]>
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. |
5dbd68a3 | 23 | * Copyright (c) 2011, 2014 by Delphix. All rights reserved. |
3a17a7a9 | 24 | * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. |
3bec585e | 25 | * Copyright 2014 Nexenta Systems, Inc. All rights reserved. |
34dc7c2f BB |
26 | */ |
27 | ||
34dc7c2f BB |
28 | /* |
29 | * DVA-based Adjustable Replacement Cache | |
30 | * | |
31 | * While much of the theory of operation used here is | |
32 | * based on the self-tuning, low overhead replacement cache | |
33 | * presented by Megiddo and Modha at FAST 2003, there are some | |
34 | * significant differences: | |
35 | * | |
36 | * 1. The Megiddo and Modha model assumes any page is evictable. | |
37 | * Pages in its cache cannot be "locked" into memory. This makes | |
38 | * the eviction algorithm simple: evict the last page in the list. | |
39 | * This also make the performance characteristics easy to reason | |
40 | * about. Our cache is not so simple. At any given moment, some | |
41 | * subset of the blocks in the cache are un-evictable because we | |
42 | * have handed out a reference to them. Blocks are only evictable | |
43 | * when there are no external references active. This makes | |
44 | * eviction far more problematic: we choose to evict the evictable | |
45 | * blocks that are the "lowest" in the list. | |
46 | * | |
47 | * There are times when it is not possible to evict the requested | |
48 | * space. In these circumstances we are unable to adjust the cache | |
49 | * size. To prevent the cache growing unbounded at these times we | |
50 | * implement a "cache throttle" that slows the flow of new data | |
51 | * into the cache until we can make space available. | |
52 | * | |
53 | * 2. The Megiddo and Modha model assumes a fixed cache size. | |
54 | * Pages are evicted when the cache is full and there is a cache | |
55 | * miss. Our model has a variable sized cache. It grows with | |
56 | * high use, but also tries to react to memory pressure from the | |
57 | * operating system: decreasing its size when system memory is | |
58 | * tight. | |
59 | * | |
60 | * 3. The Megiddo and Modha model assumes a fixed page size. All | |
d3cc8b15 | 61 | * elements of the cache are therefore exactly the same size. So |
34dc7c2f BB |
62 | * when adjusting the cache size following a cache miss, its simply |
63 | * a matter of choosing a single page to evict. In our model, we | |
64 | * have variable sized cache blocks (rangeing from 512 bytes to | |
d3cc8b15 | 65 | * 128K bytes). We therefore choose a set of blocks to evict to make |
34dc7c2f BB |
66 | * space for a cache miss that approximates as closely as possible |
67 | * the space used by the new block. | |
68 | * | |
69 | * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" | |
70 | * by N. Megiddo & D. Modha, FAST 2003 | |
71 | */ | |
72 | ||
73 | /* | |
74 | * The locking model: | |
75 | * | |
76 | * A new reference to a cache buffer can be obtained in two | |
77 | * ways: 1) via a hash table lookup using the DVA as a key, | |
78 | * or 2) via one of the ARC lists. The arc_read() interface | |
79 | * uses method 1, while the internal arc algorithms for | |
d3cc8b15 | 80 | * adjusting the cache use method 2. We therefore provide two |
34dc7c2f BB |
81 | * types of locks: 1) the hash table lock array, and 2) the |
82 | * arc list locks. | |
83 | * | |
5c839890 BC |
84 | * Buffers do not have their own mutexes, rather they rely on the |
85 | * hash table mutexes for the bulk of their protection (i.e. most | |
86 | * fields in the arc_buf_hdr_t are protected by these mutexes). | |
34dc7c2f BB |
87 | * |
88 | * buf_hash_find() returns the appropriate mutex (held) when it | |
89 | * locates the requested buffer in the hash table. It returns | |
90 | * NULL for the mutex if the buffer was not in the table. | |
91 | * | |
92 | * buf_hash_remove() expects the appropriate hash mutex to be | |
93 | * already held before it is invoked. | |
94 | * | |
95 | * Each arc state also has a mutex which is used to protect the | |
