDR_HAS_L1HDR(hdr)); |
2434 | if (GHOST_STATE(new_state)) { | |
d3c2ae1c | 2435 | ASSERT0(bufcnt); |
36da08ef PS |
2436 | |
2437 | /* | |
d3c2ae1c | 2438 | * When moving a header to a ghost state, we first |
36da08ef | 2439 | * remove all arc buffers. Thus, we'll have a |
d3c2ae1c | 2440 | * bufcnt of zero, and no arc buffer to use for |
36da08ef PS |
2441 | * the reference. As a result, we use the arc |
2442 | * header pointer for the reference. | |
2443 | */ | |
424fd7c3 | 2444 | (void) zfs_refcount_add_many(&new_state->arcs_size, |
d3c2ae1c | 2445 | HDR_GET_LSIZE(hdr), hdr); |
a6255b7f | 2446 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 2447 | ASSERT(!HDR_HAS_RABD(hdr)); |
36da08ef | 2448 | } else { |
d3c2ae1c | 2449 | uint32_t buffers = 0; |
36da08ef PS |
2450 | |
2451 | /* | |
2452 | * Each individual buffer holds a unique reference, | |
2453 | * thus we must remove each of these references one | |
2454 | * at a time. | |
2455 | */ | |
1c27024e | 2456 | for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; |
36da08ef | 2457 | buf = buf->b_next) { |
d3c2ae1c GW |
2458 | ASSERT3U(bufcnt, !=, 0); |
2459 | buffers++; | |
2460 | ||
2461 | /* | |
2462 | * When the arc_buf_t is sharing the data | |
2463 | * block with the hdr, the owner of the | |
2464 | * reference belongs to the hdr. Only | |
2465 | * add to the refcount if the arc_buf_t is | |
2466 | * not shared. | |
2467 | */ | |
2aa34383 | 2468 | if (arc_buf_is_shared(buf)) |
d3c2ae1c | 2469 | continue; |
d3c2ae1c | 2470 | |
424fd7c3 TS |
2471 | (void) zfs_refcount_add_many( |
2472 | &new_state->arcs_size, | |
2aa34383 | 2473 | arc_buf_size(buf), buf); |
d3c2ae1c GW |
2474 | } |
2475 | ASSERT3U(bufcnt, ==, buffers); | |
2476 | ||
a6255b7f | 2477 | if (hdr->b_l1hdr.b_pabd != NULL) { |
424fd7c3 TS |
2478 | (void) zfs_refcount_add_many( |
2479 | &new_state->arcs_size, | |
d3c2ae1c | 2480 | arc_hdr_size(hdr), hdr); |
b5256303 TC |
2481 | } |
2482 | ||
2483 | if (HDR_HAS_RABD(hdr)) { | |
424fd7c3 TS |
2484 | (void) zfs_refcount_add_many( |
2485 | &new_state->arcs_size, | |
b5256303 | 2486 | HDR_GET_PSIZE(hdr), hdr); |
36da08ef PS |
2487 | } |
2488 | } | |
2489 | } | |
2490 | ||
d3c2ae1c | 2491 | if (update_old && old_state != arc_l2c_only) { |
36da08ef PS |
2492 | ASSERT(HDR_HAS_L1HDR(hdr)); |
2493 | if (GHOST_STATE(old_state)) { | |
d3c2ae1c | 2494 | ASSERT0(bufcnt); |
a6255b7f | 2495 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 2496 | ASSERT(!HDR_HAS_RABD(hdr)); |
d3c2ae1c | 2497 | |
36da08ef PS |
2498 | /* |
2499 | * When moving a header off of a ghost state, | |
d3c2ae1c GW |
2500 | * the header will not contain any arc buffers. |
2501 | * We use the arc header pointer for the reference | |
2502 | * which is exactly what we did when we put the | |
2503 | * header on the ghost state. | |
36da08ef PS |
2504 | */ |
2505 | ||
424fd7c3 | 2506 | (void) zfs_refcount_remove_many(&old_state->arcs_size, |
d3c2ae1c | 2507 | HDR_GET_LSIZE(hdr), hdr); |
36da08ef | 2508 | } else { |
d3c2ae1c | 2509 | uint32_t buffers = 0; |
36da08ef PS |
2510 | |
2511 | /* | |
2512 | * Each individual buffer holds a unique reference, | |
2513 | * thus we must remove each of these references one | |
2514 | * at a time. | |
2515 | */ | |
1c27024e | 2516 | for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; |
36da08ef | 2517 | buf = buf->b_next) { |
d3c2ae1c GW |
2518 | ASSERT3U(bufcnt, !=, 0); |
2519 | buffers++; | |
2520 | ||
2521 | /* | |
2522 | * When the arc_buf_t is sharing the data | |
2523 | * block with the hdr, the owner of the | |
2524 | * reference belongs to the hdr. Only | |
2525 | * add to the refcount if the arc_buf_t is | |
2526 | * not shared. | |
2527 | */ | |
2aa34383 | 2528 | if (arc_buf_is_shared(buf)) |
d3c2ae1c | 2529 | continue; |
d3c2ae1c | 2530 | |
424fd7c3 | 2531 | (void) zfs_refcount_remove_many( |
2aa34383 | 2532 | &old_state->arcs_size, arc_buf_size(buf), |
d3c2ae1c | 2533 | buf); |
36da08ef | 2534 | } |
d3c2ae1c | 2535 | ASSERT3U(bufcnt, ==, buffers); |
b5256303 TC |
2536 | ASSERT(hdr->b_l1hdr.b_pabd != NULL || |
2537 | HDR_HAS_RABD(hdr)); | |
2538 | ||
2539 | if (hdr->b_l1hdr.b_pabd != NULL) { | |
424fd7c3 | 2540 | (void) zfs_refcount_remove_many( |
b5256303 TC |
2541 | &old_state->arcs_size, arc_hdr_size(hdr), |
2542 | hdr); | |
2543 | } | |
2544 | ||
2545 | if (HDR_HAS_RABD(hdr)) { | |
424fd7c3 | 2546 | (void) zfs_refcount_remove_many( |
b5256303 TC |
2547 | &old_state->arcs_size, HDR_GET_PSIZE(hdr), |
2548 | hdr); | |
2549 | } | |
36da08ef | 2550 | } |
34dc7c2f | 2551 | } |
36da08ef | 2552 | |
b9541d6b CW |
2553 | if (HDR_HAS_L1HDR(hdr)) |
2554 | hdr->b_l1hdr.b_state = new_state; | |
34dc7c2f | 2555 | |
b9541d6b CW |
2556 | /* |
2557 | * L2 headers should never be on the L2 state list since they don't | |
2558 | * have L1 headers allocated. | |
2559 | */ | |
64fc7762 MA |
2560 | ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && |
2561 | multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); | |
34dc7c2f BB |
2562 | } |
2563 | ||
2564 | void | |
d164b209 | 2565 | arc_space_consume(uint64_t space, arc_space_type_t type) |
34dc7c2f | 2566 | { |
d164b209 BB |
2567 | ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); |
2568 | ||
2569 | switch (type) { | |
e75c13c3 BB |
2570 | default: |
2571 | break; | |
d164b209 | 2572 | case ARC_SPACE_DATA: |
37fb3e43 | 2573 | aggsum_add(&astat_data_size, space); |
d164b209 | 2574 | break; |
cc7f677c | 2575 | case ARC_SPACE_META: |
37fb3e43 | 2576 | aggsum_add(&astat_metadata_size, space); |
cc7f677c | 2577 | break; |
25458cbe | 2578 | case ARC_SPACE_BONUS: |
37fb3e43 | 2579 | aggsum_add(&astat_bonus_size, space); |
25458cbe TC |
2580 | break; |
2581 | case ARC_SPACE_DNODE: | |
37fb3e43 | 2582 | aggsum_add(&astat_dnode_size, space); |
25458cbe TC |
2583 | break; |
2584 | case ARC_SPACE_DBUF: | |
37fb3e43 | 2585 | aggsum_add(&astat_dbuf_size, space); |
d164b209 BB |
2586 | break; |
2587 | case ARC_SPACE_HDRS: | |
37fb3e43 | 2588 | aggsum_add(&astat_hdr_size, space); |
d164b209 BB |
2589 | break; |
2590 | case ARC_SPACE_L2HDRS: | |
37fb3e43 | 2591 | aggsum_add(&astat_l2_hdr_size, space); |
d164b209 BB |
2592 | break; |
2593 | } | |
2594 | ||
500445c0 | 2595 | if (type != ARC_SPACE_DATA) |
37fb3e43 | 2596 | aggsum_add(&arc_meta_used, space); |
cc7f677c | 2597 | |
37fb3e43 | 2598 | aggsum_add(&arc_size, space); |
34dc7c2f BB |
2599 | } |
2600 | ||
2601 | void | |
d164b209 | 2602 | arc_space_return(uint64_t space, arc_space_type_t type) |
34dc7c2f | 2603 | { |
d164b209 BB |
2604 | ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); |
2605 | ||
2606 | switch (type) { | |
e75c13c3 BB |
2607 | default: |
2608 | break; | |
d164b209 | 2609 | case ARC_SPACE_DATA: |
37fb3e43 | 2610 | aggsum_add(&astat_data_size, -space); |
d164b209 | 2611 | break; |
cc7f677c | 2612 | case ARC_SPACE_META: |
37fb3e43 | 2613 | aggsum_add(&astat_metadata_size, -space); |
cc7f677c | 2614 | break; |
25458cbe | 2615 | case ARC_SPACE_BONUS: |
37fb3e43 | 2616 | aggsum_add(&astat_bonus_size, -space); |
25458cbe TC |
2617 | break; |
2618 | case ARC_SPACE_DNODE: | |
37fb3e43 | 2619 | aggsum_add(&astat_dnode_size, -space); |
25458cbe TC |
2620 | break; |
2621 | case ARC_SPACE_DBUF: | |
37fb3e43 | 2622 | aggsum_add(&astat_dbuf_size, -space); |
d164b209 BB |
2623 | break; |
2624 | case ARC_SPACE_HDRS: | |
37fb3e43 | 2625 | aggsum_add(&astat_hdr_size, -space); |
d164b209 BB |
2626 | break; |
2627 | case ARC_SPACE_L2HDRS: | |
37fb3e43 | 2628 | aggsum_add(&astat_l2_hdr_size, -space); |
d164b209 BB |
2629 | break; |
2630 | } | |
2631 | ||
cc7f677c | 2632 | if (type != ARC_SPACE_DATA) { |
37fb3e43 PD |
2633 | ASSERT(aggsum_compare(&arc_meta_used, space) >= 0); |
2634 | /* | |
2635 | * We use the upper bound here rather than the precise value | |
2636 | * because the arc_meta_max value doesn't need to be | |
2637 | * precise. It's only consumed by humans via arcstats. | |
2638 | */ | |
2639 | if (arc_meta_max < aggsum_upper_bound(&arc_meta_used)) | |
2640 | arc_meta_max = aggsum_upper_bound(&arc_meta_used); | |
2641 | aggsum_add(&arc_meta_used, -space); | |
cc7f677c PS |
2642 | } |
2643 | ||
37fb3e43 PD |
2644 | ASSERT(aggsum_compare(&arc_size, space) >= 0); |
2645 | aggsum_add(&arc_size, -space); | |
34dc7c2f BB |
2646 | } |
2647 | ||
d3c2ae1c | 2648 | /* |
524b4217 | 2649 | * Given a hdr and a buf, returns whether that buf can share its b_data buffer |
a6255b7f | 2650 | * with the hdr's b_pabd. |
d3c2ae1c | 2651 | */ |
524b4217 DK |
2652 | static boolean_t |
2653 | arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) | |
2654 | { | |
524b4217 DK |
2655 | /* |
2656 | * The criteria for sharing a hdr's data are: | |
b5256303 TC |
2657 | * 1. the buffer is not encrypted |
2658 | * 2. the hdr's compression matches the buf's compression | |
2659 | * 3. the hdr doesn't need to be byteswapped | |
2660 | * 4. the hdr isn't already being shared | |
2661 | * 5. the buf is either compressed or it is the last buf in the hdr list | |
524b4217 | 2662 | * |
b5256303 | 2663 | * Criterion #5 maintains the invariant that shared uncompressed |
524b4217 DK |
2664 | * bufs must be the final buf in the hdr's b_buf list. Reading this, you |
2665 | * might ask, "if a compressed buf is allocated first, won't that be the | |
2666 | * last thing in the list?", but in that case it's impossible to create | |
2667 | * a shared uncompressed buf anyway (because the hdr must be compressed | |
2668 | * to have the compressed buf). You might also think that #3 is | |
2669 | * sufficient to make this guarantee, however it's possible | |
2670 | * (specifically in the rare L2ARC write race mentioned in | |
2671 | * arc_buf_alloc_impl()) there will be an existing uncompressed buf that | |
e1cfd73f | 2672 | * is shareable, but wasn't at the time of its allocation. Rather than |
524b4217 DK |
2673 | * allow a new shared uncompressed buf to be created and then shuffle |
2674 | * the list around to make it the last element, this simply disallows | |
2675 | * sharing if the new buf isn't the first to be added. | |
2676 | */ | |
2677 | ASSERT3P(buf->b_hdr, ==, hdr); | |
b5256303 TC |
2678 | boolean_t hdr_compressed = |
2679 | arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF; | |
a7004725 | 2680 | boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; |
b5256303 TC |
2681 | return (!ARC_BUF_ENCRYPTED(buf) && |
2682 | buf_compressed == hdr_compressed && | |
524b4217 DK |
2683 | hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && |
2684 | !HDR_SHARED_DATA(hdr) && | |
2685 | (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); | |
2686 | } | |
2687 | ||
2688 | /* | |
2689 | * Allocate a buf for this hdr. If you care about the data that's in the hdr, | |
2690 | * or if you want a compressed buffer, pass those flags in. Returns 0 if the | |
2691 | * copy was made successfully, or an error code otherwise. | |
2692 | */ | |
2693 | static int | |
be9a5c35 TC |
2694 | arc_buf_alloc_impl(arc_buf_hdr_t *hdr, spa_t *spa, const zbookmark_phys_t *zb, |
2695 | void *tag, boolean_t encrypted, boolean_t compressed, boolean_t noauth, | |
524b4217 | 2696 | boolean_t fill, arc_buf_t **ret) |
34dc7c2f | 2697 | { |
34dc7c2f | 2698 | arc_buf_t *buf; |
b5256303 | 2699 | arc_fill_flags_t flags = ARC_FILL_LOCKED; |
34dc7c2f | 2700 | |
d3c2ae1c GW |
2701 | ASSERT(HDR_HAS_L1HDR(hdr)); |
2702 | ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); | |
2703 | VERIFY(hdr->b_type == ARC_BUFC_DATA || | |
2704 | hdr->b_type == ARC_BUFC_METADATA); | |
524b4217 DK |
2705 | ASSERT3P(ret, !=, NULL); |
2706 | ASSERT3P(*ret, ==, NULL); | |
b5256303 | 2707 | IMPLY(encrypted, compressed); |
d3c2ae1c | 2708 | |
b9541d6b CW |
2709 | hdr->b_l1hdr.b_mru_hits = 0; |
2710 | hdr->b_l1hdr.b_mru_ghost_hits = 0; | |
2711 | hdr->b_l1hdr.b_mfu_hits = 0; | |
2712 | hdr->b_l1hdr.b_mfu_ghost_hits = 0; | |
2713 | hdr->b_l1hdr.b_l2_hits = 0; | |
2714 | ||
524b4217 | 2715 | buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); |
34dc7c2f BB |
2716 | buf->b_hdr = hdr; |
2717 | buf->b_data = NULL; | |
2aa34383 | 2718 | buf->b_next = hdr->b_l1hdr.b_buf; |
524b4217 | 2719 | buf->b_flags = 0; |
b9541d6b | 2720 | |
d3c2ae1c GW |
2721 | add_reference(hdr, tag); |
2722 | ||
2723 | /* | |
2724 | * We're about to change the hdr's b_flags. We must either | |
2725 | * hold the hash_lock or be undiscoverable. | |
2726 | */ | |
ca6c7a94 | 2727 | ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); |
d3c2ae1c GW |
2728 | |
2729 | /* | |
524b4217 | 2730 | * Only honor requests for compressed bufs if the hdr is actually |
e1cfd73f | 2731 | * compressed. This must be overridden if the buffer is encrypted since |
b5256303 | 2732 | * encrypted buffers cannot be decompressed. |
524b4217 | 2733 | */ |
b5256303 TC |
2734 | if (encrypted) { |
2735 | buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; | |
2736 | buf->b_flags |= ARC_BUF_FLAG_ENCRYPTED; | |
2737 | flags |= ARC_FILL_COMPRESSED | ARC_FILL_ENCRYPTED; | |
2738 | } else if (compressed && | |
2739 | arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) { | |
524b4217 | 2740 | buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; |
b5256303 TC |
2741 | flags |= ARC_FILL_COMPRESSED; |
2742 | } | |
2743 | ||
2744 | if (noauth) { | |
2745 | ASSERT0(encrypted); | |
2746 | flags |= ARC_FILL_NOAUTH; | |
2747 | } | |
524b4217 | 2748 | |
524b4217 DK |
2749 | /* |
2750 | * If the hdr's data can be shared then we share the data buffer and | |
2751 | * set the appropriate bit in the hdr's b_flags to indicate the hdr is | |
5662fd57 MA |
2752 | * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new |
2753 | * buffer to store the buf's data. | |
524b4217 | 2754 | * |
a6255b7f DQ |
2755 | * There are two additional restrictions here because we're sharing |
2756 | * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be | |
2757 | * actively involved in an L2ARC write, because if this buf is used by | |
2758 | * an arc_write() then the hdr's data buffer will be released when the | |
524b4217 | 2759 | * write completes, even though the L2ARC write might still be using it. |
a6255b7f | 2760 | * Second, the hdr's ABD must be linear so that the buf's user doesn't |
5662fd57 MA |
2761 | * need to be ABD-aware. It must be allocated via |
2762 | * zio_[data_]buf_alloc(), not as a page, because we need to be able | |
2763 | * to abd_release_ownership_of_buf(), which isn't allowed on "linear | |
2764 | * page" buffers because the ABD code needs to handle freeing them | |
2765 | * specially. | |
2766 | */ | |
2767 | boolean_t can_share = arc_can_share(hdr, buf) && | |
2768 | !HDR_L2_WRITING(hdr) && | |
2769 | hdr->b_l1hdr.b_pabd != NULL && | |
2770 | abd_is_linear(hdr->b_l1hdr.b_pabd) && | |
2771 | !abd_is_linear_page(hdr->b_l1hdr.b_pabd); | |
524b4217 DK |
2772 | |
2773 | /* Set up b_data and sharing */ | |
2774 | if (can_share) { | |
a6255b7f | 2775 | buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd); |
524b4217 | 2776 | buf->b_flags |= ARC_BUF_FLAG_SHARED; |
d3c2ae1c GW |
2777 | arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); |
2778 | } else { | |
524b4217 DK |
2779 | buf->b_data = |
2780 | arc_get_data_buf(hdr, arc_buf_size(buf), buf); | |
2781 | ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); | |
d3c2ae1c GW |
2782 | } |
2783 | VERIFY3P(buf->b_data, !=, NULL); | |
b9541d6b CW |
2784 | |
2785 | hdr->b_l1hdr.b_buf = buf; | |
d3c2ae1c | 2786 | hdr->b_l1hdr.b_bufcnt += 1; |
b5256303 TC |
2787 | if (encrypted) |
2788 | hdr->b_crypt_hdr.b_ebufcnt += 1; | |
b9541d6b | 2789 | |
524b4217 DK |
2790 | /* |
2791 | * If the user wants the data from the hdr, we need to either copy or | |
2792 | * decompress the data. | |
2793 | */ | |
2794 | if (fill) { | |
be9a5c35 TC |
2795 | ASSERT3P(zb, !=, NULL); |
2796 | return (arc_buf_fill(buf, spa, zb, flags)); | |
524b4217 | 2797 | } |
d3c2ae1c | 2798 | |
524b4217 | 2799 | return (0); |
34dc7c2f BB |
2800 | } |
2801 | ||
9babb374 BB |
2802 | static char *arc_onloan_tag = "onloan"; |
2803 | ||
a7004725 DK |
2804 | static inline void |
2805 | arc_loaned_bytes_update(int64_t delta) | |
2806 | { | |
2807 | atomic_add_64(&arc_loaned_bytes, delta); | |
2808 | ||
2809 | /* assert that it did not wrap around */ | |
2810 | ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); | |
2811 | } | |
2812 | ||
9babb374 BB |
2813 | /* |
2814 | * Loan out an anonymous arc buffer. Loaned buffers are not counted as in | |
2815 | * flight data by arc_tempreserve_space() until they are "returned". Loaned | |
2816 | * buffers must be returned to the arc before they can be used by the DMU or | |
2817 | * freed. | |
2818 | */ | |
2819 | arc_buf_t * | |
2aa34383 | 2820 | arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) |
9babb374 | 2821 | { |
2aa34383 DK |
2822 | arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, |
2823 | is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); | |
9babb374 | 2824 | |
5152a740 | 2825 | arc_loaned_bytes_update(arc_buf_size(buf)); |
a7004725 | 2826 | |
9babb374 BB |
2827 | return (buf); |
2828 | } | |
2829 | ||
2aa34383 DK |
2830 | arc_buf_t * |
2831 | arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, | |
2832 | enum zio_compress compression_type) | |
2833 | { | |
2834 | arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, | |
2835 | psize, lsize, compression_type); | |
2836 | ||
5152a740 | 2837 | arc_loaned_bytes_update(arc_buf_size(buf)); |
a7004725 | 2838 | |
2aa34383 DK |
2839 | return (buf); |
2840 | } | |
2841 | ||
b5256303 TC |
2842 | arc_buf_t * |
2843 | arc_loan_raw_buf(spa_t *spa, uint64_t dsobj, boolean_t byteorder, | |
2844 | const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, | |
2845 | dmu_object_type_t ot, uint64_t psize, uint64_t lsize, | |
2846 | enum zio_compress compression_type) | |
2847 | { | |
2848 | arc_buf_t *buf = arc_alloc_raw_buf(spa, arc_onloan_tag, dsobj, | |
2849 | byteorder, salt, iv, mac, ot, psize, lsize, compression_type); | |
2850 | ||
2851 | atomic_add_64(&arc_loaned_bytes, psize); | |
2852 | return (buf); | |
2853 | } | |
2854 | ||
2aa34383 | 2855 | |
9babb374 BB |
2856 | /* |
2857 | * Return a loaned arc buffer to the arc. | |
2858 | */ | |
2859 | void | |
2860 | arc_return_buf(arc_buf_t *buf, void *tag) | |
2861 | { | |
2862 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
2863 | ||
d3c2ae1c | 2864 | ASSERT3P(buf->b_data, !=, NULL); |
b9541d6b | 2865 | ASSERT(HDR_HAS_L1HDR(hdr)); |
c13060e4 | 2866 | (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag); |
424fd7c3 | 2867 | (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); |
9babb374 | 2868 | |
a7004725 | 2869 | arc_loaned_bytes_update(-arc_buf_size(buf)); |
9babb374 BB |
2870 | } |
2871 | ||
428870ff BB |
2872 | /* Detach an arc_buf from a dbuf (tag) */ |
2873 | void | |
2874 | arc_loan_inuse_buf(arc_buf_t *buf, void *tag) | |
2875 | { | |
b9541d6b | 2876 | arc_buf_hdr_t *hdr = buf->b_hdr; |
428870ff | 2877 | |
d3c2ae1c | 2878 | ASSERT3P(buf->b_data, !=, NULL); |
b9541d6b | 2879 | ASSERT(HDR_HAS_L1HDR(hdr)); |
c13060e4 | 2880 | (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); |
424fd7c3 | 2881 | (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); |
428870ff | 2882 | |
a7004725 | 2883 | arc_loaned_bytes_update(arc_buf_size(buf)); |
428870ff BB |
2884 | } |
2885 | ||
d3c2ae1c | 2886 | static void |
a6255b7f | 2887 | l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type) |
34dc7c2f | 2888 | { |
d3c2ae1c | 2889 | l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); |
34dc7c2f | 2890 | |
a6255b7f | 2891 | df->l2df_abd = abd; |
d3c2ae1c GW |
2892 | df->l2df_size = size; |
2893 | df->l2df_type = type; | |
2894 | mutex_enter(&l2arc_free_on_write_mtx); | |
2895 | list_insert_head(l2arc_free_on_write, df); | |
2896 | mutex_exit(&l2arc_free_on_write_mtx); | |
2897 | } | |
428870ff | 2898 | |
d3c2ae1c | 2899 | static void |
b5256303 | 2900 | arc_hdr_free_on_write(arc_buf_hdr_t *hdr, boolean_t free_rdata) |
d3c2ae1c GW |
2901 | { |
2902 | arc_state_t *state = hdr->b_l1hdr.b_state; | |
2903 | arc_buf_contents_t type = arc_buf_type(hdr); | |
b5256303 | 2904 | uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr); |
1eb5bfa3 | 2905 | |
d3c2ae1c GW |
2906 | /* protected by hash lock, if in the hash table */ |
2907 | if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { | |
424fd7c3 | 2908 | ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
d3c2ae1c GW |
2909 | ASSERT(state != arc_anon && state != arc_l2c_only); |
2910 | ||
424fd7c3 | 2911 | (void) zfs_refcount_remove_many(&state->arcs_esize[type], |
d3c2ae1c | 2912 | size, hdr); |
1eb5bfa3 | 2913 | } |
424fd7c3 | 2914 | (void) zfs_refcount_remove_many(&state->arcs_size, size, hdr); |
423e7b62 AG |
2915 | if (type == ARC_BUFC_METADATA) { |
2916 | arc_space_return(size, ARC_SPACE_META); | |
2917 | } else { | |
2918 | ASSERT(type == ARC_BUFC_DATA); | |
2919 | arc_space_return(size, ARC_SPACE_DATA); | |
2920 | } | |
d3c2ae1c | 2921 | |
b5256303 TC |
2922 | if (free_rdata) { |
2923 | l2arc_free_abd_on_write(hdr->b_crypt_hdr.b_rabd, size, type); | |
2924 | } else { | |
2925 | l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type); | |
2926 | } | |
34dc7c2f BB |
2927 | } |
2928 | ||
d3c2ae1c GW |
2929 | /* |
2930 | * Share the arc_buf_t's data with the hdr. Whenever we are sharing the | |
2931 | * data buffer, we transfer the refcount ownership to the hdr and update | |
2932 | * the appropriate kstats. | |
2933 | */ | |
2934 | static void | |
2935 | arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) | |
34dc7c2f | 2936 | { |
524b4217 | 2937 | ASSERT(arc_can_share(hdr, buf)); |
a6255b7f | 2938 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 2939 | ASSERT(!ARC_BUF_ENCRYPTED(buf)); |
ca6c7a94 | 2940 | ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); |
34dc7c2f BB |
2941 | |
2942 | /* | |
d3c2ae1c GW |
2943 | * Start sharing the data buffer. We transfer the |
2944 | * refcount ownership to the hdr since it always owns | |
2945 | * the refcount whenever an arc_buf_t is shared. | |
34dc7c2f | 2946 | */ |
d7e4b30a BB |
2947 | zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size, |
2948 | arc_hdr_size(hdr), buf, hdr); | |
a6255b7f DQ |
2949 | hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf)); |
2950 | abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd, | |
2951 | HDR_ISTYPE_METADATA(hdr)); | |
d3c2ae1c | 2952 | arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); |
524b4217 | 2953 | buf->b_flags |= ARC_BUF_FLAG_SHARED; |
34dc7c2f | 2954 | |
d3c2ae1c GW |
2955 | /* |
2956 | * Since we've transferred ownership to the hdr we need | |
2957 | * to increment its compressed and uncompressed kstats and | |
2958 | * decrement the overhead size. | |
2959 | */ | |
2960 | ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); | |
2961 | ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); | |
2aa34383 | 2962 | ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); |
34dc7c2f BB |
2963 | } |
2964 | ||
ca0bf58d | 2965 | static void |
d3c2ae1c | 2966 | arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) |
ca0bf58d | 2967 | { |
d3c2ae1c | 2968 | ASSERT(arc_buf_is_shared(buf)); |
a6255b7f | 2969 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); |
ca6c7a94 | 2970 | ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); |
ca0bf58d | 2971 | |
d3c2ae1c GW |
2972 | /* |
2973 | * We are no longer sharing this buffer so we need | |
2974 | * to transfer its ownership to the rightful owner. | |
2975 | */ | |
d7e4b30a BB |
2976 | zfs_refcount_transfer_ownership_many(&hdr->b_l1hdr.b_state->arcs_size, |
2977 | arc_hdr_size(hdr), hdr, buf); | |
d3c2ae1c | 2978 | arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); |
a6255b7f DQ |
2979 | abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd); |
2980 | abd_put(hdr->b_l1hdr.b_pabd); | |
2981 | hdr->b_l1hdr.b_pabd = NULL; | |
524b4217 | 2982 | buf->b_flags &= ~ARC_BUF_FLAG_SHARED; |
d3c2ae1c GW |
2983 | |
2984 | /* | |
2985 | * Since the buffer is no longer shared between | |
2986 | * the arc buf and the hdr, count it as overhead. | |
2987 | */ | |
2988 | ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); | |
2989 | ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); | |
2aa34383 | 2990 | ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); |
ca0bf58d PS |
2991 | } |
2992 | ||
34dc7c2f | 2993 | /* |
2aa34383 DK |
2994 | * Remove an arc_buf_t from the hdr's buf list and return the last |
2995 | * arc_buf_t on the list. If no buffers remain on the list then return | |
2996 | * NULL. | |
2997 | */ | |
2998 | static arc_buf_t * | |
2999 | arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) | |
3000 | { | |
2aa34383 | 3001 | ASSERT(HDR_HAS_L1HDR(hdr)); |
ca6c7a94 | 3002 | ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); |
2aa34383 | 3003 | |
a7004725 DK |
3004 | arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; |
3005 | arc_buf_t *lastbuf = NULL; | |
3006 | ||
2aa34383 DK |
3007 | /* |
3008 | * Remove the buf from the hdr list and locate the last | |
3009 | * remaining buffer on the list. | |
3010 | */ | |
3011 | while (*bufp != NULL) { | |
3012 | if (*bufp == buf) | |
3013 | *bufp = buf->b_next; | |
3014 | ||
3015 | /* | |
3016 | * If we've removed a buffer in the middle of | |
3017 | * the list then update the lastbuf and update | |
3018 | * bufp. | |
3019 | */ | |
3020 | if (*bufp != NULL) { | |
3021 | lastbuf = *bufp; | |
3022 | bufp = &(*bufp)->b_next; | |
3023 | } | |
3024 | } | |
3025 | buf->b_next = NULL; | |
3026 | ASSERT3P(lastbuf, !=, buf); | |
3027 | IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); | |
3028 | IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); | |
3029 | IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); | |
3030 | ||
3031 | return (lastbuf); | |
3032 | } | |
3033 | ||
3034 | /* | |
e1cfd73f | 3035 | * Free up buf->b_data and pull the arc_buf_t off of the arc_buf_hdr_t's |
2aa34383 | 3036 | * list and free it. |
34dc7c2f BB |
3037 | */ |
3038 | static void | |
2aa34383 | 3039 | arc_buf_destroy_impl(arc_buf_t *buf) |
34dc7c2f | 3040 | { |
498877ba | 3041 | arc_buf_hdr_t *hdr = buf->b_hdr; |
ca0bf58d PS |
3042 | |
3043 | /* | |
524b4217 DK |
3044 | * Free up the data associated with the buf but only if we're not |
3045 | * sharing this with the hdr. If we are sharing it with the hdr, the | |
3046 | * hdr is responsible for doing the free. | |
ca0bf58d | 3047 | */ |
d3c2ae1c GW |
3048 | if (buf->b_data != NULL) { |
3049 | /* | |
3050 | * We're about to change the hdr's b_flags. We must either | |
3051 | * hold the hash_lock or be undiscoverable. | |
3052 | */ | |
ca6c7a94 | 3053 | ASSERT(HDR_EMPTY_OR_LOCKED(hdr)); |
d3c2ae1c | 3054 | |
524b4217 | 3055 | arc_cksum_verify(buf); |
d3c2ae1c GW |
3056 | arc_buf_unwatch(buf); |
3057 | ||
2aa34383 | 3058 | if (arc_buf_is_shared(buf)) { |
d3c2ae1c GW |
3059 | arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); |
3060 | } else { | |
2aa34383 | 3061 | uint64_t size = arc_buf_size(buf); |
d3c2ae1c GW |
3062 | arc_free_data_buf(hdr, buf->b_data, size, buf); |
3063 | ARCSTAT_INCR(arcstat_overhead_size, -size); | |
3064 | } | |
3065 | buf->b_data = NULL; | |
3066 | ||
3067 | ASSERT(hdr->b_l1hdr.b_bufcnt > 0); | |
3068 | hdr->b_l1hdr.b_bufcnt -= 1; | |
b5256303 | 3069 | |
da5d4697 | 3070 | if (ARC_BUF_ENCRYPTED(buf)) { |
b5256303 TC |
3071 | hdr->b_crypt_hdr.b_ebufcnt -= 1; |
3072 | ||
da5d4697 D |
3073 | /* |
3074 | * If we have no more encrypted buffers and we've | |
3075 | * already gotten a copy of the decrypted data we can | |
3076 | * free b_rabd to save some space. | |
3077 | */ | |
3078 | if (hdr->b_crypt_hdr.b_ebufcnt == 0 && | |
3079 | HDR_HAS_RABD(hdr) && hdr->b_l1hdr.b_pabd != NULL && | |
3080 | !HDR_IO_IN_PROGRESS(hdr)) { | |
3081 | arc_hdr_free_abd(hdr, B_TRUE); | |
3082 | } | |
440a3eb9 | 3083 | } |
d3c2ae1c GW |
3084 | } |
3085 | ||
a7004725 | 3086 | arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); |
d3c2ae1c | 3087 | |
524b4217 | 3088 | if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { |
2aa34383 | 3089 | /* |
524b4217 | 3090 | * If the current arc_buf_t is sharing its data buffer with the |
a6255b7f | 3091 | * hdr, then reassign the hdr's b_pabd to share it with the new |
524b4217 DK |
3092 | * buffer at the end of the list. The shared buffer is always |
3093 | * the last one on the hdr's buffer list. | |
3094 | * | |
3095 | * There is an equivalent case for compressed bufs, but since | |
3096 | * they aren't guaranteed to be the last buf in the list and | |
3097 | * that is an exceedingly rare case, we just allow that space be | |
b5256303 TC |
3098 | * wasted temporarily. We must also be careful not to share |
3099 | * encrypted buffers, since they cannot be shared. | |
2aa34383 | 3100 | */ |
b5256303 | 3101 | if (lastbuf != NULL && !ARC_BUF_ENCRYPTED(lastbuf)) { |
524b4217 | 3102 | /* Only one buf can be shared at once */ |
2aa34383 | 3103 | VERIFY(!arc_buf_is_shared(lastbuf)); |
524b4217 DK |
3104 | /* hdr is uncompressed so can't have compressed buf */ |
3105 | VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); | |
d3c2ae1c | 3106 | |
a6255b7f | 3107 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); |
b5256303 | 3108 | arc_hdr_free_abd(hdr, B_FALSE); |
d3c2ae1c | 3109 | |
2aa34383 DK |
3110 | /* |
3111 | * We must setup a new shared block between the | |
3112 | * last buffer and the hdr. The data would have | |
3113 | * been allocated by the arc buf so we need to transfer | |
3114 | * ownership to the hdr since it's now being shared. | |
3115 | */ | |
3116 | arc_share_buf(hdr, lastbuf); | |
3117 | } | |
3118 | } else if (HDR_SHARED_DATA(hdr)) { | |
d3c2ae1c | 3119 | /* |
2aa34383 DK |
3120 | * Uncompressed shared buffers are always at the end |
3121 | * of the list. Compressed buffers don't have the | |
3122 | * same requirements. This makes it hard to | |
3123 | * simply assert that the lastbuf is shared so | |
3124 | * we rely on the hdr's compression flags to determine | |
3125 | * if we have a compressed, shared buffer. | |
d3c2ae1c | 3126 | */ |
2aa34383 DK |
3127 | ASSERT3P(lastbuf, !=, NULL); |
3128 | ASSERT(arc_buf_is_shared(lastbuf) || | |
b5256303 | 3129 | arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF); |
ca0bf58d PS |
3130 | } |
3131 | ||
a7004725 DK |
3132 | /* |
3133 | * Free the checksum if we're removing the last uncompressed buf from | |
3134 | * this hdr. | |
3135 | */ | |
3136 | if (!arc_hdr_has_uncompressed_buf(hdr)) { | |
d3c2ae1c | 3137 | arc_cksum_free(hdr); |
a7004725 | 3138 | } |
d3c2ae1c GW |
3139 | |
3140 | /* clean up the buf */ | |
3141 | buf->b_hdr = NULL; | |
3142 | kmem_cache_free(buf_cache, buf); | |
3143 | } | |
3144 | ||
3145 | static void | |
b5256303 | 3146 | arc_hdr_alloc_abd(arc_buf_hdr_t *hdr, boolean_t alloc_rdata) |
d3c2ae1c | 3147 | { |
b5256303 TC |
3148 | uint64_t size; |
3149 | ||
d3c2ae1c GW |
3150 | ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); |
3151 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
b5256303 TC |
3152 | ASSERT(!HDR_SHARED_DATA(hdr) || alloc_rdata); |
3153 | IMPLY(alloc_rdata, HDR_PROTECTED(hdr)); | |
d3c2ae1c | 3154 | |
b5256303 TC |
3155 | if (alloc_rdata) { |
3156 | size = HDR_GET_PSIZE(hdr); | |
3157 | ASSERT3P(hdr->b_crypt_hdr.b_rabd, ==, NULL); | |
3158 | hdr->b_crypt_hdr.b_rabd = arc_get_data_abd(hdr, size, hdr); | |
3159 | ASSERT3P(hdr->b_crypt_hdr.b_rabd, !=, NULL); | |
3160 | ARCSTAT_INCR(arcstat_raw_size, size); | |
3161 | } else { | |
3162 | size = arc_hdr_size(hdr); | |
3163 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); | |
3164 | hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, size, hdr); | |
3165 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); | |
3166 | } | |
3167 | ||
3168 | ARCSTAT_INCR(arcstat_compressed_size, size); | |
d3c2ae1c GW |
3169 | ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); |
3170 | } | |
3171 | ||
3172 | static void | |
b5256303 | 3173 | arc_hdr_free_abd(arc_buf_hdr_t *hdr, boolean_t free_rdata) |
d3c2ae1c | 3174 | { |
b5256303 TC |
3175 | uint64_t size = (free_rdata) ? HDR_GET_PSIZE(hdr) : arc_hdr_size(hdr); |
3176 | ||
d3c2ae1c | 3177 | ASSERT(HDR_HAS_L1HDR(hdr)); |
b5256303 TC |
3178 | ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); |
3179 | IMPLY(free_rdata, HDR_HAS_RABD(hdr)); | |
d3c2ae1c | 3180 | |
ca0bf58d | 3181 | /* |
d3c2ae1c GW |
3182 | * If the hdr is currently being written to the l2arc then |
3183 | * we defer freeing the data by adding it to the l2arc_free_on_write | |
3184 | * list. The l2arc will free the data once it's finished | |
3185 | * writing it to the l2arc device. | |
ca0bf58d | 3186 | */ |
d3c2ae1c | 3187 | if (HDR_L2_WRITING(hdr)) { |
b5256303 | 3188 | arc_hdr_free_on_write(hdr, free_rdata); |
d3c2ae1c | 3189 | ARCSTAT_BUMP(arcstat_l2_free_on_write); |
b5256303 TC |
3190 | } else if (free_rdata) { |
3191 | arc_free_data_abd(hdr, hdr->b_crypt_hdr.b_rabd, size, hdr); | |
d3c2ae1c | 3192 | } else { |
b5256303 | 3193 | arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, size, hdr); |
ca0bf58d PS |
3194 | } |
3195 | ||
b5256303 TC |
3196 | if (free_rdata) { |
3197 | hdr->b_crypt_hdr.b_rabd = NULL; | |
3198 | ARCSTAT_INCR(arcstat_raw_size, -size); | |
3199 | } else { | |
3200 | hdr->b_l1hdr.b_pabd = NULL; | |
3201 | } | |
3202 | ||
3203 | if (hdr->b_l1hdr.b_pabd == NULL && !HDR_HAS_RABD(hdr)) | |
3204 | hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; | |
3205 | ||
3206 | ARCSTAT_INCR(arcstat_compressed_size, -size); | |
d3c2ae1c GW |
3207 | ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); |
3208 | } | |
3209 | ||
3210 | static arc_buf_hdr_t * | |
3211 | arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, | |
b5256303 TC |
3212 | boolean_t protected, enum zio_compress compression_type, |
3213 | arc_buf_contents_t type, boolean_t alloc_rdata) | |
d3c2ae1c GW |
3214 | { |
3215 | arc_buf_hdr_t *hdr; | |
3216 | ||
d3c2ae1c | 3217 | VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); |
b5256303 TC |
3218 | if (protected) { |
3219 | hdr = kmem_cache_alloc(hdr_full_crypt_cache, KM_PUSHPAGE); | |
3220 | } else { | |
3221 | hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); | |
3222 | } | |
d3c2ae1c | 3223 | |
d3c2ae1c GW |
3224 | ASSERT(HDR_EMPTY(hdr)); |
3225 | ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); | |
3226 | HDR_SET_PSIZE(hdr, psize); | |
3227 | HDR_SET_LSIZE(hdr, lsize); | |
3228 | hdr->b_spa = spa; | |
3229 | hdr->b_type = type; | |
3230 | hdr->b_flags = 0; | |
3231 | arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); | |
2aa34383 | 3232 | arc_hdr_set_compress(hdr, compression_type); |
b5256303 TC |
3233 | if (protected) |
3234 | arc_hdr_set_flags(hdr, ARC_FLAG_PROTECTED); | |
ca0bf58d | 3235 | |
d3c2ae1c GW |
3236 | hdr->b_l1hdr.b_state = arc_anon; |
3237 | hdr->b_l1hdr.b_arc_access = 0; | |
3238 | hdr->b_l1hdr.b_bufcnt = 0; | |
3239 | hdr->b_l1hdr.b_buf = NULL; | |
ca0bf58d | 3240 | |
d3c2ae1c GW |
3241 | /* |
3242 | * Allocate the hdr's buffer. This will contain either | |
3243 | * the compressed or uncompressed data depending on the block | |
3244 | * it references and compressed arc enablement. | |
3245 | */ | |
b5256303 | 3246 | arc_hdr_alloc_abd(hdr, alloc_rdata); |
424fd7c3 | 3247 | ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
ca0bf58d | 3248 | |
d3c2ae1c | 3249 | return (hdr); |
ca0bf58d PS |
3250 | } |
3251 | ||
bd089c54 | 3252 | /* |
d3c2ae1c GW |
3253 | * Transition between the two allocation states for the arc_buf_hdr struct. |
3254 | * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without | |
3255 | * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller | |
3256 | * version is used when a cache buffer is only in the L2ARC in order to reduce | |
3257 | * memory usage. | |
bd089c54 | 3258 | */ |
d3c2ae1c GW |
3259 | static arc_buf_hdr_t * |
3260 | arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) | |
34dc7c2f | 3261 | { |
1c27024e DB |
3262 | ASSERT(HDR_HAS_L2HDR(hdr)); |
3263 | ||
d3c2ae1c GW |
3264 | arc_buf_hdr_t *nhdr; |
3265 | l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; | |
34dc7c2f | 3266 | |
d3c2ae1c GW |
3267 | ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || |
3268 | (old == hdr_l2only_cache && new == hdr_full_cache)); | |
34dc7c2f | 3269 | |
b5256303 TC |
3270 | /* |
3271 | * if the caller wanted a new full header and the header is to be | |
3272 | * encrypted we will actually allocate the header from the full crypt | |
3273 | * cache instead. The same applies to freeing from the old cache. | |
3274 | */ | |
3275 | if (HDR_PROTECTED(hdr) && new == hdr_full_cache) | |
3276 | new = hdr_full_crypt_cache; | |
3277 | if (HDR_PROTECTED(hdr) && old == hdr_full_cache) | |
3278 | old = hdr_full_crypt_cache; | |
3279 | ||
d3c2ae1c | 3280 | nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); |
428870ff | 3281 | |
d3c2ae1c GW |
3282 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); |
3283 | buf_hash_remove(hdr); | |
ca0bf58d | 3284 | |
d3c2ae1c | 3285 | bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); |
34dc7c2f | 3286 | |
b5256303 | 3287 | if (new == hdr_full_cache || new == hdr_full_crypt_cache) { |
d3c2ae1c GW |
3288 | arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); |
3289 | /* | |
3290 | * arc_access and arc_change_state need to be aware that a | |
3291 | * header has just come out of L2ARC, so we set its state to | |
3292 | * l2c_only even though it's about to change. | |
3293 | */ | |
3294 | nhdr->b_l1hdr.b_state = arc_l2c_only; | |
34dc7c2f | 3295 | |
d3c2ae1c | 3296 | /* Verify previous threads set to NULL before freeing */ |
a6255b7f | 3297 | ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 3298 | ASSERT(!HDR_HAS_RABD(hdr)); |
d3c2ae1c GW |
3299 | } else { |
3300 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
3301 | ASSERT0(hdr->b_l1hdr.b_bufcnt); | |
3302 | ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); | |
36da08ef | 3303 | |
d3c2ae1c GW |
3304 | /* |
3305 | * If we've reached here, We must have been called from | |
3306 | * arc_evict_hdr(), as such we should have already been | |
3307 | * removed from any ghost list we were previously on | |
3308 | * (which protects us from racing with arc_evict_state), | |
3309 | * thus no locking is needed during this check. | |
3310 | */ | |
3311 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); | |
1eb5bfa3 GW |
3312 | |
3313 | /* | |
d3c2ae1c GW |
3314 | * A buffer must not be moved into the arc_l2c_only |
3315 | * state if it's not finished being written out to the | |
a6255b7f | 3316 | * l2arc device. Otherwise, the b_l1hdr.b_pabd field |
d3c2ae1c | 3317 | * might try to be accessed, even though it was removed. |
1eb5bfa3 | 3318 | */ |
d3c2ae1c | 3319 | VERIFY(!HDR_L2_WRITING(hdr)); |
a6255b7f | 3320 | VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 3321 | ASSERT(!HDR_HAS_RABD(hdr)); |
d3c2ae1c GW |
3322 | |
3323 | arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); | |
34dc7c2f | 3324 | } |
d3c2ae1c GW |
3325 | /* |
3326 | * The header has been reallocated so we need to re-insert it into any | |
3327 | * lists it was on. | |
3328 | */ | |
3329 | (void) buf_hash_insert(nhdr, NULL); | |
34dc7c2f | 3330 | |
d3c2ae1c | 3331 | ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); |
34dc7c2f | 3332 | |
d3c2ae1c GW |
3333 | mutex_enter(&dev->l2ad_mtx); |
3334 | ||
3335 | /* | |
3336 | * We must place the realloc'ed header back into the list at | |
3337 | * the same spot. Otherwise, if it's placed earlier in the list, | |
3338 | * l2arc_write_buffers() could find it during the function's | |
3339 | * write phase, and try to write it out to the l2arc. | |
3340 | */ | |
3341 | list_insert_after(&dev->l2ad_buflist, hdr, nhdr); | |
3342 | list_remove(&dev->l2ad_buflist, hdr); | |
34dc7c2f | 3343 | |
d3c2ae1c | 3344 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f | 3345 | |
d3c2ae1c GW |
3346 | /* |
3347 | * Since we're using the pointer address as the tag when | |
3348 | * incrementing and decrementing the l2ad_alloc refcount, we | |
3349 | * must remove the old pointer (that we're about to destroy) and | |
3350 | * add the new pointer to the refcount. Otherwise we'd remove | |
3351 | * the wrong pointer address when calling arc_hdr_destroy() later. | |
3352 | */ | |
3353 | ||
424fd7c3 TS |
3354 | (void) zfs_refcount_remove_many(&dev->l2ad_alloc, |
3355 | arc_hdr_size(hdr), hdr); | |
3356 | (void) zfs_refcount_add_many(&dev->l2ad_alloc, | |
3357 | arc_hdr_size(nhdr), nhdr); | |
d3c2ae1c GW |
3358 | |
3359 | buf_discard_identity(hdr); | |
3360 | kmem_cache_free(old, hdr); | |
3361 | ||
3362 | return (nhdr); | |
3363 | } | |
3364 | ||
b5256303 TC |
3365 | /* |
3366 | * This function allows an L1 header to be reallocated as a crypt | |
3367 | * header and vice versa. If we are going to a crypt header, the | |
3368 | * new fields will be zeroed out. | |
3369 | */ | |
3370 | static arc_buf_hdr_t * | |
3371 | arc_hdr_realloc_crypt(arc_buf_hdr_t *hdr, boolean_t need_crypt) | |
3372 | { | |
3373 | arc_buf_hdr_t *nhdr; | |
3374 | arc_buf_t *buf; | |
3375 | kmem_cache_t *ncache, *ocache; | |
b7ddeaef | 3376 | unsigned nsize, osize; |
b5256303 | 3377 | |
b7ddeaef TC |
3378 | /* |
3379 | * This function requires that hdr is in the arc_anon state. | |
3380 | * Therefore it won't have any L2ARC data for us to worry | |
3381 | * about copying. | |
3382 | */ | |
b5256303 | 3383 | ASSERT(HDR_HAS_L1HDR(hdr)); |
b7ddeaef | 3384 | ASSERT(!HDR_HAS_L2HDR(hdr)); |
b5256303 TC |
3385 | ASSERT3U(!!HDR_PROTECTED(hdr), !=, need_crypt); |
3386 | ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); | |
3387 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); | |
b7ddeaef TC |
3388 | ASSERT(!list_link_active(&hdr->b_l2hdr.b_l2node)); |
3389 | ASSERT3P(hdr->b_hash_next, ==, NULL); | |
b5256303 TC |
3390 | |
3391 | if (need_crypt) { | |
3392 | ncache = hdr_full_crypt_cache; | |
b7ddeaef | 3393 | nsize = sizeof (hdr->b_crypt_hdr); |
b5256303 | 3394 | ocache = hdr_full_cache; |
b7ddeaef | 3395 | osize = HDR_FULL_SIZE; |
b5256303 TC |
3396 | } else { |
3397 | ncache = hdr_full_cache; | |
b7ddeaef | 3398 | nsize = HDR_FULL_SIZE; |
b5256303 | 3399 | ocache = hdr_full_crypt_cache; |
b7ddeaef | 3400 | osize = sizeof (hdr->b_crypt_hdr); |
b5256303 TC |
3401 | } |
3402 | ||
3403 | nhdr = kmem_cache_alloc(ncache, KM_PUSHPAGE); | |
b7ddeaef TC |
3404 | |
3405 | /* | |
3406 | * Copy all members that aren't locks or condvars to the new header. | |
3407 | * No lists are pointing to us (as we asserted above), so we don't | |
3408 | * need to worry about the list nodes. | |
3409 | */ | |
3410 | nhdr->b_dva = hdr->b_dva; | |
3411 | nhdr->b_birth = hdr->b_birth; | |
3412 | nhdr->b_type = hdr->b_type; | |
3413 | nhdr->b_flags = hdr->b_flags; | |
3414 | nhdr->b_psize = hdr->b_psize; | |
3415 | nhdr->b_lsize = hdr->b_lsize; | |