96 | * buffer list associated with the state. When attempting to | |
97 | * obtain a hash table lock while holding an arc list lock you | |
98 | * must use: mutex_tryenter() to avoid deadlock. Also note that | |
99 | * the active state mutex must be held before the ghost state mutex. | |
100 | * | |
101 | * Arc buffers may have an associated eviction callback function. | |
102 | * This function will be invoked prior to removing the buffer (e.g. | |
103 | * in arc_do_user_evicts()). Note however that the data associated | |
104 | * with the buffer may be evicted prior to the callback. The callback | |
105 | * must be made with *no locks held* (to prevent deadlock). Additionally, | |
106 | * the users of callbacks must ensure that their private data is | |
bd089c54 | 107 | * protected from simultaneous callbacks from arc_clear_callback() |
34dc7c2f BB |
108 | * and arc_do_user_evicts(). |
109 | * | |
ab26409d BB |
110 | * It as also possible to register a callback which is run when the |
111 | * arc_meta_limit is reached and no buffers can be safely evicted. In | |
112 | * this case the arc user should drop a reference on some arc buffers so | |
113 | * they can be reclaimed and the arc_meta_limit honored. For example, | |
114 | * when using the ZPL each dentry holds a references on a znode. These | |
115 | * dentries must be pruned before the arc buffer holding the znode can | |
116 | * be safely evicted. | |
117 | * | |
34dc7c2f BB |
118 | * Note that the majority of the performance stats are manipulated |
119 | * with atomic operations. | |
120 | * | |
b9541d6b | 121 | * The L2ARC uses the l2ad_mtx on each vdev for the following: |
34dc7c2f BB |
122 | * |
123 | * - L2ARC buflist creation | |
124 | * - L2ARC buflist eviction | |
125 | * - L2ARC write completion, which walks L2ARC buflists | |
126 | * - ARC header destruction, as it removes from L2ARC buflists | |
127 | * - ARC header release, as it removes from L2ARC buflists | |
128 | */ | |
129 | ||
130 | #include <sys/spa.h> | |
131 | #include <sys/zio.h> | |
3a17a7a9 | 132 | #include <sys/zio_compress.h> |
34dc7c2f BB |
133 | #include <sys/zfs_context.h> |
134 | #include <sys/arc.h> | |
b128c09f | 135 | #include <sys/vdev.h> |
9babb374 | 136 | #include <sys/vdev_impl.h> |
e8b96c60 | 137 | #include <sys/dsl_pool.h> |
ca0bf58d | 138 | #include <sys/multilist.h> |
34dc7c2f BB |
139 | #ifdef _KERNEL |
140 | #include <sys/vmsystm.h> | |
141 | #include <vm/anon.h> | |
142 | #include <sys/fs/swapnode.h> | |
ab26409d | 143 | #include <sys/zpl.h> |
aaed7c40 | 144 | #include <linux/mm_compat.h> |
34dc7c2f BB |
145 | #endif |
146 | #include <sys/callb.h> | |
147 | #include <sys/kstat.h> | |
570827e1 | 148 | #include <sys/dmu_tx.h> |
428870ff | 149 | #include <zfs_fletcher.h> |
59ec819a | 150 | #include <sys/arc_impl.h> |
49ee64e5 | 151 | #include <sys/trace_arc.h> |
34dc7c2f | 152 | |
498877ba MA |
153 | #ifndef _KERNEL |
154 | /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ | |
155 | boolean_t arc_watch = B_FALSE; | |
156 | #endif | |
157 | ||
ca0bf58d PS |
158 | static kmutex_t arc_reclaim_lock; |
159 | static kcondvar_t arc_reclaim_thread_cv; | |
160 | static boolean_t arc_reclaim_thread_exit; | |
161 | static kcondvar_t arc_reclaim_waiters_cv; | |
162 | ||
163 | static kmutex_t arc_user_evicts_lock; | |
164 | static kcondvar_t arc_user_evicts_cv; | |
165 | static boolean_t arc_user_evicts_thread_exit; | |
34dc7c2f | 166 | |
e8b96c60 | 167 | /* |
ca0bf58d PS |
168 | * The number of headers to evict in arc_evict_state_impl() before |
169 | * dropping the sublist lock and evicting from another sublist. A lower | |
170 | * value means we're more likely to evict the "correct" header (i.e. the | |
171 | * oldest header in the arc state), but comes with higher overhead | |
172 | * (i.e. more invocations of arc_evict_state_impl()). | |
173 | */ | |
174 | int zfs_arc_evict_batch_limit = 10; | |
175 | ||
176 | /* | |
177 | * The number of sublists used for each of the arc state lists. If this | |
178 | * is not set to a suitable value by the user, it will be configured to | |