3416 | nhdr->b_spa = hdr->b_spa; | |
b5256303 TC |
3417 | nhdr->b_l1hdr.b_freeze_cksum = hdr->b_l1hdr.b_freeze_cksum; |
3418 | nhdr->b_l1hdr.b_bufcnt = hdr->b_l1hdr.b_bufcnt; | |
3419 | nhdr->b_l1hdr.b_byteswap = hdr->b_l1hdr.b_byteswap; | |
3420 | nhdr->b_l1hdr.b_state = hdr->b_l1hdr.b_state; | |
3421 | nhdr->b_l1hdr.b_arc_access = hdr->b_l1hdr.b_arc_access; | |
3422 | nhdr->b_l1hdr.b_mru_hits = hdr->b_l1hdr.b_mru_hits; | |
3423 | nhdr->b_l1hdr.b_mru_ghost_hits = hdr->b_l1hdr.b_mru_ghost_hits; | |
3424 | nhdr->b_l1hdr.b_mfu_hits = hdr->b_l1hdr.b_mfu_hits; | |
3425 | nhdr->b_l1hdr.b_mfu_ghost_hits = hdr->b_l1hdr.b_mfu_ghost_hits; | |
3426 | nhdr->b_l1hdr.b_l2_hits = hdr->b_l1hdr.b_l2_hits; | |
3427 | nhdr->b_l1hdr.b_acb = hdr->b_l1hdr.b_acb; | |
3428 | nhdr->b_l1hdr.b_pabd = hdr->b_l1hdr.b_pabd; | |
b5256303 TC |
3429 | |
3430 | /* | |
c13060e4 | 3431 | * This zfs_refcount_add() exists only to ensure that the individual |
b5256303 TC |
3432 | * arc buffers always point to a header that is referenced, avoiding |
3433 | * a small race condition that could trigger ASSERTs. | |
3434 | */ | |
c13060e4 | 3435 | (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, FTAG); |
b7ddeaef | 3436 | nhdr->b_l1hdr.b_buf = hdr->b_l1hdr.b_buf; |
b5256303 TC |
3437 | for (buf = nhdr->b_l1hdr.b_buf; buf != NULL; buf = buf->b_next) { |
3438 | mutex_enter(&buf->b_evict_lock); | |
3439 | buf->b_hdr = nhdr; | |
3440 | mutex_exit(&buf->b_evict_lock); | |
3441 | } | |
3442 | ||
424fd7c3 TS |
3443 | zfs_refcount_transfer(&nhdr->b_l1hdr.b_refcnt, &hdr->b_l1hdr.b_refcnt); |
3444 | (void) zfs_refcount_remove(&nhdr->b_l1hdr.b_refcnt, FTAG); | |
3445 | ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt)); | |
b5256303 TC |
3446 | |
3447 | if (need_crypt) { | |
3448 | arc_hdr_set_flags(nhdr, ARC_FLAG_PROTECTED); | |
3449 | } else { | |
3450 | arc_hdr_clear_flags(nhdr, ARC_FLAG_PROTECTED); | |
3451 | } | |
3452 | ||
b7ddeaef TC |
3453 | /* unset all members of the original hdr */ |
3454 | bzero(&hdr->b_dva, sizeof (dva_t)); | |
3455 | hdr->b_birth = 0; | |
3456 | hdr->b_type = ARC_BUFC_INVALID; | |
3457 | hdr->b_flags = 0; | |
3458 | hdr->b_psize = 0; | |
3459 | hdr->b_lsize = 0; | |
3460 | hdr->b_spa = 0; | |
3461 | hdr->b_l1hdr.b_freeze_cksum = NULL; | |
3462 | hdr->b_l1hdr.b_buf = NULL; | |
3463 | hdr->b_l1hdr.b_bufcnt = 0; | |
3464 | hdr->b_l1hdr.b_byteswap = 0; | |
3465 | hdr->b_l1hdr.b_state = NULL; | |
3466 | hdr->b_l1hdr.b_arc_access = 0; | |
3467 | hdr->b_l1hdr.b_mru_hits = 0; | |
3468 | hdr->b_l1hdr.b_mru_ghost_hits = 0; | |
3469 | hdr->b_l1hdr.b_mfu_hits = 0; | |
3470 | hdr->b_l1hdr.b_mfu_ghost_hits = 0; | |
3471 | hdr->b_l1hdr.b_l2_hits = 0; | |
3472 | hdr->b_l1hdr.b_acb = NULL; | |
3473 | hdr->b_l1hdr.b_pabd = NULL; | |
3474 | ||
3475 | if (ocache == hdr_full_crypt_cache) { | |
3476 | ASSERT(!HDR_HAS_RABD(hdr)); | |
3477 | hdr->b_crypt_hdr.b_ot = DMU_OT_NONE; | |
3478 | hdr->b_crypt_hdr.b_ebufcnt = 0; | |
3479 | hdr->b_crypt_hdr.b_dsobj = 0; | |
3480 | bzero(hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN); | |
3481 | bzero(hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN); | |
3482 | bzero(hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN); | |
3483 | } | |
3484 | ||
b5256303 TC |
3485 | buf_discard_identity(hdr); |
3486 | kmem_cache_free(ocache, hdr); | |
3487 | ||
3488 | return (nhdr); | |
3489 | } | |
3490 | ||
3491 | /* | |
3492 | * This function is used by the send / receive code to convert a newly | |
3493 | * allocated arc_buf_t to one that is suitable for a raw encrypted write. It | |
e1cfd73f | 3494 | * is also used to allow the root objset block to be updated without altering |
b5256303 TC |
3495 | * its embedded MACs. Both block types will always be uncompressed so we do not |
3496 | * have to worry about compression type or psize. | |
3497 | */ | |
3498 | void | |
3499 | arc_convert_to_raw(arc_buf_t *buf, uint64_t dsobj, boolean_t byteorder, | |
3500 | dmu_object_type_t ot, const uint8_t *salt, const uint8_t *iv, | |
3501 | const uint8_t *mac) | |
3502 | { | |
3503 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
3504 | ||
3505 | ASSERT(ot == DMU_OT_DNODE || ot == DMU_OT_OBJSET); | |
3506 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
3507 | ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); | |
3508 | ||
3509 | buf->b_flags |= (ARC_BUF_FLAG_COMPRESSED | ARC_BUF_FLAG_ENCRYPTED); | |
3510 | if (!HDR_PROTECTED(hdr)) | |
3511 | hdr = arc_hdr_realloc_crypt(hdr, B_TRUE); | |
3512 | hdr->b_crypt_hdr.b_dsobj = dsobj; | |
3513 | hdr->b_crypt_hdr.b_ot = ot; | |
3514 | hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ? | |
3515 | DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot); | |
3516 | if (!arc_hdr_has_uncompressed_buf(hdr)) | |
3517 | arc_cksum_free(hdr); | |
3518 | ||
3519 | if (salt != NULL) | |
3520 | bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN); | |
3521 | if (iv != NULL) | |
3522 | bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN); | |
3523 | if (mac != NULL) | |
3524 | bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN); | |
3525 | } | |
3526 | ||
d3c2ae1c GW |
3527 | /* |
3528 | * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. | |
3529 | * The buf is returned thawed since we expect the consumer to modify it. | |
3530 | */ | |
3531 | arc_buf_t * | |
2aa34383 | 3532 | arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) |
d3c2ae1c | 3533 | { |
d3c2ae1c | 3534 | arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, |
b5256303 | 3535 | B_FALSE, ZIO_COMPRESS_OFF, type, B_FALSE); |
2aa34383 | 3536 | |
a7004725 | 3537 | arc_buf_t *buf = NULL; |
be9a5c35 | 3538 | VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, B_FALSE, |
b5256303 | 3539 | B_FALSE, B_FALSE, &buf)); |
d3c2ae1c | 3540 | arc_buf_thaw(buf); |
2aa34383 DK |
3541 | |
3542 | return (buf); | |
3543 | } | |
3544 | ||
3545 | /* | |
3546 | * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this | |
3547 | * for bufs containing metadata. | |
3548 | */ | |
3549 | arc_buf_t * | |
3550 | arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, | |
3551 | enum zio_compress compression_type) | |
3552 | { | |
2aa34383 DK |
3553 | ASSERT3U(lsize, >, 0); |
3554 | ASSERT3U(lsize, >=, psize); | |
b5256303 TC |
3555 | ASSERT3U(compression_type, >, ZIO_COMPRESS_OFF); |
3556 | ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS); | |
2aa34383 | 3557 | |
a7004725 | 3558 | arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, |
b5256303 | 3559 | B_FALSE, compression_type, ARC_BUFC_DATA, B_FALSE); |
2aa34383 | 3560 | |
a7004725 | 3561 | arc_buf_t *buf = NULL; |
be9a5c35 | 3562 | VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_FALSE, |
b5256303 | 3563 | B_TRUE, B_FALSE, B_FALSE, &buf)); |
2aa34383 DK |
3564 | arc_buf_thaw(buf); |
3565 | ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); | |
3566 | ||
a6255b7f DQ |
3567 | if (!arc_buf_is_shared(buf)) { |
3568 | /* | |
3569 | * To ensure that the hdr has the correct data in it if we call | |
b5256303 | 3570 | * arc_untransform() on this buf before it's been written to |
a6255b7f DQ |
3571 | * disk, it's easiest if we just set up sharing between the |
3572 | * buf and the hdr. | |
3573 | */ | |
b5256303 | 3574 | arc_hdr_free_abd(hdr, B_FALSE); |
a6255b7f DQ |
3575 | arc_share_buf(hdr, buf); |
3576 | } | |
3577 | ||
d3c2ae1c | 3578 | return (buf); |
34dc7c2f BB |
3579 | } |
3580 | ||
b5256303 TC |
3581 | arc_buf_t * |
3582 | arc_alloc_raw_buf(spa_t *spa, void *tag, uint64_t dsobj, boolean_t byteorder, | |
3583 | const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, | |
3584 | dmu_object_type_t ot, uint64_t psize, uint64_t lsize, | |
3585 | enum zio_compress compression_type) | |
3586 | { | |
3587 | arc_buf_hdr_t *hdr; | |
3588 | arc_buf_t *buf; | |
3589 | arc_buf_contents_t type = DMU_OT_IS_METADATA(ot) ? | |
3590 | ARC_BUFC_METADATA : ARC_BUFC_DATA; | |
3591 | ||
3592 | ASSERT3U(lsize, >, 0); | |
3593 | ASSERT3U(lsize, >=, psize); | |
3594 | ASSERT3U(compression_type, >=, ZIO_COMPRESS_OFF); | |
3595 | ASSERT3U(compression_type, <, ZIO_COMPRESS_FUNCTIONS); | |
3596 | ||
3597 | hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, B_TRUE, | |
3598 | compression_type, type, B_TRUE); | |
b5256303 TC |
3599 | |
3600 | hdr->b_crypt_hdr.b_dsobj = dsobj; | |
3601 | hdr->b_crypt_hdr.b_ot = ot; | |
3602 | hdr->b_l1hdr.b_byteswap = (byteorder == ZFS_HOST_BYTEORDER) ? | |
3603 | DMU_BSWAP_NUMFUNCS : DMU_OT_BYTESWAP(ot); | |
3604 | bcopy(salt, hdr->b_crypt_hdr.b_salt, ZIO_DATA_SALT_LEN); | |
3605 | bcopy(iv, hdr->b_crypt_hdr.b_iv, ZIO_DATA_IV_LEN); | |
3606 | bcopy(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN); | |
3607 | ||
3608 | /* | |
3609 | * This buffer will be considered encrypted even if the ot is not an | |
3610 | * encrypted type. It will become authenticated instead in | |
3611 | * arc_write_ready(). | |
3612 | */ | |
3613 | buf = NULL; | |
be9a5c35 | 3614 | VERIFY0(arc_buf_alloc_impl(hdr, spa, NULL, tag, B_TRUE, B_TRUE, |
b5256303 TC |
3615 | B_FALSE, B_FALSE, &buf)); |
3616 | arc_buf_thaw(buf); | |
3617 | ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); | |
3618 | ||
3619 | return (buf); | |
3620 | } | |
3621 | ||
d962d5da PS |
3622 | static void |
3623 | arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) | |
3624 | { | |
3625 | l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; | |
3626 | l2arc_dev_t *dev = l2hdr->b_dev; | |
7558997d SD |
3627 | uint64_t psize = HDR_GET_PSIZE(hdr); |
3628 | uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, psize); | |
d962d5da PS |
3629 | |
3630 | ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); | |
3631 | ASSERT(HDR_HAS_L2HDR(hdr)); | |
3632 | ||
3633 | list_remove(&dev->l2ad_buflist, hdr); | |
3634 | ||
01850391 AG |
3635 | ARCSTAT_INCR(arcstat_l2_psize, -psize); |
3636 | ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); | |
d962d5da | 3637 | |
7558997d | 3638 | vdev_space_update(dev->l2ad_vdev, -asize, 0, 0); |
d962d5da | 3639 | |
7558997d SD |
3640 | (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), |
3641 | hdr); | |
d3c2ae1c | 3642 | arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); |
d962d5da PS |
3643 | } |
3644 | ||
34dc7c2f BB |
3645 | static void |
3646 | arc_hdr_destroy(arc_buf_hdr_t *hdr) | |
3647 | { | |
b9541d6b CW |
3648 | if (HDR_HAS_L1HDR(hdr)) { |
3649 | ASSERT(hdr->b_l1hdr.b_buf == NULL || | |
d3c2ae1c | 3650 | hdr->b_l1hdr.b_bufcnt > 0); |
424fd7c3 | 3651 | ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
b9541d6b CW |
3652 | ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); |
3653 | } | |
34dc7c2f | 3654 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); |
b9541d6b CW |
3655 | ASSERT(!HDR_IN_HASH_TABLE(hdr)); |
3656 | ||
3657 | if (HDR_HAS_L2HDR(hdr)) { | |
d962d5da PS |
3658 | l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; |
3659 | boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); | |
428870ff | 3660 | |
d962d5da PS |
3661 | if (!buflist_held) |
3662 | mutex_enter(&dev->l2ad_mtx); | |
b9541d6b | 3663 | |
ca0bf58d | 3664 | /* |
d962d5da PS |
3665 | * Even though we checked this conditional above, we |
3666 | * need to check this again now that we have the | |
3667 | * l2ad_mtx. This is because we could be racing with | |
3668 | * another thread calling l2arc_evict() which might have | |
3669 | * destroyed this header's L2 portion as we were waiting | |
3670 | * to acquire the l2ad_mtx. If that happens, we don't | |
3671 | * want to re-destroy the header's L2 portion. | |
ca0bf58d | 3672 | */ |
d962d5da PS |
3673 | if (HDR_HAS_L2HDR(hdr)) |
3674 | arc_hdr_l2hdr_destroy(hdr); | |
428870ff BB |
3675 | |
3676 | if (!buflist_held) | |
d962d5da | 3677 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f BB |
3678 | } |
3679 | ||
ca6c7a94 BB |
3680 | /* |
3681 | * The header's identify can only be safely discarded once it is no | |
3682 | * longer discoverable. This requires removing it from the hash table | |
3683 | * and the l2arc header list. After this point the hash lock can not | |
3684 | * be used to protect the header. | |
3685 | */ | |
3686 | if (!HDR_EMPTY(hdr)) | |
3687 | buf_discard_identity(hdr); | |
3688 | ||
d3c2ae1c GW |
3689 | if (HDR_HAS_L1HDR(hdr)) { |
3690 | arc_cksum_free(hdr); | |
b9541d6b | 3691 | |
d3c2ae1c | 3692 | while (hdr->b_l1hdr.b_buf != NULL) |
2aa34383 | 3693 | arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); |
34dc7c2f | 3694 | |
ca6c7a94 | 3695 | if (hdr->b_l1hdr.b_pabd != NULL) |
b5256303 | 3696 | arc_hdr_free_abd(hdr, B_FALSE); |
b5256303 | 3697 | |
440a3eb9 | 3698 | if (HDR_HAS_RABD(hdr)) |
b5256303 | 3699 | arc_hdr_free_abd(hdr, B_TRUE); |
b9541d6b CW |
3700 | } |
3701 | ||
34dc7c2f | 3702 | ASSERT3P(hdr->b_hash_next, ==, NULL); |
b9541d6b | 3703 | if (HDR_HAS_L1HDR(hdr)) { |
ca0bf58d | 3704 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); |
b9541d6b | 3705 | ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); |
b5256303 TC |
3706 | |
3707 | if (!HDR_PROTECTED(hdr)) { | |
3708 | kmem_cache_free(hdr_full_cache, hdr); | |
3709 | } else { | |
3710 | kmem_cache_free(hdr_full_crypt_cache, hdr); | |
3711 | } | |
b9541d6b CW |
3712 | } else { |
3713 | kmem_cache_free(hdr_l2only_cache, hdr); | |
3714 | } | |
34dc7c2f BB |
3715 | } |
3716 | ||
3717 | void | |
d3c2ae1c | 3718 | arc_buf_destroy(arc_buf_t *buf, void* tag) |
34dc7c2f BB |
3719 | { |
3720 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
34dc7c2f | 3721 | |
b9541d6b | 3722 | if (hdr->b_l1hdr.b_state == arc_anon) { |
d3c2ae1c GW |
3723 | ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); |
3724 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
3725 | VERIFY0(remove_reference(hdr, NULL, tag)); | |
3726 | arc_hdr_destroy(hdr); | |
3727 | return; | |
34dc7c2f BB |
3728 | } |
3729 | ||
ca6c7a94 | 3730 | kmutex_t *hash_lock = HDR_LOCK(hdr); |
34dc7c2f | 3731 | mutex_enter(hash_lock); |
ca6c7a94 | 3732 | |
d3c2ae1c GW |
3733 | ASSERT3P(hdr, ==, buf->b_hdr); |
3734 | ASSERT(hdr->b_l1hdr.b_bufcnt > 0); | |
428870ff | 3735 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); |
d3c2ae1c GW |
3736 | ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); |
3737 | ASSERT3P(buf->b_data, !=, NULL); | |
34dc7c2f BB |
3738 | |
3739 | (void) remove_reference(hdr, hash_lock, tag); | |
2aa34383 | 3740 | arc_buf_destroy_impl(buf); |
34dc7c2f | 3741 | mutex_exit(hash_lock); |
34dc7c2f BB |
3742 | } |
3743 | ||
34dc7c2f | 3744 | /* |
ca0bf58d PS |
3745 | * Evict the arc_buf_hdr that is provided as a parameter. The resultant |
3746 | * state of the header is dependent on its state prior to entering this | |
3747 | * function. The following transitions are possible: | |
34dc7c2f | 3748 | * |
ca0bf58d PS |
3749 | * - arc_mru -> arc_mru_ghost |
3750 | * - arc_mfu -> arc_mfu_ghost | |
3751 | * - arc_mru_ghost -> arc_l2c_only | |
3752 | * - arc_mru_ghost -> deleted | |
3753 | * - arc_mfu_ghost -> arc_l2c_only | |
3754 | * - arc_mfu_ghost -> deleted | |
34dc7c2f | 3755 | */ |
ca0bf58d PS |
3756 | static int64_t |
3757 | arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) | |
34dc7c2f | 3758 | { |
ca0bf58d PS |
3759 | arc_state_t *evicted_state, *state; |
3760 | int64_t bytes_evicted = 0; | |
d4a72f23 TC |
3761 | int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ? |
3762 | arc_min_prescient_prefetch_ms : arc_min_prefetch_ms; | |
34dc7c2f | 3763 | |
ca0bf58d PS |
3764 | ASSERT(MUTEX_HELD(hash_lock)); |
3765 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
e8b96c60 | 3766 | |
ca0bf58d PS |
3767 | state = hdr->b_l1hdr.b_state; |
3768 | if (GHOST_STATE(state)) { | |
3769 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
d3c2ae1c | 3770 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); |
e8b96c60 MA |
3771 | |
3772 | /* | |
ca0bf58d | 3773 | * l2arc_write_buffers() relies on a header's L1 portion |
a6255b7f | 3774 | * (i.e. its b_pabd field) during it's write phase. |
ca0bf58d PS |
3775 | * Thus, we cannot push a header onto the arc_l2c_only |
3776 | * state (removing its L1 piece) until the header is | |
3777 | * done being written to the l2arc. | |
e8b96c60 | 3778 | */ |
ca0bf58d PS |
3779 | if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { |
3780 | ARCSTAT_BUMP(arcstat_evict_l2_skip); | |
3781 | return (bytes_evicted); | |
e8b96c60 MA |
3782 | } |
3783 | ||
ca0bf58d | 3784 | ARCSTAT_BUMP(arcstat_deleted); |
d3c2ae1c | 3785 | bytes_evicted += HDR_GET_LSIZE(hdr); |
428870ff | 3786 | |
ca0bf58d | 3787 | DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); |
428870ff | 3788 | |
ca0bf58d | 3789 | if (HDR_HAS_L2HDR(hdr)) { |
a6255b7f | 3790 | ASSERT(hdr->b_l1hdr.b_pabd == NULL); |
b5256303 | 3791 | ASSERT(!HDR_HAS_RABD(hdr)); |
ca0bf58d PS |
3792 | /* |
3793 | * This buffer is cached on the 2nd Level ARC; | |
3794 | * don't destroy the header. | |
3795 | */ | |
3796 | arc_change_state(arc_l2c_only, hdr, hash_lock); | |
3797 | /* | |
3798 | * dropping from L1+L2 cached to L2-only, | |
3799 | * realloc to remove the L1 header. | |
3800 | */ | |
3801 | hdr = arc_hdr_realloc(hdr, hdr_full_cache, | |
3802 | hdr_l2only_cache); | |
34dc7c2f | 3803 | } else { |
ca0bf58d PS |
3804 | arc_change_state(arc_anon, hdr, hash_lock); |
3805 | arc_hdr_destroy(hdr); | |
34dc7c2f | 3806 | } |
ca0bf58d | 3807 | return (bytes_evicted); |
34dc7c2f BB |
3808 | } |
3809 | ||
ca0bf58d PS |
3810 | ASSERT(state == arc_mru || state == arc_mfu); |
3811 | evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; | |
34dc7c2f | 3812 | |
ca0bf58d PS |
3813 | /* prefetch buffers have a minimum lifespan */ |
3814 | if (HDR_IO_IN_PROGRESS(hdr) || | |
3815 | ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && | |
2b84817f TC |
3816 | ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < |
3817 | MSEC_TO_TICK(min_lifetime))) { | |
ca0bf58d PS |
3818 | ARCSTAT_BUMP(arcstat_evict_skip); |
3819 | return (bytes_evicted); | |
da8ccd0e PS |
3820 | } |
3821 | ||
424fd7c3 | 3822 | ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt)); |
ca0bf58d PS |
3823 | while (hdr->b_l1hdr.b_buf) { |
3824 | arc_buf_t *buf = hdr->b_l1hdr.b_buf; | |
3825 | if (!mutex_tryenter(&buf->b_evict_lock)) { | |
3826 | ARCSTAT_BUMP(arcstat_mutex_miss); | |
3827 | break; | |
3828 | } | |
3829 | if (buf->b_data != NULL) | |
d3c2ae1c GW |
3830 | bytes_evicted += HDR_GET_LSIZE(hdr); |
3831 | mutex_exit(&buf->b_evict_lock); | |
2aa34383 | 3832 | arc_buf_destroy_impl(buf); |
ca0bf58d | 3833 | } |
34dc7c2f | 3834 | |
ca0bf58d | 3835 | if (HDR_HAS_L2HDR(hdr)) { |
d3c2ae1c | 3836 | ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); |
ca0bf58d | 3837 | } else { |
d3c2ae1c GW |
3838 | if (l2arc_write_eligible(hdr->b_spa, hdr)) { |
3839 | ARCSTAT_INCR(arcstat_evict_l2_eligible, | |
3840 | HDR_GET_LSIZE(hdr)); | |
3841 | } else { | |
3842 | ARCSTAT_INCR(arcstat_evict_l2_ineligible, | |
3843 | HDR_GET_LSIZE(hdr)); | |
3844 | } | |
ca0bf58d | 3845 | } |
34dc7c2f | 3846 | |
d3c2ae1c GW |
3847 | if (hdr->b_l1hdr.b_bufcnt == 0) { |
3848 | arc_cksum_free(hdr); | |
3849 | ||
3850 | bytes_evicted += arc_hdr_size(hdr); | |
3851 | ||
3852 | /* | |
3853 | * If this hdr is being evicted and has a compressed | |
3854 | * buffer then we discard it here before we change states. | |
3855 | * This ensures that the accounting is updated correctly | |
a6255b7f | 3856 | * in arc_free_data_impl(). |
d3c2ae1c | 3857 | */ |
b5256303 TC |
3858 | if (hdr->b_l1hdr.b_pabd != NULL) |
3859 | arc_hdr_free_abd(hdr, B_FALSE); | |
3860 | ||
3861 | if (HDR_HAS_RABD(hdr)) | |
3862 | arc_hdr_free_abd(hdr, B_TRUE); | |
d3c2ae1c | 3863 | |
ca0bf58d PS |
3864 | arc_change_state(evicted_state, hdr, hash_lock); |
3865 | ASSERT(HDR_IN_HASH_TABLE(hdr)); | |
d3c2ae1c | 3866 | arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); |
ca0bf58d PS |
3867 | DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); |
3868 | } | |
34dc7c2f | 3869 | |
ca0bf58d | 3870 | return (bytes_evicted); |
34dc7c2f BB |
3871 | } |
3872 | ||
ca0bf58d PS |
3873 | static uint64_t |
3874 | arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, | |
3875 | uint64_t spa, int64_t bytes) | |
34dc7c2f | 3876 | { |
ca0bf58d PS |
3877 | multilist_sublist_t *mls; |
3878 | uint64_t bytes_evicted = 0; | |
3879 | arc_buf_hdr_t *hdr; | |
34dc7c2f | 3880 | kmutex_t *hash_lock; |
ca0bf58d | 3881 | int evict_count = 0; |
34dc7c2f | 3882 | |
ca0bf58d | 3883 | ASSERT3P(marker, !=, NULL); |
96c080cb | 3884 | IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); |
ca0bf58d PS |
3885 | |
3886 | mls = multilist_sublist_lock(ml, idx); | |
572e2857 | 3887 | |
ca0bf58d PS |
3888 | for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; |
3889 | hdr = multilist_sublist_prev(mls, marker)) { | |
3890 | if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || | |
3891 | (evict_count >= zfs_arc_evict_batch_limit)) | |
3892 | break; | |
3893 | ||
3894 | /* | |
3895 | * To keep our iteration location, move the marker | |
3896 | * forward. Since we're not holding hdr's hash lock, we | |
3897 | * must be very careful and not remove 'hdr' from the | |
3898 | * sublist. Otherwise, other consumers might mistake the | |
3899 | * 'hdr' as not being on a sublist when they call the | |
3900 | * multilist_link_active() function (they all rely on | |
3901 | * the hash lock protecting concurrent insertions and | |
3902 | * removals). multilist_sublist_move_forward() was | |
3903 | * specifically implemented to ensure this is the case | |
3904 | * (only 'marker' will be removed and re-inserted). | |
3905 | */ | |
3906 | multilist_sublist_move_forward(mls, marker); | |
3907 | ||
3908 | /* | |
3909 | * The only case where the b_spa field should ever be | |
3910 | * zero, is the marker headers inserted by | |
3911 | * arc_evict_state(). It's possible for multiple threads | |
3912 | * to be calling arc_evict_state() concurrently (e.g. | |
3913 | * dsl_pool_close() and zio_inject_fault()), so we must | |
3914 | * skip any markers we see from these other threads. | |
3915 | */ | |
2a432414 | 3916 | if (hdr->b_spa == 0) |
572e2857 BB |
3917 | continue; |
3918 | ||
ca0bf58d PS |
3919 | /* we're only interested in evicting buffers of a certain spa */ |
3920 | if (spa != 0 && hdr->b_spa != spa) { | |
3921 | ARCSTAT_BUMP(arcstat_evict_skip); | |
428870ff | 3922 | continue; |
ca0bf58d PS |
3923 | } |
3924 | ||
3925 | hash_lock = HDR_LOCK(hdr); | |
e8b96c60 MA |
3926 | |
3927 | /* | |
ca0bf58d PS |
3928 | * We aren't calling this function from any code path |
3929 | * that would already be holding a hash lock, so we're | |
3930 | * asserting on this assumption to be defensive in case | |
3931 | * this ever changes. Without this check, it would be | |
3932 | * possible to incorrectly increment arcstat_mutex_miss | |
3933 | * below (e.g. if the code changed such that we called | |
3934 | * this function with a hash lock held). | |
e8b96c60 | 3935 | */ |
ca0bf58d PS |
3936 | ASSERT(!MUTEX_HELD(hash_lock)); |
3937 | ||
34dc7c2f | 3938 | if (mutex_tryenter(hash_lock)) { |
ca0bf58d PS |
3939 | uint64_t evicted = arc_evict_hdr(hdr, hash_lock); |
3940 | mutex_exit(hash_lock); | |
34dc7c2f | 3941 | |
ca0bf58d | 3942 | bytes_evicted += evicted; |
34dc7c2f | 3943 | |
572e2857 | 3944 | /* |
ca0bf58d PS |
3945 | * If evicted is zero, arc_evict_hdr() must have |
3946 | * decided to skip this header, don't increment | |
3947 | * evict_count in this case. | |
572e2857 | 3948 | */ |
ca0bf58d PS |
3949 | if (evicted != 0) |
3950 | evict_count++; | |
3951 | ||
3952 | /* | |
3953 | * If arc_size isn't overflowing, signal any | |
3954 | * threads that might happen to be waiting. | |
3955 | * | |
3956 | * For each header evicted, we wake up a single | |
3957 | * thread. If we used cv_broadcast, we could | |
3958 | * wake up "too many" threads causing arc_size | |
3959 | * to significantly overflow arc_c; since | |
a6255b7f | 3960 | * arc_get_data_impl() doesn't check for overflow |
ca0bf58d PS |
3961 | * when it's woken up (it doesn't because it's |
3962 | * possible for the ARC to be overflowing while | |
3963 | * full of un-evictable buffers, and the | |
3964 | * function should proceed in this case). | |
3965 | * | |
3966 | * If threads are left sleeping, due to not | |
3ec34e55 BL |
3967 | * using cv_broadcast here, they will be woken |
3968 | * up via cv_broadcast in arc_adjust_cb() just | |
3969 | * before arc_adjust_zthr sleeps. | |
ca0bf58d | 3970 | */ |
3ec34e55 | 3971 | mutex_enter(&arc_adjust_lock); |
ca0bf58d | 3972 | if (!arc_is_overflowing()) |
3ec34e55 BL |
3973 | cv_signal(&arc_adjust_waiters_cv); |
3974 | mutex_exit(&arc_adjust_lock); | |
e8b96c60 | 3975 | } else { |
ca0bf58d | 3976 | ARCSTAT_BUMP(arcstat_mutex_miss); |
e8b96c60 | 3977 | } |
34dc7c2f | 3978 | } |
34dc7c2f | 3979 | |
ca0bf58d | 3980 | multilist_sublist_unlock(mls); |
34dc7c2f | 3981 | |
ca0bf58d | 3982 | return (bytes_evicted); |
34dc7c2f BB |
3983 | } |
3984 | ||
ca0bf58d PS |
3985 | /* |
3986 | * Evict buffers from the given arc state, until we've removed the | |
3987 | * specified number of bytes. Move the removed buffers to the | |
3988 | * appropriate evict state. | |
3989 | * | |
3990 | * This function makes a "best effort". It skips over any buffers | |
3991 | * it can't get a hash_lock on, and so, may not catch all candidates. | |
3992 | * It may also return without evicting as much space as requested. | |
3993 | * | |
3994 | * If bytes is specified using the special value ARC_EVICT_ALL, this | |
3995 | * will evict all available (i.e. unlocked and evictable) buffers from | |
3996 | * the given arc state; which is used by arc_flush(). | |
3997 | */ | |
3998 | static uint64_t | |
3999 | arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, | |
4000 | arc_buf_contents_t type) | |
34dc7c2f | 4001 | { |
ca0bf58d | 4002 | uint64_t total_evicted = 0; |
64fc7762 | 4003 | multilist_t *ml = state->arcs_list[type]; |
ca0bf58d PS |
4004 | int num_sublists; |
4005 | arc_buf_hdr_t **markers; | |
ca0bf58d | 4006 | |
96c080cb | 4007 | IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); |
ca0bf58d PS |
4008 | |
4009 | num_sublists = multilist_get_num_sublists(ml); | |
d164b209 BB |
4010 | |
4011 | /* | |
ca0bf58d PS |
4012 | * If we've tried to evict from each sublist, made some |
4013 | * progress, but still have not hit the target number of bytes | |
4014 | * to evict, we want to keep trying. The markers allow us to | |
4015 | * pick up where we left off for each individual sublist, rather | |
4016 | * than starting from the tail each time. | |
d164b209 | 4017 | */ |
ca0bf58d | 4018 | markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); |
1c27024e | 4019 | for (int i = 0; i < num_sublists; i++) { |
ca0bf58d | 4020 | multilist_sublist_t *mls; |
34dc7c2f | 4021 | |
ca0bf58d PS |
4022 | markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); |
4023 | ||
4024 | /* | |
4025 | * A b_spa of 0 is used to indicate that this header is | |
4026 | * a marker. This fact is used in arc_adjust_type() and | |
4027 | * arc_evict_state_impl(). | |
4028 | */ | |
4029 | markers[i]->b_spa = 0; | |
34dc7c2f | 4030 | |
ca0bf58d PS |
4031 | mls = multilist_sublist_lock(ml, i); |
4032 | multilist_sublist_insert_tail(mls, markers[i]); | |
4033 | multilist_sublist_unlock(mls); | |
34dc7c2f BB |
4034 | } |
4035 | ||
d164b209 | 4036 | /* |
ca0bf58d PS |
4037 | * While we haven't hit our target number of bytes to evict, or |
4038 | * we're evicting all available buffers. | |
d164b209 | 4039 | */ |
ca0bf58d | 4040 | while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { |
25458cbe TC |
4041 | int sublist_idx = multilist_get_random_index(ml); |
4042 | uint64_t scan_evicted = 0; | |
4043 | ||
4044 | /* | |
4045 | * Try to reduce pinned dnodes with a floor of arc_dnode_limit. | |
4046 | * Request that 10% of the LRUs be scanned by the superblock | |
4047 | * shrinker. | |
4048 | */ | |
37fb3e43 | 4049 | if (type == ARC_BUFC_DATA && aggsum_compare(&astat_dnode_size, |
03fdcb9a | 4050 | arc_dnode_size_limit) > 0) { |
37fb3e43 | 4051 | arc_prune_async((aggsum_upper_bound(&astat_dnode_size) - |
03fdcb9a | 4052 | arc_dnode_size_limit) / sizeof (dnode_t) / |
37fb3e43 PD |
4053 | zfs_arc_dnode_reduce_percent); |
4054 | } | |
25458cbe | 4055 | |
ca0bf58d PS |
4056 | /* |
4057 | * Start eviction using a randomly selected sublist, | |
4058 | * this is to try and evenly balance eviction across all | |
4059 | * sublists. Always starting at the same sublist | |
4060 | * (e.g. index 0) would cause evictions to favor certain | |
4061 | * sublists over others. | |
4062 | */ | |
1c27024e | 4063 | for (int i = 0; i < num_sublists; i++) { |
ca0bf58d PS |
4064 | uint64_t bytes_remaining; |
4065 | uint64_t bytes_evicted; | |
d164b209 | 4066 | |
ca0bf58d PS |
4067 | if (bytes == ARC_EVICT_ALL) |
4068 | bytes_remaining = ARC_EVICT_ALL; | |
4069 | else if (total_evicted < bytes) | |
4070 | bytes_remaining = bytes - total_evicted; | |
4071 | else | |
4072 | break; | |
34dc7c2f | 4073 | |
ca0bf58d PS |
4074 | bytes_evicted = arc_evict_state_impl(ml, sublist_idx, |
4075 | markers[sublist_idx], spa, bytes_remaining); | |
4076 | ||
4077 | scan_evicted += bytes_evicted; | |
4078 | total_evicted += bytes_evicted; | |
4079 | ||
4080 | /* we've reached the end, wrap to the beginning */ | |
4081 | if (++sublist_idx >= num_sublists) | |
4082 | sublist_idx = 0; | |
4083 | } | |
4084 | ||
4085 | /* | |
4086 | * If we didn't evict anything during this scan, we have | |
4087 | * no reason to believe we'll evict more during another | |
4088 | * scan, so break the loop. | |
4089 | */ | |
4090 | if (scan_evicted == 0) { | |
4091 | /* This isn't possible, let's make that obvious */ | |
4092 | ASSERT3S(bytes, !=, 0); | |
34dc7c2f | 4093 | |
ca0bf58d PS |
4094 | /* |
4095 | * When bytes is ARC_EVICT_ALL, the only way to | |
4096 | * break the loop is when scan_evicted is zero. | |
4097 | * In that case, we actually have evicted enough, | |
4098 | * so we don't want to increment the kstat. | |
4099 | */ | |
4100 | if (bytes != ARC_EVICT_ALL) { | |
4101 | ASSERT3S(total_evicted, <, bytes); | |
4102 | ARCSTAT_BUMP(arcstat_evict_not_enough); | |
4103 | } | |
d164b209 | 4104 | |
ca0bf58d PS |
4105 | break; |
4106 | } | |
d164b209 | 4107 | } |
34dc7c2f | 4108 | |
1c27024e | 4109 | for (int i = 0; i < num_sublists; i++) { |
ca0bf58d PS |
4110 | multilist_sublist_t *mls = multilist_sublist_lock(ml, i); |
4111 | multilist_sublist_remove(mls, markers[i]); | |
4112 | multilist_sublist_unlock(mls); | |
34dc7c2f | 4113 | |
ca0bf58d | 4114 | kmem_cache_free(hdr_full_cache, markers[i]); |
34dc7c2f | 4115 | } |
ca0bf58d PS |
4116 | kmem_free(markers, sizeof (*markers) * num_sublists); |
4117 | ||
4118 | return (total_evicted); | |
4119 | } | |
4120 | ||
4121 | /* | |
4122 | * Flush all "evictable" data of the given type from the arc state | |
4123 | * specified. This will not evict any "active" buffers (i.e. referenced). | |
4124 | * | |
d3c2ae1c | 4125 | * When 'retry' is set to B_FALSE, the function will make a single pass |
ca0bf58d PS |
4126 | * over the state and evict any buffers that it can. Since it doesn't |
4127 | * continually retry the eviction, it might end up leaving some buffers | |
4128 | * in the ARC due to lock misses. | |
4129 | * | |
d3c2ae1c | 4130 | * When 'retry' is set to B_TRUE, the function will continually retry the |
ca0bf58d PS |
4131 | * eviction until *all* evictable buffers have been removed from the |
4132 | * state. As a result, if concurrent insertions into the state are | |
4133 | * allowed (e.g. if the ARC isn't shutting down), this function might | |
4134 | * wind up in an infinite loop, continually trying to evict buffers. | |
4135 | */ | |
4136 | static uint64_t | |
4137 | arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, | |
4138 | boolean_t retry) | |
4139 | { | |
4140 | uint64_t evicted = 0; | |
4141 | ||
424fd7c3 | 4142 | while (zfs_refcount_count(&state->arcs_esize[type]) != 0) { |
ca0bf58d PS |
4143 | evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); |
4144 | ||
4145 | if (!retry) | |
4146 | break; | |
4147 | } | |
4148 | ||
4149 | return (evicted); | |
34dc7c2f BB |
4150 | } |
4151 | ||
ca0bf58d PS |
4152 | /* |
4153 | * Evict the specified number of bytes from the state specified, | |
4154 | * restricting eviction to the spa and type given. This function | |
4155 | * prevents us from trying to evict more from a state's list than | |
4156 | * is "evictable", and to skip evicting altogether when passed a | |
4157 | * negative value for "bytes". In contrast, arc_evict_state() will | |
4158 | * evict everything it can, when passed a negative value for "bytes". | |
4159 | */ | |
4160 | static uint64_t | |
4161 | arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, | |
4162 | arc_buf_contents_t type) | |
4163 | { | |
4164 | int64_t delta; | |
4165 | ||
424fd7c3 TS |
4166 | if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) { |
4167 | delta = MIN(zfs_refcount_count(&state->arcs_esize[type]), | |
4168 | bytes); | |
ca0bf58d PS |
4169 | return (arc_evict_state(state, spa, delta, type)); |
4170 | } | |
4171 | ||
4172 | return (0); | |
4173 | } | |
4174 | ||
4175 | /* | |
4176 | * The goal of this function is to evict enough meta data buffers from the | |
4177 | * ARC in order to enforce the arc_meta_limit. Achieving this is slightly | |
4178 | * more complicated than it appears because it is common for data buffers | |
4179 | * to have holds on meta data buffers. In addition, dnode meta data buffers | |
4180 | * will be held by the dnodes in the block preventing them from being freed. | |
4181 | * This means we can't simply traverse the ARC and expect to always find | |
4182 | * enough unheld meta data buffer to release. | |
4183 | * | |
4184 | * Therefore, this function has been updated to make alternating passes | |
4185 | * over the ARC releasing data buffers and then newly unheld meta data | |
37fb3e43 | 4186 | * buffers. This ensures forward progress is maintained and meta_used |
ca0bf58d PS |
4187 | * will decrease. Normally this is sufficient, but if required the ARC |
4188 | * will call the registered prune callbacks causing dentry and inodes to | |
4189 | * be dropped from the VFS cache. This will make dnode meta data buffers | |
4190 | * available for reclaim. | |
4191 | */ | |
4192 | static uint64_t | |
37fb3e43 | 4193 | arc_adjust_meta_balanced(uint64_t meta_used) |
ca0bf58d | 4194 | { |
25e2ab16 TC |
4195 | int64_t delta, prune = 0, adjustmnt; |
4196 | uint64_t total_evicted = 0; | |
ca0bf58d | 4197 | arc_buf_contents_t type = ARC_BUFC_DATA; |
ca67b33a | 4198 | int restarts = MAX(zfs_arc_meta_adjust_restarts, 0); |
ca0bf58d PS |
4199 | |
4200 | restart: | |
4201 | /* | |
4202 | * This slightly differs than the way we evict from the mru in | |
4203 | * arc_adjust because we don't have a "target" value (i.e. no | |
4204 | * "meta" arc_p). As a result, I think we can completely | |
4205 | * cannibalize the metadata in the MRU before we evict the | |
4206 | * metadata from the MFU. I think we probably need to implement a | |
4207 | * "metadata arc_p" value to do this properly. | |
4208 | */ | |
37fb3e43 | 4209 | adjustmnt = meta_used - arc_meta_limit; |
ca0bf58d | 4210 | |
424fd7c3 TS |
4211 | if (adjustmnt > 0 && |
4212 | zfs_refcount_count(&arc_mru->arcs_esize[type]) > 0) { | |
4213 | delta = MIN(zfs_refcount_count(&arc_mru->arcs_esize[type]), | |
d3c2ae1c | 4214 | adjustmnt); |
ca0bf58d PS |
4215 | total_evicted += arc_adjust_impl(arc_mru, 0, delta, type); |
4216 | adjustmnt -= delta; | |
4217 | } | |
4218 | ||
4219 | /* | |
4220 | * We can't afford to recalculate adjustmnt here. If we do, | |
4221 | * new metadata buffers can sneak into the MRU or ANON lists, | |
4222 | * thus penalize the MFU metadata. Although the fudge factor is | |
4223 | * small, it has been empirically shown to be significant for | |
4224 | * certain workloads (e.g. creating many empty directories). As | |
4225 | * such, we use the original calculation for adjustmnt, and | |
4226 | * simply decrement the amount of data evicted from the MRU. | |
4227 | */ | |
4228 | ||
424fd7c3 TS |
4229 | if (adjustmnt > 0 && |
4230 | zfs_refcount_count(&arc_mfu->arcs_esize[type]) > 0) { | |
4231 | delta = MIN(zfs_refcount_count(&arc_mfu->arcs_esize[type]), | |
d3c2ae1c | 4232 | adjustmnt); |
ca0bf58d PS |
4233 | total_evicted += arc_adjust_impl(arc_mfu, 0, delta, type); |
4234 | } | |
4235 | ||
37fb3e43 | 4236 | adjustmnt = meta_used - arc_meta_limit; |
ca0bf58d | 4237 | |
d3c2ae1c | 4238 | if (adjustmnt > 0 && |
424fd7c3 | 4239 | zfs_refcount_count(&arc_mru_ghost->arcs_esize[type]) > 0) { |
ca0bf58d | 4240 | delta = MIN(adjustmnt, |
424fd7c3 | 4241 | zfs_refcount_count(&arc_mru_ghost->arcs_esize[type])); |
ca0bf58d PS |
4242 | total_evicted += arc_adjust_impl(arc_mru_ghost, 0, delta, type); |
4243 | adjustmnt -= delta; | |
4244 | } | |
4245 | ||
d3c2ae1c | 4246 | if (adjustmnt > 0 && |
424fd7c3 | 4247 | zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type]) > 0) { |
ca0bf58d | 4248 | delta = MIN(adjustmnt, |
424fd7c3 | 4249 | zfs_refcount_count(&arc_mfu_ghost->arcs_esize[type])); |
ca0bf58d PS |
4250 | total_evicted += arc_adjust_impl(arc_mfu_ghost, 0, delta, type); |
4251 | } | |
4252 | ||
4253 | /* | |
4254 | * If after attempting to make the requested adjustment to the ARC | |
4255 | * the meta limit is still being exceeded then request that the | |
4256 | * higher layers drop some cached objects which have holds on ARC | |
4257 | * meta buffers. Requests to the upper layers will be made with | |
4258 | * increasingly large scan sizes until the ARC is below the limit. | |
4259 | */ | |
37fb3e43 | 4260 | if (meta_used > arc_meta_limit) { |
ca0bf58d PS |
4261 | if (type == ARC_BUFC_DATA) { |
4262 | type = ARC_BUFC_METADATA; | |
4263 | } else { | |
4264 | type = ARC_BUFC_DATA; | |
4265 | ||
4266 | if (zfs_arc_meta_prune) { | |
4267 | prune += zfs_arc_meta_prune; | |
f6046738 | 4268 | arc_prune_async(prune); |
ca0bf58d PS |
4269 | } |
4270 | } | |
4271 | ||
4272 | if (restarts > 0) { | |
4273 | restarts--; | |
4274 | goto restart; | |
4275 | } | |
4276 | } | |
4277 | return (total_evicted); | |
4278 | } | |
4279 | ||
f6046738 BB |
4280 | /* |
4281 | * Evict metadata buffers from the cache, such that arc_meta_used is | |
4282 | * capped by the arc_meta_limit tunable. | |
4283 | */ | |
4284 | static uint64_t | |
37fb3e43 | 4285 | arc_adjust_meta_only(uint64_t meta_used) |
f6046738 BB |
4286 | { |
4287 | uint64_t total_evicted = 0; | |
4288 | int64_t target; | |
4289 | ||
4290 | /* | |
4291 | * If we're over the meta limit, we want to evict enough | |
4292 | * metadata to get back under the meta limit. We don't want to | |
4293 | * evict so much that we drop the MRU below arc_p, though. If | |
4294 | * we're over the meta limit more than we're over arc_p, we | |
4295 | * evict some from the MRU here, and some from the MFU below. | |
4296 | */ | |
37fb3e43 | 4297 | target = MIN((int64_t)(meta_used - arc_meta_limit), |
424fd7c3 TS |
4298 | (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) + |
4299 | zfs_refcount_count(&arc_mru->arcs_size) - arc_p)); | |
f6046738 BB |
4300 | |
4301 | total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); | |
4302 | ||
4303 | /* | |
4304 | * Similar to the above, we want to evict enough bytes to get us | |
4305 | * below the meta limit, but not so much as to drop us below the | |
2aa34383 | 4306 | * space allotted to the MFU (which is defined as arc_c - arc_p). |
f6046738 | 4307 | */ |
37fb3e43 | 4308 | target = MIN((int64_t)(meta_used - arc_meta_limit), |
424fd7c3 | 4309 | (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) - |
37fb3e43 | 4310 | (arc_c - arc_p))); |
f6046738 BB |
4311 | |
4312 | total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); | |
4313 | ||
4314 | return (total_evicted); | |
4315 | } | |
4316 | ||
4317 | static uint64_t | |
37fb3e43 | 4318 | arc_adjust_meta(uint64_t meta_used) |
f6046738 BB |
4319 | { |
4320 | if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY) | |
37fb3e43 | 4321 | return (arc_adjust_meta_only(meta_used)); |
f6046738 | 4322 | else |
37fb3e43 | 4323 | return (arc_adjust_meta_balanced(meta_used)); |
f6046738 BB |
4324 | } |
4325 | ||
ca0bf58d PS |
4326 | /* |
4327 | * Return the type of the oldest buffer in the given arc state | |
4328 | * | |
4329 | * This function will select a random sublist of type ARC_BUFC_DATA and | |
4330 | * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist | |
4331 | * is compared, and the type which contains the "older" buffer will be | |
4332 | * returned. | |
4333 | */ | |
4334 | static arc_buf_contents_t | |
4335 | arc_adjust_type(arc_state_t *state) | |
4336 | { | |
64fc7762 MA |
4337 | multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA]; |
4338 | multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA]; | |
ca0bf58d PS |
4339 | int data_idx = multilist_get_random_index(data_ml); |
4340 | int meta_idx = multilist_get_random_index(meta_ml); | |
4341 | multilist_sublist_t *data_mls; | |
4342 | multilist_sublist_t *meta_mls; | |
4343 | arc_buf_contents_t type; | |
4344 | arc_buf_hdr_t *data_hdr; | |
4345 | arc_buf_hdr_t *meta_hdr; | |
4346 | ||
4347 | /* | |
4348 | * We keep the sublist lock until we're finished, to prevent | |
4349 | * the headers from being destroyed via arc_evict_state(). | |
4350 | */ | |
4351 | data_mls = multilist_sublist_lock(data_ml, data_idx); | |
4352 | meta_mls = multilist_sublist_lock(meta_ml, meta_idx); | |
4353 | ||
4354 | /* | |
4355 | * These two loops are to ensure we skip any markers that | |
4356 | * might be at the tail of the lists due to arc_evict_state(). | |
4357 | */ | |
4358 | ||
4359 | for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; | |
4360 | data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { | |
4361 | if (data_hdr->b_spa != 0) | |
4362 | break; | |
4363 | } | |
4364 | ||
4365 | for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; | |
4366 | meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { | |
4367 | if (meta_hdr->b_spa != 0) | |
4368 | break; | |
4369 | } | |
4370 | ||
4371 | if (data_hdr == NULL && meta_hdr == NULL) { | |
4372 | type = ARC_BUFC_DATA; | |
4373 | } else if (data_hdr == NULL) { | |
4374 | ASSERT3P(meta_hdr, !=, NULL); | |
4375 | type = ARC_BUFC_METADATA; | |
4376 | } else if (meta_hdr == NULL) { | |
4377 | ASSERT3P(data_hdr, !=, NULL); | |
4378 | type = ARC_BUFC_DATA; | |
4379 | } else { | |
4380 | ASSERT3P(data_hdr, !=, NULL); | |
4381 | ASSERT3P(meta_hdr, !=, NULL); | |
4382 | ||
4383 | /* The headers can't be on the sublist without an L1 header */ | |
4384 | ASSERT(HDR_HAS_L1HDR(data_hdr)); | |
4385 | ASSERT(HDR_HAS_L1HDR(meta_hdr)); | |
4386 | ||
4387 | if (data_hdr->b_l1hdr.b_arc_access < | |
4388 | meta_hdr->b_l1hdr.b_arc_access) { | |
4389 | type = ARC_BUFC_DATA; | |
4390 | } else { | |
4391 | type = ARC_BUFC_METADATA; | |
4392 | } | |
4393 | } | |
4394 | ||
4395 | multilist_sublist_unlock(meta_mls); | |
4396 | multilist_sublist_unlock(data_mls); | |
4397 | ||
4398 | return (type); | |
4399 | } | |
4400 | ||
4401 | /* | |
4402 | * Evict buffers from the cache, such that arc_size is capped by arc_c. | |
4403 | */ | |
4404 | static uint64_t | |
4405 | arc_adjust(void) | |
4406 | { | |
4407 | uint64_t total_evicted = 0; | |
4408 | uint64_t bytes; | |
4409 | int64_t target; | |
37fb3e43 PD |
4410 | uint64_t asize = aggsum_value(&arc_size); |
4411 | uint64_t ameta = aggsum_value(&arc_meta_used); | |
ca0bf58d PS |
4412 | |
4413 | /* | |
4414 | * If we're over arc_meta_limit, we want to correct that before | |
4415 | * potentially evicting data buffers below. | |
4416 | */ | |
37fb3e43 | 4417 | total_evicted += arc_adjust_meta(ameta); |
ca0bf58d PS |
4418 | |
4419 | /* | |
4420 | * Adjust MRU size | |
4421 | * | |
4422 | * If we're over the target cache size, we want to evict enough | |
4423 | * from the list to get back to our target size. We don't want | |
4424 | * to evict too much from the MRU, such that it drops below | |
4425 | * arc_p. So, if we're over our target cache size more than | |
4426 | * the MRU is over arc_p, we'll evict enough to get back to | |
4427 | * arc_p here, and then evict more from the MFU below. | |
4428 | */ | |
37fb3e43 | 4429 | target = MIN((int64_t)(asize - arc_c), |
424fd7c3 TS |
4430 | (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) + |
4431 | zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p)); | |
ca0bf58d PS |
4432 | |
4433 | /* | |
4434 | * If we're below arc_meta_min, always prefer to evict data. | |
4435 | * Otherwise, try to satisfy the requested number of bytes to | |
4436 | * evict from the type which contains older buffers; in an | |
4437 | * effort to keep newer buffers in the cache regardless of their | |
4438 | * type. If we cannot satisfy the number of bytes from this | |
4439 | * type, spill over into the next type. | |
4440 | */ | |
4441 | if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && | |
37fb3e43 | 4442 | ameta > arc_meta_min) { |
ca0bf58d PS |
4443 | bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); |
4444 | total_evicted += bytes; | |
4445 | ||
4446 | /* | |
4447 | * If we couldn't evict our target number of bytes from | |
4448 | * metadata, we try to get the rest from data. | |
4449 | */ | |
4450 | target -= bytes; | |
4451 | ||
4452 | total_evicted += | |
4453 | arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); | |
4454 | } else { | |
4455 | bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); | |
4456 | total_evicted += bytes; | |
4457 | ||
4458 | /* | |
4459 | * If we couldn't evict our target number of bytes from | |
4460 | * data, we try to get the rest from metadata. | |
4461 | */ | |
4462 | target -= bytes; | |
4463 | ||
4464 | total_evicted += | |
4465 | arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); | |
4466 | } | |
4467 | ||
0405eeea RE |
4468 | /* |
4469 | * Re-sum ARC stats after the first round of evictions. | |
4470 | */ | |
4471 | asize = aggsum_value(&arc_size); | |
4472 | ameta = aggsum_value(&arc_meta_used); | |
4473 | ||
4474 | ||
ca0bf58d PS |
4475 | /* |
4476 | * Adjust MFU size | |
4477 | * | |
4478 | * Now that we've tried to evict enough from the MRU to get its | |
4479 | * size back to arc_p, if we're still above the target cache | |
4480 | * size, we evict the rest from the MFU. | |
4481 | */ | |
37fb3e43 | 4482 | target = asize - arc_c; |
ca0bf58d | 4483 | |
a7b10a93 | 4484 | if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && |
37fb3e43 | 4485 | ameta > arc_meta_min) { |
ca0bf58d PS |
4486 | bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); |
4487 | total_evicted += bytes; | |
4488 | ||
4489 | /* | |
4490 | * If we couldn't evict our target number of bytes from | |
4491 | * metadata, we try to get the rest from data. | |
4492 | */ | |
4493 | target -= bytes; | |
4494 | ||
4495 | total_evicted += | |
4496 | arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); | |
4497 | } else { | |
4498 | bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); | |
4499 | total_evicted += bytes; | |
4500 | ||
4501 | /* | |
4502 | * If we couldn't evict our target number of bytes from | |
4503 | * data, we try to get the rest from data. | |
4504 | */ | |
4505 | target -= bytes; | |
4506 | ||
4507 | total_evicted += | |
4508 | arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); | |
4509 | } | |
4510 | ||
4511 | /* | |
4512 | * Adjust ghost lists | |
4513 | * | |
4514 | * In addition to the above, the ARC also defines target values | |
4515 | * for the ghost lists. The sum of the mru list and mru ghost | |
4516 | * list should never exceed the target size of the cache, and | |
4517 | * the sum of the mru list, mfu list, mru ghost list, and mfu | |
4518 | * ghost list should never exceed twice the target size of the | |
4519 | * cache. The following logic enforces these limits on the ghost | |
4520 | * caches, and evicts from them as needed. | |
4521 | */ | |
424fd7c3 TS |
4522 | target = zfs_refcount_count(&arc_mru->arcs_size) + |
4523 | zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c; | |
ca0bf58d PS |
4524 | |
4525 | bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); | |
4526 | total_evicted += bytes; | |
4527 | ||
4528 | target -= bytes; | |
4529 | ||
4530 | total_evicted += | |
4531 | arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); | |
4532 | ||
4533 | /* | |
4534 | * We assume the sum of the mru list and mfu list is less than | |
4535 | * or equal to arc_c (we enforced this above), which means we | |
4536 | * can use the simpler of the two equations below: | |
4537 | * | |
4538 | * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c | |
4539 | * mru ghost + mfu ghost <= arc_c | |
4540 | */ | |
424fd7c3 TS |
4541 | target = zfs_refcount_count(&arc_mru_ghost->arcs_size) + |
4542 | zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; | |
ca0bf58d PS |
4543 | |
4544 | bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); | |
4545 | total_evicted += bytes; | |
4546 | ||
4547 | target -= bytes; | |
4548 | ||
4549 | total_evicted += | |
4550 | arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); | |
4551 | ||
4552 | return (total_evicted); | |
4553 | } | |
4554 | ||
ca0bf58d PS |
4555 | void |
4556 | arc_flush(spa_t *spa, boolean_t retry) | |
ab26409d | 4557 | { |
ca0bf58d | 4558 | uint64_t guid = 0; |
94520ca4 | 4559 | |
bc888666 | 4560 | /* |
d3c2ae1c | 4561 | * If retry is B_TRUE, a spa must not be specified since we have |
ca0bf58d PS |
4562 | * no good way to determine if all of a spa's buffers have been |
4563 | * evicted from an arc state. | |
bc888666 | 4564 | */ |
ca0bf58d | 4565 | ASSERT(!retry || spa == 0); |
d164b209 | 4566 | |
b9541d6b | 4567 | if (spa != NULL) |
3541dc6d | 4568 | guid = spa_load_guid(spa); |
d164b209 | 4569 | |
ca0bf58d PS |
4570 | (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); |
4571 | (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); | |
4572 | ||
4573 | (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); | |
4574 | (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); | |
4575 | ||
4576 | (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); | |
4577 | (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); | |
34dc7c2f | 4578 | |
ca0bf58d PS |
4579 | (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); |
4580 | (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); | |
34dc7c2f BB |
4581 | } |
4582 | ||
c9c9c1e2 | 4583 | void |
3ec34e55 | 4584 | arc_reduce_target_size(int64_t to_free) |
34dc7c2f | 4585 | { |
37fb3e43 | 4586 | uint64_t asize = aggsum_value(&arc_size); |
1b8951b3 | 4587 | uint64_t c = arc_c; |
34dc7c2f | 4588 | |
1b8951b3 TC |
4589 | if (c > to_free && c - to_free > arc_c_min) { |
4590 | arc_c = c - to_free; | |
ca67b33a | 4591 | atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); |
34dc7c2f BB |
4592 | if (arc_p > arc_c) |
4593 | arc_p = (arc_c >> 1); | |
4594 | ASSERT(arc_c >= arc_c_min); | |
4595 | ASSERT((int64_t)arc_p >= 0); | |
1b8951b3 TC |
4596 | } else { |
4597 | arc_c = arc_c_min; | |
34dc7c2f BB |
4598 | } |
4599 | ||
3ec34e55 BL |
4600 | if (asize > arc_c) { |
4601 | /* See comment in arc_adjust_cb_check() on why lock+flag */ | |
4602 | mutex_enter(&arc_adjust_lock); | |
4603 | arc_adjust_needed = B_TRUE; | |
4604 | mutex_exit(&arc_adjust_lock); | |
4605 | zthr_wakeup(arc_adjust_zthr); | |
4606 | } | |
34dc7c2f | 4607 | } |
ca67b33a MA |
4608 | |
4609 | /* | |
4610 | * Determine if the system is under memory pressure and is asking | |
d3c2ae1c | 4611 | * to reclaim memory. A return value of B_TRUE indicates that the system |
ca67b33a MA |
4612 | * is under memory pressure and that the arc should adjust accordingly. |
4613 | */ | |
c9c9c1e2 | 4614 | boolean_t |
ca67b33a MA |
4615 | arc_reclaim_needed(void) |
4616 | { | |
4617 | return (arc_available_memory() < 0); | |
4618 | } | |
4619 | ||
c9c9c1e2 | 4620 | void |
3ec34e55 | 4621 | arc_kmem_reap_soon(void) |
34dc7c2f BB |
4622 | { |
4623 | size_t i; | |
4624 | kmem_cache_t *prev_cache = NULL; | |
4625 | kmem_cache_t *prev_data_cache = NULL; | |
4626 | extern kmem_cache_t *zio_buf_cache[]; | |
4627 | extern kmem_cache_t *zio_data_buf_cache[]; | |
34dc7c2f | 4628 | |
70f02287 | 4629 | #ifdef _KERNEL |
37fb3e43 PD |
4630 | if ((aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) && |
4631 | zfs_arc_meta_prune) { | |
f6046738 BB |
4632 | /* |
4633 | * We are exceeding our meta-data cache limit. | |
4634 | * Prune some entries to release holds on meta-data. | |
4635 | */ | |
ef5b2e10 | 4636 | arc_prune_async(zfs_arc_meta_prune); |
f6046738 | 4637 | } |
70f02287 BB |
4638 | #if defined(_ILP32) |
4639 | /* | |
4640 | * Reclaim unused memory from all kmem caches. | |
4641 | */ | |
4642 | kmem_reap(); | |
4643 | #endif | |
4644 | #endif | |
f6046738 | 4645 | |
34dc7c2f | 4646 | for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { |
70f02287 | 4647 | #if defined(_ILP32) |
d0c614ec | 4648 | /* reach upper limit of cache size on 32-bit */ |
4649 | if (zio_buf_cache[i] == NULL) | |
4650 | break; | |
4651 | #endif | |
34dc7c2f BB |
4652 | if (zio_buf_cache[i] != prev_cache) { |
4653 | prev_cache = zio_buf_cache[i]; | |
4654 | kmem_cache_reap_now(zio_buf_cache[i]); | |
4655 | } | |
4656 | if (zio_data_buf_cache[i] != prev_data_cache) { | |
4657 | prev_data_cache = zio_data_buf_cache[i]; | |
4658 | kmem_cache_reap_now(zio_data_buf_cache[i]); | |
4659 | } | |
4660 | } | |
ca0bf58d | 4661 | kmem_cache_reap_now(buf_cache); |
b9541d6b CW |
4662 | kmem_cache_reap_now(hdr_full_cache); |
4663 | kmem_cache_reap_now(hdr_l2only_cache); | |
ca577779 | 4664 | kmem_cache_reap_now(zfs_btree_leaf_cache); |
ca67b33a MA |
4665 | |
4666 | if (zio_arena != NULL) { | |
4667 | /* | |
4668 | * Ask the vmem arena to reclaim unused memory from its | |
4669 | * quantum caches. | |
4670 | */ | |
4671 | vmem_qcache_reap(zio_arena); | |
4672 | } | |
34dc7c2f BB |
4673 | } |
4674 | ||
3ec34e55 BL |
4675 | /* ARGSUSED */ |
4676 | static boolean_t | |
4677 | arc_adjust_cb_check(void *arg, zthr_t *zthr) | |
4678 | { | |
1c44a5c9 SD |
4679 | if (!arc_initialized) |
4680 | return (B_FALSE); | |
4681 | ||
cffa8372 JG |
4682 | /* |
4683 | * This is necessary so that any changes which may have been made to | |
4684 | * many of the zfs_arc_* module parameters will be propagated to | |
4685 | * their actual internal variable counterparts. Without this, | |
4686 | * changing those module params at runtime would have no effect. | |
4687 | */ | |
36a6e233 | 4688 | arc_tuning_update(B_FALSE); |
cffa8372 | 4689 | |
3ec34e55 BL |
4690 | /* |
4691 | * This is necessary in order to keep the kstat information | |
4692 | * up to date for tools that display kstat data such as the | |
4693 | * mdb ::arc dcmd and the Linux crash utility. These tools | |
4694 | * typically do not call kstat's update function, but simply | |
4695 | * dump out stats from the most recent update. Without | |
4696 | * this call, these commands may show stale stats for the | |
4697 | * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even | |
4698 | * with this change, the data might be up to 1 second | |
4699 | * out of date(the arc_adjust_zthr has a maximum sleep | |
4700 | * time of 1 second); but that should suffice. The | |
4701 | * arc_state_t structures can be queried directly if more | |
4702 | * accurate information is needed. | |
4703 | */ | |
4704 | if (arc_ksp != NULL) | |
4705 | arc_ksp->ks_update(arc_ksp, KSTAT_READ); | |
4706 | ||
4707 | /* | |
4708 | * We have to rely on arc_get_data_impl() to tell us when to adjust, | |
4709 | * rather than checking if we are overflowing here, so that we are | |
4710 | * sure to not leave arc_get_data_impl() waiting on | |
4711 | * arc_adjust_waiters_cv. If we have become "not overflowing" since | |
4712 | * arc_get_data_impl() checked, we need to wake it up. We could | |
4713 | * broadcast the CV here, but arc_get_data_impl() may have not yet | |
4714 | * gone to sleep. We would need to use a mutex to ensure that this | |
4715 | * function doesn't broadcast until arc_get_data_impl() has gone to | |
4716 | * sleep (e.g. the arc_adjust_lock). However, the lock ordering of | |
4717 | * such a lock would necessarily be incorrect with respect to the | |
4718 | * zthr_lock, which is held before this function is called, and is | |
4719 | * held by arc_get_data_impl() when it calls zthr_wakeup(). | |
4720 | */ | |
4721 | return (arc_adjust_needed); | |
4722 | } | |
4723 | ||
302f753f | 4724 | /* |
3ec34e55 BL |
4725 | * Keep arc_size under arc_c by running arc_adjust which evicts data |
4726 | * from the ARC. | |
302f753f | 4727 | */ |
867959b5 | 4728 | /* ARGSUSED */ |
61c3391a | 4729 | static void |
3ec34e55 | 4730 | arc_adjust_cb(void *arg, zthr_t *zthr) |
34dc7c2f | 4731 | { |
3ec34e55 BL |
4732 | uint64_t evicted = 0; |
4733 | fstrans_cookie_t cookie = spl_fstrans_mark(); | |
34dc7c2f | 4734 | |
3ec34e55 BL |
4735 | /* Evict from cache */ |
4736 | evicted = arc_adjust(); | |
34dc7c2f | 4737 | |
3ec34e55 BL |
4738 | /* |
4739 | * If evicted is zero, we couldn't evict anything | |
4740 | * via arc_adjust(). This could be due to hash lock | |
4741 | * collisions, but more likely due to the majority of | |
4742 | * arc buffers being unevictable. Therefore, even if | |
4743 | * arc_size is above arc_c, another pass is unlikely to | |
4744 | * be helpful and could potentially cause us to enter an | |
4745 | * infinite loop. Additionally, zthr_iscancelled() is | |
4746 | * checked here so that if the arc is shutting down, the | |
4747 | * broadcast will wake any remaining arc adjust waiters. | |
4748 | */ | |
4749 | mutex_enter(&arc_adjust_lock); | |
4750 | arc_adjust_needed = !zthr_iscancelled(arc_adjust_zthr) && | |
4751 | evicted > 0 && aggsum_compare(&arc_size, arc_c) > 0; | |
4752 | if (!arc_adjust_needed) { | |
d3c2ae1c | 4753 | /* |
3ec34e55 BL |
4754 | * We're either no longer overflowing, or we |
4755 | * can't evict anything more, so we should wake | |
4756 | * arc_get_data_impl() sooner. | |
d3c2ae1c | 4757 | */ |
3ec34e55 BL |
4758 | cv_broadcast(&arc_adjust_waiters_cv); |
4759 | arc_need_free = 0; | |
4760 | } | |
4761 | mutex_exit(&arc_adjust_lock); | |
4762 | spl_fstrans_unmark(cookie); | |
3ec34e55 BL |
4763 | } |
4764 | ||
4765 | /* ARGSUSED */ | |
4766 | static boolean_t | |
4767 | arc_reap_cb_check(void *arg, zthr_t *zthr) | |
4768 | { | |
1c44a5c9 SD |
4769 | if (!arc_initialized) |
4770 | return (B_FALSE); | |
4771 | ||
3ec34e55 BL |
4772 | int64_t free_memory = arc_available_memory(); |
4773 | ||
4774 | /* | |
4775 | * If a kmem reap is already active, don't schedule more. We must | |
4776 | * check for this because kmem_cache_reap_soon() won't actually | |
4777 | * block on the cache being reaped (this is to prevent callers from | |
4778 | * becoming implicitly blocked by a system-wide kmem reap -- which, | |
4779 | * on a system with many, many full magazines, can take minutes). | |
4780 | */ | |
4781 | if (!kmem_cache_reap_active() && free_memory < 0) { | |
34dc7c2f | 4782 | |
3ec34e55 BL |
4783 | arc_no_grow = B_TRUE; |
4784 | arc_warm = B_TRUE; | |
0a252dae | 4785 | /* |
3ec34e55 BL |
4786 | * Wait at least zfs_grow_retry (default 5) seconds |
4787 | * before considering growing. | |
0a252dae | 4788 | */ |
3ec34e55 BL |
4789 | arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry); |
4790 | return (B_TRUE); | |
4791 | } else if (free_memory < arc_c >> arc_no_grow_shift) { | |
4792 | arc_no_grow = B_TRUE; | |
4793 | } else if (gethrtime() >= arc_growtime) { | |
4794 | arc_no_grow = B_FALSE; | |
4795 | } | |
0a252dae | 4796 | |
3ec34e55 BL |
4797 | return (B_FALSE); |
4798 | } | |
34dc7c2f | 4799 | |
3ec34e55 BL |
4800 | /* |
4801 | * Keep enough free memory in the system by reaping the ARC's kmem | |
4802 | * caches. To cause more slabs to be reapable, we may reduce the | |
4803 | * target size of the cache (arc_c), causing the arc_adjust_cb() | |
4804 | * to free more buffers. | |
4805 | */ | |
4806 | /* ARGSUSED */ | |
61c3391a | 4807 | static void |
3ec34e55 BL |
4808 | arc_reap_cb(void *arg, zthr_t *zthr) |
4809 | { | |
4810 | int64_t free_memory; | |
4811 | fstrans_cookie_t cookie = spl_fstrans_mark(); | |
34dc7c2f | 4812 | |
3ec34e55 BL |
4813 | /* |
4814 | * Kick off asynchronous kmem_reap()'s of all our caches. | |
4815 | */ | |
4816 | arc_kmem_reap_soon(); | |
6a8f9b6b | 4817 | |
3ec34e55 BL |
4818 | /* |
4819 | * Wait at least arc_kmem_cache_reap_retry_ms between | |
4820 | * arc_kmem_reap_soon() calls. Without this check it is possible to | |
4821 | * end up in a situation where we spend lots of time reaping | |
4822 | * caches, while we're near arc_c_min. Waiting here also gives the | |
4823 | * subsequent free memory check a chance of finding that the | |
4824 | * asynchronous reap has already freed enough memory, and we don't | |
4825 | * need to call arc_reduce_target_size(). | |
4826 | */ | |
4827 | delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000); | |
34dc7c2f | 4828 | |
3ec34e55 BL |
4829 | /* |
4830 | * Reduce the target size as needed to maintain the amount of free | |
4831 | * memory in the system at a fraction of the arc_size (1/128th by | |
4832 | * default). If oversubscribed (free_memory < 0) then reduce the | |
4833 | * target arc_size by the deficit amount plus the fractional | |
4834 | * amount. If free memory is positive but less then the fractional | |
4835 | * amount, reduce by what is needed to hit the fractional amount. | |
4836 | */ | |
4837 | free_memory = arc_available_memory(); | |
34dc7c2f | 4838 | |
3ec34e55 BL |
4839 | int64_t to_free = |
4840 | (arc_c >> arc_shrink_shift) - free_memory; | |
4841 | if (to_free > 0) { | |
ca67b33a | 4842 | #ifdef _KERNEL |
3ec34e55 | 4843 | to_free = MAX(to_free, arc_need_free); |
ca67b33a | 4844 | #endif |
3ec34e55 | 4845 | arc_reduce_target_size(to_free); |
ca0bf58d | 4846 | } |
ca0bf58d | 4847 | spl_fstrans_unmark(cookie); |
ca0bf58d PS |
4848 | } |
4849 | ||
7cb67b45 BB |
4850 | #ifdef _KERNEL |
4851 | /* | |
302f753f BB |
4852 | * Determine the amount of memory eligible for eviction contained in the |
4853 | * ARC. All clean data reported by the ghost lists can always be safely | |
4854 | * evicted. Due to arc_c_min, the same does not hold for all clean data | |
4855 | * contained by the regular mru and mfu lists. | |
4856 | * | |
4857 | * In the case of the regular mru and mfu lists, we need to report as | |
4858 | * much clean data as possible, such that evicting that same reported | |
4859 | * data will not bring arc_size below arc_c_min. Thus, in certain | |
4860 | * circumstances, the total amount of clean data in the mru and mfu | |
4861 | * lists might not actually be evictable. | |
4862 | * | |
4863 | * The following two distinct cases are accounted for: | |
4864 | * | |
4865 | * 1. The sum of the amount of dirty data contained by both the mru and | |
4866 | * mfu lists, plus the ARC's other accounting (e.g. the anon list), | |
4867 | * is greater than or equal to arc_c_min. | |
4868 | * (i.e. amount of dirty data >= arc_c_min) | |
4869 | * | |
4870 | * This is the easy case; all clean data contained by the mru and mfu | |
4871 | * lists is evictable. Evicting all clean data can only drop arc_size | |
4872 | * to the amount of dirty data, which is greater than arc_c_min. | |
4873 | * | |
4874 | * 2. The sum of the amount of dirty data contained by both the mru and | |
4875 | * mfu lists, plus the ARC's other accounting (e.g. the anon list), | |
4876 | * is less than arc_c_min. | |
4877 | * (i.e. arc_c_min > amount of dirty data) | |
4878 | * | |
4879 | * 2.1. arc_size is greater than or equal arc_c_min. | |
4880 | * (i.e. arc_size >= arc_c_min > amount of dirty data) | |
4881 | * | |
4882 | * In this case, not all clean data from the regular mru and mfu | |
4883 | * lists is actually evictable; we must leave enough clean data | |
4884 | * to keep arc_size above arc_c_min. Thus, the maximum amount of | |
4885 | * evictable data from the two lists combined, is exactly the | |
4886 | * difference between arc_size and arc_c_min. | |
4887 | * | |
4888 | * 2.2. arc_size is less than arc_c_min | |
4889 | * (i.e. arc_c_min > arc_size > amount of dirty data) | |
4890 | * | |
4891 | * In this case, none of the data contained in the mru and mfu | |
4892 | * lists is evictable, even if it's clean. Since arc_size is | |
4893 | * already below arc_c_min, evicting any more would only | |
4894 | * increase this negative difference. | |
7cb67b45 | 4895 | */ |
7cb67b45 | 4896 | |
7cb67b45 BB |
4897 | #endif /* _KERNEL */ |
4898 | ||
34dc7c2f BB |
4899 | /* |
4900 | * Adapt arc info given the number of bytes we are trying to add and | |
4e33ba4c | 4901 | * the state that we are coming from. This function is only called |
34dc7c2f BB |
4902 | * when we are adding new content to the cache. |
4903 | */ | |
4904 | static void | |
4905 | arc_adapt(int bytes, arc_state_t *state) | |
4906 | { | |
4907 | int mult; | |
728d6ae9 | 4908 | uint64_t arc_p_min = (arc_c >> arc_p_min_shift); |
424fd7c3 TS |
4909 | int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size); |
4910 | int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size); | |
34dc7c2f BB |
4911 | |
4912 | if (state == arc_l2c_only) | |
4913 | return; | |
4914 | ||
4915 | ASSERT(bytes > 0); | |
4916 | /* | |
4917 | * Adapt the target size of the MRU list: | |
4918 | * - if we just hit in the MRU ghost list, then increase | |
4919 | * the target size of the MRU list. | |
4920 | * - if we just hit in the MFU ghost list, then increase | |
4921 | * the target size of the MFU list by decreasing the | |
4922 | * target size of the MRU list. | |
4923 | */ | |
4924 | if (state == arc_mru_ghost) { | |
36da08ef | 4925 | mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); |
62422785 PS |
4926 | if (!zfs_arc_p_dampener_disable) |
4927 | mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ | |
34dc7c2f | 4928 | |
728d6ae9 | 4929 | arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); |
34dc7c2f | 4930 | } else if (state == arc_mfu_ghost) { |
d164b209 BB |
4931 | uint64_t delta; |
4932 | ||
36da08ef | 4933 | mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); |
62422785 PS |
4934 | if (!zfs_arc_p_dampener_disable) |
4935 | mult = MIN(mult, 10); | |
34dc7c2f | 4936 | |
d164b209 | 4937 | delta = MIN(bytes * mult, arc_p); |
728d6ae9 | 4938 | arc_p = MAX(arc_p_min, arc_p - delta); |
34dc7c2f BB |
4939 | } |
4940 | ASSERT((int64_t)arc_p >= 0); | |
4941 | ||
3ec34e55 BL |
4942 | /* |
4943 | * Wake reap thread if we do not have any available memory | |
4944 | */ | |
ca67b33a | 4945 | if (arc_reclaim_needed()) { |
3ec34e55 | 4946 | zthr_wakeup(arc_reap_zthr); |
ca67b33a MA |
4947 | return; |
4948 | } | |
4949 | ||
34dc7c2f BB |
4950 | if (arc_no_grow) |
4951 | return; | |
4952 | ||
4953 | if (arc_c >= arc_c_max) | |
4954 | return; | |
4955 | ||
4956 | /* | |
4957 | * If we're within (2 * maxblocksize) bytes of the target | |
4958 | * cache size, increment the target cache size | |
4959 | */ | |
935434ef | 4960 | ASSERT3U(arc_c, >=, 2ULL << SPA_MAXBLOCKSHIFT); |
37fb3e43 PD |
4961 | if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >= |
4962 | 0) { | |
34dc7c2f BB |
4963 | atomic_add_64(&arc_c, (int64_t)bytes); |
4964 | if (arc_c > arc_c_max) | |
4965 | arc_c = arc_c_max; | |
4966 | else if (state == arc_anon) | |
4967 | atomic_add_64(&arc_p, (int64_t)bytes); | |
4968 | if (arc_p > arc_c) | |
4969 | arc_p = arc_c; | |
4970 | } | |
4971 | ASSERT((int64_t)arc_p >= 0); | |
4972 | } | |
4973 | ||
4974 | /* | |
ca0bf58d PS |
4975 | * Check if arc_size has grown past our upper threshold, determined by |
4976 | * zfs_arc_overflow_shift. | |
34dc7c2f | 4977 | */ |
ca0bf58d PS |
4978 | static boolean_t |
4979 | arc_is_overflowing(void) | |
34dc7c2f | 4980 | { |
ca0bf58d | 4981 | /* Always allow at least one block of overflow */ |
5a902f5a | 4982 | int64_t overflow = MAX(SPA_MAXBLOCKSIZE, |
ca0bf58d | 4983 | arc_c >> zfs_arc_overflow_shift); |
34dc7c2f | 4984 | |
37fb3e43 PD |
4985 | /* |
4986 | * We just compare the lower bound here for performance reasons. Our | |
4987 | * primary goals are to make sure that the arc never grows without | |
4988 | * bound, and that it can reach its maximum size. This check | |
4989 | * accomplishes both goals. The maximum amount we could run over by is | |
4990 | * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block | |
4991 | * in the ARC. In practice, that's in the tens of MB, which is low | |
4992 | * enough to be safe. | |
4993 | */ | |
5a902f5a | 4994 | return (aggsum_lower_bound(&arc_size) >= (int64_t)arc_c + overflow); |
34dc7c2f BB |
4995 | } |
4996 | ||
a6255b7f DQ |
4997 | static abd_t * |
4998 | arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag) | |
4999 | { | |
5000 | arc_buf_contents_t type = arc_buf_type(hdr); | |
5001 | ||
5002 | arc_get_data_impl(hdr, size, tag); | |
5003 | if (type == ARC_BUFC_METADATA) { | |
5004 | return (abd_alloc(size, B_TRUE)); | |
5005 | } else { | |
5006 | ASSERT(type == ARC_BUFC_DATA); | |
5007 | return (abd_alloc(size, B_FALSE)); | |
5008 | } | |
5009 | } | |
5010 | ||
5011 | static void * | |
5012 | arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) | |
5013 | { | |
5014 | arc_buf_contents_t type = arc_buf_type(hdr); | |
5015 | ||
5016 | arc_get_data_impl(hdr, size, tag); | |
5017 | if (type == ARC_BUFC_METADATA) { | |
5018 | return (zio_buf_alloc(size)); | |
5019 | } else { | |
5020 | ASSERT(type == ARC_BUFC_DATA); | |
5021 | return (zio_data_buf_alloc(size)); | |
5022 | } | |
5023 | } | |
5024 | ||
34dc7c2f | 5025 | /* |
d3c2ae1c GW |
5026 | * Allocate a block and return it to the caller. If we are hitting the |
5027 | * hard limit for the cache size, we must sleep, waiting for the eviction | |
5028 | * thread to catch up. If we're past the target size but below the hard | |
5029 | * limit, we'll only signal the reclaim thread and continue on. | |
34dc7c2f | 5030 | */ |
a6255b7f DQ |
5031 | static void |
5032 | arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) | |
34dc7c2f | 5033 | { |
a6255b7f DQ |
5034 | arc_state_t *state = hdr->b_l1hdr.b_state; |
5035 | arc_buf_contents_t type = arc_buf_type(hdr); | |
34dc7c2f BB |
5036 | |
5037 | arc_adapt(size, state); | |
5038 | ||
5039 | /* | |
ca0bf58d PS |
5040 | * If arc_size is currently overflowing, and has grown past our |
5041 | * upper limit, we must be adding data faster than the evict | |
5042 | * thread can evict. Thus, to ensure we don't compound the | |
5043 | * problem by adding more data and forcing arc_size to grow even | |
5044 | * further past it's target size, we halt and wait for the | |
5045 | * eviction thread to catch up. | |
5046 | * | |
5047 | * It's also possible that the reclaim thread is unable to evict | |
5048 | * enough buffers to get arc_size below the overflow limit (e.g. | |
5049 | * due to buffers being un-evictable, or hash lock collisions). | |
5050 | * In this case, we want to proceed regardless if we're | |
5051 | * overflowing; thus we don't use a while loop here. | |
34dc7c2f | 5052 | */ |
ca0bf58d | 5053 | if (arc_is_overflowing()) { |
3ec34e55 | 5054 | mutex_enter(&arc_adjust_lock); |
ca0bf58d PS |
5055 | |
5056 | /* | |
5057 | * Now that we've acquired the lock, we may no longer be | |
5058 | * over the overflow limit, lets check. | |
5059 | * | |
5060 | * We're ignoring the case of spurious wake ups. If that | |
5061 | * were to happen, it'd let this thread consume an ARC | |
5062 | * buffer before it should have (i.e. before we're under | |
5063 | * the overflow limit and were signalled by the reclaim | |
5064 | * thread). As long as that is a rare occurrence, it | |
5065 | * shouldn't cause any harm. | |
5066 | */ | |
5067 | if (arc_is_overflowing()) { | |
3ec34e55 BL |
5068 | arc_adjust_needed = B_TRUE; |
5069 | zthr_wakeup(arc_adjust_zthr); | |
5070 | (void) cv_wait(&arc_adjust_waiters_cv, | |
5071 | &arc_adjust_lock); | |
34dc7c2f | 5072 | } |
3ec34e55 | 5073 | mutex_exit(&arc_adjust_lock); |
34dc7c2f | 5074 | } |
ab26409d | 5075 | |
d3c2ae1c | 5076 | VERIFY3U(hdr->b_type, ==, type); |
da8ccd0e | 5077 | if (type == ARC_BUFC_METADATA) { |
ca0bf58d PS |
5078 | arc_space_consume(size, ARC_SPACE_META); |
5079 | } else { | |
ca0bf58d | 5080 | arc_space_consume(size, ARC_SPACE_DATA); |
da8ccd0e PS |
5081 | } |
5082 | ||
34dc7c2f BB |
5083 | /* |
5084 | * Update the state size. Note that ghost states have a | |
5085 | * "ghost size" and so don't need to be updated. | |
5086 | */ | |
d3c2ae1c | 5087 | if (!GHOST_STATE(state)) { |
34dc7c2f | 5088 | |
424fd7c3 | 5089 | (void) zfs_refcount_add_many(&state->arcs_size, size, tag); |
ca0bf58d PS |
5090 | |
5091 | /* | |
5092 | * If this is reached via arc_read, the link is | |
5093 | * protected by the hash lock. If reached via | |
5094 | * arc_buf_alloc, the header should not be accessed by | |
5095 | * any other thread. And, if reached via arc_read_done, | |
5096 | * the hash lock will protect it if it's found in the | |
5097 | * hash table; otherwise no other thread should be | |
5098 | * trying to [add|remove]_reference it. | |
5099 | */ | |
5100 | if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { | |
424fd7c3 TS |
5101 | ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
5102 | (void) zfs_refcount_add_many(&state->arcs_esize[type], | |
d3c2ae1c | 5103 | size, tag); |
34dc7c2f | 5104 | } |
d3c2ae1c | 5105 | |
34dc7c2f BB |
5106 | /* |
5107 | * If we are growing the cache, and we are adding anonymous | |
5108 | * data, and we have outgrown arc_p, update arc_p | |
5109 | */ | |
c1b5801b | 5110 | if (aggsum_upper_bound(&arc_size) < arc_c && |
37fb3e43 | 5111 | hdr->b_l1hdr.b_state == arc_anon && |
424fd7c3 TS |
5112 | (zfs_refcount_count(&arc_anon->arcs_size) + |
5113 | zfs_refcount_count(&arc_mru->arcs_size) > arc_p)) | |
34dc7c2f BB |
5114 | arc_p = MIN(arc_c, arc_p + size); |
5115 | } | |
a6255b7f DQ |
5116 | } |
5117 | ||
5118 | static void | |
5119 | arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag) | |
5120 | { | |
5121 | arc_free_data_impl(hdr, size, tag); | |
5122 | abd_free(abd); | |
5123 | } | |
5124 | ||
5125 | static void | |
5126 | arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag) | |
5127 | { | |
5128 | arc_buf_contents_t type = arc_buf_type(hdr); | |
5129 | ||
5130 | arc_free_data_impl(hdr, size, tag); | |
5131 | if (type == ARC_BUFC_METADATA) { | |
5132 | zio_buf_free(buf, size); | |
5133 | } else { | |
5134 | ASSERT(type == ARC_BUFC_DATA); | |
5135 | zio_data_buf_free(buf, size); | |
5136 | } | |
d3c2ae1c GW |
5137 | } |
5138 | ||
5139 | /* | |
5140 | * Free the arc data buffer. | |
5141 | */ | |
5142 | static void | |
a6255b7f | 5143 | arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) |
d3c2ae1c GW |
5144 | { |
5145 | arc_state_t *state = hdr->b_l1hdr.b_state; | |
5146 | arc_buf_contents_t type = arc_buf_type(hdr); | |
5147 | ||
5148 | /* protected by hash lock, if in the hash table */ | |
5149 | if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { | |
424fd7c3 | 5150 | ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
d3c2ae1c GW |
5151 | ASSERT(state != arc_anon && state != arc_l2c_only); |
5152 | ||
424fd7c3 | 5153 | (void) zfs_refcount_remove_many(&state->arcs_esize[type], |
d3c2ae1c GW |
5154 | size, tag); |
5155 | } | |
424fd7c3 | 5156 | (void) zfs_refcount_remove_many(&state->arcs_size, size, tag); |
d3c2ae1c GW |
5157 | |
5158 | VERIFY3U(hdr->b_type, ==, type); | |
5159 | if (type == ARC_BUFC_METADATA) { | |
d3c2ae1c GW |
5160 | arc_space_return(size, ARC_SPACE_META); |
5161 | } else { | |
5162 | ASSERT(type == ARC_BUFC_DATA); | |
d3c2ae1c GW |
5163 | arc_space_return(size, ARC_SPACE_DATA); |
5164 | } | |
34dc7c2f BB |
5165 | } |
5166 | ||
5167 | /* | |
5168 | * This routine is called whenever a buffer is accessed. | |
5169 | * NOTE: the hash lock is dropped in this function. | |
5170 | */ | |
5171 | static void | |
2a432414 | 5172 | arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) |
34dc7c2f | 5173 | { |
428870ff BB |
5174 | clock_t now; |
5175 | ||
34dc7c2f | 5176 | ASSERT(MUTEX_HELD(hash_lock)); |
b9541d6b | 5177 | ASSERT(HDR_HAS_L1HDR(hdr)); |
34dc7c2f | 5178 | |
b9541d6b | 5179 | if (hdr->b_l1hdr.b_state == arc_anon) { |
34dc7c2f BB |
5180 | /* |
5181 | * This buffer is not in the cache, and does not | |
5182 | * appear in our "ghost" list. Add the new buffer | |
5183 | * to the MRU state. | |
5184 | */ | |
5185 | ||
b9541d6b CW |
5186 | ASSERT0(hdr->b_l1hdr.b_arc_access); |
5187 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); | |
2a432414 GW |
5188 | DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); |
5189 | arc_change_state(arc_mru, hdr, hash_lock); | |
34dc7c2f | 5190 | |
b9541d6b | 5191 | } else if (hdr->b_l1hdr.b_state == arc_mru) { |
428870ff BB |
5192 | now = ddi_get_lbolt(); |
5193 | ||
34dc7c2f BB |
5194 | /* |
5195 | * If this buffer is here because of a prefetch, then either: | |
5196 | * - clear the flag if this is a "referencing" read | |
5197 | * (any subsequent access will bump this into the MFU state). | |
5198 | * or | |
5199 | * - move the buffer to the head of the list if this is | |
5200 | * another prefetch (to make it less likely to be evicted). | |
5201 | */ | |
d4a72f23 | 5202 | if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) { |
424fd7c3 | 5203 | if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { |
ca0bf58d PS |
5204 | /* link protected by hash lock */ |
5205 | ASSERT(multilist_link_active( | |
b9541d6b | 5206 | &hdr->b_l1hdr.b_arc_node)); |
34dc7c2f | 5207 | } else { |
d4a72f23 TC |
5208 | arc_hdr_clear_flags(hdr, |
5209 | ARC_FLAG_PREFETCH | | |
5210 | ARC_FLAG_PRESCIENT_PREFETCH); | |
b9541d6b | 5211 | atomic_inc_32(&hdr->b_l1hdr.b_mru_hits); |
34dc7c2f BB |
5212 | ARCSTAT_BUMP(arcstat_mru_hits); |
5213 | } | |
b9541d6b | 5214 | hdr->b_l1hdr.b_arc_access = now; |
34dc7c2f BB |
5215 | return; |
5216 | } | |
5217 | ||
5218 | /* | |
5219 | * This buffer has been "accessed" only once so far, | |
5220 | * but it is still in the cache. Move it to the MFU | |
5221 | * state. | |
5222 | */ | |
b9541d6b CW |
5223 | if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access + |
5224 | ARC_MINTIME)) { | |
34dc7c2f BB |
5225 | /* |
5226 | * More than 125ms have passed since we | |
5227 | * instantiated this buffer. Move it to the | |
5228 | * most frequently used state. | |
5229 | */ | |
b9541d6b | 5230 | hdr->b_l1hdr.b_arc_access = now; |
2a432414 GW |
5231 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); |
5232 | arc_change_state(arc_mfu, hdr, hash_lock); | |
34dc7c2f | 5233 | } |
b9541d6b | 5234 | atomic_inc_32(&hdr->b_l1hdr.b_mru_hits); |
34dc7c2f | 5235 | ARCSTAT_BUMP(arcstat_mru_hits); |
b9541d6b | 5236 | } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { |
34dc7c2f BB |
5237 | arc_state_t *new_state; |
5238 | /* | |
5239 | * This buffer has been "accessed" recently, but | |
5240 | * was evicted from the cache. Move it to the | |
5241 | * MFU state. | |
5242 | */ | |
5243 | ||
d4a72f23 | 5244 | if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) { |
34dc7c2f | 5245 | new_state = arc_mru; |
424fd7c3 | 5246 | if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) { |
d4a72f23 TC |
5247 | arc_hdr_clear_flags(hdr, |
5248 | ARC_FLAG_PREFETCH | | |
5249 | ARC_FLAG_PRESCIENT_PREFETCH); | |
5250 | } | |
2a432414 | 5251 | DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); |
34dc7c2f BB |
5252 | } else { |
5253 | new_state = arc_mfu; | |
2a432414 | 5254 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); |
34dc7c2f BB |
5255 | } |
5256 | ||
b9541d6b | 5257 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); |
2a432414 | 5258 | arc_change_state(new_state, hdr, hash_lock); |
34dc7c2f | 5259 | |
b9541d6b | 5260 | atomic_inc_32(&hdr->b_l1hdr.b_mru_ghost_hits); |
34dc7c2f | 5261 | ARCSTAT_BUMP(arcstat_mru_ghost_hits); |
b9541d6b | 5262 | } else if (hdr->b_l1hdr.b_state == arc_mfu) { |
34dc7c2f BB |
5263 | /* |
5264 | * This buffer has been accessed more than once and is | |
5265 | * still in the cache. Keep it in the MFU state. | |
5266 | * | |
5267 | * NOTE: an add_reference() that occurred when we did | |
5268 | * the arc_read() will have kicked this off the list. | |
5269 | * If it was a prefetch, we will explicitly move it to | |
5270 | * the head of the list now. | |
5271 | */ | |
d4a72f23 | 5272 | |
b9541d6b | 5273 | atomic_inc_32(&hdr->b_l1hdr.b_mfu_hits); |
34dc7c2f | 5274 | ARCSTAT_BUMP(arcstat_mfu_hits); |
b9541d6b CW |
5275 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); |
5276 | } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { | |
34dc7c2f BB |
5277 | arc_state_t *new_state = arc_mfu; |
5278 | /* | |
5279 | * This buffer has been accessed more than once but has | |
5280 | * been evicted from the cache. Move it back to the | |
5281 | * MFU state. | |
5282 | */ | |
5283 | ||
d4a72f23 | 5284 | if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) { |
34dc7c2f BB |
5285 | /* |
5286 | * This is a prefetch access... | |
5287 | * move this block back to the MRU state. | |
5288 | */ | |
34dc7c2f BB |
5289 | new_state = arc_mru; |
5290 | } | |
5291 | ||
b9541d6b | 5292 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); |
2a432414 GW |
5293 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); |
5294 | arc_change_state(new_state, hdr, hash_lock); | |
34dc7c2f | 5295 | |
b9541d6b | 5296 | atomic_inc_32(&hdr->b_l1hdr.b_mfu_ghost_hits); |
34dc7c2f | 5297 | ARCSTAT_BUMP(arcstat_mfu_ghost_hits); |
b9541d6b | 5298 | } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { |
34dc7c2f BB |
5299 | /* |
5300 | * This buffer is on the 2nd Level ARC. | |
5301 | */ | |
5302 | ||
b9541d6b | 5303 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); |
2a432414 GW |
5304 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); |
5305 | arc_change_state(arc_mfu, hdr, hash_lock); | |
34dc7c2f | 5306 | } else { |
b9541d6b CW |
5307 | cmn_err(CE_PANIC, "invalid arc state 0x%p", |
5308 | hdr->b_l1hdr.b_state); | |
34dc7c2f BB |
5309 | } |
5310 | } | |
5311 | ||
0873bb63 BB |
5312 | /* |
5313 | * This routine is called by dbuf_hold() to update the arc_access() state | |
5314 | * which otherwise would be skipped for entries in the dbuf cache. | |
5315 | */ | |
5316 | void | |
5317 | arc_buf_access(arc_buf_t *buf) | |
5318 | { | |
5319 | mutex_enter(&buf->b_evict_lock); | |
5320 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
5321 | ||
5322 | /* | |
5323 | * Avoid taking the hash_lock when possible as an optimization. | |
5324 | * The header must be checked again under the hash_lock in order | |
5325 | * to handle the case where it is concurrently being released. | |
5326 | */ | |
5327 | if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { | |
5328 | mutex_exit(&buf->b_evict_lock); | |
5329 | return; | |
5330 | } | |
5331 | ||
5332 | kmutex_t *hash_lock = HDR_LOCK(hdr); | |
5333 | mutex_enter(hash_lock); | |
5334 | ||
5335 | if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { | |
5336 | mutex_exit(hash_lock); | |
5337 | mutex_exit(&buf->b_evict_lock); | |
5338 | ARCSTAT_BUMP(arcstat_access_skip); | |
5339 | return; | |
5340 | } | |
5341 | ||
5342 | mutex_exit(&buf->b_evict_lock); | |
5343 | ||
5344 | ASSERT(hdr->b_l1hdr.b_state == arc_mru || | |
5345 | hdr->b_l1hdr.b_state == arc_mfu); | |
5346 | ||
5347 | DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); | |
5348 | arc_access(hdr, hash_lock); | |
5349 | mutex_exit(hash_lock); | |
5350 | ||
5351 | ARCSTAT_BUMP(arcstat_hits); | |
5352 | ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr) && !HDR_PRESCIENT_PREFETCH(hdr), | |
5353 | demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits); | |
5354 | } | |
5355 | ||
b5256303 | 5356 | /* a generic arc_read_done_func_t which you can use */ |
34dc7c2f BB |
5357 | /* ARGSUSED */ |
5358 | void | |
d4a72f23 TC |
5359 | arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, |
5360 | arc_buf_t *buf, void *arg) | |
34dc7c2f | 5361 | { |
d4a72f23 TC |
5362 | if (buf == NULL) |
5363 | return; | |
5364 | ||
5365 | bcopy(buf->b_data, arg, arc_buf_size(buf)); | |
d3c2ae1c | 5366 | arc_buf_destroy(buf, arg); |
34dc7c2f BB |
5367 | } |
5368 | ||
b5256303 | 5369 | /* a generic arc_read_done_func_t */ |
d4a72f23 | 5370 | /* ARGSUSED */ |
34dc7c2f | 5371 | void |
d4a72f23 TC |
5372 | arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, |
5373 | arc_buf_t *buf, void *arg) | |
34dc7c2f BB |
5374 | { |
5375 | arc_buf_t **bufp = arg; | |
d4a72f23 TC |
5376 | |
5377 | if (buf == NULL) { | |
c3bd3fb4 | 5378 | ASSERT(zio == NULL || zio->io_error != 0); |
34dc7c2f BB |
5379 | *bufp = NULL; |
5380 | } else { | |
c3bd3fb4 | 5381 | ASSERT(zio == NULL || zio->io_error == 0); |
34dc7c2f | 5382 | *bufp = buf; |
c3bd3fb4 | 5383 | ASSERT(buf->b_data != NULL); |
34dc7c2f BB |
5384 | } |
5385 | } | |
5386 | ||
d3c2ae1c GW |
5387 | static void |
5388 | arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) | |
5389 | { | |
5390 | if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { | |
5391 | ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); | |
b5256303 | 5392 | ASSERT3U(arc_hdr_get_compress(hdr), ==, ZIO_COMPRESS_OFF); |
d3c2ae1c GW |
5393 | } else { |
5394 | if (HDR_COMPRESSION_ENABLED(hdr)) { | |
b5256303 | 5395 | ASSERT3U(arc_hdr_get_compress(hdr), ==, |
d3c2ae1c GW |
5396 | BP_GET_COMPRESS(bp)); |
5397 | } | |
5398 | ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); | |
5399 | ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); | |
b5256303 | 5400 | ASSERT3U(!!HDR_PROTECTED(hdr), ==, BP_IS_PROTECTED(bp)); |
d3c2ae1c GW |
5401 | } |
5402 | } | |
5403 | ||
34dc7c2f BB |
5404 | static void |
5405 | arc_read_done(zio_t *zio) | |
5406 | { | |
b5256303 | 5407 | blkptr_t *bp = zio->io_bp; |
d3c2ae1c | 5408 | arc_buf_hdr_t *hdr = zio->io_private; |
9b67f605 | 5409 | kmutex_t *hash_lock = NULL; |
524b4217 DK |
5410 | arc_callback_t *callback_list; |
5411 | arc_callback_t *acb; | |
2aa34383 | 5412 | boolean_t freeable = B_FALSE; |
a7004725 | 5413 | |
34dc7c2f BB |
5414 | /* |
5415 | * The hdr was inserted into hash-table and removed from lists | |
5416 | * prior to starting I/O. We should find this header, since | |
5417 | * it's in the hash table, and it should be legit since it's | |
5418 | * not possible to evict it during the I/O. The only possible | |
5419 | * reason for it not to be found is if we were freed during the | |
5420 | * read. | |
5421 | */ | |
9b67f605 | 5422 | if (HDR_IN_HASH_TABLE(hdr)) { |
31df97cd DB |
5423 | arc_buf_hdr_t *found; |
5424 | ||
9b67f605 MA |
5425 | ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); |
5426 | ASSERT3U(hdr->b_dva.dva_word[0], ==, | |
5427 | BP_IDENTITY(zio->io_bp)->dva_word[0]); | |
5428 | ASSERT3U(hdr->b_dva.dva_word[1], ==, | |
5429 | BP_IDENTITY(zio->io_bp)->dva_word[1]); | |
5430 | ||
31df97cd | 5431 | found = buf_hash_find(hdr->b_spa, zio->io_bp, &hash_lock); |
9b67f605 | 5432 | |
d3c2ae1c | 5433 | ASSERT((found == hdr && |
9b67f605 MA |
5434 | DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || |
5435 | (found == hdr && HDR_L2_READING(hdr))); | |
d3c2ae1c GW |
5436 | ASSERT3P(hash_lock, !=, NULL); |
5437 | } | |
5438 | ||
b5256303 TC |
5439 | if (BP_IS_PROTECTED(bp)) { |
5440 | hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp); | |
5441 | hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset; | |
5442 | zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt, | |
5443 | hdr->b_crypt_hdr.b_iv); | |
5444 | ||
5445 | if (BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG) { | |
5446 | void *tmpbuf; | |
5447 | ||
5448 | tmpbuf = abd_borrow_buf_copy(zio->io_abd, | |
5449 | sizeof (zil_chain_t)); | |
5450 | zio_crypt_decode_mac_zil(tmpbuf, | |
5451 | hdr->b_crypt_hdr.b_mac); | |
5452 | abd_return_buf(zio->io_abd, tmpbuf, | |
5453 | sizeof (zil_chain_t)); | |
5454 | } else { | |
5455 | zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac); | |
5456 | } | |
5457 | } | |
5458 | ||
d4a72f23 | 5459 | if (zio->io_error == 0) { |
d3c2ae1c GW |
5460 | /* byteswap if necessary */ |
5461 | if (BP_SHOULD_BYTESWAP(zio->io_bp)) { | |
5462 | if (BP_GET_LEVEL(zio->io_bp) > 0) { | |
5463 | hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; | |
5464 | } else { | |
5465 | hdr->b_l1hdr.b_byteswap = | |
5466 | DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); | |
5467 | } | |
5468 | } else { | |
5469 | hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; | |
5470 | } | |
9b67f605 | 5471 | } |
34dc7c2f | 5472 | |
d3c2ae1c | 5473 | arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); |
b9541d6b | 5474 | if (l2arc_noprefetch && HDR_PREFETCH(hdr)) |
d3c2ae1c | 5475 | arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); |
34dc7c2f | 5476 | |
b9541d6b | 5477 | callback_list = hdr->b_l1hdr.b_acb; |
d3c2ae1c | 5478 | ASSERT3P(callback_list, !=, NULL); |
34dc7c2f | 5479 | |
d4a72f23 TC |
5480 | if (hash_lock && zio->io_error == 0 && |
5481 | hdr->b_l1hdr.b_state == arc_anon) { | |
428870ff BB |
5482 | /* |
5483 | * Only call arc_access on anonymous buffers. This is because | |
5484 | * if we've issued an I/O for an evicted buffer, we've already | |
5485 | * called arc_access (to prevent any simultaneous readers from | |
5486 | * getting confused). | |
5487 | */ | |
5488 | arc_access(hdr, hash_lock); | |
5489 | } | |
5490 | ||
524b4217 DK |
5491 | /* |
5492 | * If a read request has a callback (i.e. acb_done is not NULL), then we | |
5493 | * make a buf containing the data according to the parameters which were | |
5494 | * passed in. The implementation of arc_buf_alloc_impl() ensures that we | |
5495 | * aren't needlessly decompressing the data multiple times. | |
5496 | */ | |
a7004725 | 5497 | int callback_cnt = 0; |
2aa34383 DK |
5498 | for (acb = callback_list; acb != NULL; acb = acb->acb_next) { |
5499 | if (!acb->acb_done) | |
5500 | continue; | |
5501 | ||
2aa34383 | 5502 | callback_cnt++; |
524b4217 | 5503 | |
d4a72f23 TC |
5504 | if (zio->io_error != 0) |
5505 | continue; | |
5506 | ||
b5256303 | 5507 | int error = arc_buf_alloc_impl(hdr, zio->io_spa, |
be9a5c35 | 5508 | &acb->acb_zb, acb->acb_private, acb->acb_encrypted, |
d4a72f23 | 5509 | acb->acb_compressed, acb->acb_noauth, B_TRUE, |
440a3eb9 | 5510 | &acb->acb_buf); |
b5256303 TC |
5511 | |
5512 | /* | |
440a3eb9 | 5513 | * Assert non-speculative zios didn't fail because an |
b5256303 TC |
5514 | * encryption key wasn't loaded |
5515 | */ | |
a2c2ed1b | 5516 | ASSERT((zio->io_flags & ZIO_FLAG_SPECULATIVE) || |
be9a5c35 | 5517 | error != EACCES); |
b5256303 TC |
5518 | |
5519 | /* | |
5520 | * If we failed to decrypt, report an error now (as the zio | |
5521 | * layer would have done if it had done the transforms). | |
5522 | */ | |
5523 | if (error == ECKSUM) { | |
5524 | ASSERT(BP_IS_PROTECTED(bp)); | |
5525 | error = SET_ERROR(EIO); | |
b5256303 | 5526 | if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) { |
be9a5c35 | 5527 | spa_log_error(zio->io_spa, &acb->acb_zb); |
b5256303 | 5528 | zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, |
be9a5c35 | 5529 | zio->io_spa, NULL, &acb->acb_zb, zio, 0, 0); |
b5256303 TC |
5530 | } |
5531 | } | |
5532 | ||
c3bd3fb4 TC |
5533 | if (error != 0) { |
5534 | /* | |
5535 | * Decompression or decryption failed. Set | |
5536 | * io_error so that when we call acb_done | |
5537 | * (below), we will indicate that the read | |
5538 | * failed. Note that in the unusual case | |
5539 | * where one callback is compressed and another | |
5540 | * uncompressed, we will mark all of them | |
5541 | * as failed, even though the uncompressed | |
5542 | * one can't actually fail. In this case, | |
5543 | * the hdr will not be anonymous, because | |
5544 | * if there are multiple callbacks, it's | |
5545 | * because multiple threads found the same | |
5546 | * arc buf in the hash table. | |
5547 | */ | |
524b4217 | 5548 | zio->io_error = error; |
c3bd3fb4 | 5549 | } |
34dc7c2f | 5550 | } |
c3bd3fb4 TC |
5551 | |
5552 | /* | |
5553 | * If there are multiple callbacks, we must have the hash lock, | |
5554 | * because the only way for multiple threads to find this hdr is | |
5555 | * in the hash table. This ensures that if there are multiple | |
5556 | * callbacks, the hdr is not anonymous. If it were anonymous, | |
5557 | * we couldn't use arc_buf_destroy() in the error case below. | |
5558 | */ | |
5559 | ASSERT(callback_cnt < 2 || hash_lock != NULL); | |
5560 | ||
b9541d6b | 5561 | hdr->b_l1hdr.b_acb = NULL; |
d3c2ae1c | 5562 | arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); |
440a3eb9 | 5563 | if (callback_cnt == 0) |
b5256303 | 5564 | ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); |
34dc7c2f | 5565 | |
424fd7c3 | 5566 | ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || |
b9541d6b | 5567 | callback_list != NULL); |
34dc7c2f | 5568 | |
d4a72f23 | 5569 | if (zio->io_error == 0) { |
d3c2ae1c GW |
5570 | arc_hdr_verify(hdr, zio->io_bp); |
5571 | } else { | |
5572 | arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); | |
b9541d6b | 5573 | if (hdr->b_l1hdr.b_state != arc_anon) |
34dc7c2f BB |
5574 | arc_change_state(arc_anon, hdr, hash_lock); |
5575 | if (HDR_IN_HASH_TABLE(hdr)) | |
5576 | buf_hash_remove(hdr); | |
424fd7c3 | 5577 | freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt); |
34dc7c2f BB |
5578 | } |
5579 | ||
5580 | /* | |
5581 | * Broadcast before we drop the hash_lock to avoid the possibility | |
5582 | * that the hdr (and hence the cv) might be freed before we get to | |
5583 | * the cv_broadcast(). | |
5584 | */ | |
b9541d6b | 5585 | cv_broadcast(&hdr->b_l1hdr.b_cv); |
34dc7c2f | 5586 | |
b9541d6b | 5587 | if (hash_lock != NULL) { |
34dc7c2f BB |
5588 | mutex_exit(hash_lock); |
5589 | } else { | |
5590 | /* | |
5591 | * This block was freed while we waited for the read to | |
5592 | * complete. It has been removed from the hash table and | |
5593 | * moved to the anonymous state (so that it won't show up | |
5594 | * in the cache). | |
5595 | */ | |
b9541d6b | 5596 | ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); |
424fd7c3 | 5597 | freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt); |
34dc7c2f BB |
5598 | } |
5599 | ||
5600 | /* execute each callback and free its structure */ | |
5601 | while ((acb = callback_list) != NULL) { | |
c3bd3fb4 TC |
5602 | if (acb->acb_done != NULL) { |
5603 | if (zio->io_error != 0 && acb->acb_buf != NULL) { | |
5604 | /* | |
5605 | * If arc_buf_alloc_impl() fails during | |
5606 | * decompression, the buf will still be | |
5607 | * allocated, and needs to be freed here. | |
5608 | */ | |
5609 | arc_buf_destroy(acb->acb_buf, | |
5610 | acb->acb_private); | |
5611 | acb->acb_buf = NULL; | |
5612 | } | |
d4a72f23 TC |
5613 | acb->acb_done(zio, &zio->io_bookmark, zio->io_bp, |
5614 | acb->acb_buf, acb->acb_private); | |
b5256303 | 5615 | } |
34dc7c2f BB |
5616 | |
5617 | if (acb->acb_zio_dummy != NULL) { | |
5618 | acb->acb_zio_dummy->io_error = zio->io_error; | |
5619 | zio_nowait(acb->acb_zio_dummy); | |
5620 | } | |
5621 | ||
5622 | callback_list = acb->acb_next; | |
5623 | kmem_free(acb, sizeof (arc_callback_t)); | |
5624 | } | |
5625 | ||
5626 | if (freeable) | |
5627 | arc_hdr_destroy(hdr); | |
5628 | } | |
5629 | ||
5630 | /* | |
5c839890 | 5631 | * "Read" the block at the specified DVA (in bp) via the |
34dc7c2f BB |
5632 | * cache. If the block is found in the cache, invoke the provided |
5633 | * callback immediately and return. Note that the `zio' parameter | |
5634 | * in the callback will be NULL in this case, since no IO was | |
5635 | * required. If the block is not in the cache pass the read request | |
5636 | * on to the spa with a substitute callback function, so that the | |
5637 | * requested block will be added to the cache. | |
5638 | * | |
5639 | * If a read request arrives for a block that has a read in-progress, | |
5640 | * either wait for the in-progress read to complete (and return the | |
5641 | * results); or, if this is a read with a "done" func, add a record | |
5642 | * to the read to invoke the "done" func when the read completes, | |
5643 | * and return; or just return. | |
5644 | * | |
5645 | * arc_read_done() will invoke all the requested "done" functions | |
5646 | * for readers of this block. | |
5647 | */ | |
5648 | int | |
b5256303 TC |
5649 | arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, |
5650 | arc_read_done_func_t *done, void *private, zio_priority_t priority, | |
5651 | int zio_flags, arc_flags_t *arc_flags, const zbookmark_phys_t *zb) | |
34dc7c2f | 5652 | { |
9b67f605 | 5653 | arc_buf_hdr_t *hdr = NULL; |
9b67f605 | 5654 | kmutex_t *hash_lock = NULL; |
34dc7c2f | 5655 | zio_t *rzio; |
3541dc6d | 5656 | uint64_t guid = spa_load_guid(spa); |
b5256303 TC |
5657 | boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW_COMPRESS) != 0; |
5658 | boolean_t encrypted_read = BP_IS_ENCRYPTED(bp) && | |
5659 | (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0; | |
5660 | boolean_t noauth_read = BP_IS_AUTHENTICATED(bp) && | |
5661 | (zio_flags & ZIO_FLAG_RAW_ENCRYPT) != 0; | |
0902c457 | 5662 | boolean_t embedded_bp = !!BP_IS_EMBEDDED(bp); |
1421c891 | 5663 | int rc = 0; |
34dc7c2f | 5664 | |
0902c457 | 5665 | ASSERT(!embedded_bp || |
9b67f605 | 5666 | BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); |
30af21b0 PD |
5667 | ASSERT(!BP_IS_HOLE(bp)); |
5668 | ASSERT(!BP_IS_REDACTED(bp)); | |
9b67f605 | 5669 | |
1e9231ad MR |
5670 | /* |
5671 | * Normally SPL_FSTRANS will already be set since kernel threads which | |
5672 | * expect to call the DMU interfaces will set it when created. System | |
5673 | * calls are similarly handled by setting/cleaning the bit in the | |
5674 | * registered callback (module/os/.../zfs/zpl_*). | |
5675 | * | |
5676 | * External consumers such as Lustre which call the exported DMU | |
5677 | * interfaces may not have set SPL_FSTRANS. To avoid a deadlock | |
5678 | * on the hash_lock always set and clear the bit. | |
5679 | */ | |
5680 | fstrans_cookie_t cookie = spl_fstrans_mark(); | |
34dc7c2f | 5681 | top: |
0902c457 | 5682 | if (!embedded_bp) { |
9b67f605 MA |
5683 | /* |
5684 | * Embedded BP's have no DVA and require no I/O to "read". | |
5685 | * Create an anonymous arc buf to back it. | |
5686 | */ | |
5687 | hdr = buf_hash_find(guid, bp, &hash_lock); | |
5688 | } | |
5689 | ||
b5256303 TC |
5690 | /* |
5691 | * Determine if we have an L1 cache hit or a cache miss. For simplicity | |
e1cfd73f | 5692 | * we maintain encrypted data separately from compressed / uncompressed |
b5256303 TC |
5693 | * data. If the user is requesting raw encrypted data and we don't have |
5694 | * that in the header we will read from disk to guarantee that we can | |
5695 | * get it even if the encryption keys aren't loaded. | |
5696 | */ | |
5697 | if (hdr != NULL && HDR_HAS_L1HDR(hdr) && (HDR_HAS_RABD(hdr) || | |
5698 | (hdr->b_l1hdr.b_pabd != NULL && !encrypted_read))) { | |
d3c2ae1c | 5699 | arc_buf_t *buf = NULL; |
2a432414 | 5700 | *arc_flags |= ARC_FLAG_CACHED; |
34dc7c2f BB |
5701 | |
5702 | if (HDR_IO_IN_PROGRESS(hdr)) { | |
a8b2e306 | 5703 | zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head; |
34dc7c2f | 5704 | |
1dc32a67 MA |
5705 | if (*arc_flags & ARC_FLAG_CACHED_ONLY) { |
5706 | mutex_exit(hash_lock); | |
5707 | ARCSTAT_BUMP(arcstat_cached_only_in_progress); | |
5708 | rc = SET_ERROR(ENOENT); | |
5709 | goto out; | |
5710 | } | |
5711 | ||
a8b2e306 | 5712 | ASSERT3P(head_zio, !=, NULL); |
7f60329a MA |
5713 | if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && |
5714 | priority == ZIO_PRIORITY_SYNC_READ) { | |
5715 | /* | |
a8b2e306 TC |
5716 | * This is a sync read that needs to wait for |
5717 | * an in-flight async read. Request that the | |
5718 | * zio have its priority upgraded. | |
7f60329a | 5719 | */ |
a8b2e306 TC |
5720 | zio_change_priority(head_zio, priority); |
5721 | DTRACE_PROBE1(arc__async__upgrade__sync, | |
7f60329a | 5722 | arc_buf_hdr_t *, hdr); |
a8b2e306 | 5723 | ARCSTAT_BUMP(arcstat_async_upgrade_sync); |
7f60329a MA |
5724 | } |
5725 | if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { | |
d3c2ae1c GW |
5726 | arc_hdr_clear_flags(hdr, |
5727 | ARC_FLAG_PREDICTIVE_PREFETCH); | |
7f60329a MA |
5728 | } |
5729 | ||
2a432414 | 5730 | if (*arc_flags & ARC_FLAG_WAIT) { |
b9541d6b | 5731 | cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); |
34dc7c2f BB |
5732 | mutex_exit(hash_lock); |
5733 | goto top; | |
5734 | } | |
2a432414 | 5735 | ASSERT(*arc_flags & ARC_FLAG_NOWAIT); |
34dc7c2f BB |
5736 | |
5737 | if (done) { | |
7f60329a | 5738 | arc_callback_t *acb = NULL; |
34dc7c2f BB |
5739 | |
5740 | acb = kmem_zalloc(sizeof (arc_callback_t), | |
79c76d5b | 5741 | KM_SLEEP); |
34dc7c2f BB |
5742 | acb->acb_done = done; |
5743 | acb->acb_private = private; | |
a7004725 | 5744 | acb->acb_compressed = compressed_read; |
440a3eb9 TC |
5745 | acb->acb_encrypted = encrypted_read; |
5746 | acb->acb_noauth = noauth_read; | |
be9a5c35 | 5747 | acb->acb_zb = *zb; |
34dc7c2f BB |
5748 | if (pio != NULL) |
5749 | acb->acb_zio_dummy = zio_null(pio, | |
d164b209 | 5750 | spa, NULL, NULL, NULL, zio_flags); |
34dc7c2f | 5751 | |
d3c2ae1c | 5752 | ASSERT3P(acb->acb_done, !=, NULL); |
a8b2e306 | 5753 | acb->acb_zio_head = head_zio; |
b9541d6b CW |
5754 | acb->acb_next = hdr->b_l1hdr.b_acb; |
5755 | hdr->b_l1hdr.b_acb = acb; | |
34dc7c2f | 5756 | mutex_exit(hash_lock); |
1421c891 | 5757 | goto out; |
34dc7c2f BB |
5758 | } |
5759 | mutex_exit(hash_lock); | |
1421c891 | 5760 | goto out; |
34dc7c2f BB |
5761 | } |
5762 | ||
b9541d6b CW |
5763 | ASSERT(hdr->b_l1hdr.b_state == arc_mru || |
5764 | hdr->b_l1hdr.b_state == arc_mfu); | |
34dc7c2f BB |
5765 | |
5766 | if (done) { | |
7f60329a MA |
5767 | if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { |
5768 | /* | |
5769 | * This is a demand read which does not have to | |
5770 | * wait for i/o because we did a predictive | |
5771 | * prefetch i/o for it, which has completed. | |
5772 | */ | |
5773 | DTRACE_PROBE1( | |
5774 | arc__demand__hit__predictive__prefetch, | |
5775 | arc_buf_hdr_t *, hdr); | |
5776 | ARCSTAT_BUMP( | |
5777 | arcstat_demand_hit_predictive_prefetch); | |
d3c2ae1c GW |
5778 | arc_hdr_clear_flags(hdr, |
5779 | ARC_FLAG_PREDICTIVE_PREFETCH); | |
7f60329a | 5780 | } |
d4a72f23 TC |
5781 | |
5782 | if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) { | |
5783 | ARCSTAT_BUMP( | |
5784 | arcstat_demand_hit_prescient_prefetch); | |
5785 | arc_hdr_clear_flags(hdr, | |
5786 | ARC_FLAG_PRESCIENT_PREFETCH); | |
5787 | } | |
5788 | ||
0902c457 | 5789 | ASSERT(!embedded_bp || !BP_IS_HOLE(bp)); |
d3c2ae1c | 5790 | |
524b4217 | 5791 | /* Get a buf with the desired data in it. */ |
be9a5c35 TC |
5792 | rc = arc_buf_alloc_impl(hdr, spa, zb, private, |
5793 | encrypted_read, compressed_read, noauth_read, | |
5794 | B_TRUE, &buf); | |
a2c2ed1b TC |
5795 | if (rc == ECKSUM) { |
5796 | /* | |
5797 | * Convert authentication and decryption errors | |
be9a5c35 TC |
5798 | * to EIO (and generate an ereport if needed) |
5799 | * before leaving the ARC. | |
a2c2ed1b TC |
5800 | */ |
5801 | rc = SET_ERROR(EIO); | |
be9a5c35 TC |
5802 | if ((zio_flags & ZIO_FLAG_SPECULATIVE) == 0) { |
5803 | spa_log_error(spa, zb); | |
5804 | zfs_ereport_post( | |
5805 | FM_EREPORT_ZFS_AUTHENTICATION, | |
5806 | spa, NULL, zb, NULL, 0, 0); | |
5807 | } | |
a2c2ed1b | 5808 | } |
d4a72f23 | 5809 | if (rc != 0) { |
2c24b5b1 TC |
5810 | (void) remove_reference(hdr, hash_lock, |
5811 | private); | |
5812 | arc_buf_destroy_impl(buf); | |
d4a72f23 TC |
5813 | buf = NULL; |
5814 | } | |
5815 | ||
a2c2ed1b TC |
5816 | /* assert any errors weren't due to unloaded keys */ |
5817 | ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) || | |
be9a5c35 | 5818 | rc != EACCES); |
2a432414 | 5819 | } else if (*arc_flags & ARC_FLAG_PREFETCH && |
424fd7c3 | 5820 | zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { |
d3c2ae1c | 5821 | arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); |
34dc7c2f BB |
5822 | } |
5823 | DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); | |
5824 | arc_access(hdr, hash_lock); | |
d4a72f23 TC |
5825 | if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH) |
5826 | arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH); | |
2a432414 | 5827 | if (*arc_flags & ARC_FLAG_L2CACHE) |
d3c2ae1c | 5828 | arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); |
34dc7c2f BB |
5829 | mutex_exit(hash_lock); |
5830 | ARCSTAT_BUMP(arcstat_hits); | |
b9541d6b CW |
5831 | ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), |
5832 | demand, prefetch, !HDR_ISTYPE_METADATA(hdr), | |
34dc7c2f BB |
5833 | data, metadata, hits); |
5834 | ||
5835 | if (done) | |
d4a72f23 | 5836 | done(NULL, zb, bp, buf, private); |
34dc7c2f | 5837 | } else { |
d3c2ae1c GW |
5838 | uint64_t lsize = BP_GET_LSIZE(bp); |
5839 | uint64_t psize = BP_GET_PSIZE(bp); | |
9b67f605 | 5840 | arc_callback_t *acb; |
b128c09f | 5841 | vdev_t *vd = NULL; |
a117a6d6 | 5842 | uint64_t addr = 0; |
d164b209 | 5843 | boolean_t devw = B_FALSE; |
d3c2ae1c | 5844 | uint64_t size; |
440a3eb9 | 5845 | abd_t *hdr_abd; |
34dc7c2f | 5846 | |
1dc32a67 MA |
5847 | if (*arc_flags & ARC_FLAG_CACHED_ONLY) { |
5848 | rc = SET_ERROR(ENOENT); | |
5849 | if (hash_lock != NULL) | |
5850 | mutex_exit(hash_lock); | |
5851 | goto out; | |
5852 | } | |
5853 | ||
5f6d0b6f BB |
5854 | /* |
5855 | * Gracefully handle a damaged logical block size as a | |
1cdb86cb | 5856 | * checksum error. |
5f6d0b6f | 5857 | */ |
d3c2ae1c | 5858 | if (lsize > spa_maxblocksize(spa)) { |
1cdb86cb | 5859 | rc = SET_ERROR(ECKSUM); |
1dc32a67 MA |
5860 | if (hash_lock != NULL) |
5861 | mutex_exit(hash_lock); | |
5f6d0b6f BB |
5862 | goto out; |
5863 | } | |
5864 | ||
34dc7c2f | 5865 | if (hdr == NULL) { |
0902c457 TC |
5866 | /* |
5867 | * This block is not in the cache or it has | |
5868 | * embedded data. | |
5869 | */ | |
9b67f605 | 5870 | arc_buf_hdr_t *exists = NULL; |
34dc7c2f | 5871 | arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); |
d3c2ae1c | 5872 | hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, |
b5256303 TC |
5873 | BP_IS_PROTECTED(bp), BP_GET_COMPRESS(bp), type, |
5874 | encrypted_read); | |
d3c2ae1c | 5875 | |
0902c457 | 5876 | if (!embedded_bp) { |
9b67f605 MA |
5877 | hdr->b_dva = *BP_IDENTITY(bp); |
5878 | hdr->b_birth = BP_PHYSICAL_BIRTH(bp); | |
9b67f605 MA |
5879 | exists = buf_hash_insert(hdr, &hash_lock); |
5880 | } | |
5881 | if (exists != NULL) { | |
34dc7c2f BB |
5882 | /* somebody beat us to the hash insert */ |
5883 | mutex_exit(hash_lock); | |
428870ff | 5884 | buf_discard_identity(hdr); |
d3c2ae1c | 5885 | arc_hdr_destroy(hdr); |
34dc7c2f BB |
5886 | goto top; /* restart the IO request */ |
5887 | } | |
34dc7c2f | 5888 | } else { |
b9541d6b | 5889 | /* |
b5256303 TC |
5890 | * This block is in the ghost cache or encrypted data |
5891 | * was requested and we didn't have it. If it was | |
5892 | * L2-only (and thus didn't have an L1 hdr), | |
5893 | * we realloc the header to add an L1 hdr. | |
b9541d6b CW |
5894 | */ |
5895 | if (!HDR_HAS_L1HDR(hdr)) { | |
5896 | hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, | |
5897 | hdr_full_cache); | |
5898 | } | |
5899 | ||
b5256303 TC |
5900 | if (GHOST_STATE(hdr->b_l1hdr.b_state)) { |
5901 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); | |
5902 | ASSERT(!HDR_HAS_RABD(hdr)); | |
5903 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
424fd7c3 TS |
5904 | ASSERT0(zfs_refcount_count( |
5905 | &hdr->b_l1hdr.b_refcnt)); | |
b5256303 TC |
5906 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); |
5907 | ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); | |
5908 | } else if (HDR_IO_IN_PROGRESS(hdr)) { | |
5909 | /* | |
5910 | * If this header already had an IO in progress | |
5911 | * and we are performing another IO to fetch | |
5912 | * encrypted data we must wait until the first | |
5913 | * IO completes so as not to confuse | |
5914 | * arc_read_done(). This should be very rare | |
5915 | * and so the performance impact shouldn't | |
5916 | * matter. | |
5917 | */ | |
5918 | cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); | |
5919 | mutex_exit(hash_lock); | |
5920 | goto top; | |
5921 | } | |
34dc7c2f | 5922 | |
7f60329a | 5923 | /* |
d3c2ae1c | 5924 | * This is a delicate dance that we play here. |
b5256303 TC |
5925 | * This hdr might be in the ghost list so we access |
5926 | * it to move it out of the ghost list before we | |
d3c2ae1c GW |
5927 | * initiate the read. If it's a prefetch then |
5928 | * it won't have a callback so we'll remove the | |
5929 | * reference that arc_buf_alloc_impl() created. We | |
5930 | * do this after we've called arc_access() to | |
5931 | * avoid hitting an assert in remove_reference(). | |
7f60329a | 5932 | */ |
428870ff | 5933 | arc_access(hdr, hash_lock); |
b5256303 | 5934 | arc_hdr_alloc_abd(hdr, encrypted_read); |
d3c2ae1c | 5935 | } |
d3c2ae1c | 5936 | |
b5256303 TC |
5937 | if (encrypted_read) { |
5938 | ASSERT(HDR_HAS_RABD(hdr)); | |
5939 | size = HDR_GET_PSIZE(hdr); | |
5940 | hdr_abd = hdr->b_crypt_hdr.b_rabd; | |
d3c2ae1c | 5941 | zio_flags |= ZIO_FLAG_RAW; |
b5256303 TC |
5942 | } else { |
5943 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); | |
5944 | size = arc_hdr_size(hdr); | |
5945 | hdr_abd = hdr->b_l1hdr.b_pabd; | |
5946 | ||
5947 | if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF) { | |
5948 | zio_flags |= ZIO_FLAG_RAW_COMPRESS; | |
5949 | } | |
5950 | ||
5951 | /* | |
5952 | * For authenticated bp's, we do not ask the ZIO layer | |
5953 | * to authenticate them since this will cause the entire | |
5954 | * IO to fail if the key isn't loaded. Instead, we | |
5955 | * defer authentication until arc_buf_fill(), which will | |
5956 | * verify the data when the key is available. | |
5957 | */ | |
5958 | if (BP_IS_AUTHENTICATED(bp)) | |
5959 | zio_flags |= ZIO_FLAG_RAW_ENCRYPT; | |
34dc7c2f BB |
5960 | } |
5961 | ||
b5256303 | 5962 | if (*arc_flags & ARC_FLAG_PREFETCH && |
424fd7c3 | 5963 | zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) |
d3c2ae1c | 5964 | arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); |
d4a72f23 TC |
5965 | if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH) |
5966 | arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH); | |
d3c2ae1c GW |
5967 | if (*arc_flags & ARC_FLAG_L2CACHE) |
5968 | arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); | |
b5256303 TC |
5969 | if (BP_IS_AUTHENTICATED(bp)) |
5970 | arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH); | |
d3c2ae1c GW |
5971 | if (BP_GET_LEVEL(bp) > 0) |
5972 | arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); | |
7f60329a | 5973 | if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) |
d3c2ae1c | 5974 | arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); |
b9541d6b | 5975 | ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); |
428870ff | 5976 | |
79c76d5b | 5977 | acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); |
34dc7c2f BB |
5978 | acb->acb_done = done; |
5979 | acb->acb_private = private; | |
2aa34383 | 5980 | acb->acb_compressed = compressed_read; |
b5256303 TC |
5981 | acb->acb_encrypted = encrypted_read; |
5982 | acb->acb_noauth = noauth_read; | |
be9a5c35 | 5983 | acb->acb_zb = *zb; |
34dc7c2f | 5984 | |
d3c2ae1c | 5985 | ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); |
b9541d6b | 5986 | hdr->b_l1hdr.b_acb = acb; |
d3c2ae1c | 5987 | arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); |
34dc7c2f | 5988 | |
b9541d6b CW |
5989 | if (HDR_HAS_L2HDR(hdr) && |
5990 | (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { | |
5991 | devw = hdr->b_l2hdr.b_dev->l2ad_writing; | |
5992 | addr = hdr->b_l2hdr.b_daddr; | |
b128c09f | 5993 | /* |
a1d477c2 | 5994 | * Lock out L2ARC device removal. |
b128c09f BB |
5995 | */ |
5996 | if (vdev_is_dead(vd) || | |
5997 | !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) | |
5998 | vd = NULL; | |
5999 | } | |
6000 | ||
a8b2e306 TC |
6001 | /* |
6002 | * We count both async reads and scrub IOs as asynchronous so | |
6003 | * that both can be upgraded in the event of a cache hit while | |
6004 | * the read IO is still in-flight. | |
6005 | */ | |
6006 | if (priority == ZIO_PRIORITY_ASYNC_READ || | |
6007 | priority == ZIO_PRIORITY_SCRUB) | |
d3c2ae1c GW |
6008 | arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); |
6009 | else | |
6010 | arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); | |
6011 | ||
e49f1e20 | 6012 | /* |
0902c457 TC |
6013 | * At this point, we have a level 1 cache miss or a blkptr |
6014 | * with embedded data. Try again in L2ARC if possible. | |
e49f1e20 | 6015 | */ |
d3c2ae1c GW |
6016 | ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); |
6017 | ||
0902c457 TC |
6018 | /* |
6019 | * Skip ARC stat bump for block pointers with embedded | |
6020 | * data. The data are read from the blkptr itself via | |
6021 | * decode_embedded_bp_compressed(). | |
6022 | */ | |
6023 | if (!embedded_bp) { | |
6024 | DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, | |
6025 | blkptr_t *, bp, uint64_t, lsize, | |
6026 | zbookmark_phys_t *, zb); | |
6027 | ARCSTAT_BUMP(arcstat_misses); | |
6028 | ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), | |
6029 | demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, | |
6030 | metadata, misses); | |
6031 | } | |
34dc7c2f | 6032 | |
d164b209 | 6033 | if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { |
34dc7c2f BB |
6034 | /* |
6035 | * Read from the L2ARC if the following are true: | |
b128c09f BB |
6036 | * 1. The L2ARC vdev was previously cached. |
6037 | * 2. This buffer still has L2ARC metadata. | |
6038 | * 3. This buffer isn't currently writing to the L2ARC. | |
6039 | * 4. The L2ARC entry wasn't evicted, which may | |
6040 | * also have invalidated the vdev. | |
d164b209 | 6041 | * 5. This isn't prefetch and l2arc_noprefetch is set. |
34dc7c2f | 6042 | */ |
b9541d6b | 6043 | if (HDR_HAS_L2HDR(hdr) && |
d164b209 BB |
6044 | !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && |
6045 | !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { | |
34dc7c2f | 6046 | l2arc_read_callback_t *cb; |
82710e99 GDN |
6047 | abd_t *abd; |
6048 | uint64_t asize; | |
34dc7c2f BB |
6049 | |
6050 | DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); | |
6051 | ARCSTAT_BUMP(arcstat_l2_hits); | |
b9541d6b | 6052 | atomic_inc_32(&hdr->b_l2hdr.b_hits); |
34dc7c2f | 6053 | |
34dc7c2f | 6054 | cb = kmem_zalloc(sizeof (l2arc_read_callback_t), |
79c76d5b | 6055 | KM_SLEEP); |
d3c2ae1c | 6056 | cb->l2rcb_hdr = hdr; |
34dc7c2f BB |
6057 | cb->l2rcb_bp = *bp; |
6058 | cb->l2rcb_zb = *zb; | |
b128c09f | 6059 | cb->l2rcb_flags = zio_flags; |
34dc7c2f | 6060 | |
82710e99 GDN |
6061 | asize = vdev_psize_to_asize(vd, size); |
6062 | if (asize != size) { | |
6063 | abd = abd_alloc_for_io(asize, | |
6064 | HDR_ISTYPE_METADATA(hdr)); | |
6065 | cb->l2rcb_abd = abd; | |
6066 | } else { | |
b5256303 | 6067 | abd = hdr_abd; |
82710e99 GDN |
6068 | } |
6069 | ||
a117a6d6 | 6070 | ASSERT(addr >= VDEV_LABEL_START_SIZE && |
82710e99 | 6071 | addr + asize <= vd->vdev_psize - |
a117a6d6 GW |
6072 | VDEV_LABEL_END_SIZE); |
6073 | ||
34dc7c2f | 6074 | /* |
b128c09f BB |
6075 | * l2arc read. The SCL_L2ARC lock will be |
6076 | * released by l2arc_read_done(). | |
3a17a7a9 SK |
6077 | * Issue a null zio if the underlying buffer |
6078 | * was squashed to zero size by compression. | |
34dc7c2f | 6079 | */ |
b5256303 | 6080 | ASSERT3U(arc_hdr_get_compress(hdr), !=, |
d3c2ae1c GW |
6081 | ZIO_COMPRESS_EMPTY); |
6082 | rzio = zio_read_phys(pio, vd, addr, | |
82710e99 | 6083 | asize, abd, |
d3c2ae1c GW |
6084 | ZIO_CHECKSUM_OFF, |
6085 | l2arc_read_done, cb, priority, | |
6086 | zio_flags | ZIO_FLAG_DONT_CACHE | | |
6087 | ZIO_FLAG_CANFAIL | | |
6088 | ZIO_FLAG_DONT_PROPAGATE | | |
6089 | ZIO_FLAG_DONT_RETRY, B_FALSE); | |
a8b2e306 TC |
6090 | acb->acb_zio_head = rzio; |
6091 | ||
6092 | if (hash_lock != NULL) | |
6093 | mutex_exit(hash_lock); | |
d3c2ae1c | 6094 | |
34dc7c2f BB |
6095 | DTRACE_PROBE2(l2arc__read, vdev_t *, vd, |
6096 | zio_t *, rzio); | |
b5256303 TC |
6097 | ARCSTAT_INCR(arcstat_l2_read_bytes, |
6098 | HDR_GET_PSIZE(hdr)); | |
34dc7c2f | 6099 | |
2a432414 | 6100 | if (*arc_flags & ARC_FLAG_NOWAIT) { |
b128c09f | 6101 | zio_nowait(rzio); |
1421c891 | 6102 | goto out; |
b128c09f | 6103 | } |
34dc7c2f | 6104 | |
2a432414 | 6105 | ASSERT(*arc_flags & ARC_FLAG_WAIT); |
b128c09f | 6106 | if (zio_wait(rzio) == 0) |
1421c891 | 6107 | goto out; |
b128c09f BB |
6108 | |
6109 | /* l2arc read error; goto zio_read() */ | |
a8b2e306 TC |
6110 | if (hash_lock != NULL) |
6111 | mutex_enter(hash_lock); | |
34dc7c2f BB |
6112 | } else { |
6113 | DTRACE_PROBE1(l2arc__miss, | |
6114 | arc_buf_hdr_t *, hdr); | |
6115 | ARCSTAT_BUMP(arcstat_l2_misses); | |
6116 | if (HDR_L2_WRITING(hdr)) | |
6117 | ARCSTAT_BUMP(arcstat_l2_rw_clash); | |
b128c09f | 6118 | spa_config_exit(spa, SCL_L2ARC, vd); |
34dc7c2f | 6119 | } |
d164b209 BB |
6120 | } else { |
6121 | if (vd != NULL) | |
6122 | spa_config_exit(spa, SCL_L2ARC, vd); | |
0902c457 TC |
6123 | /* |
6124 | * Skip ARC stat bump for block pointers with | |
6125 | * embedded data. The data are read from the blkptr | |
6126 | * itself via decode_embedded_bp_compressed(). | |
6127 | */ | |
6128 | if (l2arc_ndev != 0 && !embedded_bp) { | |
d164b209 BB |
6129 | DTRACE_PROBE1(l2arc__miss, |
6130 | arc_buf_hdr_t *, hdr); | |
6131 | ARCSTAT_BUMP(arcstat_l2_misses); | |
6132 | } | |
34dc7c2f | 6133 | } |
34dc7c2f | 6134 | |
b5256303 | 6135 | rzio = zio_read(pio, spa, bp, hdr_abd, size, |
d3c2ae1c | 6136 | arc_read_done, hdr, priority, zio_flags, zb); |
a8b2e306 TC |
6137 | acb->acb_zio_head = rzio; |
6138 | ||
6139 | if (hash_lock != NULL) | |
6140 | mutex_exit(hash_lock); | |
34dc7c2f | 6141 | |
2a432414 | 6142 | if (*arc_flags & ARC_FLAG_WAIT) { |
1421c891 PS |
6143 | rc = zio_wait(rzio); |
6144 | goto out; | |
6145 | } | |
34dc7c2f | 6146 | |
2a432414 | 6147 | ASSERT(*arc_flags & ARC_FLAG_NOWAIT); |
34dc7c2f BB |
6148 | zio_nowait(rzio); |
6149 | } | |
1421c891 PS |
6150 | |
6151 | out: | |
157ef7f6 | 6152 | /* embedded bps don't actually go to disk */ |
0902c457 | 6153 | if (!embedded_bp) |
157ef7f6 | 6154 | spa_read_history_add(spa, zb, *arc_flags); |
1e9231ad | 6155 | spl_fstrans_unmark(cookie); |
1421c891 | 6156 | return (rc); |
34dc7c2f BB |
6157 | } |
6158 | ||
ab26409d BB |
6159 | arc_prune_t * |
6160 | arc_add_prune_callback(arc_prune_func_t *func, void *private) | |
6161 | { | |
6162 | arc_prune_t *p; | |
6163 | ||
d1d7e268 | 6164 | p = kmem_alloc(sizeof (*p), KM_SLEEP); |
ab26409d BB |
6165 | p->p_pfunc = func; |
6166 | p->p_private = private; | |
6167 | list_link_init(&p->p_node); | |
424fd7c3 | 6168 | zfs_refcount_create(&p->p_refcnt); |
ab26409d BB |
6169 | |
6170 | mutex_enter(&arc_prune_mtx); | |
c13060e4 | 6171 | zfs_refcount_add(&p->p_refcnt, &arc_prune_list); |
ab26409d BB |
6172 | list_insert_head(&arc_prune_list, p); |
6173 | mutex_exit(&arc_prune_mtx); | |
6174 | ||
6175 | return (p); | |
6176 | } | |
6177 | ||
6178 | void | |
6179 | arc_remove_prune_callback(arc_prune_t *p) | |
6180 | { | |
4442f60d | 6181 | boolean_t wait = B_FALSE; |
ab26409d BB |
6182 | mutex_enter(&arc_prune_mtx); |
6183 | list_remove(&arc_prune_list, p); | |
424fd7c3 | 6184 | if (zfs_refcount_remove(&p->p_refcnt, &arc_prune_list) > 0) |
4442f60d | 6185 | wait = B_TRUE; |
ab26409d | 6186 | mutex_exit(&arc_prune_mtx); |
4442f60d CC |
6187 | |
6188 | /* wait for arc_prune_task to finish */ | |
6189 | if (wait) | |
6190 | taskq_wait_outstanding(arc_prune_taskq, 0); | |
424fd7c3 TS |
6191 | ASSERT0(zfs_refcount_count(&p->p_refcnt)); |
6192 | zfs_refcount_destroy(&p->p_refcnt); | |
4442f60d | 6193 | kmem_free(p, sizeof (*p)); |
ab26409d BB |
6194 | } |
6195 | ||
df4474f9 MA |
6196 | /* |
6197 | * Notify the arc that a block was freed, and thus will never be used again. | |
6198 | */ | |
6199 | void | |
6200 | arc_freed(spa_t *spa, const blkptr_t *bp) | |
6201 | { | |
6202 | arc_buf_hdr_t *hdr; | |
6203 | kmutex_t *hash_lock; | |
6204 | uint64_t guid = spa_load_guid(spa); | |
6205 | ||
9b67f605 MA |
6206 | ASSERT(!BP_IS_EMBEDDED(bp)); |
6207 | ||
6208 | hdr = buf_hash_find(guid, bp, &hash_lock); | |
df4474f9 MA |
6209 | if (hdr == NULL) |
6210 | return; | |
df4474f9 | 6211 | |
d3c2ae1c GW |
6212 | /* |
6213 | * We might be trying to free a block that is still doing I/O | |
6214 | * (i.e. prefetch) or has a reference (i.e. a dedup-ed, | |
6215 | * dmu_sync-ed block). If this block is being prefetched, then it | |
6216 | * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr | |
6217 | * until the I/O completes. A block may also have a reference if it is | |
6218 | * part of a dedup-ed, dmu_synced write. The dmu_sync() function would | |
6219 | * have written the new block to its final resting place on disk but | |
6220 | * without the dedup flag set. This would have left the hdr in the MRU | |
6221 | * state and discoverable. When the txg finally syncs it detects that | |
6222 | * the block was overridden in open context and issues an override I/O. | |
6223 | * Since this is a dedup block, the override I/O will determine if the | |
6224 | * block is already in the DDT. If so, then it will replace the io_bp | |
6225 | * with the bp from the DDT and allow the I/O to finish. When the I/O | |
6226 | * reaches the done callback, dbuf_write_override_done, it will | |
6227 | * check to see if the io_bp and io_bp_override are identical. | |
6228 | * If they are not, then it indicates that the bp was replaced with | |
6229 | * the bp in the DDT and the override bp is freed. This allows | |
6230 | * us to arrive here with a reference on a block that is being | |
6231 | * freed. So if we have an I/O in progress, or a reference to | |
6232 | * this hdr, then we don't destroy the hdr. | |
6233 | */ | |
6234 | if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && | |
424fd7c3 | 6235 | zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { |
d3c2ae1c GW |
6236 | arc_change_state(arc_anon, hdr, hash_lock); |
6237 | arc_hdr_destroy(hdr); | |
df4474f9 | 6238 | mutex_exit(hash_lock); |
bd089c54 | 6239 | } else { |
d3c2ae1c | 6240 | mutex_exit(hash_lock); |
34dc7c2f | 6241 | } |
34dc7c2f | 6242 | |
34dc7c2f BB |
6243 | } |
6244 | ||
6245 | /* | |
e49f1e20 WA |
6246 | * Release this buffer from the cache, making it an anonymous buffer. This |
6247 | * must be done after a read and prior to modifying the buffer contents. | |
34dc7c2f | 6248 | * If the buffer has more than one reference, we must make |
b128c09f | 6249 | * a new hdr for the buffer. |
34dc7c2f BB |
6250 | */ |
6251 | void | |
6252 | arc_release(arc_buf_t *buf, void *tag) | |
6253 | { | |
b9541d6b | 6254 | arc_buf_hdr_t *hdr = buf->b_hdr; |
34dc7c2f | 6255 | |
428870ff | 6256 | /* |
ca0bf58d | 6257 | * It would be nice to assert that if its DMU metadata (level > |
428870ff BB |
6258 | * 0 || it's the dnode file), then it must be syncing context. |
6259 | * But we don't know that information at this level. | |
6260 | */ | |
6261 | ||
6262 | mutex_enter(&buf->b_evict_lock); | |
b128c09f | 6263 | |
ca0bf58d PS |
6264 | ASSERT(HDR_HAS_L1HDR(hdr)); |
6265 | ||
b9541d6b CW |
6266 | /* |
6267 | * We don't grab the hash lock prior to this check, because if | |
6268 | * the buffer's header is in the arc_anon state, it won't be | |
6269 | * linked into the hash table. | |
6270 | */ | |
6271 | if (hdr->b_l1hdr.b_state == arc_anon) { | |
6272 | mutex_exit(&buf->b_evict_lock); | |
6273 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
6274 | ASSERT(!HDR_IN_HASH_TABLE(hdr)); | |
6275 | ASSERT(!HDR_HAS_L2HDR(hdr)); | |
d3c2ae1c | 6276 | ASSERT(HDR_EMPTY(hdr)); |
34dc7c2f | 6277 | |
d3c2ae1c | 6278 | ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); |
424fd7c3 | 6279 | ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); |
b9541d6b CW |
6280 | ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); |
6281 | ||
b9541d6b | 6282 | hdr->b_l1hdr.b_arc_access = 0; |
d3c2ae1c GW |
6283 | |
6284 | /* | |
6285 | * If the buf is being overridden then it may already | |
6286 | * have a hdr that is not empty. | |
6287 | */ | |
6288 | buf_discard_identity(hdr); | |
b9541d6b CW |
6289 | arc_buf_thaw(buf); |
6290 | ||
6291 | return; | |
34dc7c2f BB |
6292 | } |
6293 | ||
1c27024e | 6294 | kmutex_t *hash_lock = HDR_LOCK(hdr); |
b9541d6b CW |
6295 | mutex_enter(hash_lock); |
6296 | ||
6297 | /* | |
6298 | * This assignment is only valid as long as the hash_lock is | |
6299 | * held, we must be careful not to reference state or the | |
6300 | * b_state field after dropping the lock. | |
6301 | */ | |
1c27024e | 6302 | arc_state_t *state = hdr->b_l1hdr.b_state; |
b9541d6b CW |
6303 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); |
6304 | ASSERT3P(state, !=, arc_anon); | |
6305 | ||
6306 | /* this buffer is not on any list */ | |
424fd7c3 | 6307 | ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); |
b9541d6b CW |
6308 | |
6309 | if (HDR_HAS_L2HDR(hdr)) { | |
b9541d6b | 6310 | mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); |
ca0bf58d PS |
6311 | |
6312 | /* | |
d962d5da PS |
6313 | * We have to recheck this conditional again now that |
6314 | * we're holding the l2ad_mtx to prevent a race with | |
6315 | * another thread which might be concurrently calling | |
6316 | * l2arc_evict(). In that case, l2arc_evict() might have | |
6317 | * destroyed the header's L2 portion as we were waiting | |
6318 | * to acquire the l2ad_mtx. | |
ca0bf58d | 6319 | */ |
d962d5da PS |
6320 | if (HDR_HAS_L2HDR(hdr)) |
6321 | arc_hdr_l2hdr_destroy(hdr); | |
ca0bf58d | 6322 | |
b9541d6b | 6323 | mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); |
b128c09f BB |
6324 | } |
6325 | ||
34dc7c2f BB |
6326 | /* |
6327 | * Do we have more than one buf? | |
6328 | */ | |
d3c2ae1c | 6329 | if (hdr->b_l1hdr.b_bufcnt > 1) { |
34dc7c2f | 6330 | arc_buf_hdr_t *nhdr; |
d164b209 | 6331 | uint64_t spa = hdr->b_spa; |
d3c2ae1c GW |
6332 | uint64_t psize = HDR_GET_PSIZE(hdr); |
6333 | uint64_t lsize = HDR_GET_LSIZE(hdr); | |
b5256303 TC |
6334 | boolean_t protected = HDR_PROTECTED(hdr); |
6335 | enum zio_compress compress = arc_hdr_get_compress(hdr); | |
b9541d6b | 6336 | arc_buf_contents_t type = arc_buf_type(hdr); |
d3c2ae1c | 6337 | VERIFY3U(hdr->b_type, ==, type); |
34dc7c2f | 6338 | |
b9541d6b | 6339 | ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); |
d3c2ae1c GW |
6340 | (void) remove_reference(hdr, hash_lock, tag); |
6341 | ||
524b4217 | 6342 | if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { |
d3c2ae1c | 6343 | ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); |
524b4217 DK |
6344 | ASSERT(ARC_BUF_LAST(buf)); |
6345 | } | |
d3c2ae1c | 6346 | |
34dc7c2f | 6347 | /* |
428870ff | 6348 | * Pull the data off of this hdr and attach it to |
d3c2ae1c GW |
6349 | * a new anonymous hdr. Also find the last buffer |
6350 | * in the hdr's buffer list. | |
34dc7c2f | 6351 | */ |
a7004725 | 6352 | arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); |
d3c2ae1c | 6353 | ASSERT3P(lastbuf, !=, NULL); |
34dc7c2f | 6354 | |
d3c2ae1c GW |
6355 | /* |
6356 | * If the current arc_buf_t and the hdr are sharing their data | |
524b4217 | 6357 | * buffer, then we must stop sharing that block. |
d3c2ae1c GW |
6358 | */ |
6359 | if (arc_buf_is_shared(buf)) { | |
6360 | ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); | |
d3c2ae1c GW |
6361 | VERIFY(!arc_buf_is_shared(lastbuf)); |
6362 | ||
6363 | /* | |
6364 | * First, sever the block sharing relationship between | |
a7004725 | 6365 | * buf and the arc_buf_hdr_t. |
d3c2ae1c GW |
6366 | */ |
6367 | arc_unshare_buf(hdr, buf); | |
2aa34383 DK |
6368 | |
6369 | /* | |
a6255b7f | 6370 | * Now we need to recreate the hdr's b_pabd. Since we |
524b4217 | 6371 | * have lastbuf handy, we try to share with it, but if |
a6255b7f | 6372 | * we can't then we allocate a new b_pabd and copy the |
524b4217 | 6373 | * data from buf into it. |
2aa34383 | 6374 | */ |
524b4217 DK |
6375 | if (arc_can_share(hdr, lastbuf)) { |
6376 | arc_share_buf(hdr, lastbuf); | |
6377 | } else { | |
b5256303 | 6378 | arc_hdr_alloc_abd(hdr, B_FALSE); |
a6255b7f DQ |
6379 | abd_copy_from_buf(hdr->b_l1hdr.b_pabd, |
6380 | buf->b_data, psize); | |
2aa34383 | 6381 | } |
d3c2ae1c GW |
6382 | VERIFY3P(lastbuf->b_data, !=, NULL); |
6383 | } else if (HDR_SHARED_DATA(hdr)) { | |
2aa34383 DK |
6384 | /* |
6385 | * Uncompressed shared buffers are always at the end | |
6386 | * of the list. Compressed buffers don't have the | |
6387 | * same requirements. This makes it hard to | |
6388 | * simply assert that the lastbuf is shared so | |
6389 | * we rely on the hdr's compression flags to determine | |
6390 | * if we have a compressed, shared buffer. | |
6391 | */ | |
6392 | ASSERT(arc_buf_is_shared(lastbuf) || | |
b5256303 | 6393 | arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF); |
2aa34383 | 6394 | ASSERT(!ARC_BUF_SHARED(buf)); |
d3c2ae1c | 6395 | } |
b5256303 TC |
6396 | |
6397 | ASSERT(hdr->b_l1hdr.b_pabd != NULL || HDR_HAS_RABD(hdr)); | |
b9541d6b | 6398 | ASSERT3P(state, !=, arc_l2c_only); |
36da08ef | 6399 | |
424fd7c3 | 6400 | (void) zfs_refcount_remove_many(&state->arcs_size, |
2aa34383 | 6401 | arc_buf_size(buf), buf); |
36da08ef | 6402 | |
424fd7c3 | 6403 | if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { |
b9541d6b | 6404 | ASSERT3P(state, !=, arc_l2c_only); |
424fd7c3 TS |
6405 | (void) zfs_refcount_remove_many( |
6406 | &state->arcs_esize[type], | |
2aa34383 | 6407 | arc_buf_size(buf), buf); |
34dc7c2f | 6408 | } |
1eb5bfa3 | 6409 | |
d3c2ae1c | 6410 | hdr->b_l1hdr.b_bufcnt -= 1; |
b5256303 TC |
6411 | if (ARC_BUF_ENCRYPTED(buf)) |
6412 | hdr->b_crypt_hdr.b_ebufcnt -= 1; | |
6413 | ||
34dc7c2f | 6414 | arc_cksum_verify(buf); |
498877ba | 6415 | arc_buf_unwatch(buf); |
34dc7c2f | 6416 | |
f486f584 TC |
6417 | /* if this is the last uncompressed buf free the checksum */ |
6418 | if (!arc_hdr_has_uncompressed_buf(hdr)) | |
6419 | arc_cksum_free(hdr); | |
6420 | ||
34dc7c2f BB |
6421 | mutex_exit(hash_lock); |
6422 | ||
d3c2ae1c | 6423 | /* |
a6255b7f | 6424 | * Allocate a new hdr. The new hdr will contain a b_pabd |
d3c2ae1c GW |
6425 | * buffer which will be freed in arc_write(). |
6426 | */ | |
b5256303 TC |
6427 | nhdr = arc_hdr_alloc(spa, psize, lsize, protected, |
6428 | compress, type, HDR_HAS_RABD(hdr)); | |
d3c2ae1c GW |
6429 | ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); |
6430 | ASSERT0(nhdr->b_l1hdr.b_bufcnt); | |
424fd7c3 | 6431 | ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt)); |
d3c2ae1c GW |
6432 | VERIFY3U(nhdr->b_type, ==, type); |
6433 | ASSERT(!HDR_SHARED_DATA(nhdr)); | |
b9541d6b | 6434 | |
d3c2ae1c GW |
6435 | nhdr->b_l1hdr.b_buf = buf; |
6436 | nhdr->b_l1hdr.b_bufcnt = 1; | |
b5256303 TC |
6437 | if (ARC_BUF_ENCRYPTED(buf)) |
6438 | nhdr->b_crypt_hdr.b_ebufcnt = 1; | |
b9541d6b CW |
6439 | nhdr->b_l1hdr.b_mru_hits = 0; |
6440 | nhdr->b_l1hdr.b_mru_ghost_hits = 0; | |
6441 | nhdr->b_l1hdr.b_mfu_hits = 0; | |
6442 | nhdr->b_l1hdr.b_mfu_ghost_hits = 0; | |
6443 | nhdr->b_l1hdr.b_l2_hits = 0; | |
c13060e4 | 6444 | (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); |
34dc7c2f | 6445 | buf->b_hdr = nhdr; |
d3c2ae1c | 6446 | |
428870ff | 6447 | mutex_exit(&buf->b_evict_lock); |
424fd7c3 | 6448 | (void) zfs_refcount_add_many(&arc_anon->arcs_size, |
5e8ff256 | 6449 | arc_buf_size(buf), buf); |
34dc7c2f | 6450 | } else { |
428870ff | 6451 | mutex_exit(&buf->b_evict_lock); |
424fd7c3 | 6452 | ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); |
ca0bf58d PS |
6453 | /* protected by hash lock, or hdr is on arc_anon */ |
6454 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); | |
34dc7c2f | 6455 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); |
b9541d6b CW |
6456 | hdr->b_l1hdr.b_mru_hits = 0; |
6457 | hdr->b_l1hdr.b_mru_ghost_hits = 0; | |
6458 | hdr->b_l1hdr.b_mfu_hits = 0; | |
6459 | hdr->b_l1hdr.b_mfu_ghost_hits = 0; | |
6460 | hdr->b_l1hdr.b_l2_hits = 0; | |
6461 | arc_change_state(arc_anon, hdr, hash_lock); | |
6462 | hdr->b_l1hdr.b_arc_access = 0; | |
34dc7c2f | 6463 | |
b5256303 | 6464 | mutex_exit(hash_lock); |
428870ff | 6465 | buf_discard_identity(hdr); |
34dc7c2f BB |
6466 | arc_buf_thaw(buf); |
6467 | } | |
34dc7c2f BB |
6468 | } |
6469 | ||
6470 | int | |
6471 | arc_released(arc_buf_t *buf) | |
6472 | { | |
b128c09f BB |
6473 | int released; |
6474 | ||
428870ff | 6475 | mutex_enter(&buf->b_evict_lock); |
b9541d6b CW |
6476 | released = (buf->b_data != NULL && |
6477 | buf->b_hdr->b_l1hdr.b_state == arc_anon); | |
428870ff | 6478 | mutex_exit(&buf->b_evict_lock); |
b128c09f | 6479 | return (released); |
34dc7c2f BB |
6480 | } |
6481 | ||
34dc7c2f BB |
6482 | #ifdef ZFS_DEBUG |
6483 | int | |
6484 | arc_referenced(arc_buf_t *buf) | |
6485 | { | |
b128c09f BB |
6486 | int referenced; |
6487 | ||
428870ff | 6488 | mutex_enter(&buf->b_evict_lock); |
424fd7c3 | 6489 | referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); |
428870ff | 6490 | mutex_exit(&buf->b_evict_lock); |
b128c09f | 6491 | return (referenced); |
34dc7c2f BB |
6492 | } |
6493 | #endif | |
6494 | ||
6495 | static void | |
6496 | arc_write_ready(zio_t *zio) | |
6497 | { | |
6498 | arc_write_callback_t *callback = zio->io_private; | |
6499 | arc_buf_t *buf = callback->awcb_buf; | |
6500 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
b5256303 TC |
6501 | blkptr_t *bp = zio->io_bp; |
6502 | uint64_t psize = BP_IS_HOLE(bp) ? 0 : BP_GET_PSIZE(bp); | |
a6255b7f | 6503 | fstrans_cookie_t cookie = spl_fstrans_mark(); |
34dc7c2f | 6504 | |
b9541d6b | 6505 | ASSERT(HDR_HAS_L1HDR(hdr)); |
424fd7c3 | 6506 | ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); |
d3c2ae1c | 6507 | ASSERT(hdr->b_l1hdr.b_bufcnt > 0); |
b128c09f | 6508 | |
34dc7c2f | 6509 | /* |
d3c2ae1c GW |
6510 | * If we're reexecuting this zio because the pool suspended, then |
6511 | * cleanup any state that was previously set the first time the | |
2aa34383 | 6512 | * callback was invoked. |
34dc7c2f | 6513 | */ |
d3c2ae1c GW |
6514 | if (zio->io_flags & ZIO_FLAG_REEXECUTED) { |
6515 | arc_cksum_free(hdr); | |
6516 | arc_buf_unwatch(buf); | |
a6255b7f | 6517 | if (hdr->b_l1hdr.b_pabd != NULL) { |
d3c2ae1c | 6518 | if (arc_buf_is_shared(buf)) { |
d3c2ae1c GW |
6519 | arc_unshare_buf(hdr, buf); |
6520 | } else { | |
b5256303 | 6521 | arc_hdr_free_abd(hdr, B_FALSE); |
d3c2ae1c | 6522 | } |
34dc7c2f | 6523 | } |
b5256303 TC |
6524 | |
6525 | if (HDR_HAS_RABD(hdr)) | |
6526 | arc_hdr_free_abd(hdr, B_TRUE); | |
34dc7c2f | 6527 | } |
a6255b7f | 6528 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
b5256303 | 6529 | ASSERT(!HDR_HAS_RABD(hdr)); |
d3c2ae1c GW |
6530 | ASSERT(!HDR_SHARED_DATA(hdr)); |
6531 | ASSERT(!arc_buf_is_shared(buf)); | |
6532 | ||
6533 | callback->awcb_ready(zio, buf, callback->awcb_private); | |
6534 | ||
6535 | if (HDR_IO_IN_PROGRESS(hdr)) | |
6536 | ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); | |
6537 | ||
d3c2ae1c GW |
6538 | arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); |
6539 | ||
b5256303 TC |
6540 | if (BP_IS_PROTECTED(bp) != !!HDR_PROTECTED(hdr)) |
6541 | hdr = arc_hdr_realloc_crypt(hdr, BP_IS_PROTECTED(bp)); | |
6542 | ||
6543 | if (BP_IS_PROTECTED(bp)) { | |
6544 | /* ZIL blocks are written through zio_rewrite */ | |
6545 | ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG); | |
6546 | ASSERT(HDR_PROTECTED(hdr)); | |
6547 | ||
ae76f45c TC |
6548 | if (BP_SHOULD_BYTESWAP(bp)) { |
6549 | if (BP_GET_LEVEL(bp) > 0) { | |
6550 | hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; | |
6551 | } else { | |
6552 | hdr->b_l1hdr.b_byteswap = | |
6553 | DMU_OT_BYTESWAP(BP_GET_TYPE(bp)); | |
6554 | } | |
6555 | } else { | |
6556 | hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; | |
6557 | } | |
6558 | ||
b5256303 TC |
6559 | hdr->b_crypt_hdr.b_ot = BP_GET_TYPE(bp); |
6560 | hdr->b_crypt_hdr.b_dsobj = zio->io_bookmark.zb_objset; | |
6561 | zio_crypt_decode_params_bp(bp, hdr->b_crypt_hdr.b_salt, | |
6562 | hdr->b_crypt_hdr.b_iv); | |
6563 | zio_crypt_decode_mac_bp(bp, hdr->b_crypt_hdr.b_mac); | |
6564 | } | |
6565 | ||
6566 | /* | |
6567 | * If this block was written for raw encryption but the zio layer | |
6568 | * ended up only authenticating it, adjust the buffer flags now. | |
6569 | */ | |
6570 | if (BP_IS_AUTHENTICATED(bp) && ARC_BUF_ENCRYPTED(buf)) { | |
6571 | arc_hdr_set_flags(hdr, ARC_FLAG_NOAUTH); | |
6572 | buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED; | |
6573 | if (BP_GET_COMPRESS(bp) == ZIO_COMPRESS_OFF) | |
6574 | buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; | |
b1d21733 TC |
6575 | } else if (BP_IS_HOLE(bp) && ARC_BUF_ENCRYPTED(buf)) { |
6576 | buf->b_flags &= ~ARC_BUF_FLAG_ENCRYPTED; | |
6577 | buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; | |
b5256303 TC |
6578 | } |
6579 | ||
6580 | /* this must be done after the buffer flags are adjusted */ | |
6581 | arc_cksum_compute(buf); | |
6582 | ||
1c27024e | 6583 | enum zio_compress compress; |
b5256303 | 6584 | if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { |
d3c2ae1c GW |
6585 | compress = ZIO_COMPRESS_OFF; |
6586 | } else { | |
b5256303 TC |
6587 | ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); |
6588 | compress = BP_GET_COMPRESS(bp); | |
d3c2ae1c GW |
6589 | } |
6590 | HDR_SET_PSIZE(hdr, psize); | |
6591 | arc_hdr_set_compress(hdr, compress); | |
6592 | ||
4807c0ba TC |
6593 | if (zio->io_error != 0 || psize == 0) |
6594 | goto out; | |
6595 | ||
d3c2ae1c | 6596 | /* |
b5256303 TC |
6597 | * Fill the hdr with data. If the buffer is encrypted we have no choice |
6598 | * but to copy the data into b_radb. If the hdr is compressed, the data | |
6599 | * we want is available from the zio, otherwise we can take it from | |
6600 | * the buf. | |
a6255b7f DQ |
6601 | * |
6602 | * We might be able to share the buf's data with the hdr here. However, | |
6603 | * doing so would cause the ARC to be full of linear ABDs if we write a | |
6604 | * lot of shareable data. As a compromise, we check whether scattered | |
6605 | * ABDs are allowed, and assume that if they are then the user wants | |
6606 | * the ARC to be primarily filled with them regardless of the data being | |
6607 | * written. Therefore, if they're allowed then we allocate one and copy | |
6608 | * the data into it; otherwise, we share the data directly if we can. | |
d3c2ae1c | 6609 | */ |
b5256303 | 6610 | if (ARC_BUF_ENCRYPTED(buf)) { |
4807c0ba | 6611 | ASSERT3U(psize, >, 0); |
b5256303 TC |
6612 | ASSERT(ARC_BUF_COMPRESSED(buf)); |
6613 | arc_hdr_alloc_abd(hdr, B_TRUE); | |
6614 | abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize); | |
6615 | } else if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) { | |
a6255b7f DQ |
6616 | /* |
6617 | * Ideally, we would always copy the io_abd into b_pabd, but the | |
6618 | * user may have disabled compressed ARC, thus we must check the | |
6619 | * hdr's compression setting rather than the io_bp's. | |
6620 | */ | |
b5256303 | 6621 | if (BP_IS_ENCRYPTED(bp)) { |
a6255b7f | 6622 | ASSERT3U(psize, >, 0); |
b5256303 TC |
6623 | arc_hdr_alloc_abd(hdr, B_TRUE); |
6624 | abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize); | |
6625 | } else if (arc_hdr_get_compress(hdr) != ZIO_COMPRESS_OFF && | |
6626 | !ARC_BUF_COMPRESSED(buf)) { | |
6627 | ASSERT3U(psize, >, 0); | |
6628 | arc_hdr_alloc_abd(hdr, B_FALSE); | |
a6255b7f DQ |
6629 | abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize); |
6630 | } else { | |
6631 | ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr)); | |
b5256303 | 6632 | arc_hdr_alloc_abd(hdr, B_FALSE); |
a6255b7f DQ |
6633 | abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, |
6634 | arc_buf_size(buf)); | |
6635 | } | |
d3c2ae1c | 6636 | } else { |
a6255b7f | 6637 | ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd)); |
2aa34383 | 6638 | ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); |
d3c2ae1c | 6639 | ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); |
d3c2ae1c | 6640 | |
d3c2ae1c | 6641 | arc_share_buf(hdr, buf); |
d3c2ae1c | 6642 | } |
a6255b7f | 6643 | |
4807c0ba | 6644 | out: |
b5256303 | 6645 | arc_hdr_verify(hdr, bp); |
a6255b7f | 6646 | spl_fstrans_unmark(cookie); |
34dc7c2f BB |
6647 | } |
6648 | ||
bc77ba73 PD |
6649 | static void |
6650 | arc_write_children_ready(zio_t *zio) | |
6651 | { | |
6652 | arc_write_callback_t *callback = zio->io_private; | |
6653 | arc_buf_t *buf = callback->awcb_buf; | |
6654 | ||
6655 | callback->awcb_children_ready(zio, buf, callback->awcb_private); | |
6656 | } | |
6657 | ||
e8b96c60 MA |
6658 | /* |
6659 | * The SPA calls this callback for each physical write that happens on behalf | |
6660 | * of a logical write. See the comment in dbuf_write_physdone() for details. | |
6661 | */ | |
6662 | static void | |
6663 | arc_write_physdone(zio_t *zio) | |
6664 | { | |
6665 | arc_write_callback_t *cb = zio->io_private; | |
6666 | if (cb->awcb_physdone != NULL) | |
6667 | cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); | |
6668 | } | |
6669 | ||
34dc7c2f BB |
6670 | static void |
6671 | arc_write_done(zio_t *zio) | |
6672 | { | |
6673 | arc_write_callback_t *callback = zio->io_private; | |
6674 | arc_buf_t *buf = callback->awcb_buf; | |
6675 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
6676 | ||
d3c2ae1c | 6677 | ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); |
428870ff BB |
6678 | |
6679 | if (zio->io_error == 0) { | |
d3c2ae1c GW |
6680 | arc_hdr_verify(hdr, zio->io_bp); |
6681 | ||
9b67f605 | 6682 | if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { |
b0bc7a84 MG |
6683 | buf_discard_identity(hdr); |
6684 | } else { | |
6685 | hdr->b_dva = *BP_IDENTITY(zio->io_bp); | |
6686 | hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); | |
b0bc7a84 | 6687 | } |
428870ff | 6688 | } else { |
d3c2ae1c | 6689 | ASSERT(HDR_EMPTY(hdr)); |
428870ff | 6690 | } |
34dc7c2f | 6691 | |
34dc7c2f | 6692 | /* |
9b67f605 MA |
6693 | * If the block to be written was all-zero or compressed enough to be |
6694 | * embedded in the BP, no write was performed so there will be no | |
6695 | * dva/birth/checksum. The buffer must therefore remain anonymous | |
6696 | * (and uncached). | |
34dc7c2f | 6697 | */ |
d3c2ae1c | 6698 | if (!HDR_EMPTY(hdr)) { |
34dc7c2f BB |
6699 | arc_buf_hdr_t *exists; |
6700 | kmutex_t *hash_lock; | |
6701 | ||
524b4217 | 6702 | ASSERT3U(zio->io_error, ==, 0); |
428870ff | 6703 | |
34dc7c2f BB |
6704 | arc_cksum_verify(buf); |
6705 | ||
6706 | exists = buf_hash_insert(hdr, &hash_lock); | |
b9541d6b | 6707 | if (exists != NULL) { |
34dc7c2f BB |
6708 | /* |
6709 | * This can only happen if we overwrite for | |
6710 | * sync-to-convergence, because we remove | |
6711 | * buffers from the hash table when we arc_free(). | |
6712 | */ | |
428870ff BB |
6713 | if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { |
6714 | if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) | |
6715 | panic("bad overwrite, hdr=%p exists=%p", | |
6716 | (void *)hdr, (void *)exists); | |
424fd7c3 | 6717 | ASSERT(zfs_refcount_is_zero( |
b9541d6b | 6718 | &exists->b_l1hdr.b_refcnt)); |
428870ff | 6719 | arc_change_state(arc_anon, exists, hash_lock); |
428870ff | 6720 | arc_hdr_destroy(exists); |
ca6c7a94 | 6721 | mutex_exit(hash_lock); |
428870ff BB |
6722 | exists = buf_hash_insert(hdr, &hash_lock); |
6723 | ASSERT3P(exists, ==, NULL); | |
03c6040b GW |
6724 | } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { |
6725 | /* nopwrite */ | |
6726 | ASSERT(zio->io_prop.zp_nopwrite); | |
6727 | if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) | |
6728 | panic("bad nopwrite, hdr=%p exists=%p", | |
6729 | (void *)hdr, (void *)exists); | |
428870ff BB |
6730 | } else { |
6731 | /* Dedup */ | |
d3c2ae1c | 6732 | ASSERT(hdr->b_l1hdr.b_bufcnt == 1); |
b9541d6b | 6733 | ASSERT(hdr->b_l1hdr.b_state == arc_anon); |
428870ff BB |
6734 | ASSERT(BP_GET_DEDUP(zio->io_bp)); |
6735 | ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); | |
6736 | } | |
34dc7c2f | 6737 | } |
d3c2ae1c | 6738 | arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); |
b128c09f | 6739 | /* if it's not anon, we are doing a scrub */ |
b9541d6b | 6740 | if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) |
b128c09f | 6741 | arc_access(hdr, hash_lock); |
34dc7c2f | 6742 | mutex_exit(hash_lock); |
34dc7c2f | 6743 | } else { |
d3c2ae1c | 6744 | arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); |
34dc7c2f BB |
6745 | } |
6746 | ||
424fd7c3 | 6747 | ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
428870ff | 6748 | callback->awcb_done(zio, buf, callback->awcb_private); |
34dc7c2f | 6749 | |
a6255b7f | 6750 | abd_put(zio->io_abd); |
34dc7c2f BB |
6751 | kmem_free(callback, sizeof (arc_write_callback_t)); |
6752 | } | |
6753 | ||
6754 | zio_t * | |
428870ff | 6755 | arc_write(zio_t *pio, spa_t *spa, uint64_t txg, |
d3c2ae1c | 6756 | blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, |
b5256303 TC |
6757 | const zio_prop_t *zp, arc_write_done_func_t *ready, |
6758 | arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone, | |
6759 | arc_write_done_func_t *done, void *private, zio_priority_t priority, | |
5dbd68a3 | 6760 | int zio_flags, const zbookmark_phys_t *zb) |
34dc7c2f BB |
6761 | { |
6762 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
6763 | arc_write_callback_t *callback; | |
b128c09f | 6764 | zio_t *zio; |
82644107 | 6765 | zio_prop_t localprop = *zp; |
34dc7c2f | 6766 | |
d3c2ae1c GW |
6767 | ASSERT3P(ready, !=, NULL); |
6768 | ASSERT3P(done, !=, NULL); | |
34dc7c2f | 6769 | ASSERT(!HDR_IO_ERROR(hdr)); |
b9541d6b | 6770 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); |
d3c2ae1c GW |
6771 | ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); |
6772 | ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); | |
b128c09f | 6773 | if (l2arc) |
d3c2ae1c | 6774 | arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); |
82644107 | 6775 | |
b5256303 TC |
6776 | if (ARC_BUF_ENCRYPTED(buf)) { |
6777 | ASSERT(ARC_BUF_COMPRESSED(buf)); | |
6778 | localprop.zp_encrypt = B_TRUE; | |
6779 | localprop.zp_compress = HDR_GET_COMPRESS(hdr); | |
6780 | localprop.zp_byteorder = | |
6781 | (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS) ? | |
6782 | ZFS_HOST_BYTEORDER : !ZFS_HOST_BYTEORDER; | |
6783 | bcopy(hdr->b_crypt_hdr.b_salt, localprop.zp_salt, | |
6784 | ZIO_DATA_SALT_LEN); | |
6785 | bcopy(hdr->b_crypt_hdr.b_iv, localprop.zp_iv, | |
6786 | ZIO_DATA_IV_LEN); | |
6787 | bcopy(hdr->b_crypt_hdr.b_mac, localprop.zp_mac, | |
6788 | ZIO_DATA_MAC_LEN); | |
6789 | if (DMU_OT_IS_ENCRYPTED(localprop.zp_type)) { | |
6790 | localprop.zp_nopwrite = B_FALSE; | |
6791 | localprop.zp_copies = | |
6792 | MIN(localprop.zp_copies, SPA_DVAS_PER_BP - 1); | |
6793 | } | |
2aa34383 | 6794 | zio_flags |= ZIO_FLAG_RAW; |
b5256303 TC |
6795 | } else if (ARC_BUF_COMPRESSED(buf)) { |
6796 | ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); | |
6797 | localprop.zp_compress = HDR_GET_COMPRESS(hdr); | |
6798 | zio_flags |= ZIO_FLAG_RAW_COMPRESS; | |
2aa34383 | 6799 | } |
79c76d5b | 6800 | callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); |
34dc7c2f | 6801 | callback->awcb_ready = ready; |
bc77ba73 | 6802 | callback->awcb_children_ready = children_ready; |
e8b96c60 | 6803 | callback->awcb_physdone = physdone; |
34dc7c2f BB |
6804 | callback->awcb_done = done; |
6805 | callback->awcb_private = private; | |
6806 | callback->awcb_buf = buf; | |
b128c09f | 6807 | |
d3c2ae1c | 6808 | /* |
a6255b7f | 6809 | * The hdr's b_pabd is now stale, free it now. A new data block |
d3c2ae1c GW |
6810 | * will be allocated when the zio pipeline calls arc_write_ready(). |
6811 | */ | |
a6255b7f | 6812 | if (hdr->b_l1hdr.b_pabd != NULL) { |
d3c2ae1c GW |
6813 | /* |
6814 | * If the buf is currently sharing the data block with | |
6815 | * the hdr then we need to break that relationship here. | |
6816 | * The hdr will remain with a NULL data pointer and the | |
6817 | * buf will take sole ownership of the block. | |
6818 | */ | |
6819 | if (arc_buf_is_shared(buf)) { | |
d3c2ae1c GW |
6820 | arc_unshare_buf(hdr, buf); |
6821 | } else { | |
b5256303 | 6822 | arc_hdr_free_abd(hdr, B_FALSE); |
d3c2ae1c GW |
6823 | } |
6824 | VERIFY3P(buf->b_data, !=, NULL); | |
d3c2ae1c | 6825 | } |
b5256303 TC |
6826 | |
6827 | if (HDR_HAS_RABD(hdr)) | |
6828 | arc_hdr_free_abd(hdr, B_TRUE); | |
6829 | ||
71a24c3c TC |
6830 | if (!(zio_flags & ZIO_FLAG_RAW)) |
6831 | arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); | |
b5256303 | 6832 | |
d3c2ae1c | 6833 | ASSERT(!arc_buf_is_shared(buf)); |
a6255b7f | 6834 | ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); |
d3c2ae1c | 6835 | |
a6255b7f DQ |
6836 | zio = zio_write(pio, spa, txg, bp, |
6837 | abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)), | |
82644107 | 6838 | HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready, |
bc77ba73 PD |
6839 | (children_ready != NULL) ? arc_write_children_ready : NULL, |
6840 | arc_write_physdone, arc_write_done, callback, | |
e8b96c60 | 6841 | priority, zio_flags, zb); |
34dc7c2f BB |
6842 | |
6843 | return (zio); | |
6844 | } | |
6845 | ||
34dc7c2f BB |
6846 | void |
6847 | arc_tempreserve_clear(uint64_t reserve) | |
6848 | { | |
6849 | atomic_add_64(&arc_tempreserve, -reserve); | |
6850 | ASSERT((int64_t)arc_tempreserve >= 0); | |
6851 | } | |
6852 | ||
6853 | int | |
dae3e9ea | 6854 | arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg) |
34dc7c2f BB |
6855 | { |
6856 | int error; | |
9babb374 | 6857 | uint64_t anon_size; |
34dc7c2f | 6858 | |
1b8951b3 TC |
6859 | if (!arc_no_grow && |
6860 | reserve > arc_c/4 && | |
6861 | reserve * 4 > (2ULL << SPA_MAXBLOCKSHIFT)) | |
34dc7c2f | 6862 | arc_c = MIN(arc_c_max, reserve * 4); |
12f9a6a3 BB |
6863 | |
6864 | /* | |
6865 | * Throttle when the calculated memory footprint for the TXG | |
6866 | * exceeds the target ARC size. | |
6867 | */ | |
570827e1 BB |
6868 | if (reserve > arc_c) { |
6869 | DMU_TX_STAT_BUMP(dmu_tx_memory_reserve); | |
12f9a6a3 | 6870 | return (SET_ERROR(ERESTART)); |
570827e1 | 6871 | } |
34dc7c2f | 6872 | |
9babb374 BB |
6873 | /* |
6874 | * Don't count loaned bufs as in flight dirty data to prevent long | |
6875 | * network delays from blocking transactions that are ready to be | |
6876 | * assigned to a txg. | |
6877 | */ | |
a7004725 DK |
6878 | |
6879 | /* assert that it has not wrapped around */ | |
6880 | ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); | |
6881 | ||
424fd7c3 | 6882 | anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) - |
36da08ef | 6883 | arc_loaned_bytes), 0); |
9babb374 | 6884 | |
34dc7c2f BB |
6885 | /* |
6886 | * Writes will, almost always, require additional memory allocations | |
d3cc8b15 | 6887 | * in order to compress/encrypt/etc the data. We therefore need to |
34dc7c2f BB |
6888 | * make sure that there is sufficient available memory for this. |
6889 | */ | |
dae3e9ea | 6890 | error = arc_memory_throttle(spa, reserve, txg); |
e8b96c60 | 6891 | if (error != 0) |
34dc7c2f BB |
6892 | return (error); |
6893 | ||
6894 | /* | |
6895 | * Throttle writes when the amount of dirty data in the cache | |
6896 | * gets too large. We try to keep the cache less than half full | |
6897 | * of dirty blocks so that our sync times don't grow too large. | |
dae3e9ea DB |
6898 | * |
6899 | * In the case of one pool being built on another pool, we want | |
6900 | * to make sure we don't end up throttling the lower (backing) | |
6901 | * pool when the upper pool is the majority contributor to dirty | |
6902 | * data. To insure we make forward progress during throttling, we | |
6903 | * also check the current pool's net dirty data and only throttle | |
6904 | * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty | |
6905 | * data in the cache. | |
6906 | * | |
34dc7c2f BB |
6907 | * Note: if two requests come in concurrently, we might let them |
6908 | * both succeed, when one of them should fail. Not a huge deal. | |
6909 | */ | |
dae3e9ea DB |
6910 | uint64_t total_dirty = reserve + arc_tempreserve + anon_size; |
6911 | uint64_t spa_dirty_anon = spa_dirty_data(spa); | |
9babb374 | 6912 | |
dae3e9ea DB |
6913 | if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 && |
6914 | anon_size > arc_c * zfs_arc_anon_limit_percent / 100 && | |
6915 | spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) { | |
2fd92c3d | 6916 | #ifdef ZFS_DEBUG |
424fd7c3 TS |
6917 | uint64_t meta_esize = zfs_refcount_count( |
6918 | &arc_anon->arcs_esize[ARC_BUFC_METADATA]); | |
d3c2ae1c | 6919 | uint64_t data_esize = |
424fd7c3 | 6920 | zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); |
34dc7c2f BB |
6921 | dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " |
6922 | "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", | |
d3c2ae1c GW |
6923 | arc_tempreserve >> 10, meta_esize >> 10, |
6924 | data_esize >> 10, reserve >> 10, arc_c >> 10); | |
2fd92c3d | 6925 | #endif |
570827e1 | 6926 | DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle); |
2e528b49 | 6927 | return (SET_ERROR(ERESTART)); |
34dc7c2f BB |
6928 | } |
6929 | atomic_add_64(&arc_tempreserve, reserve); | |
6930 | return (0); | |
6931 | } | |
6932 | ||
13be560d BB |
6933 | static void |
6934 | arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, | |
6935 | kstat_named_t *evict_data, kstat_named_t *evict_metadata) | |
6936 | { | |
424fd7c3 | 6937 | size->value.ui64 = zfs_refcount_count(&state->arcs_size); |
d3c2ae1c | 6938 | evict_data->value.ui64 = |
424fd7c3 | 6939 | zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); |
d3c2ae1c | 6940 | evict_metadata->value.ui64 = |
424fd7c3 | 6941 | zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); |
13be560d BB |
6942 | } |
6943 | ||
6944 | static int | |
6945 | arc_kstat_update(kstat_t *ksp, int rw) | |
6946 | { | |
6947 | arc_stats_t *as = ksp->ks_data; | |
6948 | ||
6949 | if (rw == KSTAT_WRITE) { | |
ecb2b7dc | 6950 | return (SET_ERROR(EACCES)); |
13be560d BB |
6951 | } else { |
6952 | arc_kstat_update_state(arc_anon, | |
6953 | &as->arcstat_anon_size, | |
500445c0 PS |
6954 | &as->arcstat_anon_evictable_data, |
6955 | &as->arcstat_anon_evictable_metadata); | |
13be560d BB |
6956 | arc_kstat_update_state(arc_mru, |
6957 | &as->arcstat_mru_size, | |
500445c0 PS |
6958 | &as->arcstat_mru_evictable_data, |
6959 | &as->arcstat_mru_evictable_metadata); | |
13be560d BB |
6960 | arc_kstat_update_state(arc_mru_ghost, |
6961 | &as->arcstat_mru_ghost_size, | |
500445c0 PS |
6962 | &as->arcstat_mru_ghost_evictable_data, |
6963 | &as->arcstat_mru_ghost_evictable_metadata); | |
13be560d BB |
6964 | arc_kstat_update_state(arc_mfu, |
6965 | &as->arcstat_mfu_size, | |
500445c0 PS |
6966 | &as->arcstat_mfu_evictable_data, |
6967 | &as->arcstat_mfu_evictable_metadata); | |
fc41c640 | 6968 | arc_kstat_update_state(arc_mfu_ghost, |
13be560d | 6969 | &as->arcstat_mfu_ghost_size, |
500445c0 PS |
6970 | &as->arcstat_mfu_ghost_evictable_data, |
6971 | &as->arcstat_mfu_ghost_evictable_metadata); | |
70f02287 | 6972 | |
37fb3e43 PD |
6973 | ARCSTAT(arcstat_size) = aggsum_value(&arc_size); |
6974 | ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used); | |
6975 | ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size); | |
6976 | ARCSTAT(arcstat_metadata_size) = | |
6977 | aggsum_value(&astat_metadata_size); | |
6978 | ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size); | |
6979 | ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size); | |
6980 | ARCSTAT(arcstat_dbuf_size) = aggsum_value(&astat_dbuf_size); | |
6981 | ARCSTAT(arcstat_dnode_size) = aggsum_value(&astat_dnode_size); | |
6982 | ARCSTAT(arcstat_bonus_size) = aggsum_value(&astat_bonus_size); | |
6983 | ||
70f02287 BB |
6984 | as->arcstat_memory_all_bytes.value.ui64 = |
6985 | arc_all_memory(); | |
6986 | as->arcstat_memory_free_bytes.value.ui64 = | |
6987 | arc_free_memory(); | |
6988 | as->arcstat_memory_available_bytes.value.i64 = | |
6989 | arc_available_memory(); | |
13be560d BB |
6990 | } |
6991 | ||
6992 | return (0); | |
6993 | } | |
6994 | ||
ca0bf58d PS |
6995 | /* |
6996 | * This function *must* return indices evenly distributed between all | |
6997 | * sublists of the multilist. This is needed due to how the ARC eviction | |
6998 | * code is laid out; arc_evict_state() assumes ARC buffers are evenly | |
6999 | * distributed between all sublists and uses this assumption when | |
7000 | * deciding which sublist to evict from and how much to evict from it. | |
7001 | */ | |
7002 | unsigned int | |
7003 | arc_state_multilist_index_func(multilist_t *ml, void *obj) | |
7004 | { | |
7005 | arc_buf_hdr_t *hdr = obj; | |
7006 | ||
7007 | /* | |
7008 | * We rely on b_dva to generate evenly distributed index | |
7009 | * numbers using buf_hash below. So, as an added precaution, | |
7010 | * let's make sure we never add empty buffers to the arc lists. | |
7011 | */ | |
d3c2ae1c | 7012 | ASSERT(!HDR_EMPTY(hdr)); |
ca0bf58d PS |
7013 | |
7014 | /* | |
7015 | * The assumption here, is the hash value for a given | |
7016 | * arc_buf_hdr_t will remain constant throughout its lifetime | |
7017 | * (i.e. its b_spa, b_dva, and b_birth fields don't change). | |
7018 | * Thus, we don't need to store the header's sublist index | |
7019 | * on insertion, as this index can be recalculated on removal. | |
7020 | * | |
7021 | * Also, the low order bits of the hash value are thought to be | |
7022 | * distributed evenly. Otherwise, in the case that the multilist | |
7023 | * has a power of two number of sublists, each sublists' usage | |
7024 | * would not be evenly distributed. | |
7025 | */ | |
7026 | return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % | |
7027 | multilist_get_num_sublists(ml)); | |
7028 | } | |
7029 | ||
36a6e233 RM |
7030 | #define WARN_IF_TUNING_IGNORED(tuning, value, do_warn) do { \ |
7031 | if ((do_warn) && (tuning) && ((tuning) != (value))) { \ | |
7032 | cmn_err(CE_WARN, \ | |
7033 | "ignoring tunable %s (using %llu instead)", \ | |
7034 | (#tuning), (value)); \ | |
7035 | } \ | |
7036 | } while (0) | |
7037 | ||
ca67b33a MA |
7038 | /* |
7039 | * Called during module initialization and periodically thereafter to | |
e3570464 | 7040 | * apply reasonable changes to the exposed performance tunings. Can also be |
7041 | * called explicitly by param_set_arc_*() functions when ARC tunables are | |
7042 | * updated manually. Non-zero zfs_* values which differ from the currently set | |
7043 | * values will be applied. | |
ca67b33a | 7044 | */ |
e3570464 | 7045 | void |
36a6e233 | 7046 | arc_tuning_update(boolean_t verbose) |
ca67b33a | 7047 | { |
b8a97fb1 | 7048 | uint64_t allmem = arc_all_memory(); |
7049 | unsigned long limit; | |
9edb3695 | 7050 | |
36a6e233 RM |
7051 | /* Valid range: 32M - <arc_c_max> */ |
7052 | if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) && | |
7053 | (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) && | |
7054 | (zfs_arc_min <= arc_c_max)) { | |
7055 | arc_c_min = zfs_arc_min; | |
7056 | arc_c = MAX(arc_c, arc_c_min); | |
7057 | } | |
7058 | WARN_IF_TUNING_IGNORED(zfs_arc_min, arc_c_min, verbose); | |
7059 | ||
ca67b33a MA |
7060 | /* Valid range: 64M - <all physical memory> */ |
7061 | if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) && | |
7403d074 | 7062 | (zfs_arc_max >= 64 << 20) && (zfs_arc_max < allmem) && |
ca67b33a MA |
7063 | (zfs_arc_max > arc_c_min)) { |
7064 | arc_c_max = zfs_arc_max; | |
7065 | arc_c = arc_c_max; | |
7066 | arc_p = (arc_c >> 1); | |
b8a97fb1 | 7067 | if (arc_meta_limit > arc_c_max) |
7068 | arc_meta_limit = arc_c_max; | |
03fdcb9a MM |
7069 | if (arc_dnode_size_limit > arc_meta_limit) |
7070 | arc_dnode_size_limit = arc_meta_limit; | |
ca67b33a | 7071 | } |
36a6e233 | 7072 | WARN_IF_TUNING_IGNORED(zfs_arc_max, arc_c_max, verbose); |
ca67b33a MA |
7073 | |
7074 | /* Valid range: 16M - <arc_c_max> */ | |
7075 | if ((zfs_arc_meta_min) && (zfs_arc_meta_min != arc_meta_min) && | |
7076 | (zfs_arc_meta_min >= 1ULL << SPA_MAXBLOCKSHIFT) && | |
7077 | (zfs_arc_meta_min <= arc_c_max)) { | |
7078 | arc_meta_min = zfs_arc_meta_min; | |
b8a97fb1 | 7079 | if (arc_meta_limit < arc_meta_min) |
7080 | arc_meta_limit = arc_meta_min; | |
03fdcb9a MM |
7081 | if (arc_dnode_size_limit < arc_meta_min) |
7082 | arc_dnode_size_limit = arc_meta_min; | |
ca67b33a | 7083 | } |
36a6e233 | 7084 | WARN_IF_TUNING_IGNORED(zfs_arc_meta_min, arc_meta_min, verbose); |
ca67b33a MA |
7085 | |
7086 | /* Valid range: <arc_meta_min> - <arc_c_max> */ | |
b8a97fb1 | 7087 | limit = zfs_arc_meta_limit ? zfs_arc_meta_limit : |
7088 | MIN(zfs_arc_meta_limit_percent, 100) * arc_c_max / 100; | |
7089 | if ((limit != arc_meta_limit) && | |
7090 | (limit >= arc_meta_min) && | |
7091 | (limit <= arc_c_max)) | |
7092 | arc_meta_limit = limit; | |
36a6e233 | 7093 | WARN_IF_TUNING_IGNORED(zfs_arc_meta_limit, arc_meta_limit, verbose); |
b8a97fb1 | 7094 | |
7095 | /* Valid range: <arc_meta_min> - <arc_meta_limit> */ | |
7096 | limit = zfs_arc_dnode_limit ? zfs_arc_dnode_limit : | |
7097 | MIN(zfs_arc_dnode_limit_percent, 100) * arc_meta_limit / 100; | |
03fdcb9a | 7098 | if ((limit != arc_dnode_size_limit) && |
b8a97fb1 | 7099 | (limit >= arc_meta_min) && |
7100 | (limit <= arc_meta_limit)) | |
03fdcb9a | 7101 | arc_dnode_size_limit = limit; |
36a6e233 RM |
7102 | WARN_IF_TUNING_IGNORED(zfs_arc_dnode_limit, arc_dnode_size_limit, |
7103 | verbose); | |
25458cbe | 7104 | |
ca67b33a MA |
7105 | /* Valid range: 1 - N */ |
7106 | if (zfs_arc_grow_retry) | |
7107 | arc_grow_retry = zfs_arc_grow_retry; | |
7108 | ||
7109 | /* Valid range: 1 - N */ | |
7110 | if (zfs_arc_shrink_shift) { | |
7111 | arc_shrink_shift = zfs_arc_shrink_shift; | |
7112 | arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1); | |
7113 | } | |
7114 | ||
728d6ae9 BB |
7115 | /* Valid range: 1 - N */ |
7116 | if (zfs_arc_p_min_shift) | |
7117 | arc_p_min_shift = zfs_arc_p_min_shift; | |
7118 | ||
d4a72f23 TC |
7119 | /* Valid range: 1 - N ms */ |
7120 | if (zfs_arc_min_prefetch_ms) | |
7121 | arc_min_prefetch_ms = zfs_arc_min_prefetch_ms; | |
7122 | ||
7123 | /* Valid range: 1 - N ms */ | |
7124 | if (zfs_arc_min_prescient_prefetch_ms) { | |
7125 | arc_min_prescient_prefetch_ms = | |
7126 | zfs_arc_min_prescient_prefetch_ms; | |
7127 | } | |
11f552fa | 7128 | |
7e8bddd0 BB |
7129 | /* Valid range: 0 - 100 */ |
7130 | if ((zfs_arc_lotsfree_percent >= 0) && | |
7131 | (zfs_arc_lotsfree_percent <= 100)) | |
7132 | arc_lotsfree_percent = zfs_arc_lotsfree_percent; | |
36a6e233 RM |
7133 | WARN_IF_TUNING_IGNORED(zfs_arc_lotsfree_percent, arc_lotsfree_percent, |
7134 | verbose); | |
7e8bddd0 | 7135 | |
11f552fa BB |
7136 | /* Valid range: 0 - <all physical memory> */ |
7137 | if ((zfs_arc_sys_free) && (zfs_arc_sys_free != arc_sys_free)) | |
9edb3695 | 7138 | arc_sys_free = MIN(MAX(zfs_arc_sys_free, 0), allmem); |
36a6e233 | 7139 | WARN_IF_TUNING_IGNORED(zfs_arc_sys_free, arc_sys_free, verbose); |
ca67b33a MA |
7140 | } |
7141 | ||
d3c2ae1c GW |
7142 | static void |
7143 | arc_state_init(void) | |
7144 | { | |
7145 | arc_anon = &ARC_anon; | |
7146 | arc_mru = &ARC_mru; | |
7147 | arc_mru_ghost = &ARC_mru_ghost; | |
7148 | arc_mfu = &ARC_mfu; | |
7149 | arc_mfu_ghost = &ARC_mfu_ghost; | |
7150 | arc_l2c_only = &ARC_l2c_only; | |
7151 | ||
64fc7762 MA |
7152 | arc_mru->arcs_list[ARC_BUFC_METADATA] = |
7153 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7154 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7155 | arc_state_multilist_index_func); |
64fc7762 MA |
7156 | arc_mru->arcs_list[ARC_BUFC_DATA] = |
7157 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7158 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7159 | arc_state_multilist_index_func); |
64fc7762 MA |
7160 | arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] = |
7161 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7162 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7163 | arc_state_multilist_index_func); |
64fc7762 MA |
7164 | arc_mru_ghost->arcs_list[ARC_BUFC_DATA] = |
7165 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7166 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7167 | arc_state_multilist_index_func); |
64fc7762 MA |
7168 | arc_mfu->arcs_list[ARC_BUFC_METADATA] = |
7169 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7170 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7171 | arc_state_multilist_index_func); |
64fc7762 MA |
7172 | arc_mfu->arcs_list[ARC_BUFC_DATA] = |
7173 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7174 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7175 | arc_state_multilist_index_func); |
64fc7762 MA |
7176 | arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] = |
7177 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7178 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7179 | arc_state_multilist_index_func); |
64fc7762 MA |
7180 | arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] = |
7181 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7182 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7183 | arc_state_multilist_index_func); |
64fc7762 MA |
7184 | arc_l2c_only->arcs_list[ARC_BUFC_METADATA] = |
7185 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7186 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7187 | arc_state_multilist_index_func); |
64fc7762 MA |
7188 | arc_l2c_only->arcs_list[ARC_BUFC_DATA] = |
7189 | multilist_create(sizeof (arc_buf_hdr_t), | |
d3c2ae1c | 7190 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
c30e58c4 | 7191 | arc_state_multilist_index_func); |
d3c2ae1c | 7192 | |
424fd7c3 TS |
7193 | zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); |
7194 | zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); | |
7195 | zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); | |
7196 | zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); | |
7197 | zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); | |
7198 | zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); | |
7199 | zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); | |
7200 | zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); | |
7201 | zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); | |
7202 | zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); | |
7203 | zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); | |
7204 | zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); | |
7205 | ||
7206 | zfs_refcount_create(&arc_anon->arcs_size); | |
7207 | zfs_refcount_create(&arc_mru->arcs_size); | |
7208 | zfs_refcount_create(&arc_mru_ghost->arcs_size); | |
7209 | zfs_refcount_create(&arc_mfu->arcs_size); | |
7210 | zfs_refcount_create(&arc_mfu_ghost->arcs_size); | |
7211 | zfs_refcount_create(&arc_l2c_only->arcs_size); | |
d3c2ae1c | 7212 | |
37fb3e43 PD |
7213 | aggsum_init(&arc_meta_used, 0); |
7214 | aggsum_init(&arc_size, 0); | |
7215 | aggsum_init(&astat_data_size, 0); | |
7216 | aggsum_init(&astat_metadata_size, 0); | |
7217 | aggsum_init(&astat_hdr_size, 0); | |
7218 | aggsum_init(&astat_l2_hdr_size, 0); | |
7219 | aggsum_init(&astat_bonus_size, 0); | |
7220 | aggsum_init(&astat_dnode_size, 0); | |
7221 | aggsum_init(&astat_dbuf_size, 0); | |
7222 | ||
d3c2ae1c GW |
7223 | arc_anon->arcs_state = ARC_STATE_ANON; |
7224 | arc_mru->arcs_state = ARC_STATE_MRU; | |
7225 | arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST; | |
7226 | arc_mfu->arcs_state = ARC_STATE_MFU; | |
7227 | arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST; | |
7228 | arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY; | |
7229 | } | |
7230 | ||
7231 | static void | |
7232 | arc_state_fini(void) | |
7233 | { | |
424fd7c3 TS |
7234 | zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); |
7235 | zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); | |
7236 | zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); | |
7237 | zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); | |
7238 | zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); | |
7239 | zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); | |
7240 | zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); | |
7241 | zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); | |
7242 | zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); | |
7243 | zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); | |
7244 | zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); | |
7245 | zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); | |
7246 | ||
7247 | zfs_refcount_destroy(&arc_anon->arcs_size); | |
7248 | zfs_refcount_destroy(&arc_mru->arcs_size); | |
7249 | zfs_refcount_destroy(&arc_mru_ghost->arcs_size); | |
7250 | zfs_refcount_destroy(&arc_mfu->arcs_size); | |
7251 | zfs_refcount_destroy(&arc_mfu_ghost->arcs_size); | |
7252 | zfs_refcount_destroy(&arc_l2c_only->arcs_size); | |
d3c2ae1c | 7253 | |
64fc7762 MA |
7254 | multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]); |
7255 | multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); | |
7256 | multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]); | |
7257 | multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); | |
7258 | multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]); | |
7259 | multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); | |
7260 | multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]); | |
7261 | multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); | |
7262 | multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]); | |
7263 | multilist_destroy(arc_l2c_only->arcs_list[ARC_BUFC_DATA]); | |
37fb3e43 PD |
7264 | |
7265 | aggsum_fini(&arc_meta_used); | |
7266 | aggsum_fini(&arc_size); | |
7267 | aggsum_fini(&astat_data_size); | |
7268 | aggsum_fini(&astat_metadata_size); | |
7269 | aggsum_fini(&astat_hdr_size); | |
7270 | aggsum_fini(&astat_l2_hdr_size); | |
7271 | aggsum_fini(&astat_bonus_size); | |
7272 | aggsum_fini(&astat_dnode_size); | |
7273 | aggsum_fini(&astat_dbuf_size); | |
d3c2ae1c GW |
7274 | } |
7275 | ||
7276 | uint64_t | |
e71cade6 | 7277 | arc_target_bytes(void) |
d3c2ae1c | 7278 | { |
e71cade6 | 7279 | return (arc_c); |
d3c2ae1c GW |
7280 | } |
7281 | ||
34dc7c2f BB |
7282 | void |
7283 | arc_init(void) | |
7284 | { | |
9edb3695 | 7285 | uint64_t percent, allmem = arc_all_memory(); |
3ec34e55 BL |
7286 | mutex_init(&arc_adjust_lock, NULL, MUTEX_DEFAULT, NULL); |
7287 | cv_init(&arc_adjust_waiters_cv, NULL, CV_DEFAULT, NULL); | |
ca0bf58d | 7288 | |
2b84817f TC |
7289 | arc_min_prefetch_ms = 1000; |
7290 | arc_min_prescient_prefetch_ms = 6000; | |
34dc7c2f | 7291 | |
c9c9c1e2 MM |
7292 | #if defined(_KERNEL) |
7293 | arc_lowmem_init(); | |
34dc7c2f BB |
7294 | #endif |
7295 | ||
9a51738b | 7296 | /* Set min cache to 1/32 of all memory, or 32MB, whichever is more. */ |
4ce3c45a | 7297 | arc_c_min = MAX(allmem / 32, 2ULL << SPA_MAXBLOCKSHIFT); |
9a51738b RM |
7298 | |
7299 | /* How to set default max varies by platform. */ | |
7300 | arc_c_max = arc_default_max(arc_c_min, allmem); | |
7301 | ||
7302 | #ifndef _KERNEL | |
ab5cbbd1 BB |
7303 | /* |
7304 | * In userland, there's only the memory pressure that we artificially | |
7305 | * create (see arc_available_memory()). Don't let arc_c get too | |
7306 | * small, because it can cause transactions to be larger than | |
7307 | * arc_c, causing arc_tempreserve_space() to fail. | |
7308 | */ | |
0a1f8cd9 | 7309 | arc_c_min = MAX(arc_c_max / 2, 2ULL << SPA_MAXBLOCKSHIFT); |
ab5cbbd1 BB |
7310 | #endif |
7311 | ||
34dc7c2f BB |
7312 | arc_c = arc_c_max; |
7313 | arc_p = (arc_c >> 1); | |
7314 | ||
ca67b33a MA |
7315 | /* Set min to 1/2 of arc_c_min */ |
7316 | arc_meta_min = 1ULL << SPA_MAXBLOCKSHIFT; | |
7317 | /* Initialize maximum observed usage to zero */ | |
1834f2d8 | 7318 | arc_meta_max = 0; |
9907cc1c G |
7319 | /* |
7320 | * Set arc_meta_limit to a percent of arc_c_max with a floor of | |
7321 | * arc_meta_min, and a ceiling of arc_c_max. | |
7322 | */ | |
7323 | percent = MIN(zfs_arc_meta_limit_percent, 100); | |
7324 | arc_meta_limit = MAX(arc_meta_min, (percent * arc_c_max) / 100); | |
7325 | percent = MIN(zfs_arc_dnode_limit_percent, 100); | |
03fdcb9a | 7326 | arc_dnode_size_limit = (percent * arc_meta_limit) / 100; |
34dc7c2f | 7327 | |
ca67b33a | 7328 | /* Apply user specified tunings */ |
36a6e233 | 7329 | arc_tuning_update(B_TRUE); |
c52fca13 | 7330 | |
34dc7c2f BB |
7331 | /* if kmem_flags are set, lets try to use less memory */ |
7332 | if (kmem_debugging()) | |
7333 | arc_c = arc_c / 2; | |
7334 | if (arc_c < arc_c_min) | |
7335 | arc_c = arc_c_min; | |
7336 | ||
d3c2ae1c | 7337 | arc_state_init(); |
3ec34e55 BL |
7338 | |
7339 | /* | |
7340 | * The arc must be "uninitialized", so that hdr_recl() (which is | |
7341 | * registered by buf_init()) will not access arc_reap_zthr before | |
7342 | * it is created. | |
7343 | */ | |
7344 | ASSERT(!arc_initialized); | |
34dc7c2f BB |
7345 | buf_init(); |
7346 | ||
ab26409d BB |
7347 | list_create(&arc_prune_list, sizeof (arc_prune_t), |
7348 | offsetof(arc_prune_t, p_node)); | |
ab26409d | 7349 | mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f | 7350 | |
57434aba D |
7351 | arc_prune_taskq = taskq_create("arc_prune", boot_ncpus, defclsyspri, |
7352 | boot_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC); | |
f6046738 | 7353 | |
34dc7c2f BB |
7354 | arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, |
7355 | sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); | |
7356 | ||
7357 | if (arc_ksp != NULL) { | |
7358 | arc_ksp->ks_data = &arc_stats; | |
13be560d | 7359 | arc_ksp->ks_update = arc_kstat_update; |
34dc7c2f BB |
7360 | kstat_install(arc_ksp); |
7361 | } | |
7362 | ||
3ec34e55 BL |
7363 | arc_adjust_zthr = zthr_create(arc_adjust_cb_check, |
7364 | arc_adjust_cb, NULL); | |
7365 | arc_reap_zthr = zthr_create_timer(arc_reap_cb_check, | |
7366 | arc_reap_cb, NULL, SEC2NSEC(1)); | |
34dc7c2f | 7367 | |
3ec34e55 | 7368 | arc_initialized = B_TRUE; |
b128c09f | 7369 | arc_warm = B_FALSE; |
34dc7c2f | 7370 | |
e8b96c60 MA |
7371 | /* |
7372 | * Calculate maximum amount of dirty data per pool. | |
7373 | * | |
7374 | * If it has been set by a module parameter, take that. | |
7375 | * Otherwise, use a percentage of physical memory defined by | |
7376 | * zfs_dirty_data_max_percent (default 10%) with a cap at | |
e99932f7 | 7377 | * zfs_dirty_data_max_max (default 4G or 25% of physical memory). |
e8b96c60 MA |
7378 | */ |
7379 | if (zfs_dirty_data_max_max == 0) | |
e99932f7 BB |
7380 | zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024, |
7381 | allmem * zfs_dirty_data_max_max_percent / 100); | |
e8b96c60 MA |
7382 | |
7383 | if (zfs_dirty_data_max == 0) { | |
9edb3695 | 7384 | zfs_dirty_data_max = allmem * |
e8b96c60 MA |
7385 | zfs_dirty_data_max_percent / 100; |
7386 | zfs_dirty_data_max = MIN(zfs_dirty_data_max, | |
7387 | zfs_dirty_data_max_max); | |
7388 | } | |
34dc7c2f BB |
7389 | } |
7390 | ||
7391 | void | |
7392 | arc_fini(void) | |
7393 | { | |
ab26409d BB |
7394 | arc_prune_t *p; |
7395 | ||
7cb67b45 | 7396 | #ifdef _KERNEL |
c9c9c1e2 | 7397 | arc_lowmem_fini(); |
7cb67b45 BB |
7398 | #endif /* _KERNEL */ |
7399 | ||
d3c2ae1c GW |
7400 | /* Use B_TRUE to ensure *all* buffers are evicted */ |
7401 | arc_flush(NULL, B_TRUE); | |
34dc7c2f | 7402 | |
3ec34e55 | 7403 | arc_initialized = B_FALSE; |
34dc7c2f BB |
7404 | |
7405 | if (arc_ksp != NULL) { | |
7406 | kstat_delete(arc_ksp); | |
7407 | arc_ksp = NULL; | |
7408 | } | |
7409 | ||
f6046738 BB |
7410 | taskq_wait(arc_prune_taskq); |
7411 | taskq_destroy(arc_prune_taskq); | |
7412 | ||
ab26409d BB |
7413 | mutex_enter(&arc_prune_mtx); |
7414 | while ((p = list_head(&arc_prune_list)) != NULL) { | |
7415 | list_remove(&arc_prune_list, p); | |
424fd7c3 TS |
7416 | zfs_refcount_remove(&p->p_refcnt, &arc_prune_list); |
7417 | zfs_refcount_destroy(&p->p_refcnt); | |
ab26409d BB |
7418 | kmem_free(p, sizeof (*p)); |
7419 | } | |
7420 | mutex_exit(&arc_prune_mtx); | |
7421 | ||
7422 | list_destroy(&arc_prune_list); | |
7423 | mutex_destroy(&arc_prune_mtx); | |
3ec34e55 | 7424 | |
1c44a5c9 | 7425 | (void) zthr_cancel(arc_adjust_zthr); |
3ec34e55 | 7426 | (void) zthr_cancel(arc_reap_zthr); |
3ec34e55 BL |
7427 | |
7428 | mutex_destroy(&arc_adjust_lock); | |
7429 | cv_destroy(&arc_adjust_waiters_cv); | |
ca0bf58d | 7430 | |
ae3d8491 PD |
7431 | /* |
7432 | * buf_fini() must proceed arc_state_fini() because buf_fin() may | |
7433 | * trigger the release of kmem magazines, which can callback to | |
7434 | * arc_space_return() which accesses aggsums freed in act_state_fini(). | |
7435 | */ | |
34dc7c2f | 7436 | buf_fini(); |
ae3d8491 | 7437 | arc_state_fini(); |
9babb374 | 7438 | |
1c44a5c9 SD |
7439 | /* |
7440 | * We destroy the zthrs after all the ARC state has been | |
7441 | * torn down to avoid the case of them receiving any | |
7442 | * wakeup() signals after they are destroyed. | |
7443 | */ | |
7444 | zthr_destroy(arc_adjust_zthr); | |
7445 | zthr_destroy(arc_reap_zthr); | |
7446 | ||
b9541d6b | 7447 | ASSERT0(arc_loaned_bytes); |
34dc7c2f BB |
7448 | } |
7449 | ||
7450 | /* | |
7451 | * Level 2 ARC | |
7452 | * | |
7453 | * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. | |
7454 | * It uses dedicated storage devices to hold cached data, which are populated | |
7455 | * using large infrequent writes. The main role of this cache is to boost | |
7456 | * the performance of random read workloads. The intended L2ARC devices | |
7457 | * include short-stroked disks, solid state disks, and other media with | |
7458 | * substantially faster read latency than disk. | |
7459 | * | |
7460 | * +-----------------------+ | |
7461 | * | ARC | | |
7462 | * +-----------------------+ | |
7463 | * | ^ ^ | |
7464 | * | | | | |
7465 | * l2arc_feed_thread() arc_read() | |
7466 | * | | | | |
7467 | * | l2arc read | | |
7468 | * V | | | |
7469 | * +---------------+ | | |
7470 | * | L2ARC | | | |
7471 | * +---------------+ | | |
7472 | * | ^ | | |
7473 | * l2arc_write() | | | |
7474 | * | | | | |
7475 | * V | | | |
7476 | * +-------+ +-------+ | |
7477 | * | vdev | | vdev | | |
7478 | * | cache | | cache | | |
7479 | * +-------+ +-------+ | |
7480 | * +=========+ .-----. | |
7481 | * : L2ARC : |-_____-| | |
7482 | * : devices : | Disks | | |
7483 | * +=========+ `-_____-' | |
7484 | * | |
7485 | * Read requests are satisfied from the following sources, in order: | |
7486 | * | |
7487 | * 1) ARC | |
7488 | * 2) vdev cache of L2ARC devices | |
7489 | * 3) L2ARC devices | |
7490 | * 4) vdev cache of disks | |
7491 | * 5) disks | |
7492 | * | |
7493 | * Some L2ARC device types exhibit extremely slow write performance. | |
7494 | * To accommodate for this there are some significant differences between | |
7495 | * the L2ARC and traditional cache design: | |
7496 | * | |
7497 | * 1. There is no eviction path from the ARC to the L2ARC. Evictions from | |
7498 | * the ARC behave as usual, freeing buffers and placing headers on ghost | |
7499 | * lists. The ARC does not send buffers to the L2ARC during eviction as | |
7500 | * this would add inflated write latencies for all ARC memory pressure. | |
7501 | * | |
7502 | * 2. The L2ARC attempts to cache data from the ARC before it is evicted. | |
7503 | * It does this by periodically scanning buffers from the eviction-end of | |
7504 | * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are | |
3a17a7a9 SK |
7505 | * not already there. It scans until a headroom of buffers is satisfied, |
7506 | * which itself is a buffer for ARC eviction. If a compressible buffer is | |
7507 | * found during scanning and selected for writing to an L2ARC device, we | |
7508 | * temporarily boost scanning headroom during the next scan cycle to make | |
7509 | * sure we adapt to compression effects (which might significantly reduce | |
7510 | * the data volume we write to L2ARC). The thread that does this is | |
34dc7c2f BB |
7511 | * l2arc_feed_thread(), illustrated below; example sizes are included to |
7512 | * provide a better sense of ratio than this diagram: | |
7513 | * | |
7514 | * head --> tail | |
7515 | * +---------------------+----------+ | |
7516 | * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC | |
7517 | * +---------------------+----------+ | o L2ARC eligible | |
7518 | * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer | |
7519 | * +---------------------+----------+ | | |
7520 | * 15.9 Gbytes ^ 32 Mbytes | | |
7521 | * headroom | | |
7522 | * l2arc_feed_thread() | |
7523 | * | | |
7524 | * l2arc write hand <--[oooo]--' | |
7525 | * | 8 Mbyte | |
7526 | * | write max | |
7527 | * V | |
7528 | * +==============================+ | |
7529 | * L2ARC dev |####|#|###|###| |####| ... | | |
7530 | * +==============================+ | |
7531 | * 32 Gbytes | |
7532 | * | |
7533 | * 3. If an ARC buffer is copied to the L2ARC but then hit instead of | |
7534 | * evicted, then the L2ARC has cached a buffer much sooner than it probably | |
7535 | * needed to, potentially wasting L2ARC device bandwidth and storage. It is | |
7536 | * safe to say that this is an uncommon case, since buffers at the end of | |
7537 | * the ARC lists have moved there due to inactivity. | |
7538 | * | |
7539 | * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, | |
7540 | * then the L2ARC simply misses copying some buffers. This serves as a | |
7541 | * pressure valve to prevent heavy read workloads from both stalling the ARC | |
7542 | * with waits and clogging the L2ARC with writes. This also helps prevent | |
7543 | * the potential for the L2ARC to churn if it attempts to cache content too | |
7544 | * quickly, such as during backups of the entire pool. | |
7545 | * | |
b128c09f BB |
7546 | * 5. After system boot and before the ARC has filled main memory, there are |
7547 | * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru | |
7548 | * lists can remain mostly static. Instead of searching from tail of these | |
7549 | * lists as pictured, the l2arc_feed_thread() will search from the list heads | |
7550 | * for eligible buffers, greatly increasing its chance of finding them. | |
7551 | * | |
7552 | * The L2ARC device write speed is also boosted during this time so that | |
7553 | * the L2ARC warms up faster. Since there have been no ARC evictions yet, | |
7554 | * there are no L2ARC reads, and no fear of degrading read performance | |
7555 | * through increased writes. | |
7556 | * | |
7557 | * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that | |
34dc7c2f BB |
7558 | * the vdev queue can aggregate them into larger and fewer writes. Each |
7559 | * device is written to in a rotor fashion, sweeping writes through | |
7560 | * available space then repeating. | |
7561 | * | |
b128c09f | 7562 | * 7. The L2ARC does not store dirty content. It never needs to flush |
34dc7c2f BB |
7563 | * write buffers back to disk based storage. |
7564 | * | |
b128c09f | 7565 | * 8. If an ARC buffer is written (and dirtied) which also exists in the |
34dc7c2f BB |
7566 | * L2ARC, the now stale L2ARC buffer is immediately dropped. |
7567 | * | |
7568 | * The performance of the L2ARC can be tweaked by a number of tunables, which | |
7569 | * may be necessary for different workloads: | |
7570 | * | |
7571 | * l2arc_write_max max write bytes per interval | |
b128c09f | 7572 | * l2arc_write_boost extra write bytes during device warmup |
34dc7c2f BB |
7573 | * l2arc_noprefetch skip caching prefetched buffers |
7574 | * l2arc_headroom number of max device writes to precache | |
3a17a7a9 SK |
7575 | * l2arc_headroom_boost when we find compressed buffers during ARC |
7576 | * scanning, we multiply headroom by this | |
7577 | * percentage factor for the next scan cycle, | |
7578 | * since more compressed buffers are likely to | |
7579 | * be present | |
34dc7c2f BB |
7580 | * l2arc_feed_secs seconds between L2ARC writing |
7581 | * | |
7582 | * Tunables may be removed or added as future performance improvements are | |
7583 | * integrated, and also may become zpool properties. | |
d164b209 BB |
7584 | * |
7585 | * There are three key functions that control how the L2ARC warms up: | |
7586 | * | |
7587 | * l2arc_write_eligible() check if a buffer is eligible to cache | |
7588 | * l2arc_write_size() calculate how much to write | |
7589 | * l2arc_write_interval() calculate sleep delay between writes | |
7590 | * | |
7591 | * These three functions determine what to write, how much, and how quickly | |
7592 | * to send writes. | |
77f6826b GA |
7593 | * |
7594 | * L2ARC persistence: | |
7595 | * | |
7596 | * When writing buffers to L2ARC, we periodically add some metadata to | |
7597 | * make sure we can pick them up after reboot, thus dramatically reducing | |
7598 | * the impact that any downtime has on the performance of storage systems | |
7599 | * with large caches. | |
7600 | * | |
7601 | * The implementation works fairly simply by integrating the following two | |
7602 | * modifications: | |
7603 | * | |
7604 | * *) When writing to the L2ARC, we occasionally write a "l2arc log block", | |
7605 | * which is an additional piece of metadata which describes what's been | |
7606 | * written. This allows us to rebuild the arc_buf_hdr_t structures of the | |
7607 | * main ARC buffers. There are 2 linked-lists of log blocks headed by | |
7608 | * dh_start_lbps[2]. We alternate which chain we append to, so they are | |
7609 | * time-wise and offset-wise interleaved, but that is an optimization rather | |
7610 | * than for correctness. The log block also includes a pointer to the | |
7611 | * previous block in its chain. | |
7612 | * | |
7613 | * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device | |
7614 | * for our header bookkeeping purposes. This contains a device header, | |
7615 | * which contains our top-level reference structures. We update it each | |
7616 | * time we write a new log block, so that we're able to locate it in the | |
7617 | * L2ARC device. If this write results in an inconsistent device header | |
7618 | * (e.g. due to power failure), we detect this by verifying the header's | |
7619 | * checksum and simply fail to reconstruct the L2ARC after reboot. | |
7620 | * | |
7621 | * Implementation diagram: | |
7622 | * | |
7623 | * +=== L2ARC device (not to scale) ======================================+ | |
7624 | * | ___two newest log block pointers__.__________ | | |
7625 | * | / \dh_start_lbps[1] | | |
7626 | * | / \ \dh_start_lbps[0]| | |
7627 | * |.___/__. V V | | |
7628 | * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---| | |
7629 | * || hdr| ^ /^ /^ / / | | |
7630 | * |+------+ ...--\-------/ \-----/--\------/ / | | |
7631 | * | \--------------/ \--------------/ | | |
7632 | * +======================================================================+ | |
7633 | * | |
7634 | * As can be seen on the diagram, rather than using a simple linked list, | |
7635 | * we use a pair of linked lists with alternating elements. This is a | |
7636 | * performance enhancement due to the fact that we only find out the | |
7637 | * address of the next log block access once the current block has been | |
7638 | * completely read in. Obviously, this hurts performance, because we'd be | |
7639 | * keeping the device's I/O queue at only a 1 operation deep, thus | |
7640 | * incurring a large amount of I/O round-trip latency. Having two lists | |
7641 | * allows us to fetch two log blocks ahead of where we are currently | |
7642 | * rebuilding L2ARC buffers. | |
7643 | * | |
7644 | * On-device data structures: | |
7645 | * | |
7646 | * L2ARC device header: l2arc_dev_hdr_phys_t | |
7647 | * L2ARC log block: l2arc_log_blk_phys_t | |
7648 | * | |
7649 | * L2ARC reconstruction: | |
7650 | * | |
7651 | * When writing data, we simply write in the standard rotary fashion, | |
7652 | * evicting buffers as we go and simply writing new data over them (writing | |
7653 | * a new log block every now and then). This obviously means that once we | |
7654 | * loop around the end of the device, we will start cutting into an already | |
7655 | * committed log block (and its referenced data buffers), like so: | |
7656 | * | |
7657 | * current write head__ __old tail | |
7658 | * \ / | |
7659 | * V V | |
7660 | * <--|bufs |lb |bufs |lb | |bufs |lb |bufs |lb |--> | |
7661 | * ^ ^^^^^^^^^___________________________________ | |
7662 | * | \ | |
7663 | * <<nextwrite>> may overwrite this blk and/or its bufs --' | |
7664 | * | |
7665 | * When importing the pool, we detect this situation and use it to stop | |
7666 | * our scanning process (see l2arc_rebuild). | |
7667 | * | |
7668 | * There is one significant caveat to consider when rebuilding ARC contents | |
7669 | * from an L2ARC device: what about invalidated buffers? Given the above | |
7670 | * construction, we cannot update blocks which we've already written to amend | |
7671 | * them to remove buffers which were invalidated. Thus, during reconstruction, | |
7672 | * we might be populating the cache with buffers for data that's not on the | |
7673 | * main pool anymore, or may have been overwritten! | |
7674 | * | |
7675 | * As it turns out, this isn't a problem. Every arc_read request includes | |
7676 | * both the DVA and, crucially, the birth TXG of the BP the caller is | |
7677 | * looking for. So even if the cache were populated by completely rotten | |
7678 | * blocks for data that had been long deleted and/or overwritten, we'll | |
7679 | * never actually return bad data from the cache, since the DVA with the | |
7680 | * birth TXG uniquely identify a block in space and time - once created, | |
7681 | * a block is immutable on disk. The worst thing we have done is wasted | |
7682 | * some time and memory at l2arc rebuild to reconstruct outdated ARC | |
7683 | * entries that will get dropped from the l2arc as it is being updated | |
7684 | * with new blocks. | |
7685 | * | |
7686 | * L2ARC buffers that have been evicted by l2arc_evict() ahead of the write | |
7687 | * hand are not restored. This is done by saving the offset (in bytes) | |
7688 | * l2arc_evict() has evicted to in the L2ARC device header and taking it | |
7689 | * into account when restoring buffers. | |
34dc7c2f BB |
7690 | */ |
7691 | ||
d164b209 | 7692 | static boolean_t |
2a432414 | 7693 | l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) |
d164b209 BB |
7694 | { |
7695 | /* | |
7696 | * A buffer is *not* eligible for the L2ARC if it: | |
7697 | * 1. belongs to a different spa. | |
428870ff BB |
7698 | * 2. is already cached on the L2ARC. |
7699 | * 3. has an I/O in progress (it may be an incomplete read). | |
7700 | * 4. is flagged not eligible (zfs property). | |
d164b209 | 7701 | */ |
b9541d6b | 7702 | if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) || |
2a432414 | 7703 | HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr)) |
d164b209 BB |
7704 | return (B_FALSE); |
7705 | ||
7706 | return (B_TRUE); | |
7707 | } | |
7708 | ||
7709 | static uint64_t | |
37c22948 | 7710 | l2arc_write_size(l2arc_dev_t *dev) |
d164b209 | 7711 | { |
37c22948 | 7712 | uint64_t size, dev_size; |
d164b209 | 7713 | |
3a17a7a9 SK |
7714 | /* |
7715 | * Make sure our globals have meaningful values in case the user | |
7716 | * altered them. | |
7717 | */ | |
7718 | size = l2arc_write_max; | |
7719 | if (size == 0) { | |
7720 | cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " | |
7721 | "be greater than zero, resetting it to the default (%d)", | |
7722 | L2ARC_WRITE_SIZE); | |
7723 | size = l2arc_write_max = L2ARC_WRITE_SIZE; | |
7724 | } | |
d164b209 BB |
7725 | |
7726 | if (arc_warm == B_FALSE) | |
3a17a7a9 | 7727 | size += l2arc_write_boost; |
d164b209 | 7728 | |
37c22948 GA |
7729 | /* |
7730 | * Make sure the write size does not exceed the size of the cache | |
7731 | * device. This is important in l2arc_evict(), otherwise infinite | |
7732 | * iteration can occur. | |
7733 | */ | |
7734 | dev_size = dev->l2ad_end - dev->l2ad_start; | |
77f6826b | 7735 | if ((size + l2arc_log_blk_overhead(size, dev)) >= dev_size) { |
37c22948 | 7736 | cmn_err(CE_NOTE, "l2arc_write_max or l2arc_write_boost " |
77f6826b GA |
7737 | "plus the overhead of log blocks (persistent L2ARC, " |
7738 | "%llu bytes) exceeds the size of the cache device " | |
7739 | "(guid %llu), resetting them to the default (%d)", | |
7740 | l2arc_log_blk_overhead(size, dev), | |
37c22948 GA |
7741 | dev->l2ad_vdev->vdev_guid, L2ARC_WRITE_SIZE); |
7742 | size = l2arc_write_max = l2arc_write_boost = L2ARC_WRITE_SIZE; | |
7743 | ||
7744 | if (arc_warm == B_FALSE) | |
7745 | size += l2arc_write_boost; | |
7746 | } | |
7747 | ||
d164b209 BB |
7748 | return (size); |
7749 | ||
7750 | } | |
7751 | ||
7752 | static clock_t | |
7753 | l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) | |
7754 | { | |
428870ff | 7755 | clock_t interval, next, now; |
d164b209 BB |
7756 | |
7757 | /* | |
7758 | * If the ARC lists are busy, increase our write rate; if the | |
7759 | * lists are stale, idle back. This is achieved by checking | |
7760 | * how much we previously wrote - if it was more than half of | |
7761 | * what we wanted, schedule the next write much sooner. | |
7762 | */ | |
7763 | if (l2arc_feed_again && wrote > (wanted / 2)) | |
7764 | interval = (hz * l2arc_feed_min_ms) / 1000; | |
7765 | else | |
7766 | interval = hz * l2arc_feed_secs; | |
7767 | ||
428870ff BB |
7768 | now = ddi_get_lbolt(); |
7769 | next = MAX(now, MIN(now + interval, began + interval)); | |
d164b209 BB |
7770 | |
7771 | return (next); | |
7772 | } | |
7773 | ||
34dc7c2f BB |
7774 | /* |
7775 | * Cycle through L2ARC devices. This is how L2ARC load balances. | |
b128c09f | 7776 | * If a device is returned, this also returns holding the spa config lock. |
34dc7c2f BB |
7777 | */ |
7778 | static l2arc_dev_t * | |
7779 | l2arc_dev_get_next(void) | |
7780 | { | |
b128c09f | 7781 | l2arc_dev_t *first, *next = NULL; |
34dc7c2f | 7782 | |
b128c09f BB |
7783 | /* |
7784 | * Lock out the removal of spas (spa_namespace_lock), then removal | |
7785 | * of cache devices (l2arc_dev_mtx). Once a device has been selected, | |
7786 | * both locks will be dropped and a spa config lock held instead. | |
7787 | */ | |
7788 | mutex_enter(&spa_namespace_lock); | |
7789 | mutex_enter(&l2arc_dev_mtx); | |
7790 | ||
7791 | /* if there are no vdevs, there is nothing to do */ | |
7792 | if (l2arc_ndev == 0) | |
7793 | goto out; | |
7794 | ||
7795 | first = NULL; | |
7796 | next = l2arc_dev_last; | |
7797 | do { | |
7798 | /* loop around the list looking for a non-faulted vdev */ | |
7799 | if (next == NULL) { | |
34dc7c2f | 7800 | next = list_head(l2arc_dev_list); |
b128c09f BB |
7801 | } else { |
7802 | next = list_next(l2arc_dev_list, next); | |
7803 | if (next == NULL) | |
7804 | next = list_head(l2arc_dev_list); | |
7805 | } | |
7806 | ||
7807 | /* if we have come back to the start, bail out */ | |
7808 | if (first == NULL) | |
7809 | first = next; | |
7810 | else if (next == first) | |
7811 | break; | |
7812 | ||
77f6826b | 7813 | } while (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild); |
b128c09f BB |
7814 | |
7815 | /* if we were unable to find any usable vdevs, return NULL */ | |
77f6826b | 7816 | if (vdev_is_dead(next->l2ad_vdev) || next->l2ad_rebuild) |
b128c09f | 7817 | next = NULL; |
34dc7c2f BB |
7818 | |
7819 | l2arc_dev_last = next; | |
7820 | ||
b128c09f BB |
7821 | out: |
7822 | mutex_exit(&l2arc_dev_mtx); | |
7823 | ||
7824 | /* | |
7825 | * Grab the config lock to prevent the 'next' device from being | |
7826 | * removed while we are writing to it. | |
7827 | */ | |
7828 | if (next != NULL) | |
7829 | spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); | |
7830 | mutex_exit(&spa_namespace_lock); | |
7831 | ||
34dc7c2f BB |
7832 | return (next); |
7833 | } | |
7834 | ||
b128c09f BB |
7835 | /* |
7836 | * Free buffers that were tagged for destruction. | |
7837 | */ | |
7838 | static void | |
0bc8fd78 | 7839 | l2arc_do_free_on_write(void) |
b128c09f BB |
7840 | { |
7841 | list_t *buflist; | |
7842 | l2arc_data_free_t *df, *df_prev; | |
7843 | ||
7844 | mutex_enter(&l2arc_free_on_write_mtx); | |
7845 | buflist = l2arc_free_on_write; | |
7846 | ||
7847 | for (df = list_tail(buflist); df; df = df_prev) { | |
7848 | df_prev = list_prev(buflist, df); | |
a6255b7f DQ |
7849 | ASSERT3P(df->l2df_abd, !=, NULL); |
7850 | abd_free(df->l2df_abd); | |
b128c09f BB |
7851 | list_remove(buflist, df); |
7852 | kmem_free(df, sizeof (l2arc_data_free_t)); | |
7853 | } | |
7854 | ||
7855 | mutex_exit(&l2arc_free_on_write_mtx); | |
7856 | } | |
7857 | ||
34dc7c2f BB |
7858 | /* |
7859 | * A write to a cache device has completed. Update all headers to allow | |
7860 | * reads from these buffers to begin. | |
7861 | */ | |
7862 | static void | |
7863 | l2arc_write_done(zio_t *zio) | |
7864 | { | |
77f6826b GA |
7865 | l2arc_write_callback_t *cb; |
7866 | l2arc_lb_abd_buf_t *abd_buf; | |
7867 | l2arc_lb_ptr_buf_t *lb_ptr_buf; | |
7868 | l2arc_dev_t *dev; | |
657fd33b | 7869 | l2arc_dev_hdr_phys_t *l2dhdr; |
77f6826b GA |
7870 | list_t *buflist; |
7871 | arc_buf_hdr_t *head, *hdr, *hdr_prev; | |
7872 | kmutex_t *hash_lock; | |
7873 | int64_t bytes_dropped = 0; | |
34dc7c2f BB |
7874 | |
7875 | cb = zio->io_private; | |
d3c2ae1c | 7876 | ASSERT3P(cb, !=, NULL); |
34dc7c2f | 7877 | dev = cb->l2wcb_dev; |
657fd33b | 7878 | l2dhdr = dev->l2ad_dev_hdr; |
d3c2ae1c | 7879 | ASSERT3P(dev, !=, NULL); |
34dc7c2f | 7880 | head = cb->l2wcb_head; |
d3c2ae1c | 7881 | ASSERT3P(head, !=, NULL); |
b9541d6b | 7882 | buflist = &dev->l2ad_buflist; |
d3c2ae1c | 7883 | ASSERT3P(buflist, !=, NULL); |
34dc7c2f BB |
7884 | DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, |
7885 | l2arc_write_callback_t *, cb); | |
7886 | ||
7887 | if (zio->io_error != 0) | |
7888 | ARCSTAT_BUMP(arcstat_l2_writes_error); | |
7889 | ||
34dc7c2f BB |
7890 | /* |
7891 | * All writes completed, or an error was hit. | |
7892 | */ | |
ca0bf58d PS |
7893 | top: |
7894 | mutex_enter(&dev->l2ad_mtx); | |
2a432414 GW |
7895 | for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { |
7896 | hdr_prev = list_prev(buflist, hdr); | |
34dc7c2f | 7897 | |
2a432414 | 7898 | hash_lock = HDR_LOCK(hdr); |
ca0bf58d PS |
7899 | |
7900 | /* | |
7901 | * We cannot use mutex_enter or else we can deadlock | |
7902 | * with l2arc_write_buffers (due to swapping the order | |
7903 | * the hash lock and l2ad_mtx are taken). | |
7904 | */ | |
34dc7c2f BB |
7905 | if (!mutex_tryenter(hash_lock)) { |
7906 | /* | |
ca0bf58d PS |
7907 | * Missed the hash lock. We must retry so we |
7908 | * don't leave the ARC_FLAG_L2_WRITING bit set. | |
34dc7c2f | 7909 | */ |
ca0bf58d PS |
7910 | ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); |
7911 | ||
7912 | /* | |
7913 | * We don't want to rescan the headers we've | |
7914 | * already marked as having been written out, so | |
7915 | * we reinsert the head node so we can pick up | |
7916 | * where we left off. | |
7917 | */ | |
7918 | list_remove(buflist, head); | |
7919 | list_insert_after(buflist, hdr, head); | |
7920 | ||
7921 | mutex_exit(&dev->l2ad_mtx); | |
7922 | ||
7923 | /* | |
7924 | * We wait for the hash lock to become available | |
7925 | * to try and prevent busy waiting, and increase | |
7926 | * the chance we'll be able to acquire the lock | |
7927 | * the next time around. | |
7928 | */ | |
7929 | mutex_enter(hash_lock); | |
7930 | mutex_exit(hash_lock); | |
7931 | goto top; | |
34dc7c2f BB |
7932 | } |
7933 | ||
b9541d6b | 7934 | /* |
ca0bf58d PS |
7935 | * We could not have been moved into the arc_l2c_only |
7936 | * state while in-flight due to our ARC_FLAG_L2_WRITING | |
7937 | * bit being set. Let's just ensure that's being enforced. | |
7938 | */ | |
7939 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
7940 | ||
8a09d5fd BB |
7941 | /* |
7942 | * Skipped - drop L2ARC entry and mark the header as no | |
7943 | * longer L2 eligibile. | |
7944 | */ | |
d3c2ae1c | 7945 | if (zio->io_error != 0) { |
34dc7c2f | 7946 | /* |
b128c09f | 7947 | * Error - drop L2ARC entry. |
34dc7c2f | 7948 | */ |
2a432414 | 7949 | list_remove(buflist, hdr); |
d3c2ae1c | 7950 | arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); |
b9541d6b | 7951 | |
7558997d SD |
7952 | uint64_t psize = HDR_GET_PSIZE(hdr); |
7953 | ARCSTAT_INCR(arcstat_l2_psize, -psize); | |
01850391 | 7954 | ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); |
d962d5da | 7955 | |
7558997d SD |
7956 | bytes_dropped += |
7957 | vdev_psize_to_asize(dev->l2ad_vdev, psize); | |
424fd7c3 | 7958 | (void) zfs_refcount_remove_many(&dev->l2ad_alloc, |
d3c2ae1c | 7959 | arc_hdr_size(hdr), hdr); |
34dc7c2f BB |
7960 | } |
7961 | ||
7962 | /* | |
ca0bf58d PS |
7963 | * Allow ARC to begin reads and ghost list evictions to |
7964 | * this L2ARC entry. | |
34dc7c2f | 7965 | */ |
d3c2ae1c | 7966 | arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); |
34dc7c2f BB |
7967 | |
7968 | mutex_exit(hash_lock); | |
7969 | } | |
7970 | ||
77f6826b GA |
7971 | /* |
7972 | * Free the allocated abd buffers for writing the log blocks. | |
7973 | * If the zio failed reclaim the allocated space and remove the | |
7974 | * pointers to these log blocks from the log block pointer list | |
7975 | * of the L2ARC device. | |
7976 | */ | |
7977 | while ((abd_buf = list_remove_tail(&cb->l2wcb_abd_list)) != NULL) { | |
7978 | abd_free(abd_buf->abd); | |
7979 | zio_buf_free(abd_buf, sizeof (*abd_buf)); | |
7980 | if (zio->io_error != 0) { | |
7981 | lb_ptr_buf = list_remove_head(&dev->l2ad_lbptr_list); | |
657fd33b GA |
7982 | /* |
7983 | * L2BLK_GET_PSIZE returns aligned size for log | |
7984 | * blocks. | |
7985 | */ | |
7986 | uint64_t asize = | |
77f6826b | 7987 | L2BLK_GET_PSIZE((lb_ptr_buf->lb_ptr)->lbp_prop); |
657fd33b GA |
7988 | bytes_dropped += asize; |
7989 | ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize); | |
7990 | ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count); | |
7991 | zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize, | |
7992 | lb_ptr_buf); | |
7993 | zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf); | |
77f6826b GA |
7994 | kmem_free(lb_ptr_buf->lb_ptr, |
7995 | sizeof (l2arc_log_blkptr_t)); | |
7996 | kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t)); | |
7997 | } | |
7998 | } | |
7999 | list_destroy(&cb->l2wcb_abd_list); | |
8000 | ||
657fd33b GA |
8001 | if (zio->io_error != 0) { |
8002 | /* restore the lbps array in the header to its previous state */ | |
8003 | lb_ptr_buf = list_head(&dev->l2ad_lbptr_list); | |
8004 | for (int i = 0; i < 2; i++) { | |
8005 | bcopy(lb_ptr_buf->lb_ptr, &l2dhdr->dh_start_lbps[i], | |
8006 | sizeof (l2arc_log_blkptr_t)); | |
8007 | lb_ptr_buf = list_next(&dev->l2ad_lbptr_list, | |
8008 | lb_ptr_buf); | |
8009 | } | |
8010 | } | |
8011 | ||
34dc7c2f BB |
8012 | atomic_inc_64(&l2arc_writes_done); |
8013 | list_remove(buflist, head); | |
b9541d6b CW |
8014 | ASSERT(!HDR_HAS_L1HDR(head)); |
8015 | kmem_cache_free(hdr_l2only_cache, head); | |
8016 | mutex_exit(&dev->l2ad_mtx); | |
34dc7c2f | 8017 | |
77f6826b | 8018 | ASSERT(dev->l2ad_vdev != NULL); |
3bec585e SK |
8019 | vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); |
8020 | ||
b128c09f | 8021 | l2arc_do_free_on_write(); |
34dc7c2f BB |
8022 | |
8023 | kmem_free(cb, sizeof (l2arc_write_callback_t)); | |
8024 | } | |
8025 | ||
b5256303 TC |
8026 | static int |
8027 | l2arc_untransform(zio_t *zio, l2arc_read_callback_t *cb) | |
8028 | { | |
8029 | int ret; | |
8030 | spa_t *spa = zio->io_spa; | |
8031 | arc_buf_hdr_t *hdr = cb->l2rcb_hdr; | |
8032 | blkptr_t *bp = zio->io_bp; | |
b5256303 TC |
8033 | uint8_t salt[ZIO_DATA_SALT_LEN]; |
8034 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
8035 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
8036 | boolean_t no_crypt = B_FALSE; | |
8037 | ||
8038 | /* | |
8039 | * ZIL data is never be written to the L2ARC, so we don't need | |
8040 | * special handling for its unique MAC storage. | |
8041 | */ | |
8042 | ASSERT3U(BP_GET_TYPE(bp), !=, DMU_OT_INTENT_LOG); | |
8043 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); | |
440a3eb9 | 8044 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); |
b5256303 | 8045 | |
440a3eb9 TC |
8046 | /* |
8047 | * If the data was encrypted, decrypt it now. Note that | |
8048 | * we must check the bp here and not the hdr, since the | |
8049 | * hdr does not have its encryption parameters updated | |
8050 | * until arc_read_done(). | |
8051 | */ | |
8052 | if (BP_IS_ENCRYPTED(bp)) { | |
be9a5c35 | 8053 | abd_t *eabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); |
b5256303 TC |
8054 | |
8055 | zio_crypt_decode_params_bp(bp, salt, iv); | |
8056 | zio_crypt_decode_mac_bp(bp, mac); | |
8057 | ||
be9a5c35 TC |
8058 | ret = spa_do_crypt_abd(B_FALSE, spa, &cb->l2rcb_zb, |
8059 | BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), | |
8060 | salt, iv, mac, HDR_GET_PSIZE(hdr), eabd, | |
8061 | hdr->b_l1hdr.b_pabd, &no_crypt); | |
b5256303 TC |
8062 | if (ret != 0) { |
8063 | arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr); | |
b5256303 TC |
8064 | goto error; |
8065 | } | |
8066 | ||
b5256303 TC |
8067 | /* |
8068 | * If we actually performed decryption, replace b_pabd | |
8069 | * with the decrypted data. Otherwise we can just throw | |
8070 | * our decryption buffer away. | |
8071 | */ | |
8072 | if (!no_crypt) { | |
8073 | arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, | |
8074 | arc_hdr_size(hdr), hdr); | |
8075 | hdr->b_l1hdr.b_pabd = eabd; | |
8076 | zio->io_abd = eabd; | |
8077 | } else { | |
8078 | arc_free_data_abd(hdr, eabd, arc_hdr_size(hdr), hdr); | |
8079 | } | |
8080 | } | |
8081 | ||
8082 | /* | |
8083 | * If the L2ARC block was compressed, but ARC compression | |
8084 | * is disabled we decompress the data into a new buffer and | |
8085 | * replace the existing data. | |
8086 | */ | |
8087 | if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && | |
8088 | !HDR_COMPRESSION_ENABLED(hdr)) { | |
8089 | abd_t *cabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); | |
8090 | void *tmp = abd_borrow_buf(cabd, arc_hdr_size(hdr)); | |
8091 | ||
8092 | ret = zio_decompress_data(HDR_GET_COMPRESS(hdr), | |
8093 | hdr->b_l1hdr.b_pabd, tmp, HDR_GET_PSIZE(hdr), | |
8094 | HDR_GET_LSIZE(hdr)); | |
8095 | if (ret != 0) { | |
8096 | abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr)); | |
8097 | arc_free_data_abd(hdr, cabd, arc_hdr_size(hdr), hdr); | |
8098 | goto error; | |
8099 | } | |
8100 | ||
8101 | abd_return_buf_copy(cabd, tmp, arc_hdr_size(hdr)); | |
8102 | arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, | |
8103 | arc_hdr_size(hdr), hdr); | |
8104 | hdr->b_l1hdr.b_pabd = cabd; | |
8105 | zio->io_abd = cabd; | |
8106 | zio->io_size = HDR_GET_LSIZE(hdr); | |
8107 | } | |
8108 | ||
8109 | return (0); | |
8110 | ||
8111 | error: | |
8112 | return (ret); | |
8113 | } | |
8114 | ||
8115 | ||
34dc7c2f BB |
8116 | /* |
8117 | * A read to a cache device completed. Validate buffer contents before | |
8118 | * handing over to the regular ARC routines. | |
8119 | */ | |
8120 | static void | |
8121 | l2arc_read_done(zio_t *zio) | |
8122 | { | |
b5256303 | 8123 | int tfm_error = 0; |
b405837a | 8124 | l2arc_read_callback_t *cb = zio->io_private; |
34dc7c2f | 8125 | arc_buf_hdr_t *hdr; |
34dc7c2f | 8126 | kmutex_t *hash_lock; |
b405837a TC |
8127 | boolean_t valid_cksum; |
8128 | boolean_t using_rdata = (BP_IS_ENCRYPTED(&cb->l2rcb_bp) && | |
8129 | (cb->l2rcb_flags & ZIO_FLAG_RAW_ENCRYPT)); | |
b128c09f | 8130 | |
d3c2ae1c | 8131 | ASSERT3P(zio->io_vd, !=, NULL); |
b128c09f BB |
8132 | ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); |
8133 | ||
8134 | spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); | |
34dc7c2f | 8135 | |
d3c2ae1c GW |
8136 | ASSERT3P(cb, !=, NULL); |
8137 | hdr = cb->l2rcb_hdr; | |
8138 | ASSERT3P(hdr, !=, NULL); | |
34dc7c2f | 8139 | |
d3c2ae1c | 8140 | hash_lock = HDR_LOCK(hdr); |
34dc7c2f | 8141 | mutex_enter(hash_lock); |
428870ff | 8142 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); |
34dc7c2f | 8143 | |
82710e99 GDN |
8144 | /* |
8145 | * If the data was read into a temporary buffer, | |
8146 | * move it and free the buffer. | |
8147 | */ | |
8148 | if (cb->l2rcb_abd != NULL) { | |
8149 | ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); | |
8150 | if (zio->io_error == 0) { | |
b405837a TC |
8151 | if (using_rdata) { |
8152 | abd_copy(hdr->b_crypt_hdr.b_rabd, | |
8153 | cb->l2rcb_abd, arc_hdr_size(hdr)); | |
8154 | } else { | |
8155 | abd_copy(hdr->b_l1hdr.b_pabd, | |
8156 | cb->l2rcb_abd, arc_hdr_size(hdr)); | |
8157 | } | |
82710e99 GDN |
8158 | } |
8159 | ||
8160 | /* | |
8161 | * The following must be done regardless of whether | |
8162 | * there was an error: | |
8163 | * - free the temporary buffer | |
8164 | * - point zio to the real ARC buffer | |
8165 | * - set zio size accordingly | |
8166 | * These are required because zio is either re-used for | |
8167 | * an I/O of the block in the case of the error | |
8168 | * or the zio is passed to arc_read_done() and it | |
8169 | * needs real data. | |
8170 | */ | |
8171 | abd_free(cb->l2rcb_abd); | |
8172 | zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); | |
440a3eb9 | 8173 | |
b405837a | 8174 | if (using_rdata) { |
440a3eb9 TC |
8175 | ASSERT(HDR_HAS_RABD(hdr)); |
8176 | zio->io_abd = zio->io_orig_abd = | |
8177 | hdr->b_crypt_hdr.b_rabd; | |
8178 | } else { | |
8179 | ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); | |
8180 | zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd; | |
8181 | } | |
82710e99 GDN |
8182 | } |
8183 | ||
a6255b7f | 8184 | ASSERT3P(zio->io_abd, !=, NULL); |
3a17a7a9 | 8185 | |
34dc7c2f BB |
8186 | /* |
8187 | * Check this survived the L2ARC journey. | |
8188 | */ | |
b5256303 TC |
8189 | ASSERT(zio->io_abd == hdr->b_l1hdr.b_pabd || |
8190 | (HDR_HAS_RABD(hdr) && zio->io_abd == hdr->b_crypt_hdr.b_rabd)); | |
d3c2ae1c GW |
8191 | zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ |
8192 | zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ | |
8193 | ||
8194 | valid_cksum = arc_cksum_is_equal(hdr, zio); | |
b5256303 TC |
8195 | |
8196 | /* | |
8197 | * b_rabd will always match the data as it exists on disk if it is | |
8198 | * being used. Therefore if we are reading into b_rabd we do not | |
8199 | * attempt to untransform the data. | |
8200 | */ | |
8201 | if (valid_cksum && !using_rdata) | |
8202 | tfm_error = l2arc_untransform(zio, cb); | |
8203 | ||
8204 | if (valid_cksum && tfm_error == 0 && zio->io_error == 0 && | |
8205 | !HDR_L2_EVICTED(hdr)) { | |
34dc7c2f | 8206 | mutex_exit(hash_lock); |
d3c2ae1c | 8207 | zio->io_private = hdr; |
34dc7c2f BB |
8208 | arc_read_done(zio); |
8209 | } else { | |
34dc7c2f BB |
8210 | /* |
8211 | * Buffer didn't survive caching. Increment stats and | |
8212 | * reissue to the original storage device. | |
8213 | */ | |
b128c09f | 8214 | if (zio->io_error != 0) { |
34dc7c2f | 8215 | ARCSTAT_BUMP(arcstat_l2_io_error); |
b128c09f | 8216 | } else { |
2e528b49 | 8217 | zio->io_error = SET_ERROR(EIO); |
b128c09f | 8218 | } |
b5256303 | 8219 | if (!valid_cksum || tfm_error != 0) |
34dc7c2f BB |
8220 | ARCSTAT_BUMP(arcstat_l2_cksum_bad); |
8221 | ||
34dc7c2f | 8222 | /* |
b128c09f BB |
8223 | * If there's no waiter, issue an async i/o to the primary |
8224 | * storage now. If there *is* a waiter, the caller must | |
8225 | * issue the i/o in a context where it's OK to block. | |
34dc7c2f | 8226 | */ |
d164b209 BB |
8227 | if (zio->io_waiter == NULL) { |
8228 | zio_t *pio = zio_unique_parent(zio); | |
b5256303 TC |
8229 | void *abd = (using_rdata) ? |
8230 | hdr->b_crypt_hdr.b_rabd : hdr->b_l1hdr.b_pabd; | |
d164b209 BB |
8231 | |
8232 | ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); | |
8233 | ||
5ff2249f | 8234 | zio = zio_read(pio, zio->io_spa, zio->io_bp, |
b5256303 | 8235 | abd, zio->io_size, arc_read_done, |
d3c2ae1c | 8236 | hdr, zio->io_priority, cb->l2rcb_flags, |
5ff2249f AM |
8237 | &cb->l2rcb_zb); |
8238 | ||
8239 | /* | |
8240 | * Original ZIO will be freed, so we need to update | |
8241 | * ARC header with the new ZIO pointer to be used | |
8242 | * by zio_change_priority() in arc_read(). | |
8243 | */ | |
8244 | for (struct arc_callback *acb = hdr->b_l1hdr.b_acb; | |
8245 | acb != NULL; acb = acb->acb_next) | |
8246 | acb->acb_zio_head = zio; | |
8247 | ||
8248 | mutex_exit(hash_lock); | |
8249 | zio_nowait(zio); | |
8250 | } else { | |
8251 | mutex_exit(hash_lock); | |
d164b209 | 8252 | } |
34dc7c2f BB |
8253 | } |
8254 | ||
8255 | kmem_free(cb, sizeof (l2arc_read_callback_t)); | |
8256 | } | |
8257 | ||
8258 | /* | |
8259 | * This is the list priority from which the L2ARC will search for pages to | |
8260 | * cache. This is used within loops (0..3) to cycle through lists in the | |
8261 | * desired order. This order can have a significant effect on cache | |
8262 | * performance. | |
8263 | * | |
8264 | * Currently the metadata lists are hit first, MFU then MRU, followed by | |
8265 | * the data lists. This function returns a locked list, and also returns | |
8266 | * the lock pointer. | |
8267 | */ | |
ca0bf58d PS |
8268 | static multilist_sublist_t * |
8269 | l2arc_sublist_lock(int list_num) | |
34dc7c2f | 8270 | { |
ca0bf58d PS |
8271 | multilist_t *ml = NULL; |
8272 | unsigned int idx; | |
34dc7c2f | 8273 | |
4aafab91 | 8274 | ASSERT(list_num >= 0 && list_num < L2ARC_FEED_TYPES); |
34dc7c2f BB |
8275 | |
8276 | switch (list_num) { | |
8277 | case 0: | |
64fc7762 | 8278 | ml = arc_mfu->arcs_list[ARC_BUFC_METADATA]; |
34dc7c2f BB |
8279 | break; |
8280 | case 1: | |
64fc7762 | 8281 | ml = arc_mru->arcs_list[ARC_BUFC_METADATA]; |
34dc7c2f BB |
8282 | break; |
8283 | case 2: | |
64fc7762 | 8284 | ml = arc_mfu->arcs_list[ARC_BUFC_DATA]; |
34dc7c2f BB |
8285 | break; |
8286 | case 3: | |
64fc7762 | 8287 | ml = arc_mru->arcs_list[ARC_BUFC_DATA]; |
34dc7c2f | 8288 | break; |
4aafab91 G |
8289 | default: |
8290 | return (NULL); | |
34dc7c2f BB |
8291 | } |
8292 | ||
ca0bf58d PS |
8293 | /* |
8294 | * Return a randomly-selected sublist. This is acceptable | |
8295 | * because the caller feeds only a little bit of data for each | |
8296 | * call (8MB). Subsequent calls will result in different | |
8297 | * sublists being selected. | |
8298 | */ | |
8299 | idx = multilist_get_random_index(ml); | |
8300 | return (multilist_sublist_lock(ml, idx)); | |
34dc7c2f BB |
8301 | } |
8302 | ||
77f6826b GA |
8303 | /* |
8304 | * Calculates the maximum overhead of L2ARC metadata log blocks for a given | |
657fd33b | 8305 | * L2ARC write size. l2arc_evict and l2arc_write_size need to include this |
77f6826b GA |
8306 | * overhead in processing to make sure there is enough headroom available |
8307 | * when writing buffers. | |
8308 | */ | |
8309 | static inline uint64_t | |
8310 | l2arc_log_blk_overhead(uint64_t write_sz, l2arc_dev_t *dev) | |
8311 | { | |
657fd33b | 8312 | if (dev->l2ad_log_entries == 0) { |
77f6826b GA |
8313 | return (0); |
8314 | } else { | |
8315 | uint64_t log_entries = write_sz >> SPA_MINBLOCKSHIFT; | |
8316 | ||
8317 | uint64_t log_blocks = (log_entries + | |
657fd33b GA |
8318 | dev->l2ad_log_entries - 1) / |
8319 | dev->l2ad_log_entries; | |
77f6826b GA |
8320 | |
8321 | return (vdev_psize_to_asize(dev->l2ad_vdev, | |
8322 | sizeof (l2arc_log_blk_phys_t)) * log_blocks); | |
8323 | } | |
8324 | } | |
8325 | ||
34dc7c2f BB |
8326 | /* |
8327 | * Evict buffers from the device write hand to the distance specified in | |
77f6826b | 8328 | * bytes. This distance may span populated buffers, it may span nothing. |
34dc7c2f BB |
8329 | * This is clearing a region on the L2ARC device ready for writing. |
8330 | * If the 'all' boolean is set, every buffer is evicted. | |
8331 | */ | |
8332 | static void | |
8333 | l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) | |
8334 | { | |
8335 | list_t *buflist; | |
2a432414 | 8336 | arc_buf_hdr_t *hdr, *hdr_prev; |
34dc7c2f BB |
8337 | kmutex_t *hash_lock; |
8338 | uint64_t taddr; | |
37c22948 | 8339 | boolean_t rerun; |
77f6826b | 8340 | l2arc_lb_ptr_buf_t *lb_ptr_buf, *lb_ptr_buf_prev; |
34dc7c2f | 8341 | |
b9541d6b | 8342 | buflist = &dev->l2ad_buflist; |
34dc7c2f | 8343 | |
77f6826b GA |
8344 | /* |
8345 | * We need to add in the worst case scenario of log block overhead. | |
8346 | */ | |
8347 | distance += l2arc_log_blk_overhead(distance, dev); | |
8348 | ||
37c22948 GA |
8349 | top: |
8350 | rerun = B_FALSE; | |
8351 | if (dev->l2ad_hand >= (dev->l2ad_end - distance)) { | |
34dc7c2f | 8352 | /* |
37c22948 | 8353 | * When there is no space to accomodate upcoming writes, |
77f6826b GA |
8354 | * evict to the end. Then bump the write and evict hands |
8355 | * to the start and iterate. This iteration does not | |
8356 | * happen indefinitely as we make sure in | |
8357 | * l2arc_write_size() that when the write hand is reset, | |
8358 | * the write size does not exceed the end of the device. | |
34dc7c2f | 8359 | */ |
37c22948 | 8360 | rerun = B_TRUE; |
34dc7c2f BB |
8361 | taddr = dev->l2ad_end; |
8362 | } else { | |
8363 | taddr = dev->l2ad_hand + distance; | |
8364 | } | |
8365 | DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, | |
8366 | uint64_t, taddr, boolean_t, all); | |
8367 | ||
77f6826b GA |
8368 | /* |
8369 | * This check has to be placed after deciding whether to iterate | |
8370 | * (rerun). | |
8371 | */ | |
37c22948 GA |
8372 | if (!all && dev->l2ad_first) { |
8373 | /* | |
77f6826b | 8374 | * This is the first sweep through the device. There is |
37c22948 GA |
8375 | * nothing to evict. |
8376 | */ | |
8377 | goto out; | |
8378 | } | |
8379 | ||
77f6826b GA |
8380 | /* |
8381 | * When rebuilding L2ARC we retrieve the evict hand from the header of | |
8382 | * the device. Of note, l2arc_evict() does not actually delete buffers | |
8383 | * from the cache device, but keeping track of the evict hand will be | |
8384 | * useful when TRIM is implemented. | |
8385 | */ | |
8386 | dev->l2ad_evict = MAX(dev->l2ad_evict, taddr); | |
8387 | ||
37c22948 | 8388 | retry: |
b9541d6b | 8389 | mutex_enter(&dev->l2ad_mtx); |
77f6826b GA |
8390 | /* |
8391 | * We have to account for evicted log blocks. Run vdev_space_update() | |
8392 | * on log blocks whose offset (in bytes) is before the evicted offset | |
8393 | * (in bytes) by searching in the list of pointers to log blocks | |
8394 | * present in the L2ARC device. | |
8395 | */ | |
8396 | for (lb_ptr_buf = list_tail(&dev->l2ad_lbptr_list); lb_ptr_buf; | |
8397 | lb_ptr_buf = lb_ptr_buf_prev) { | |
8398 | ||
8399 | lb_ptr_buf_prev = list_prev(&dev->l2ad_lbptr_list, lb_ptr_buf); | |
8400 | ||
657fd33b GA |
8401 | /* L2BLK_GET_PSIZE returns aligned size for log blocks */ |
8402 | uint64_t asize = L2BLK_GET_PSIZE( | |
8403 | (lb_ptr_buf->lb_ptr)->lbp_prop); | |
8404 | ||
77f6826b GA |
8405 | /* |
8406 | * We don't worry about log blocks left behind (ie | |
657fd33b | 8407 | * lbp_payload_start < l2ad_hand) because l2arc_write_buffers() |
77f6826b GA |
8408 | * will never write more than l2arc_evict() evicts. |
8409 | */ | |
8410 | if (!all && l2arc_log_blkptr_valid(dev, lb_ptr_buf->lb_ptr)) { | |
8411 | break; | |
8412 | } else { | |
657fd33b GA |
8413 | vdev_space_update(dev->l2ad_vdev, -asize, 0, 0); |
8414 | ARCSTAT_INCR(arcstat_l2_log_blk_asize, -asize); | |
8415 | ARCSTAT_BUMPDOWN(arcstat_l2_log_blk_count); | |
8416 | zfs_refcount_remove_many(&dev->l2ad_lb_asize, asize, | |
8417 | lb_ptr_buf); | |
8418 | zfs_refcount_remove(&dev->l2ad_lb_count, lb_ptr_buf); | |
77f6826b GA |
8419 | list_remove(&dev->l2ad_lbptr_list, lb_ptr_buf); |
8420 | kmem_free(lb_ptr_buf->lb_ptr, | |
8421 | sizeof (l2arc_log_blkptr_t)); | |
8422 | kmem_free(lb_ptr_buf, sizeof (l2arc_lb_ptr_buf_t)); | |
8423 | } | |
8424 | } | |
8425 | ||
2a432414 GW |
8426 | for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { |
8427 | hdr_prev = list_prev(buflist, hdr); | |
34dc7c2f | 8428 | |
ca6c7a94 | 8429 | ASSERT(!HDR_EMPTY(hdr)); |
2a432414 | 8430 | hash_lock = HDR_LOCK(hdr); |
ca0bf58d PS |
8431 | |
8432 | /* | |
8433 | * We cannot use mutex_enter or else we can deadlock | |
8434 | * with l2arc_write_buffers (due to swapping the order | |
8435 | * the hash lock and l2ad_mtx are taken). | |
8436 | */ | |
34dc7c2f BB |
8437 | if (!mutex_tryenter(hash_lock)) { |
8438 | /* | |
8439 | * Missed the hash lock. Retry. | |
8440 | */ | |
8441 | ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); | |
b9541d6b | 8442 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f BB |
8443 | mutex_enter(hash_lock); |
8444 | mutex_exit(hash_lock); | |
37c22948 | 8445 | goto retry; |
34dc7c2f BB |
8446 | } |
8447 | ||
f06f53fa AG |
8448 | /* |
8449 | * A header can't be on this list if it doesn't have L2 header. | |
8450 | */ | |
8451 | ASSERT(HDR_HAS_L2HDR(hdr)); | |
34dc7c2f | 8452 | |
f06f53fa AG |
8453 | /* Ensure this header has finished being written. */ |
8454 | ASSERT(!HDR_L2_WRITING(hdr)); | |
8455 | ASSERT(!HDR_L2_WRITE_HEAD(hdr)); | |
8456 | ||
77f6826b | 8457 | if (!all && (hdr->b_l2hdr.b_daddr >= dev->l2ad_evict || |
b9541d6b | 8458 | hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { |
34dc7c2f BB |
8459 | /* |
8460 | * We've evicted to the target address, | |
8461 | * or the end of the device. | |
8462 | */ | |
8463 | mutex_exit(hash_lock); | |
8464 | break; | |
8465 | } | |
8466 | ||
b9541d6b | 8467 | if (!HDR_HAS_L1HDR(hdr)) { |
2a432414 | 8468 | ASSERT(!HDR_L2_READING(hdr)); |
34dc7c2f BB |
8469 | /* |
8470 | * This doesn't exist in the ARC. Destroy. | |
8471 | * arc_hdr_destroy() will call list_remove() | |
01850391 | 8472 | * and decrement arcstat_l2_lsize. |
34dc7c2f | 8473 | */ |
2a432414 GW |
8474 | arc_change_state(arc_anon, hdr, hash_lock); |
8475 | arc_hdr_destroy(hdr); | |
34dc7c2f | 8476 | } else { |
b9541d6b CW |
8477 | ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); |
8478 | ARCSTAT_BUMP(arcstat_l2_evict_l1cached); | |
b128c09f BB |
8479 | /* |
8480 | * Invalidate issued or about to be issued | |
8481 | * reads, since we may be about to write | |
8482 | * over this location. | |
8483 | */ | |
2a432414 | 8484 | if (HDR_L2_READING(hdr)) { |
b128c09f | 8485 | ARCSTAT_BUMP(arcstat_l2_evict_reading); |
d3c2ae1c | 8486 | arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); |
b128c09f BB |
8487 | } |
8488 | ||
d962d5da | 8489 | arc_hdr_l2hdr_destroy(hdr); |
34dc7c2f BB |
8490 | } |
8491 | mutex_exit(hash_lock); | |
8492 | } | |
b9541d6b | 8493 | mutex_exit(&dev->l2ad_mtx); |
37c22948 GA |
8494 | |
8495 | out: | |
77f6826b GA |
8496 | /* |
8497 | * We need to check if we evict all buffers, otherwise we may iterate | |
8498 | * unnecessarily. | |
8499 | */ | |
8500 | if (!all && rerun) { | |
37c22948 GA |
8501 | /* |
8502 | * Bump device hand to the device start if it is approaching the | |
8503 | * end. l2arc_evict() has already evicted ahead for this case. | |
8504 | */ | |
8505 | dev->l2ad_hand = dev->l2ad_start; | |
77f6826b | 8506 | dev->l2ad_evict = dev->l2ad_start; |
37c22948 GA |
8507 | dev->l2ad_first = B_FALSE; |
8508 | goto top; | |
8509 | } | |
657fd33b GA |
8510 | |
8511 | ASSERT3U(dev->l2ad_hand + distance, <, dev->l2ad_end); | |
8512 | if (!dev->l2ad_first) | |
8513 | ASSERT3U(dev->l2ad_hand, <, dev->l2ad_evict); | |
34dc7c2f BB |
8514 | } |
8515 | ||
b5256303 TC |
8516 | /* |
8517 | * Handle any abd transforms that might be required for writing to the L2ARC. | |
8518 | * If successful, this function will always return an abd with the data | |
8519 | * transformed as it is on disk in a new abd of asize bytes. | |
8520 | */ | |
8521 | static int | |
8522 | l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize, | |
8523 | abd_t **abd_out) | |
8524 | { | |
8525 | int ret; | |
8526 | void *tmp = NULL; | |
8527 | abd_t *cabd = NULL, *eabd = NULL, *to_write = hdr->b_l1hdr.b_pabd; | |
8528 | enum zio_compress compress = HDR_GET_COMPRESS(hdr); | |
8529 | uint64_t psize = HDR_GET_PSIZE(hdr); | |
8530 | uint64_t size = arc_hdr_size(hdr); | |
8531 | boolean_t ismd = HDR_ISTYPE_METADATA(hdr); | |
8532 | boolean_t bswap = (hdr->b_l1hdr.b_byteswap != DMU_BSWAP_NUMFUNCS); | |
8533 | dsl_crypto_key_t *dck = NULL; | |
8534 | uint8_t mac[ZIO_DATA_MAC_LEN] = { 0 }; | |
4807c0ba | 8535 | boolean_t no_crypt = B_FALSE; |
b5256303 TC |
8536 | |
8537 | ASSERT((HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && | |
8538 | !HDR_COMPRESSION_ENABLED(hdr)) || | |
8539 | HDR_ENCRYPTED(hdr) || HDR_SHARED_DATA(hdr) || psize != asize); | |
8540 | ASSERT3U(psize, <=, asize); | |
8541 | ||
8542 | /* | |
8543 | * If this data simply needs its own buffer, we simply allocate it | |
b7109a41 | 8544 | * and copy the data. This may be done to eliminate a dependency on a |
b5256303 TC |
8545 | * shared buffer or to reallocate the buffer to match asize. |
8546 | */ | |
4807c0ba | 8547 | if (HDR_HAS_RABD(hdr) && asize != psize) { |
10adee27 | 8548 | ASSERT3U(asize, >=, psize); |
4807c0ba | 8549 | to_write = abd_alloc_for_io(asize, ismd); |
10adee27 TC |
8550 | abd_copy(to_write, hdr->b_crypt_hdr.b_rabd, psize); |
8551 | if (psize != asize) | |
8552 | abd_zero_off(to_write, psize, asize - psize); | |
4807c0ba TC |
8553 | goto out; |
8554 | } | |
8555 | ||
b5256303 TC |
8556 | if ((compress == ZIO_COMPRESS_OFF || HDR_COMPRESSION_ENABLED(hdr)) && |
8557 | !HDR_ENCRYPTED(hdr)) { | |
8558 | ASSERT3U(size, ==, psize); | |
8559 | to_write = abd_alloc_for_io(asize, ismd); | |
8560 | abd_copy(to_write, hdr->b_l1hdr.b_pabd, size); | |
8561 | if (size != asize) | |
8562 | abd_zero_off(to_write, size, asize - size); | |
8563 | goto out; | |
8564 | } | |
8565 | ||
8566 | if (compress != ZIO_COMPRESS_OFF && !HDR_COMPRESSION_ENABLED(hdr)) { | |
8567 | cabd = abd_alloc_for_io(asize, ismd); | |
8568 | tmp = abd_borrow_buf(cabd, asize); | |
8569 | ||
8570 | psize = zio_compress_data(compress, to_write, tmp, size); | |
8571 | ASSERT3U(psize, <=, HDR_GET_PSIZE(hdr)); | |
8572 | if (psize < asize) | |
8573 | bzero((char *)tmp + psize, asize - psize); | |
8574 | psize = HDR_GET_PSIZE(hdr); | |
8575 | abd_return_buf_copy(cabd, tmp, asize); | |
8576 | to_write = cabd; | |
8577 | } | |
8578 | ||
8579 | if (HDR_ENCRYPTED(hdr)) { | |
8580 | eabd = abd_alloc_for_io(asize, ismd); | |
8581 | ||
8582 | /* | |
8583 | * If the dataset was disowned before the buffer | |
8584 | * made it to this point, the key to re-encrypt | |
8585 | * it won't be available. In this case we simply | |
8586 | * won't write the buffer to the L2ARC. | |
8587 | */ | |
8588 | ret = spa_keystore_lookup_key(spa, hdr->b_crypt_hdr.b_dsobj, | |
8589 | FTAG, &dck); | |
8590 | if (ret != 0) | |
8591 | goto error; | |
8592 | ||
10fa2545 | 8593 | ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key, |
be9a5c35 TC |
8594 | hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt, |
8595 | hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd, | |
8596 | &no_crypt); | |
b5256303 TC |
8597 | if (ret != 0) |
8598 | goto error; | |
8599 | ||
4807c0ba TC |
8600 | if (no_crypt) |
8601 | abd_copy(eabd, to_write, psize); | |
b5256303 TC |
8602 | |
8603 | if (psize != asize) | |
8604 | abd_zero_off(eabd, psize, asize - psize); | |
8605 | ||
8606 | /* assert that the MAC we got here matches the one we saved */ | |
8607 | ASSERT0(bcmp(mac, hdr->b_crypt_hdr.b_mac, ZIO_DATA_MAC_LEN)); | |
8608 | spa_keystore_dsl_key_rele(spa, dck, FTAG); | |
8609 | ||
8610 | if (to_write == cabd) | |
8611 | abd_free(cabd); | |
8612 | ||
8613 | to_write = eabd; | |
8614 | } | |
8615 | ||
8616 | out: | |
8617 | ASSERT3P(to_write, !=, hdr->b_l1hdr.b_pabd); | |
8618 | *abd_out = to_write; | |
8619 | return (0); | |
8620 | ||
8621 | error: | |
8622 | if (dck != NULL) | |
8623 | spa_keystore_dsl_key_rele(spa, dck, FTAG); | |
8624 | if (cabd != NULL) | |
8625 | abd_free(cabd); | |
8626 | if (eabd != NULL) | |
8627 | abd_free(eabd); | |
8628 | ||
8629 | *abd_out = NULL; | |
8630 | return (ret); | |
8631 | } | |
8632 | ||
77f6826b GA |
8633 | static void |
8634 | l2arc_blk_fetch_done(zio_t *zio) | |
8635 | { | |
8636 | l2arc_read_callback_t *cb; | |
8637 | ||
8638 | cb = zio->io_private; | |
8639 | if (cb->l2rcb_abd != NULL) | |
8640 | abd_put(cb->l2rcb_abd); | |
8641 | kmem_free(cb, sizeof (l2arc_read_callback_t)); | |
8642 | } | |
8643 | ||
34dc7c2f BB |
8644 | /* |
8645 | * Find and write ARC buffers to the L2ARC device. | |
8646 | * | |
2a432414 | 8647 | * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid |
34dc7c2f | 8648 | * for reading until they have completed writing. |
3a17a7a9 SK |
8649 | * The headroom_boost is an in-out parameter used to maintain headroom boost |
8650 | * state between calls to this function. | |
8651 | * | |
8652 | * Returns the number of bytes actually written (which may be smaller than | |
77f6826b GA |
8653 | * the delta by which the device hand has changed due to alignment and the |
8654 | * writing of log blocks). | |
34dc7c2f | 8655 | */ |
d164b209 | 8656 | static uint64_t |
d3c2ae1c | 8657 | l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) |
34dc7c2f | 8658 | { |
77f6826b GA |
8659 | arc_buf_hdr_t *hdr, *hdr_prev, *head; |
8660 | uint64_t write_asize, write_psize, write_lsize, headroom; | |
8661 | boolean_t full; | |
8662 | l2arc_write_callback_t *cb = NULL; | |
8663 | zio_t *pio, *wzio; | |
8664 | uint64_t guid = spa_load_guid(spa); | |
34dc7c2f | 8665 | |
d3c2ae1c | 8666 | ASSERT3P(dev->l2ad_vdev, !=, NULL); |
3a17a7a9 | 8667 | |
34dc7c2f | 8668 | pio = NULL; |
01850391 | 8669 | write_lsize = write_asize = write_psize = 0; |
34dc7c2f | 8670 | full = B_FALSE; |
b9541d6b | 8671 | head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); |
d3c2ae1c | 8672 | arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); |
3a17a7a9 | 8673 | |
34dc7c2f BB |
8674 | /* |
8675 | * Copy buffers for L2ARC writing. | |
8676 | */ | |
1c27024e | 8677 | for (int try = 0; try < L2ARC_FEED_TYPES; try++) { |
ca0bf58d | 8678 | multilist_sublist_t *mls = l2arc_sublist_lock(try); |
3a17a7a9 SK |
8679 | uint64_t passed_sz = 0; |
8680 | ||
4aafab91 G |
8681 | VERIFY3P(mls, !=, NULL); |
8682 | ||
b128c09f BB |
8683 | /* |
8684 | * L2ARC fast warmup. | |
8685 | * | |
8686 | * Until the ARC is warm and starts to evict, read from the | |
8687 | * head of the ARC lists rather than the tail. | |
8688 | */ | |
b128c09f | 8689 | if (arc_warm == B_FALSE) |
ca0bf58d | 8690 | hdr = multilist_sublist_head(mls); |
b128c09f | 8691 | else |
ca0bf58d | 8692 | hdr = multilist_sublist_tail(mls); |
b128c09f | 8693 | |
3a17a7a9 | 8694 | headroom = target_sz * l2arc_headroom; |
d3c2ae1c | 8695 | if (zfs_compressed_arc_enabled) |
3a17a7a9 SK |
8696 | headroom = (headroom * l2arc_headroom_boost) / 100; |
8697 | ||
2a432414 | 8698 | for (; hdr; hdr = hdr_prev) { |
3a17a7a9 | 8699 | kmutex_t *hash_lock; |
b5256303 | 8700 | abd_t *to_write = NULL; |
3a17a7a9 | 8701 | |
b128c09f | 8702 | if (arc_warm == B_FALSE) |
ca0bf58d | 8703 | hdr_prev = multilist_sublist_next(mls, hdr); |
b128c09f | 8704 | else |
ca0bf58d | 8705 | hdr_prev = multilist_sublist_prev(mls, hdr); |
34dc7c2f | 8706 | |
2a432414 | 8707 | hash_lock = HDR_LOCK(hdr); |
3a17a7a9 | 8708 | if (!mutex_tryenter(hash_lock)) { |
34dc7c2f BB |
8709 | /* |
8710 | * Skip this buffer rather than waiting. | |
8711 | */ | |
8712 | continue; | |
8713 | } | |
8714 | ||
d3c2ae1c | 8715 | passed_sz += HDR_GET_LSIZE(hdr); |
77f6826b | 8716 | if (l2arc_headroom != 0 && passed_sz > headroom) { |
34dc7c2f BB |
8717 | /* |
8718 | * Searched too far. | |
8719 | */ | |
8720 | mutex_exit(hash_lock); | |
8721 | break; | |
8722 | } | |
8723 | ||
2a432414 | 8724 | if (!l2arc_write_eligible(guid, hdr)) { |
34dc7c2f BB |
8725 | mutex_exit(hash_lock); |
8726 | continue; | |
8727 | } | |
8728 | ||
01850391 AG |
8729 | /* |
8730 | * We rely on the L1 portion of the header below, so | |
8731 | * it's invalid for this header to have been evicted out | |
8732 | * of the ghost cache, prior to being written out. The | |
8733 | * ARC_FLAG_L2_WRITING bit ensures this won't happen. | |
8734 | */ | |
8735 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
8736 | ||
8737 | ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); | |
01850391 | 8738 | ASSERT3U(arc_hdr_size(hdr), >, 0); |
b5256303 TC |
8739 | ASSERT(hdr->b_l1hdr.b_pabd != NULL || |
8740 | HDR_HAS_RABD(hdr)); | |
8741 | uint64_t psize = HDR_GET_PSIZE(hdr); | |
01850391 AG |
8742 | uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, |
8743 | psize); | |
8744 | ||
8745 | if ((write_asize + asize) > target_sz) { | |
34dc7c2f BB |
8746 | full = B_TRUE; |
8747 | mutex_exit(hash_lock); | |
8748 | break; | |
8749 | } | |
8750 | ||
b5256303 TC |
8751 | /* |
8752 | * We rely on the L1 portion of the header below, so | |
8753 | * it's invalid for this header to have been evicted out | |
8754 | * of the ghost cache, prior to being written out. The | |
8755 | * ARC_FLAG_L2_WRITING bit ensures this won't happen. | |
8756 | */ | |
8757 | arc_hdr_set_flags(hdr, ARC_FLAG_L2_WRITING); | |
8758 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
8759 | ||
8760 | ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); | |
8761 | ASSERT(hdr->b_l1hdr.b_pabd != NULL || | |
8762 | HDR_HAS_RABD(hdr)); | |
8763 | ASSERT3U(arc_hdr_size(hdr), >, 0); | |
8764 | ||
8765 | /* | |
8766 | * If this header has b_rabd, we can use this since it | |
8767 | * must always match the data exactly as it exists on | |
8768 | * disk. Otherwise, the L2ARC can normally use the | |
8769 | * hdr's data, but if we're sharing data between the | |
8770 | * hdr and one of its bufs, L2ARC needs its own copy of | |
8771 | * the data so that the ZIO below can't race with the | |
8772 | * buf consumer. To ensure that this copy will be | |
8773 | * available for the lifetime of the ZIO and be cleaned | |
8774 | * up afterwards, we add it to the l2arc_free_on_write | |
8775 | * queue. If we need to apply any transforms to the | |
8776 | * data (compression, encryption) we will also need the | |
8777 | * extra buffer. | |
8778 | */ | |
8779 | if (HDR_HAS_RABD(hdr) && psize == asize) { | |
8780 | to_write = hdr->b_crypt_hdr.b_rabd; | |
8781 | } else if ((HDR_COMPRESSION_ENABLED(hdr) || | |
8782 | HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) && | |
8783 | !HDR_ENCRYPTED(hdr) && !HDR_SHARED_DATA(hdr) && | |
8784 | psize == asize) { | |
8785 | to_write = hdr->b_l1hdr.b_pabd; | |
8786 | } else { | |
8787 | int ret; | |
8788 | arc_buf_contents_t type = arc_buf_type(hdr); | |
8789 | ||
8790 | ret = l2arc_apply_transforms(spa, hdr, asize, | |
8791 | &to_write); | |
8792 | if (ret != 0) { | |
8793 | arc_hdr_clear_flags(hdr, | |
8794 | ARC_FLAG_L2_WRITING); | |
8795 | mutex_exit(hash_lock); | |
8796 | continue; | |
8797 | } | |
8798 | ||
8799 | l2arc_free_abd_on_write(to_write, asize, type); | |
8800 | } | |
8801 | ||
34dc7c2f BB |
8802 | if (pio == NULL) { |
8803 | /* | |
8804 | * Insert a dummy header on the buflist so | |
8805 | * l2arc_write_done() can find where the | |
8806 | * write buffers begin without searching. | |
8807 | */ | |
ca0bf58d | 8808 | mutex_enter(&dev->l2ad_mtx); |
b9541d6b | 8809 | list_insert_head(&dev->l2ad_buflist, head); |
ca0bf58d | 8810 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f | 8811 | |
96c080cb BB |
8812 | cb = kmem_alloc( |
8813 | sizeof (l2arc_write_callback_t), KM_SLEEP); | |
34dc7c2f BB |
8814 | cb->l2wcb_dev = dev; |
8815 | cb->l2wcb_head = head; | |
657fd33b GA |
8816 | /* |
8817 | * Create a list to save allocated abd buffers | |
8818 | * for l2arc_log_blk_commit(). | |
8819 | */ | |
77f6826b GA |
8820 | list_create(&cb->l2wcb_abd_list, |
8821 | sizeof (l2arc_lb_abd_buf_t), | |
8822 | offsetof(l2arc_lb_abd_buf_t, node)); | |
34dc7c2f BB |
8823 | pio = zio_root(spa, l2arc_write_done, cb, |
8824 | ZIO_FLAG_CANFAIL); | |
8825 | } | |
8826 | ||
b9541d6b | 8827 | hdr->b_l2hdr.b_dev = dev; |
b9541d6b | 8828 | hdr->b_l2hdr.b_hits = 0; |
3a17a7a9 | 8829 | |
d3c2ae1c | 8830 | hdr->b_l2hdr.b_daddr = dev->l2ad_hand; |
b5256303 | 8831 | arc_hdr_set_flags(hdr, ARC_FLAG_HAS_L2HDR); |
3a17a7a9 | 8832 | |
ca0bf58d | 8833 | mutex_enter(&dev->l2ad_mtx); |
b9541d6b | 8834 | list_insert_head(&dev->l2ad_buflist, hdr); |
ca0bf58d | 8835 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f | 8836 | |
424fd7c3 | 8837 | (void) zfs_refcount_add_many(&dev->l2ad_alloc, |
b5256303 | 8838 | arc_hdr_size(hdr), hdr); |
3a17a7a9 | 8839 | |
34dc7c2f | 8840 | wzio = zio_write_phys(pio, dev->l2ad_vdev, |
82710e99 | 8841 | hdr->b_l2hdr.b_daddr, asize, to_write, |
d3c2ae1c GW |
8842 | ZIO_CHECKSUM_OFF, NULL, hdr, |
8843 | ZIO_PRIORITY_ASYNC_WRITE, | |
34dc7c2f BB |
8844 | ZIO_FLAG_CANFAIL, B_FALSE); |
8845 | ||
01850391 | 8846 | write_lsize += HDR_GET_LSIZE(hdr); |
34dc7c2f BB |
8847 | DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, |
8848 | zio_t *, wzio); | |
d962d5da | 8849 | |
01850391 AG |
8850 | write_psize += psize; |
8851 | write_asize += asize; | |
d3c2ae1c | 8852 | dev->l2ad_hand += asize; |
7558997d | 8853 | vdev_space_update(dev->l2ad_vdev, asize, 0, 0); |
d3c2ae1c GW |
8854 | |
8855 | mutex_exit(hash_lock); | |
8856 | ||
77f6826b GA |
8857 | /* |
8858 | * Append buf info to current log and commit if full. | |
8859 | * arcstat_l2_{size,asize} kstats are updated | |
8860 | * internally. | |
8861 | */ | |
8862 | if (l2arc_log_blk_insert(dev, hdr)) | |
8863 | l2arc_log_blk_commit(dev, pio, cb); | |
8864 | ||
9cdf7b1f | 8865 | zio_nowait(wzio); |
34dc7c2f | 8866 | } |
d3c2ae1c GW |
8867 | |
8868 | multilist_sublist_unlock(mls); | |
8869 | ||
8870 | if (full == B_TRUE) | |
8871 | break; | |
34dc7c2f | 8872 | } |
34dc7c2f | 8873 | |
d3c2ae1c GW |
8874 | /* No buffers selected for writing? */ |
8875 | if (pio == NULL) { | |
01850391 | 8876 | ASSERT0(write_lsize); |
d3c2ae1c GW |
8877 | ASSERT(!HDR_HAS_L1HDR(head)); |
8878 | kmem_cache_free(hdr_l2only_cache, head); | |
77f6826b GA |
8879 | |
8880 | /* | |
8881 | * Although we did not write any buffers l2ad_evict may | |
8882 | * have advanced. | |
8883 | */ | |
8884 | l2arc_dev_hdr_update(dev); | |
8885 | ||
d3c2ae1c GW |
8886 | return (0); |
8887 | } | |
34dc7c2f | 8888 | |
657fd33b GA |
8889 | if (!dev->l2ad_first) |
8890 | ASSERT3U(dev->l2ad_hand, <=, dev->l2ad_evict); | |
8891 | ||
3a17a7a9 | 8892 | ASSERT3U(write_asize, <=, target_sz); |
34dc7c2f | 8893 | ARCSTAT_BUMP(arcstat_l2_writes_sent); |
01850391 AG |
8894 | ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize); |
8895 | ARCSTAT_INCR(arcstat_l2_lsize, write_lsize); | |
8896 | ARCSTAT_INCR(arcstat_l2_psize, write_psize); | |
34dc7c2f | 8897 | |
77f6826b GA |
8898 | l2arc_dev_hdr_update(dev); |
8899 | ||
d164b209 | 8900 | dev->l2ad_writing = B_TRUE; |
34dc7c2f | 8901 | (void) zio_wait(pio); |
d164b209 BB |
8902 | dev->l2ad_writing = B_FALSE; |
8903 | ||
3a17a7a9 SK |
8904 | return (write_asize); |
8905 | } | |
8906 | ||
34dc7c2f BB |
8907 | /* |
8908 | * This thread feeds the L2ARC at regular intervals. This is the beating | |
8909 | * heart of the L2ARC. | |
8910 | */ | |
867959b5 | 8911 | /* ARGSUSED */ |
34dc7c2f | 8912 | static void |
c25b8f99 | 8913 | l2arc_feed_thread(void *unused) |
34dc7c2f BB |
8914 | { |
8915 | callb_cpr_t cpr; | |
8916 | l2arc_dev_t *dev; | |
8917 | spa_t *spa; | |
d164b209 | 8918 | uint64_t size, wrote; |
428870ff | 8919 | clock_t begin, next = ddi_get_lbolt(); |
40d06e3c | 8920 | fstrans_cookie_t cookie; |
34dc7c2f BB |
8921 | |
8922 | CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); | |
8923 | ||
8924 | mutex_enter(&l2arc_feed_thr_lock); | |
8925 | ||
40d06e3c | 8926 | cookie = spl_fstrans_mark(); |
34dc7c2f | 8927 | while (l2arc_thread_exit == 0) { |
34dc7c2f | 8928 | CALLB_CPR_SAFE_BEGIN(&cpr); |
b64ccd6c | 8929 | (void) cv_timedwait_sig(&l2arc_feed_thr_cv, |
5b63b3eb | 8930 | &l2arc_feed_thr_lock, next); |
34dc7c2f | 8931 | CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); |
428870ff | 8932 | next = ddi_get_lbolt() + hz; |
34dc7c2f BB |
8933 | |
8934 | /* | |
b128c09f | 8935 | * Quick check for L2ARC devices. |
34dc7c2f BB |
8936 | */ |
8937 | mutex_enter(&l2arc_dev_mtx); | |
8938 | if (l2arc_ndev == 0) { | |
8939 | mutex_exit(&l2arc_dev_mtx); | |
8940 | continue; | |
8941 | } | |
b128c09f | 8942 | mutex_exit(&l2arc_dev_mtx); |
428870ff | 8943 | begin = ddi_get_lbolt(); |
34dc7c2f BB |
8944 | |
8945 | /* | |
b128c09f BB |
8946 | * This selects the next l2arc device to write to, and in |
8947 | * doing so the next spa to feed from: dev->l2ad_spa. This | |
8948 | * will return NULL if there are now no l2arc devices or if | |
8949 | * they are all faulted. | |
8950 | * | |
8951 | * If a device is returned, its spa's config lock is also | |
8952 | * held to prevent device removal. l2arc_dev_get_next() | |
8953 | * will grab and release l2arc_dev_mtx. | |
34dc7c2f | 8954 | */ |
b128c09f | 8955 | if ((dev = l2arc_dev_get_next()) == NULL) |
34dc7c2f | 8956 | continue; |
b128c09f BB |
8957 | |
8958 | spa = dev->l2ad_spa; | |
d3c2ae1c | 8959 | ASSERT3P(spa, !=, NULL); |
34dc7c2f | 8960 | |
572e2857 BB |
8961 | /* |
8962 | * If the pool is read-only then force the feed thread to | |
8963 | * sleep a little longer. | |
8964 | */ | |
8965 | if (!spa_writeable(spa)) { | |
8966 | next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; | |
8967 | spa_config_exit(spa, SCL_L2ARC, dev); | |
8968 | continue; | |
8969 | } | |
8970 | ||
34dc7c2f | 8971 | /* |
b128c09f | 8972 | * Avoid contributing to memory pressure. |
34dc7c2f | 8973 | */ |
ca67b33a | 8974 | if (arc_reclaim_needed()) { |
b128c09f BB |
8975 | ARCSTAT_BUMP(arcstat_l2_abort_lowmem); |
8976 | spa_config_exit(spa, SCL_L2ARC, dev); | |
34dc7c2f BB |
8977 | continue; |
8978 | } | |
b128c09f | 8979 | |
34dc7c2f BB |
8980 | ARCSTAT_BUMP(arcstat_l2_feeds); |
8981 | ||
37c22948 | 8982 | size = l2arc_write_size(dev); |
b128c09f | 8983 | |
34dc7c2f BB |
8984 | /* |
8985 | * Evict L2ARC buffers that will be overwritten. | |
8986 | */ | |
b128c09f | 8987 | l2arc_evict(dev, size, B_FALSE); |
34dc7c2f BB |
8988 | |
8989 | /* | |
8990 | * Write ARC buffers. | |
8991 | */ | |
d3c2ae1c | 8992 | wrote = l2arc_write_buffers(spa, dev, size); |
d164b209 BB |
8993 | |
8994 | /* | |
8995 | * Calculate interval between writes. | |
8996 | */ | |
8997 | next = l2arc_write_interval(begin, size, wrote); | |
b128c09f | 8998 | spa_config_exit(spa, SCL_L2ARC, dev); |
34dc7c2f | 8999 | } |
40d06e3c | 9000 | spl_fstrans_unmark(cookie); |
34dc7c2f BB |
9001 | |
9002 | l2arc_thread_exit = 0; | |
9003 | cv_broadcast(&l2arc_feed_thr_cv); | |
9004 | CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ | |
9005 | thread_exit(); | |
9006 | } | |
9007 | ||
b128c09f BB |
9008 | boolean_t |
9009 | l2arc_vdev_present(vdev_t *vd) | |
9010 | { | |
77f6826b GA |
9011 | return (l2arc_vdev_get(vd) != NULL); |
9012 | } | |
9013 | ||
9014 | /* | |
9015 | * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if | |
9016 | * the vdev_t isn't an L2ARC device. | |
9017 | */ | |
9018 | static l2arc_dev_t * | |
9019 | l2arc_vdev_get(vdev_t *vd) | |
9020 | { | |
9021 | l2arc_dev_t *dev; | |
b128c09f BB |
9022 | |
9023 | mutex_enter(&l2arc_dev_mtx); | |
9024 | for (dev = list_head(l2arc_dev_list); dev != NULL; | |
9025 | dev = list_next(l2arc_dev_list, dev)) { | |
9026 | if (dev->l2ad_vdev == vd) | |
9027 | break; | |
9028 | } | |
9029 | mutex_exit(&l2arc_dev_mtx); | |
9030 | ||
77f6826b | 9031 | return (dev); |
b128c09f BB |
9032 | } |
9033 | ||
34dc7c2f BB |
9034 | /* |
9035 | * Add a vdev for use by the L2ARC. By this point the spa has already | |
9036 | * validated the vdev and opened it. | |
9037 | */ | |
9038 | void | |
9babb374 | 9039 | l2arc_add_vdev(spa_t *spa, vdev_t *vd) |
34dc7c2f | 9040 | { |
77f6826b GA |
9041 | l2arc_dev_t *adddev; |
9042 | uint64_t l2dhdr_asize; | |
34dc7c2f | 9043 | |
b128c09f BB |
9044 | ASSERT(!l2arc_vdev_present(vd)); |
9045 | ||
34dc7c2f BB |
9046 | /* |
9047 | * Create a new l2arc device entry. | |
9048 | */ | |
77f6826b | 9049 | adddev = vmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); |
34dc7c2f BB |
9050 | adddev->l2ad_spa = spa; |
9051 | adddev->l2ad_vdev = vd; | |
77f6826b GA |
9052 | /* leave extra size for an l2arc device header */ |
9053 | l2dhdr_asize = adddev->l2ad_dev_hdr_asize = | |
9054 | MAX(sizeof (*adddev->l2ad_dev_hdr), 1 << vd->vdev_ashift); | |
9055 | adddev->l2ad_start = VDEV_LABEL_START_SIZE + l2dhdr_asize; | |
9babb374 | 9056 | adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); |
77f6826b | 9057 | ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end); |
34dc7c2f | 9058 | adddev->l2ad_hand = adddev->l2ad_start; |
77f6826b | 9059 | adddev->l2ad_evict = adddev->l2ad_start; |
34dc7c2f | 9060 | adddev->l2ad_first = B_TRUE; |
d164b209 | 9061 | adddev->l2ad_writing = B_FALSE; |
98f72a53 | 9062 | list_link_init(&adddev->l2ad_node); |
77f6826b | 9063 | adddev->l2ad_dev_hdr = kmem_zalloc(l2dhdr_asize, KM_SLEEP); |
34dc7c2f | 9064 | |
b9541d6b | 9065 | mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f BB |
9066 | /* |
9067 | * This is a list of all ARC buffers that are still valid on the | |
9068 | * device. | |
9069 | */ | |
b9541d6b CW |
9070 | list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), |
9071 | offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); | |
34dc7c2f | 9072 | |
77f6826b GA |
9073 | /* |
9074 | * This is a list of pointers to log blocks that are still present | |
9075 | * on the device. | |
9076 | */ | |
9077 | list_create(&adddev->l2ad_lbptr_list, sizeof (l2arc_lb_ptr_buf_t), | |
9078 | offsetof(l2arc_lb_ptr_buf_t, node)); | |
9079 | ||
428870ff | 9080 | vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); |
424fd7c3 | 9081 | zfs_refcount_create(&adddev->l2ad_alloc); |
657fd33b GA |
9082 | zfs_refcount_create(&adddev->l2ad_lb_asize); |
9083 | zfs_refcount_create(&adddev->l2ad_lb_count); | |
34dc7c2f BB |
9084 | |
9085 | /* | |
9086 | * Add device to global list | |
9087 | */ | |
9088 | mutex_enter(&l2arc_dev_mtx); | |
9089 | list_insert_head(l2arc_dev_list, adddev); | |
9090 | atomic_inc_64(&l2arc_ndev); | |
9091 | mutex_exit(&l2arc_dev_mtx); | |
77f6826b GA |
9092 | |
9093 | /* | |
9094 | * Decide if vdev is eligible for L2ARC rebuild | |
9095 | */ | |
9096 | l2arc_rebuild_vdev(adddev->l2ad_vdev, B_FALSE); | |
9097 | } | |
9098 | ||
9099 | void | |
9100 | l2arc_rebuild_vdev(vdev_t *vd, boolean_t reopen) | |
9101 | { | |
9102 | l2arc_dev_t *dev = NULL; | |
9103 | l2arc_dev_hdr_phys_t *l2dhdr; | |
9104 | uint64_t l2dhdr_asize; | |
9105 | spa_t *spa; | |
9106 | int err; | |
657fd33b | 9107 | boolean_t l2dhdr_valid = B_TRUE; |
77f6826b GA |
9108 | |
9109 | dev = l2arc_vdev_get(vd); | |
9110 | ASSERT3P(dev, !=, NULL); | |
9111 | spa = dev->l2ad_spa; | |
9112 | l2dhdr = dev->l2ad_dev_hdr; | |
9113 | l2dhdr_asize = dev->l2ad_dev_hdr_asize; | |
9114 | ||
9115 | /* | |
9116 | * The L2ARC has to hold at least the payload of one log block for | |
9117 | * them to be restored (persistent L2ARC). The payload of a log block | |
9118 | * depends on the amount of its log entries. We always write log blocks | |
9119 | * with 1022 entries. How many of them are committed or restored depends | |
9120 | * on the size of the L2ARC device. Thus the maximum payload of | |
9121 | * one log block is 1022 * SPA_MAXBLOCKSIZE = 16GB. If the L2ARC device | |
9122 | * is less than that, we reduce the amount of committed and restored | |
9123 | * log entries per block so as to enable persistence. | |
9124 | */ | |
9125 | if (dev->l2ad_end < l2arc_rebuild_blocks_min_l2size) { | |
9126 | dev->l2ad_log_entries = 0; | |
9127 | } else { | |
9128 | dev->l2ad_log_entries = MIN((dev->l2ad_end - | |
9129 | dev->l2ad_start) >> SPA_MAXBLOCKSHIFT, | |
9130 | L2ARC_LOG_BLK_MAX_ENTRIES); | |
9131 | } | |
9132 | ||
9133 | /* | |
9134 | * Read the device header, if an error is returned do not rebuild L2ARC. | |
9135 | */ | |
9136 | if ((err = l2arc_dev_hdr_read(dev)) != 0) | |
657fd33b | 9137 | l2dhdr_valid = B_FALSE; |
77f6826b | 9138 | |
657fd33b | 9139 | if (l2dhdr_valid && dev->l2ad_log_entries > 0) { |
77f6826b GA |
9140 | /* |
9141 | * If we are onlining a cache device (vdev_reopen) that was | |
9142 | * still present (l2arc_vdev_present()) and rebuild is enabled, | |
9143 | * we should evict all ARC buffers and pointers to log blocks | |
9144 | * and reclaim their space before restoring its contents to | |
9145 | * L2ARC. | |
9146 | */ | |
9147 | if (reopen) { | |
9148 | if (!l2arc_rebuild_enabled) { | |
9149 | return; | |
9150 | } else { | |
9151 | l2arc_evict(dev, 0, B_TRUE); | |
9152 | /* start a new log block */ | |
9153 | dev->l2ad_log_ent_idx = 0; | |
9154 | dev->l2ad_log_blk_payload_asize = 0; | |
9155 | dev->l2ad_log_blk_payload_start = 0; | |
9156 | } | |
9157 | } | |
9158 | /* | |
9159 | * Just mark the device as pending for a rebuild. We won't | |
9160 | * be starting a rebuild in line here as it would block pool | |
9161 | * import. Instead spa_load_impl will hand that off to an | |
9162 | * async task which will call l2arc_spa_rebuild_start. | |
9163 | */ | |
9164 | dev->l2ad_rebuild = B_TRUE; | |
657fd33b | 9165 | } else if (spa_writeable(spa)) { |
77f6826b | 9166 | /* |
657fd33b GA |
9167 | * In this case create a new header. We zero out the memory |
9168 | * holding the header to reset dh_start_lbps. | |
77f6826b GA |
9169 | */ |
9170 | bzero(l2dhdr, l2dhdr_asize); | |
9171 | l2arc_dev_hdr_update(dev); | |
9172 | } | |
34dc7c2f BB |
9173 | } |
9174 | ||
9175 | /* | |
9176 | * Remove a vdev from the L2ARC. | |
9177 | */ | |
9178 | void | |
9179 | l2arc_remove_vdev(vdev_t *vd) | |
9180 | { | |
77f6826b | 9181 | l2arc_dev_t *remdev = NULL; |
34dc7c2f | 9182 | |
34dc7c2f BB |
9183 | /* |
9184 | * Find the device by vdev | |
9185 | */ | |
77f6826b | 9186 | remdev = l2arc_vdev_get(vd); |
d3c2ae1c | 9187 | ASSERT3P(remdev, !=, NULL); |
34dc7c2f | 9188 | |
77f6826b GA |
9189 | /* |
9190 | * Cancel any ongoing or scheduled rebuild. | |
9191 | */ | |
9192 | mutex_enter(&l2arc_rebuild_thr_lock); | |
9193 | if (remdev->l2ad_rebuild_began == B_TRUE) { | |
9194 | remdev->l2ad_rebuild_cancel = B_TRUE; | |
9195 | while (remdev->l2ad_rebuild == B_TRUE) | |
9196 | cv_wait(&l2arc_rebuild_thr_cv, &l2arc_rebuild_thr_lock); | |
9197 | } | |
9198 | mutex_exit(&l2arc_rebuild_thr_lock); | |
9199 | ||
34dc7c2f BB |
9200 | /* |
9201 | * Remove device from global list | |
9202 | */ | |
77f6826b | 9203 | mutex_enter(&l2arc_dev_mtx); |
34dc7c2f BB |
9204 | list_remove(l2arc_dev_list, remdev); |
9205 | l2arc_dev_last = NULL; /* may have been invalidated */ | |
b128c09f BB |
9206 | atomic_dec_64(&l2arc_ndev); |
9207 | mutex_exit(&l2arc_dev_mtx); | |
34dc7c2f BB |
9208 | |
9209 | /* | |
9210 | * Clear all buflists and ARC references. L2ARC device flush. | |
9211 | */ | |
9212 | l2arc_evict(remdev, 0, B_TRUE); | |
b9541d6b | 9213 | list_destroy(&remdev->l2ad_buflist); |
77f6826b GA |
9214 | ASSERT(list_is_empty(&remdev->l2ad_lbptr_list)); |
9215 | list_destroy(&remdev->l2ad_lbptr_list); | |
b9541d6b | 9216 | mutex_destroy(&remdev->l2ad_mtx); |
424fd7c3 | 9217 | zfs_refcount_destroy(&remdev->l2ad_alloc); |
657fd33b GA |
9218 | zfs_refcount_destroy(&remdev->l2ad_lb_asize); |
9219 | zfs_refcount_destroy(&remdev->l2ad_lb_count); | |
77f6826b GA |
9220 | kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize); |
9221 | vmem_free(remdev, sizeof (l2arc_dev_t)); | |
34dc7c2f BB |
9222 | } |
9223 | ||
9224 | void | |
b128c09f | 9225 | l2arc_init(void) |
34dc7c2f BB |
9226 | { |
9227 | l2arc_thread_exit = 0; | |
9228 | l2arc_ndev = 0; | |
9229 | l2arc_writes_sent = 0; | |
9230 | l2arc_writes_done = 0; | |
9231 | ||
9232 | mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); | |
9233 | cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); | |
77f6826b GA |
9234 | mutex_init(&l2arc_rebuild_thr_lock, NULL, MUTEX_DEFAULT, NULL); |
9235 | cv_init(&l2arc_rebuild_thr_cv, NULL, CV_DEFAULT, NULL); | |
34dc7c2f | 9236 | mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f BB |
9237 | mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); |
9238 | ||
9239 | l2arc_dev_list = &L2ARC_dev_list; | |
9240 | l2arc_free_on_write = &L2ARC_free_on_write; | |
9241 | list_create(l2arc_dev_list, sizeof (l2arc_dev_t), | |
9242 | offsetof(l2arc_dev_t, l2ad_node)); | |
9243 | list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), | |
9244 | offsetof(l2arc_data_free_t, l2df_list_node)); | |
34dc7c2f BB |
9245 | } |
9246 | ||
9247 | void | |
b128c09f | 9248 | l2arc_fini(void) |
34dc7c2f | 9249 | { |
b128c09f BB |
9250 | /* |
9251 | * This is called from dmu_fini(), which is called from spa_fini(); | |
9252 | * Because of this, we can assume that all l2arc devices have | |
9253 | * already been removed when the pools themselves were removed. | |
9254 | */ | |
9255 | ||
9256 | l2arc_do_free_on_write(); | |
34dc7c2f BB |
9257 | |
9258 | mutex_destroy(&l2arc_feed_thr_lock); | |
9259 | cv_destroy(&l2arc_feed_thr_cv); | |
77f6826b GA |
9260 | mutex_destroy(&l2arc_rebuild_thr_lock); |
9261 | cv_destroy(&l2arc_rebuild_thr_cv); | |
34dc7c2f | 9262 | mutex_destroy(&l2arc_dev_mtx); |
34dc7c2f BB |
9263 | mutex_destroy(&l2arc_free_on_write_mtx); |
9264 | ||
9265 | list_destroy(l2arc_dev_list); | |
9266 | list_destroy(l2arc_free_on_write); | |
9267 | } | |
b128c09f BB |
9268 | |
9269 | void | |
9270 | l2arc_start(void) | |
9271 | { | |
da92d5cb | 9272 | if (!(spa_mode_global & SPA_MODE_WRITE)) |
b128c09f BB |
9273 | return; |
9274 | ||
9275 | (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, | |
1229323d | 9276 | TS_RUN, defclsyspri); |
b128c09f BB |
9277 | } |
9278 | ||
9279 | void | |
9280 | l2arc_stop(void) | |
9281 | { | |
da92d5cb | 9282 | if (!(spa_mode_global & SPA_MODE_WRITE)) |
b128c09f BB |
9283 | return; |
9284 | ||
9285 | mutex_enter(&l2arc_feed_thr_lock); | |
9286 | cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ | |
9287 | l2arc_thread_exit = 1; | |
9288 | while (l2arc_thread_exit != 0) | |
9289 | cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); | |
9290 | mutex_exit(&l2arc_feed_thr_lock); | |
9291 | } | |
c28b2279 | 9292 | |
77f6826b GA |
9293 | /* |
9294 | * Punches out rebuild threads for the L2ARC devices in a spa. This should | |
9295 | * be called after pool import from the spa async thread, since starting | |
9296 | * these threads directly from spa_import() will make them part of the | |
9297 | * "zpool import" context and delay process exit (and thus pool import). | |
9298 | */ | |
9299 | void | |
9300 | l2arc_spa_rebuild_start(spa_t *spa) | |
9301 | { | |
9302 | ASSERT(MUTEX_HELD(&spa_namespace_lock)); | |
9303 | ||
9304 | /* | |
9305 | * Locate the spa's l2arc devices and kick off rebuild threads. | |
9306 | */ | |
9307 | for (int i = 0; i < spa->spa_l2cache.sav_count; i++) { | |
9308 | l2arc_dev_t *dev = | |
9309 | l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]); | |
9310 | if (dev == NULL) { | |
9311 | /* Don't attempt a rebuild if the vdev is UNAVAIL */ | |
9312 | continue; | |
9313 | } | |
9314 | mutex_enter(&l2arc_rebuild_thr_lock); | |
9315 | if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) { | |
9316 | dev->l2ad_rebuild_began = B_TRUE; | |
9317 | (void) thread_create(NULL, 0, | |
9318 | (void (*)(void *))l2arc_dev_rebuild_start, | |
9319 | dev, 0, &p0, TS_RUN, minclsyspri); | |
9320 | } | |
9321 | mutex_exit(&l2arc_rebuild_thr_lock); | |
9322 | } | |
9323 | } | |
9324 | ||
9325 | /* | |
9326 | * Main entry point for L2ARC rebuilding. | |
9327 | */ | |
9328 | static void | |
9329 | l2arc_dev_rebuild_start(l2arc_dev_t *dev) | |
9330 | { | |
9331 | VERIFY(!dev->l2ad_rebuild_cancel); | |
9332 | VERIFY(dev->l2ad_rebuild); | |
9333 | (void) l2arc_rebuild(dev); | |
9334 | mutex_enter(&l2arc_rebuild_thr_lock); | |
9335 | dev->l2ad_rebuild_began = B_FALSE; | |
9336 | dev->l2ad_rebuild = B_FALSE; | |
9337 | mutex_exit(&l2arc_rebuild_thr_lock); | |
9338 | ||
9339 | thread_exit(); | |
9340 | } | |
9341 | ||
9342 | /* | |
9343 | * This function implements the actual L2ARC metadata rebuild. It: | |
9344 | * starts reading the log block chain and restores each block's contents | |
9345 | * to memory (reconstructing arc_buf_hdr_t's). | |
9346 | * | |
9347 | * Operation stops under any of the following conditions: | |
9348 | * | |
9349 | * 1) We reach the end of the log block chain. | |
9350 | * 2) We encounter *any* error condition (cksum errors, io errors) | |
9351 | */ | |
9352 | static int | |
9353 | l2arc_rebuild(l2arc_dev_t *dev) | |
9354 | { | |
9355 | vdev_t *vd = dev->l2ad_vdev; | |
9356 | spa_t *spa = vd->vdev_spa; | |
657fd33b | 9357 | int err = 0; |
77f6826b GA |
9358 | l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr; |
9359 | l2arc_log_blk_phys_t *this_lb, *next_lb; | |
9360 | zio_t *this_io = NULL, *next_io = NULL; | |
9361 | l2arc_log_blkptr_t lbps[2]; | |
9362 | l2arc_lb_ptr_buf_t *lb_ptr_buf; | |
9363 | boolean_t lock_held; | |
9364 | ||
9365 | this_lb = vmem_zalloc(sizeof (*this_lb), KM_SLEEP); | |
9366 | next_lb = vmem_zalloc(sizeof (*next_lb), KM_SLEEP); | |
9367 | ||
9368 | /* | |
9369 | * We prevent device removal while issuing reads to the device, | |
9370 | * then during the rebuilding phases we drop this lock again so | |
9371 | * that a spa_unload or device remove can be initiated - this is | |
9372 | * safe, because the spa will signal us to stop before removing | |
9373 | * our device and wait for us to stop. | |
9374 | */ | |
9375 | spa_config_enter(spa, SCL_L2ARC, vd, RW_READER); | |
9376 | lock_held = B_TRUE; | |
9377 | ||
9378 | /* | |
9379 | * Retrieve the persistent L2ARC device state. | |
657fd33b | 9380 | * L2BLK_GET_PSIZE returns aligned size for log blocks. |
77f6826b GA |
9381 | */ |
9382 | dev->l2ad_evict = MAX(l2dhdr->dh_evict, dev->l2ad_start); | |
9383 | dev->l2ad_hand = MAX(l2dhdr->dh_start_lbps[0].lbp_daddr + | |
9384 | L2BLK_GET_PSIZE((&l2dhdr->dh_start_lbps[0])->lbp_prop), | |
9385 | dev->l2ad_start); | |
9386 | dev->l2ad_first = !!(l2dhdr->dh_flags & L2ARC_DEV_HDR_EVICT_FIRST); | |
9387 | ||
9388 | /* | |
9389 | * In case the zfs module parameter l2arc_rebuild_enabled is false | |
9390 | * we do not start the rebuild process. | |
9391 | */ | |
9392 | if (!l2arc_rebuild_enabled) | |
9393 | goto out; | |
9394 | ||
9395 | /* Prepare the rebuild process */ | |
9396 | bcopy(l2dhdr->dh_start_lbps, lbps, sizeof (lbps)); | |
9397 | ||
9398 | /* Start the rebuild process */ | |
9399 | for (;;) { | |
9400 | if (!l2arc_log_blkptr_valid(dev, &lbps[0])) | |
9401 | break; | |
9402 | ||
9403 | if ((err = l2arc_log_blk_read(dev, &lbps[0], &lbps[1], | |
9404 | this_lb, next_lb, this_io, &next_io)) != 0) | |
9405 | goto out; | |
9406 | ||
9407 | /* | |
9408 | * Our memory pressure valve. If the system is running low | |
9409 | * on memory, rather than swamping memory with new ARC buf | |
9410 | * hdrs, we opt not to rebuild the L2ARC. At this point, | |
9411 | * however, we have already set up our L2ARC dev to chain in | |
9412 | * new metadata log blocks, so the user may choose to offline/ | |
9413 | * online the L2ARC dev at a later time (or re-import the pool) | |
9414 | * to reconstruct it (when there's less memory pressure). | |
9415 | */ | |
9416 | if (arc_reclaim_needed()) { | |
9417 | ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem); | |
9418 | cmn_err(CE_NOTE, "System running low on memory, " | |
9419 | "aborting L2ARC rebuild."); | |
9420 | err = SET_ERROR(ENOMEM); | |
9421 | goto out; | |
9422 | } | |
9423 | ||
9424 | spa_config_exit(spa, SCL_L2ARC, vd); | |
9425 | lock_held = B_FALSE; | |
9426 | ||
9427 | /* | |
9428 | * Now that we know that the next_lb checks out alright, we | |
9429 | * can start reconstruction from this log block. | |
657fd33b | 9430 | * L2BLK_GET_PSIZE returns aligned size for log blocks. |
77f6826b | 9431 | */ |
657fd33b GA |
9432 | uint64_t asize = L2BLK_GET_PSIZE((&lbps[0])->lbp_prop); |
9433 | l2arc_log_blk_restore(dev, this_lb, asize, lbps[0].lbp_daddr); | |
77f6826b GA |
9434 | |
9435 | /* | |
9436 | * log block restored, include its pointer in the list of | |
9437 | * pointers to log blocks present in the L2ARC device. | |
9438 | */ | |
9439 | lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP); | |
9440 | lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), | |
9441 | KM_SLEEP); | |
9442 | bcopy(&lbps[0], lb_ptr_buf->lb_ptr, | |
9443 | sizeof (l2arc_log_blkptr_t)); | |
9444 | mutex_enter(&dev->l2ad_mtx); | |
9445 | list_insert_tail(&dev->l2ad_lbptr_list, lb_ptr_buf); | |
657fd33b GA |
9446 | ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize); |
9447 | ARCSTAT_BUMP(arcstat_l2_log_blk_count); | |
9448 | zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf); | |
9449 | zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf); | |
77f6826b | 9450 | mutex_exit(&dev->l2ad_mtx); |
657fd33b | 9451 | vdev_space_update(vd, asize, 0, 0); |
77f6826b GA |
9452 | |
9453 | /* | |
9454 | * Protection against loops of log blocks: | |
9455 | * | |
9456 | * l2ad_hand l2ad_evict | |
9457 | * V V | |
9458 | * l2ad_start |=======================================| l2ad_end | |
9459 | * -----|||----|||---|||----||| | |
9460 | * (3) (2) (1) (0) | |
9461 | * ---|||---|||----|||---||| | |
9462 | * (7) (6) (5) (4) | |
9463 | * | |
9464 | * In this situation the pointer of log block (4) passes | |
9465 | * l2arc_log_blkptr_valid() but the log block should not be | |
9466 | * restored as it is overwritten by the payload of log block | |
9467 | * (0). Only log blocks (0)-(3) should be restored. We check | |
657fd33b GA |
9468 | * whether l2ad_evict lies in between the payload starting |
9469 | * offset of the next log block (lbps[1].lbp_payload_start) | |
9470 | * and the payload starting offset of the present log block | |
9471 | * (lbps[0].lbp_payload_start). If true and this isn't the | |
9472 | * first pass, we are looping from the beginning and we should | |
9473 | * stop. | |
77f6826b | 9474 | */ |
657fd33b GA |
9475 | if (l2arc_range_check_overlap(lbps[1].lbp_payload_start, |
9476 | lbps[0].lbp_payload_start, dev->l2ad_evict) && | |
9477 | !dev->l2ad_first) | |
77f6826b GA |
9478 | goto out; |
9479 | ||
9480 | for (;;) { | |
9481 | mutex_enter(&l2arc_rebuild_thr_lock); | |
9482 | if (dev->l2ad_rebuild_cancel) { | |
9483 | dev->l2ad_rebuild = B_FALSE; | |
9484 | cv_signal(&l2arc_rebuild_thr_cv); | |
9485 | mutex_exit(&l2arc_rebuild_thr_lock); | |
9486 | err = SET_ERROR(ECANCELED); | |
9487 | goto out; | |
9488 | } | |
9489 | mutex_exit(&l2arc_rebuild_thr_lock); | |
9490 | if (spa_config_tryenter(spa, SCL_L2ARC, vd, | |
9491 | RW_READER)) { | |
9492 | lock_held = B_TRUE; | |
9493 | break; | |
9494 | } | |
9495 | /* | |
9496 | * L2ARC config lock held by somebody in writer, | |
9497 | * possibly due to them trying to remove us. They'll | |
9498 | * likely to want us to shut down, so after a little | |
9499 | * delay, we check l2ad_rebuild_cancel and retry | |
9500 | * the lock again. | |
9501 | */ | |
9502 | delay(1); | |
9503 | } | |
9504 | ||
9505 | /* | |
9506 | * Continue with the next log block. | |
9507 | */ | |
9508 | lbps[0] = lbps[1]; | |
9509 | lbps[1] = this_lb->lb_prev_lbp; | |
9510 | PTR_SWAP(this_lb, next_lb); | |
9511 | this_io = next_io; | |
9512 | next_io = NULL; | |
9513 | } | |
9514 | ||
9515 | if (this_io != NULL) | |
9516 | l2arc_log_blk_fetch_abort(this_io); | |
9517 | out: | |
9518 | if (next_io != NULL) | |
9519 | l2arc_log_blk_fetch_abort(next_io); | |
9520 | vmem_free(this_lb, sizeof (*this_lb)); | |
9521 | vmem_free(next_lb, sizeof (*next_lb)); | |
9522 | ||
9523 | if (!l2arc_rebuild_enabled) { | |
657fd33b GA |
9524 | spa_history_log_internal(spa, "L2ARC rebuild", NULL, |
9525 | "disabled"); | |
9526 | } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) > 0) { | |
77f6826b | 9527 | ARCSTAT_BUMP(arcstat_l2_rebuild_success); |
657fd33b GA |
9528 | spa_history_log_internal(spa, "L2ARC rebuild", NULL, |
9529 | "successful, restored %llu blocks", | |
9530 | (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count)); | |
9531 | } else if (err == 0 && zfs_refcount_count(&dev->l2ad_lb_count) == 0) { | |
9532 | /* | |
9533 | * No error but also nothing restored, meaning the lbps array | |
9534 | * in the device header points to invalid/non-present log | |
9535 | * blocks. Reset the header. | |
9536 | */ | |
9537 | spa_history_log_internal(spa, "L2ARC rebuild", NULL, | |
9538 | "no valid log blocks"); | |
9539 | bzero(l2dhdr, dev->l2ad_dev_hdr_asize); | |
9540 | l2arc_dev_hdr_update(dev); | |
77f6826b | 9541 | } else if (err != 0) { |
657fd33b GA |
9542 | spa_history_log_internal(spa, "L2ARC rebuild", NULL, |
9543 | "aborted, restored %llu blocks", | |
9544 | (u_longlong_t)zfs_refcount_count(&dev->l2ad_lb_count)); | |
77f6826b GA |
9545 | } |
9546 | ||
9547 | if (lock_held) | |
9548 | spa_config_exit(spa, SCL_L2ARC, vd); | |
9549 | ||
9550 | return (err); | |
9551 | } | |
9552 | ||
9553 | /* | |
9554 | * Attempts to read the device header on the provided L2ARC device and writes | |
9555 | * it to `hdr'. On success, this function returns 0, otherwise the appropriate | |
9556 | * error code is returned. | |
9557 | */ | |
9558 | static int | |
9559 | l2arc_dev_hdr_read(l2arc_dev_t *dev) | |
9560 | { | |
9561 | int err; | |
9562 | uint64_t guid; | |
9563 | l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr; | |
9564 | const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize; | |
9565 | abd_t *abd; | |
9566 | ||
9567 | guid = spa_guid(dev->l2ad_vdev->vdev_spa); | |
9568 | ||
9569 | abd = abd_get_from_buf(l2dhdr, l2dhdr_asize); | |
9570 | ||
9571 | err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev, | |
9572 | VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, | |
9573 | ZIO_CHECKSUM_LABEL, NULL, NULL, ZIO_PRIORITY_ASYNC_READ, | |
9574 | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | | |
9575 | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY | | |
9576 | ZIO_FLAG_SPECULATIVE, B_FALSE)); | |
9577 | ||
9578 | abd_put(abd); | |
9579 | ||
9580 | if (err != 0) { | |
9581 | ARCSTAT_BUMP(arcstat_l2_rebuild_abort_dh_errors); | |
9582 | zfs_dbgmsg("L2ARC IO error (%d) while reading device header, " | |
9583 | "vdev guid: %llu", err, dev->l2ad_vdev->vdev_guid); | |
9584 | return (err); | |
9585 | } | |
9586 | ||
9587 | if (l2dhdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC)) | |
9588 | byteswap_uint64_array(l2dhdr, sizeof (*l2dhdr)); | |
9589 | ||
9590 | if (l2dhdr->dh_magic != L2ARC_DEV_HDR_MAGIC || | |
9591 | l2dhdr->dh_spa_guid != guid || | |
9592 | l2dhdr->dh_vdev_guid != dev->l2ad_vdev->vdev_guid || | |
9593 | l2dhdr->dh_version != L2ARC_PERSISTENT_VERSION || | |
657fd33b | 9594 | l2dhdr->dh_log_entries != dev->l2ad_log_entries || |
77f6826b GA |
9595 | l2dhdr->dh_end != dev->l2ad_end || |
9596 | !l2arc_range_check_overlap(dev->l2ad_start, dev->l2ad_end, | |
9597 | l2dhdr->dh_evict)) { | |
9598 | /* | |
9599 | * Attempt to rebuild a device containing no actual dev hdr | |
9600 | * or containing a header from some other pool or from another | |
9601 | * version of persistent L2ARC. | |
9602 | */ | |
9603 | ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported); | |
9604 | return (SET_ERROR(ENOTSUP)); | |
9605 | } | |
9606 | ||
9607 | return (0); | |
9608 | } | |
9609 | ||
9610 | /* | |
9611 | * Reads L2ARC log blocks from storage and validates their contents. | |
9612 | * | |
9613 | * This function implements a simple fetcher to make sure that while | |
9614 | * we're processing one buffer the L2ARC is already fetching the next | |
9615 | * one in the chain. | |
9616 | * | |
9617 | * The arguments this_lp and next_lp point to the current and next log block | |
9618 | * address in the block chain. Similarly, this_lb and next_lb hold the | |
9619 | * l2arc_log_blk_phys_t's of the current and next L2ARC blk. | |
9620 | * | |
9621 | * The `this_io' and `next_io' arguments are used for block fetching. | |
9622 | * When issuing the first blk IO during rebuild, you should pass NULL for | |
9623 | * `this_io'. This function will then issue a sync IO to read the block and | |
9624 | * also issue an async IO to fetch the next block in the block chain. The | |
9625 | * fetched IO is returned in `next_io'. On subsequent calls to this | |
9626 | * function, pass the value returned in `next_io' from the previous call | |
9627 | * as `this_io' and a fresh `next_io' pointer to hold the next fetch IO. | |
9628 | * Prior to the call, you should initialize your `next_io' pointer to be | |
9629 | * NULL. If no fetch IO was issued, the pointer is left set at NULL. | |
9630 | * | |
9631 | * On success, this function returns 0, otherwise it returns an appropriate | |
9632 | * error code. On error the fetching IO is aborted and cleared before | |
9633 | * returning from this function. Therefore, if we return `success', the | |
9634 | * caller can assume that we have taken care of cleanup of fetch IOs. | |
9635 | */ | |
9636 | static int | |
9637 | l2arc_log_blk_read(l2arc_dev_t *dev, | |
9638 | const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp, | |
9639 | l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb, | |
9640 | zio_t *this_io, zio_t **next_io) | |
9641 | { | |
9642 | int err = 0; | |
9643 | zio_cksum_t cksum; | |
9644 | abd_t *abd = NULL; | |
657fd33b | 9645 | uint64_t asize; |
77f6826b GA |
9646 | |
9647 | ASSERT(this_lbp != NULL && next_lbp != NULL); | |
9648 | ASSERT(this_lb != NULL && next_lb != NULL); | |
9649 | ASSERT(next_io != NULL && *next_io == NULL); | |
9650 | ASSERT(l2arc_log_blkptr_valid(dev, this_lbp)); | |
9651 | ||
9652 | /* | |
9653 | * Check to see if we have issued the IO for this log block in a | |
9654 | * previous run. If not, this is the first call, so issue it now. | |
9655 | */ | |
9656 | if (this_io == NULL) { | |
9657 | this_io = l2arc_log_blk_fetch(dev->l2ad_vdev, this_lbp, | |
9658 | this_lb); | |
9659 | } | |
9660 | ||
9661 | /* | |
9662 | * Peek to see if we can start issuing the next IO immediately. | |
9663 | */ | |
9664 | if (l2arc_log_blkptr_valid(dev, next_lbp)) { | |
9665 | /* | |
9666 | * Start issuing IO for the next log block early - this | |
9667 | * should help keep the L2ARC device busy while we | |
9668 | * decompress and restore this log block. | |
9669 | */ | |
9670 | *next_io = l2arc_log_blk_fetch(dev->l2ad_vdev, next_lbp, | |
9671 | next_lb); | |
9672 | } | |
9673 | ||
9674 | /* Wait for the IO to read this log block to complete */ | |
9675 | if ((err = zio_wait(this_io)) != 0) { | |
9676 | ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors); | |
9677 | zfs_dbgmsg("L2ARC IO error (%d) while reading log block, " | |
9678 | "offset: %llu, vdev guid: %llu", err, this_lbp->lbp_daddr, | |
9679 | dev->l2ad_vdev->vdev_guid); | |
9680 | goto cleanup; | |
9681 | } | |
9682 | ||
657fd33b GA |
9683 | /* |
9684 | * Make sure the buffer checks out. | |
9685 | * L2BLK_GET_PSIZE returns aligned size for log blocks. | |
9686 | */ | |
9687 | asize = L2BLK_GET_PSIZE((this_lbp)->lbp_prop); | |
9688 | fletcher_4_native(this_lb, asize, NULL, &cksum); | |
77f6826b GA |
9689 | if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) { |
9690 | ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_lb_errors); | |
9691 | zfs_dbgmsg("L2ARC log block cksum failed, offset: %llu, " | |
9692 | "vdev guid: %llu, l2ad_hand: %llu, l2ad_evict: %llu", | |
9693 | this_lbp->lbp_daddr, dev->l2ad_vdev->vdev_guid, | |
9694 | dev->l2ad_hand, dev->l2ad_evict); | |
9695 | err = SET_ERROR(ECKSUM); | |
9696 | goto cleanup; | |
9697 | } | |
9698 | ||
9699 | /* Now we can take our time decoding this buffer */ | |
9700 | switch (L2BLK_GET_COMPRESS((this_lbp)->lbp_prop)) { | |
9701 | case ZIO_COMPRESS_OFF: | |
9702 | break; | |
9703 | case ZIO_COMPRESS_LZ4: | |
657fd33b GA |
9704 | abd = abd_alloc_for_io(asize, B_TRUE); |
9705 | abd_copy_from_buf_off(abd, this_lb, 0, asize); | |
77f6826b GA |
9706 | if ((err = zio_decompress_data( |
9707 | L2BLK_GET_COMPRESS((this_lbp)->lbp_prop), | |
657fd33b | 9708 | abd, this_lb, asize, sizeof (*this_lb))) != 0) { |
77f6826b GA |
9709 | err = SET_ERROR(EINVAL); |
9710 | goto cleanup; | |
9711 | } | |
9712 | break; | |
9713 | default: | |
9714 | err = SET_ERROR(EINVAL); | |
9715 | goto cleanup; | |
9716 | } | |
9717 | if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC)) | |
9718 | byteswap_uint64_array(this_lb, sizeof (*this_lb)); | |
9719 | if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) { | |
9720 | err = SET_ERROR(EINVAL); | |
9721 | goto cleanup; | |
9722 | } | |
9723 | cleanup: | |
9724 | /* Abort an in-flight fetch I/O in case of error */ | |
9725 | if (err != 0 && *next_io != NULL) { | |
9726 | l2arc_log_blk_fetch_abort(*next_io); | |
9727 | *next_io = NULL; | |
9728 | } | |
9729 | if (abd != NULL) | |
9730 | abd_free(abd); | |
9731 | return (err); | |
9732 | } | |
9733 | ||
9734 | /* | |
9735 | * Restores the payload of a log block to ARC. This creates empty ARC hdr | |
9736 | * entries which only contain an l2arc hdr, essentially restoring the | |
9737 | * buffers to their L2ARC evicted state. This function also updates space | |
9738 | * usage on the L2ARC vdev to make sure it tracks restored buffers. | |
9739 | */ | |
9740 | static void | |
9741 | l2arc_log_blk_restore(l2arc_dev_t *dev, const l2arc_log_blk_phys_t *lb, | |
657fd33b | 9742 | uint64_t lb_asize, uint64_t lb_daddr) |
77f6826b | 9743 | { |
657fd33b GA |
9744 | uint64_t size = 0, asize = 0; |
9745 | uint64_t log_entries = dev->l2ad_log_entries; | |
77f6826b GA |
9746 | |
9747 | for (int i = log_entries - 1; i >= 0; i--) { | |
9748 | /* | |
9749 | * Restore goes in the reverse temporal direction to preserve | |
9750 | * correct temporal ordering of buffers in the l2ad_buflist. | |
9751 | * l2arc_hdr_restore also does a list_insert_tail instead of | |
9752 | * list_insert_head on the l2ad_buflist: | |
9753 | * | |
9754 | * LIST l2ad_buflist LIST | |
9755 | * HEAD <------ (time) ------ TAIL | |
9756 | * direction +-----+-----+-----+-----+-----+ direction | |
9757 | * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild | |
9758 | * fill +-----+-----+-----+-----+-----+ | |
9759 | * ^ ^ | |
9760 | * | | | |
9761 | * | | | |
657fd33b GA |
9762 | * l2arc_feed_thread l2arc_rebuild |
9763 | * will place new bufs here restores bufs here | |
77f6826b | 9764 | * |
657fd33b GA |
9765 | * During l2arc_rebuild() the device is not used by |
9766 | * l2arc_feed_thread() as dev->l2ad_rebuild is set to true. | |
77f6826b GA |
9767 | */ |
9768 | size += L2BLK_GET_LSIZE((&lb->lb_entries[i])->le_prop); | |
657fd33b GA |
9769 | asize += vdev_psize_to_asize(dev->l2ad_vdev, |
9770 | L2BLK_GET_PSIZE((&lb->lb_entries[i])->le_prop)); | |
77f6826b GA |
9771 | l2arc_hdr_restore(&lb->lb_entries[i], dev); |
9772 | } | |
9773 | ||
9774 | /* | |
9775 | * Record rebuild stats: | |
9776 | * size Logical size of restored buffers in the L2ARC | |
657fd33b | 9777 | * asize Aligned size of restored buffers in the L2ARC |
77f6826b GA |
9778 | */ |
9779 | ARCSTAT_INCR(arcstat_l2_rebuild_size, size); | |
657fd33b | 9780 | ARCSTAT_INCR(arcstat_l2_rebuild_asize, asize); |
77f6826b | 9781 | ARCSTAT_INCR(arcstat_l2_rebuild_bufs, log_entries); |
657fd33b GA |
9782 | ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, lb_asize); |
9783 | ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, asize / lb_asize); | |
77f6826b GA |
9784 | ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks); |
9785 | } | |
9786 | ||
9787 | /* | |
9788 | * Restores a single ARC buf hdr from a log entry. The ARC buffer is put | |
9789 | * into a state indicating that it has been evicted to L2ARC. | |
9790 | */ | |
9791 | static void | |
9792 | l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev) | |
9793 | { | |
9794 | arc_buf_hdr_t *hdr, *exists; | |
9795 | kmutex_t *hash_lock; | |
9796 | arc_buf_contents_t type = L2BLK_GET_TYPE((le)->le_prop); | |
9797 | uint64_t asize; | |
9798 | ||
9799 | /* | |
9800 | * Do all the allocation before grabbing any locks, this lets us | |
9801 | * sleep if memory is full and we don't have to deal with failed | |
9802 | * allocations. | |
9803 | */ | |
9804 | hdr = arc_buf_alloc_l2only(L2BLK_GET_LSIZE((le)->le_prop), type, | |
9805 | dev, le->le_dva, le->le_daddr, | |
9806 | L2BLK_GET_PSIZE((le)->le_prop), le->le_birth, | |
9807 | L2BLK_GET_COMPRESS((le)->le_prop), | |
9808 | L2BLK_GET_PROTECTED((le)->le_prop), | |
9809 | L2BLK_GET_PREFETCH((le)->le_prop)); | |
9810 | asize = vdev_psize_to_asize(dev->l2ad_vdev, | |
9811 | L2BLK_GET_PSIZE((le)->le_prop)); | |
9812 | ||
9813 | /* | |
9814 | * vdev_space_update() has to be called before arc_hdr_destroy() to | |
9815 | * avoid underflow since the latter also calls the former. | |
9816 | */ | |
9817 | vdev_space_update(dev->l2ad_vdev, asize, 0, 0); | |
9818 | ||
9819 | ARCSTAT_INCR(arcstat_l2_lsize, HDR_GET_LSIZE(hdr)); | |
9820 | ARCSTAT_INCR(arcstat_l2_psize, HDR_GET_PSIZE(hdr)); | |
9821 | ||
9822 | mutex_enter(&dev->l2ad_mtx); | |
9823 | list_insert_tail(&dev->l2ad_buflist, hdr); | |
9824 | (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); | |
9825 | mutex_exit(&dev->l2ad_mtx); | |
9826 | ||
9827 | exists = buf_hash_insert(hdr, &hash_lock); | |
9828 | if (exists) { | |
9829 | /* Buffer was already cached, no need to restore it. */ | |
9830 | arc_hdr_destroy(hdr); | |
9831 | /* | |
9832 | * If the buffer is already cached, check whether it has | |
9833 | * L2ARC metadata. If not, enter them and update the flag. | |
9834 | * This is important is case of onlining a cache device, since | |
9835 | * we previously evicted all L2ARC metadata from ARC. | |
9836 | */ | |
9837 | if (!HDR_HAS_L2HDR(exists)) { | |
9838 | arc_hdr_set_flags(exists, ARC_FLAG_HAS_L2HDR); | |
9839 | exists->b_l2hdr.b_dev = dev; | |
9840 | exists->b_l2hdr.b_daddr = le->le_daddr; | |
9841 | mutex_enter(&dev->l2ad_mtx); | |
9842 | list_insert_tail(&dev->l2ad_buflist, exists); | |
9843 | (void) zfs_refcount_add_many(&dev->l2ad_alloc, | |
9844 | arc_hdr_size(exists), exists); | |
9845 | mutex_exit(&dev->l2ad_mtx); | |
9846 | vdev_space_update(dev->l2ad_vdev, asize, 0, 0); | |
9847 | ARCSTAT_INCR(arcstat_l2_lsize, HDR_GET_LSIZE(exists)); | |
9848 | ARCSTAT_INCR(arcstat_l2_psize, HDR_GET_PSIZE(exists)); | |
9849 | } | |
9850 | ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached); | |
9851 | } | |
9852 | ||
9853 | mutex_exit(hash_lock); | |
9854 | } | |
9855 | ||
9856 | /* | |
9857 | * Starts an asynchronous read IO to read a log block. This is used in log | |
9858 | * block reconstruction to start reading the next block before we are done | |
9859 | * decoding and reconstructing the current block, to keep the l2arc device | |
9860 | * nice and hot with read IO to process. | |
9861 | * The returned zio will contain a newly allocated memory buffers for the IO | |
9862 | * data which should then be freed by the caller once the zio is no longer | |
9863 | * needed (i.e. due to it having completed). If you wish to abort this | |
9864 | * zio, you should do so using l2arc_log_blk_fetch_abort, which takes | |
9865 | * care of disposing of the allocated buffers correctly. | |
9866 | */ | |
9867 | static zio_t * | |
9868 | l2arc_log_blk_fetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp, | |
9869 | l2arc_log_blk_phys_t *lb) | |
9870 | { | |
657fd33b | 9871 | uint32_t asize; |
77f6826b GA |
9872 | zio_t *pio; |
9873 | l2arc_read_callback_t *cb; | |
9874 | ||
657fd33b GA |
9875 | /* L2BLK_GET_PSIZE returns aligned size for log blocks */ |
9876 | asize = L2BLK_GET_PSIZE((lbp)->lbp_prop); | |
9877 | ASSERT(asize <= sizeof (l2arc_log_blk_phys_t)); | |
9878 | ||
77f6826b | 9879 | cb = kmem_zalloc(sizeof (l2arc_read_callback_t), KM_SLEEP); |
657fd33b | 9880 | cb->l2rcb_abd = abd_get_from_buf(lb, asize); |
77f6826b GA |
9881 | pio = zio_root(vd->vdev_spa, l2arc_blk_fetch_done, cb, |
9882 | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | | |
9883 | ZIO_FLAG_DONT_RETRY); | |
657fd33b | 9884 | (void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, asize, |
77f6826b GA |
9885 | cb->l2rcb_abd, ZIO_CHECKSUM_OFF, NULL, NULL, |
9886 | ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL | | |
9887 | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE)); | |
9888 | ||
9889 | return (pio); | |
9890 | } | |
9891 | ||
9892 | /* | |
9893 | * Aborts a zio returned from l2arc_log_blk_fetch and frees the data | |
9894 | * buffers allocated for it. | |
9895 | */ | |
9896 | static void | |
9897 | l2arc_log_blk_fetch_abort(zio_t *zio) | |
9898 | { | |
9899 | (void) zio_wait(zio); | |
9900 | } | |
9901 | ||
9902 | /* | |
9903 | * Creates a zio to update the device header on an l2arc device. The zio is | |
9904 | * initiated as a child of `pio'. | |
9905 | */ | |
9906 | static void | |
9907 | l2arc_dev_hdr_update(l2arc_dev_t *dev) | |
9908 | { | |
9909 | l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr; | |
9910 | const uint64_t l2dhdr_asize = dev->l2ad_dev_hdr_asize; | |
9911 | abd_t *abd; | |
9912 | int err; | |
9913 | ||
657fd33b GA |
9914 | VERIFY(spa_config_held(dev->l2ad_spa, SCL_STATE_ALL, RW_READER)); |
9915 | ||
77f6826b GA |
9916 | l2dhdr->dh_magic = L2ARC_DEV_HDR_MAGIC; |
9917 | l2dhdr->dh_version = L2ARC_PERSISTENT_VERSION; | |
9918 | l2dhdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa); | |
9919 | l2dhdr->dh_vdev_guid = dev->l2ad_vdev->vdev_guid; | |
657fd33b | 9920 | l2dhdr->dh_log_entries = dev->l2ad_log_entries; |
77f6826b GA |
9921 | l2dhdr->dh_evict = dev->l2ad_evict; |
9922 | l2dhdr->dh_start = dev->l2ad_start; | |
9923 | l2dhdr->dh_end = dev->l2ad_end; | |
657fd33b GA |
9924 | l2dhdr->dh_lb_asize = zfs_refcount_count(&dev->l2ad_lb_asize); |
9925 | l2dhdr->dh_lb_count = zfs_refcount_count(&dev->l2ad_lb_count); | |
77f6826b GA |
9926 | l2dhdr->dh_flags = 0; |
9927 | if (dev->l2ad_first) | |
9928 | l2dhdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST; | |
9929 | ||
9930 | abd = abd_get_from_buf(l2dhdr, l2dhdr_asize); | |
9931 | ||
9932 | err = zio_wait(zio_write_phys(NULL, dev->l2ad_vdev, | |
9933 | VDEV_LABEL_START_SIZE, l2dhdr_asize, abd, ZIO_CHECKSUM_LABEL, NULL, | |
9934 | NULL, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE)); | |
9935 | ||
9936 | abd_put(abd); | |
9937 | ||
9938 | if (err != 0) { | |
9939 | zfs_dbgmsg("L2ARC IO error (%d) while writing device header, " | |
9940 | "vdev guid: %llu", err, dev->l2ad_vdev->vdev_guid); | |
9941 | } | |
9942 | } | |
9943 | ||
9944 | /* | |
9945 | * Commits a log block to the L2ARC device. This routine is invoked from | |
9946 | * l2arc_write_buffers when the log block fills up. | |
9947 | * This function allocates some memory to temporarily hold the serialized | |
9948 | * buffer to be written. This is then released in l2arc_write_done. | |
9949 | */ | |
9950 | static void | |
9951 | l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio, l2arc_write_callback_t *cb) | |
9952 | { | |
9953 | l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk; | |
9954 | l2arc_dev_hdr_phys_t *l2dhdr = dev->l2ad_dev_hdr; | |
9955 | uint64_t psize, asize; | |
9956 | zio_t *wzio; | |
9957 | l2arc_lb_abd_buf_t *abd_buf; | |
9958 | uint8_t *tmpbuf; | |
9959 | l2arc_lb_ptr_buf_t *lb_ptr_buf; | |
9960 | ||
657fd33b | 9961 | VERIFY3S(dev->l2ad_log_ent_idx, ==, dev->l2ad_log_entries); |
77f6826b GA |
9962 | |
9963 | tmpbuf = zio_buf_alloc(sizeof (*lb)); | |
9964 | abd_buf = zio_buf_alloc(sizeof (*abd_buf)); | |
9965 | abd_buf->abd = abd_get_from_buf(lb, sizeof (*lb)); | |
9966 | lb_ptr_buf = kmem_zalloc(sizeof (l2arc_lb_ptr_buf_t), KM_SLEEP); | |
9967 | lb_ptr_buf->lb_ptr = kmem_zalloc(sizeof (l2arc_log_blkptr_t), KM_SLEEP); | |
9968 | ||
9969 | /* link the buffer into the block chain */ | |
9970 | lb->lb_prev_lbp = l2dhdr->dh_start_lbps[1]; | |
9971 | lb->lb_magic = L2ARC_LOG_BLK_MAGIC; | |
9972 | ||
657fd33b GA |
9973 | /* |
9974 | * l2arc_log_blk_commit() may be called multiple times during a single | |
9975 | * l2arc_write_buffers() call. Save the allocated abd buffers in a list | |
9976 | * so we can free them in l2arc_write_done() later on. | |
9977 | */ | |
77f6826b | 9978 | list_insert_tail(&cb->l2wcb_abd_list, abd_buf); |
657fd33b GA |
9979 | |
9980 | /* try to compress the buffer */ | |
77f6826b GA |
9981 | psize = zio_compress_data(ZIO_COMPRESS_LZ4, |
9982 | abd_buf->abd, tmpbuf, sizeof (*lb)); | |
9983 | ||
9984 | /* a log block is never entirely zero */ | |
9985 | ASSERT(psize != 0); | |
9986 | asize = vdev_psize_to_asize(dev->l2ad_vdev, psize); | |
9987 | ASSERT(asize <= sizeof (*lb)); | |
9988 | ||
9989 | /* | |
9990 | * Update the start log block pointer in the device header to point | |
9991 | * to the log block we're about to write. | |
9992 | */ | |
9993 | l2dhdr->dh_start_lbps[1] = l2dhdr->dh_start_lbps[0]; | |
9994 | l2dhdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand; | |
9995 | l2dhdr->dh_start_lbps[0].lbp_payload_asize = | |
9996 | dev->l2ad_log_blk_payload_asize; | |
9997 | l2dhdr->dh_start_lbps[0].lbp_payload_start = | |
9998 | dev->l2ad_log_blk_payload_start; | |
9999 | _NOTE(CONSTCOND) | |
10000 | L2BLK_SET_LSIZE( | |
10001 | (&l2dhdr->dh_start_lbps[0])->lbp_prop, sizeof (*lb)); | |
10002 | L2BLK_SET_PSIZE( | |
10003 | (&l2dhdr->dh_start_lbps[0])->lbp_prop, asize); | |
10004 | L2BLK_SET_CHECKSUM( | |
10005 | (&l2dhdr->dh_start_lbps[0])->lbp_prop, | |
10006 | ZIO_CHECKSUM_FLETCHER_4); | |
10007 | if (asize < sizeof (*lb)) { | |
10008 | /* compression succeeded */ | |
10009 | bzero(tmpbuf + psize, asize - psize); | |
10010 | L2BLK_SET_COMPRESS( | |
10011 | (&l2dhdr->dh_start_lbps[0])->lbp_prop, | |
10012 | ZIO_COMPRESS_LZ4); | |
10013 | } else { | |
10014 | /* compression failed */ | |
10015 | bcopy(lb, tmpbuf, sizeof (*lb)); | |
10016 | L2BLK_SET_COMPRESS( | |
10017 | (&l2dhdr->dh_start_lbps[0])->lbp_prop, | |
10018 | ZIO_COMPRESS_OFF); | |
10019 | } | |
10020 | ||
10021 | /* checksum what we're about to write */ | |
10022 | fletcher_4_native(tmpbuf, asize, NULL, | |
10023 | &l2dhdr->dh_start_lbps[0].lbp_cksum); | |
10024 | ||
10025 | abd_put(abd_buf->abd); | |
10026 | ||
10027 | /* perform the write itself */ | |
10028 | abd_buf->abd = abd_get_from_buf(tmpbuf, sizeof (*lb)); | |
10029 | abd_take_ownership_of_buf(abd_buf->abd, B_TRUE); | |
10030 | wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand, | |
10031 | asize, abd_buf->abd, ZIO_CHECKSUM_OFF, NULL, NULL, | |
10032 | ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE); | |
10033 | DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio); | |
10034 | (void) zio_nowait(wzio); | |
10035 | ||
10036 | dev->l2ad_hand += asize; | |
10037 | /* | |
10038 | * Include the committed log block's pointer in the list of pointers | |
10039 | * to log blocks present in the L2ARC device. | |
10040 | */ | |
10041 | bcopy(&l2dhdr->dh_start_lbps[0], lb_ptr_buf->lb_ptr, | |
10042 | sizeof (l2arc_log_blkptr_t)); | |
10043 | mutex_enter(&dev->l2ad_mtx); | |
10044 | list_insert_head(&dev->l2ad_lbptr_list, lb_ptr_buf); | |
657fd33b GA |
10045 | ARCSTAT_INCR(arcstat_l2_log_blk_asize, asize); |
10046 | ARCSTAT_BUMP(arcstat_l2_log_blk_count); | |
10047 | zfs_refcount_add_many(&dev->l2ad_lb_asize, asize, lb_ptr_buf); | |
10048 | zfs_refcount_add(&dev->l2ad_lb_count, lb_ptr_buf); | |
77f6826b GA |
10049 | mutex_exit(&dev->l2ad_mtx); |
10050 | vdev_space_update(dev->l2ad_vdev, asize, 0, 0); | |
10051 | ||
10052 | /* bump the kstats */ | |
10053 | ARCSTAT_INCR(arcstat_l2_write_bytes, asize); | |
10054 | ARCSTAT_BUMP(arcstat_l2_log_blk_writes); | |
657fd33b | 10055 | ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_asize, asize); |
77f6826b GA |
10056 | ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, |
10057 | dev->l2ad_log_blk_payload_asize / asize); | |
10058 | ||
10059 | /* start a new log block */ | |
10060 | dev->l2ad_log_ent_idx = 0; | |
10061 | dev->l2ad_log_blk_payload_asize = 0; | |
10062 | dev->l2ad_log_blk_payload_start = 0; | |
10063 | } | |
10064 | ||
10065 | /* | |
10066 | * Validates an L2ARC log block address to make sure that it can be read | |
10067 | * from the provided L2ARC device. | |
10068 | */ | |
10069 | boolean_t | |
10070 | l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp) | |
10071 | { | |
657fd33b GA |
10072 | /* L2BLK_GET_PSIZE returns aligned size for log blocks */ |
10073 | uint64_t asize = L2BLK_GET_PSIZE((lbp)->lbp_prop); | |
10074 | uint64_t end = lbp->lbp_daddr + asize - 1; | |
77f6826b GA |
10075 | uint64_t start = lbp->lbp_payload_start; |
10076 | boolean_t evicted = B_FALSE; | |
10077 | ||
10078 | /* | |
10079 | * A log block is valid if all of the following conditions are true: | |
10080 | * - it fits entirely (including its payload) between l2ad_start and | |
10081 | * l2ad_end | |
10082 | * - it has a valid size | |
10083 | * - neither the log block itself nor part of its payload was evicted | |
10084 | * by l2arc_evict(): | |
10085 | * | |
10086 | * l2ad_hand l2ad_evict | |
10087 | * | | lbp_daddr | |
10088 | * | start | | end | |
10089 | * | | | | | | |
10090 | * V V V V V | |
10091 | * l2ad_start ============================================ l2ad_end | |
10092 | * --------------------------|||| | |
10093 | * ^ ^ | |
10094 | * | log block | |
10095 | * payload | |
10096 | */ | |
10097 | ||
10098 | evicted = | |
10099 | l2arc_range_check_overlap(start, end, dev->l2ad_hand) || | |
10100 | l2arc_range_check_overlap(start, end, dev->l2ad_evict) || | |
10101 | l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, start) || | |
10102 | l2arc_range_check_overlap(dev->l2ad_hand, dev->l2ad_evict, end); | |
10103 | ||
10104 | return (start >= dev->l2ad_start && end <= dev->l2ad_end && | |
657fd33b | 10105 | asize > 0 && asize <= sizeof (l2arc_log_blk_phys_t) && |
77f6826b GA |
10106 | (!evicted || dev->l2ad_first)); |
10107 | } | |
10108 | ||
10109 | /* | |
10110 | * Inserts ARC buffer header `hdr' into the current L2ARC log block on | |
10111 | * the device. The buffer being inserted must be present in L2ARC. | |
10112 | * Returns B_TRUE if the L2ARC log block is full and needs to be committed | |
10113 | * to L2ARC, or B_FALSE if it still has room for more ARC buffers. | |
10114 | */ | |
10115 | static boolean_t | |
10116 | l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *hdr) | |
10117 | { | |
10118 | l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk; | |
10119 | l2arc_log_ent_phys_t *le; | |
77f6826b | 10120 | |
657fd33b | 10121 | if (dev->l2ad_log_entries == 0) |
77f6826b GA |
10122 | return (B_FALSE); |
10123 | ||
10124 | int index = dev->l2ad_log_ent_idx++; | |
10125 | ||
657fd33b | 10126 | ASSERT3S(index, <, dev->l2ad_log_entries); |
77f6826b GA |
10127 | ASSERT(HDR_HAS_L2HDR(hdr)); |
10128 | ||
10129 | le = &lb->lb_entries[index]; | |
10130 | bzero(le, sizeof (*le)); | |
10131 | le->le_dva = hdr->b_dva; | |
10132 | le->le_birth = hdr->b_birth; | |
10133 | le->le_daddr = hdr->b_l2hdr.b_daddr; | |
10134 | if (index == 0) | |
10135 | dev->l2ad_log_blk_payload_start = le->le_daddr; | |
10136 | L2BLK_SET_LSIZE((le)->le_prop, HDR_GET_LSIZE(hdr)); | |
10137 | L2BLK_SET_PSIZE((le)->le_prop, HDR_GET_PSIZE(hdr)); | |
10138 | L2BLK_SET_COMPRESS((le)->le_prop, HDR_GET_COMPRESS(hdr)); | |
10139 | L2BLK_SET_TYPE((le)->le_prop, hdr->b_type); | |
10140 | L2BLK_SET_PROTECTED((le)->le_prop, !!(HDR_PROTECTED(hdr))); | |
10141 | L2BLK_SET_PREFETCH((le)->le_prop, !!(HDR_PREFETCH(hdr))); | |
10142 | ||
10143 | dev->l2ad_log_blk_payload_asize += vdev_psize_to_asize(dev->l2ad_vdev, | |
10144 | HDR_GET_PSIZE(hdr)); | |
10145 | ||
657fd33b | 10146 | return (dev->l2ad_log_ent_idx == dev->l2ad_log_entries); |
77f6826b GA |
10147 | } |
10148 | ||
10149 | /* | |
10150 | * Checks whether a given L2ARC device address sits in a time-sequential | |
10151 | * range. The trick here is that the L2ARC is a rotary buffer, so we can't | |
10152 | * just do a range comparison, we need to handle the situation in which the | |
10153 | * range wraps around the end of the L2ARC device. Arguments: | |
10154 | * bottom -- Lower end of the range to check (written to earlier). | |
10155 | * top -- Upper end of the range to check (written to later). | |
10156 | * check -- The address for which we want to determine if it sits in | |
10157 | * between the top and bottom. | |
10158 | * | |
10159 | * The 3-way conditional below represents the following cases: | |
10160 | * | |
10161 | * bottom < top : Sequentially ordered case: | |
10162 | * <check>--------+-------------------+ | |
10163 | * | (overlap here?) | | |
10164 | * L2ARC dev V V | |
10165 | * |---------------<bottom>============<top>--------------| | |
10166 | * | |
10167 | * bottom > top: Looped-around case: | |
10168 | * <check>--------+------------------+ | |
10169 | * | (overlap here?) | | |
10170 | * L2ARC dev V V | |
10171 | * |===============<top>---------------<bottom>===========| | |
10172 | * ^ ^ | |
10173 | * | (or here?) | | |
10174 | * +---------------+---------<check> | |
10175 | * | |
10176 | * top == bottom : Just a single address comparison. | |
10177 | */ | |
10178 | boolean_t | |
10179 | l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check) | |
10180 | { | |
10181 | if (bottom < top) | |
10182 | return (bottom <= check && check <= top); | |
10183 | else if (bottom > top) | |
10184 | return (check <= top || bottom <= check); | |
10185 | else | |
10186 | return (check == top); | |
10187 | } | |
10188 | ||
0f699108 AZ |
10189 | EXPORT_SYMBOL(arc_buf_size); |
10190 | EXPORT_SYMBOL(arc_write); | |
c28b2279 | 10191 | EXPORT_SYMBOL(arc_read); |
e0b0ca98 | 10192 | EXPORT_SYMBOL(arc_buf_info); |
c28b2279 | 10193 | EXPORT_SYMBOL(arc_getbuf_func); |
ab26409d BB |
10194 | EXPORT_SYMBOL(arc_add_prune_callback); |
10195 | EXPORT_SYMBOL(arc_remove_prune_callback); | |
c28b2279 | 10196 | |
02730c33 | 10197 | /* BEGIN CSTYLED */ |
e3570464 | 10198 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min, param_set_arc_long, |
10199 | param_get_long, ZMOD_RW, "Min arc size"); | |
c28b2279 | 10200 | |
e3570464 | 10201 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, max, param_set_arc_long, |
10202 | param_get_long, ZMOD_RW, "Max arc size"); | |
c28b2279 | 10203 | |
e3570464 | 10204 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_limit, param_set_arc_long, |
10205 | param_get_long, ZMOD_RW, "Metadata limit for arc size"); | |
6a8f9b6b | 10206 | |
e3570464 | 10207 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_limit_percent, |
10208 | param_set_arc_long, param_get_long, ZMOD_RW, | |
9907cc1c G |
10209 | "Percent of arc size for arc meta limit"); |
10210 | ||
e3570464 | 10211 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, meta_min, param_set_arc_long, |
10212 | param_get_long, ZMOD_RW, "Min arc metadata"); | |
ca0bf58d | 10213 | |
03fdcb9a MM |
10214 | ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_prune, INT, ZMOD_RW, |
10215 | "Meta objects to scan for prune"); | |
c409e464 | 10216 | |
03fdcb9a | 10217 | ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_adjust_restarts, INT, ZMOD_RW, |
bc888666 BB |
10218 | "Limit number of restarts in arc_adjust_meta"); |
10219 | ||
03fdcb9a MM |
10220 | ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, meta_strategy, INT, ZMOD_RW, |
10221 | "Meta reclaim strategy"); | |
f6046738 | 10222 | |
e3570464 | 10223 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, grow_retry, param_set_arc_int, |
10224 | param_get_int, ZMOD_RW, "Seconds before growing arc size"); | |
c409e464 | 10225 | |
03fdcb9a MM |
10226 | ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, p_dampener_disable, INT, ZMOD_RW, |
10227 | "Disable arc_p adapt dampener"); | |
62422785 | 10228 | |
e3570464 | 10229 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, shrink_shift, param_set_arc_int, |
10230 | param_get_int, ZMOD_RW, "log2(fraction of arc to reclaim)"); | |
c409e464 | 10231 | |
03fdcb9a | 10232 | ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, pc_percent, UINT, ZMOD_RW, |
03b60eee DB |
10233 | "Percent of pagecache to reclaim arc to"); |
10234 | ||
e3570464 | 10235 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, p_min_shift, param_set_arc_int, |
10236 | param_get_int, ZMOD_RW, "arc_c shift to calc min/max arc_p"); | |
728d6ae9 | 10237 | |
03fdcb9a MM |
10238 | ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, average_blocksize, INT, ZMOD_RD, |
10239 | "Target average block size"); | |
49ddb315 | 10240 | |
03fdcb9a MM |
10241 | ZFS_MODULE_PARAM(zfs, zfs_, compressed_arc_enabled, INT, ZMOD_RW, |
10242 | "Disable compressed arc buffers"); | |
d3c2ae1c | 10243 | |
e3570464 | 10244 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prefetch_ms, param_set_arc_int, |
10245 | param_get_int, ZMOD_RW, "Min life of prefetch block in ms"); | |
d4a72f23 | 10246 | |
e3570464 | 10247 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, min_prescient_prefetch_ms, |
10248 | param_set_arc_int, param_get_int, ZMOD_RW, | |
d4a72f23 | 10249 | "Min life of prescient prefetched block in ms"); |
bce45ec9 | 10250 | |
03fdcb9a MM |
10251 | ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_max, ULONG, ZMOD_RW, |
10252 | "Max write bytes per interval"); | |
abd8610c | 10253 | |
03fdcb9a MM |
10254 | ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, write_boost, ULONG, ZMOD_RW, |
10255 | "Extra write bytes during device warmup"); | |
abd8610c | 10256 | |
03fdcb9a MM |
10257 | ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom, ULONG, ZMOD_RW, |
10258 | "Number of max device writes to precache"); | |
abd8610c | 10259 | |
03fdcb9a MM |
10260 | ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, headroom_boost, ULONG, ZMOD_RW, |
10261 | "Compressed l2arc_headroom multiplier"); | |
3a17a7a9 | 10262 | |
03fdcb9a MM |
10263 | ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_secs, ULONG, ZMOD_RW, |
10264 | "Seconds between L2ARC writing"); | |
abd8610c | 10265 | |
03fdcb9a MM |
10266 | ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_min_ms, ULONG, ZMOD_RW, |
10267 | "Min feed interval in milliseconds"); | |
abd8610c | 10268 | |
03fdcb9a MM |
10269 | ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, noprefetch, INT, ZMOD_RW, |
10270 | "Skip caching prefetched buffers"); | |
abd8610c | 10271 | |
03fdcb9a MM |
10272 | ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, feed_again, INT, ZMOD_RW, |
10273 | "Turbo L2ARC warmup"); | |
abd8610c | 10274 | |
03fdcb9a MM |
10275 | ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, norw, INT, ZMOD_RW, |
10276 | "No reads during writes"); | |
abd8610c | 10277 | |
77f6826b GA |
10278 | ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_enabled, INT, ZMOD_RW, |
10279 | "Rebuild the L2ARC when importing a pool"); | |
10280 | ||
10281 | ZFS_MODULE_PARAM(zfs_l2arc, l2arc_, rebuild_blocks_min_l2size, ULONG, ZMOD_RW, | |
10282 | "Min size in bytes to write rebuild log blocks in L2ARC"); | |
10283 | ||
e3570464 | 10284 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, lotsfree_percent, param_set_arc_int, |
10285 | param_get_int, ZMOD_RW, "System free memory I/O throttle in bytes"); | |
7e8bddd0 | 10286 | |
e3570464 | 10287 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, sys_free, param_set_arc_long, |
10288 | param_get_long, ZMOD_RW, "System free memory target size in bytes"); | |
11f552fa | 10289 | |
e3570464 | 10290 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit, param_set_arc_long, |
10291 | param_get_long, ZMOD_RW, "Minimum bytes of dnodes in arc"); | |
25458cbe | 10292 | |
e3570464 | 10293 | ZFS_MODULE_PARAM_CALL(zfs_arc, zfs_arc_, dnode_limit_percent, |
10294 | param_set_arc_long, param_get_long, ZMOD_RW, | |
9907cc1c G |
10295 | "Percent of ARC meta buffers for dnodes"); |
10296 | ||
03fdcb9a | 10297 | ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, dnode_reduce_percent, ULONG, ZMOD_RW, |
25458cbe | 10298 | "Percentage of excess dnodes to try to unpin"); |
02730c33 | 10299 | /* END CSTYLED */ |