179 | * the number of CPUs on the system in arc_init(). | |
e8b96c60 | 180 | */ |
ca0bf58d | 181 | int zfs_arc_num_sublists_per_state = 0; |
e8b96c60 | 182 | |
34dc7c2f | 183 | /* number of seconds before growing cache again */ |
ca67b33a | 184 | static int arc_grow_retry = 5; |
34dc7c2f | 185 | |
ca0bf58d | 186 | /* shift of arc_c for calculating overflow limit in arc_get_data_buf */ |
ca67b33a | 187 | int zfs_arc_overflow_shift = 8; |
62422785 | 188 | |
d164b209 | 189 | /* log2(fraction of arc to reclaim) */ |
ca67b33a | 190 | static int arc_shrink_shift = 7; |
d164b209 | 191 | |
34dc7c2f | 192 | /* |
ca67b33a MA |
193 | * log2(fraction of ARC which must be free to allow growing). |
194 | * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, | |
195 | * when reading a new block into the ARC, we will evict an equal-sized block | |
196 | * from the ARC. | |
197 | * | |
198 | * This must be less than arc_shrink_shift, so that when we shrink the ARC, | |
199 | * we will still not allow it to grow. | |
34dc7c2f | 200 | */ |
ca67b33a | 201 | int arc_no_grow_shift = 5; |
bce45ec9 | 202 | |
49ddb315 | 203 | |
ca0bf58d PS |
204 | /* |
205 | * minimum lifespan of a prefetch block in clock ticks | |
206 | * (initialized in arc_init()) | |
207 | */ | |
ca67b33a | 208 | static int arc_min_prefetch_lifespan; |
ca0bf58d | 209 | |
e8b96c60 MA |
210 | /* |
211 | * If this percent of memory is free, don't throttle. | |
212 | */ | |
213 | int arc_lotsfree_percent = 10; | |
214 | ||
34dc7c2f BB |
215 | static int arc_dead; |
216 | ||
b128c09f BB |
217 | /* |
218 | * The arc has filled available memory and has now warmed up. | |
219 | */ | |
220 | static boolean_t arc_warm; | |
221 | ||
34dc7c2f BB |
222 | /* |
223 | * These tunables are for performance analysis. | |
224 | */ | |
c28b2279 BB |
225 | unsigned long zfs_arc_max = 0; |
226 | unsigned long zfs_arc_min = 0; | |
227 | unsigned long zfs_arc_meta_limit = 0; | |
ca0bf58d | 228 | unsigned long zfs_arc_meta_min = 0; |
ca67b33a MA |
229 | int zfs_arc_grow_retry = 0; |
230 | int zfs_arc_shrink_shift = 0; | |
231 | int zfs_disable_dup_eviction = 0; | |
232 | int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ | |
34dc7c2f | 233 | |
bc888666 | 234 | /* |
ca67b33a | 235 | * These tunables are Linux specific |
bc888666 | 236 | */ |
ca67b33a MA |
237 | int zfs_arc_memory_throttle_disable = 1; |
238 | int zfs_arc_min_prefetch_lifespan = 0; | |
239 | int zfs_arc_p_aggressive_disable = 1; | |
240 | int zfs_arc_p_dampener_disable = 1; | |
241 | int zfs_arc_meta_prune = 10000; | |
242 | int zfs_arc_meta_strategy = ARC_STRATEGY_META_BALANCED; | |
243 | int zfs_arc_meta_adjust_restarts = 4096; | |
bc888666 | 244 | |
34dc7c2f BB |
245 | /* The 6 states: */ |
246 | static arc_state_t ARC_anon; | |
247 | static arc_state_t ARC_mru; | |
248 | static arc_state_t ARC_mru_ghost; | |
249 | static arc_state_t ARC_mfu; | |
250 | static arc_state_t ARC_mfu_ghost; | |
251 | static arc_state_t ARC_l2c_only; | |
252 | ||
253 | typedef struct arc_stats { | |
254 | kstat_named_t arcstat_hits; | |
255 | kstat_named_t arcstat_misses; | |
256 | kstat_named_t arcstat_demand_data_hits; | |
257 | kstat_named_t arcstat_demand_data_misses; | |
258 | kstat_named_t arcstat_demand_metadata_hits; | |
259 | kstat_named_t arcstat_demand_metadata_misses; | |
260 | kstat_named_t arcstat_prefetch_data_hits; | |
261 | kstat_named_t arcstat_prefetch_data_misses; | |
262 | kstat_named_t arcstat_prefetch_metadata_hits; | |
263 | kstat_named_t arcstat_prefetch_metadata_misses; | |
264 | kstat_named_t arcstat_mru_hits; | |
265 | kstat_named_t arcstat_mru_ghost_hits; | |
266 | kstat_named_t arcstat_mfu_hits; | |
267 | kstat_named_t arcstat_mfu_ghost_hits; | |
268 | kstat_named_t arcstat_deleted; | |
e49f1e20 WA |
269 | /* |
270 | * Number of buffers that could not be evicted because the hash lock | |
271 | * was held by another thread. The lock may not necessarily be held | |
272 | * by something using the same buffer, since hash locks are shared | |
273 | * by multiple buffers. | |
274 | */ | |
34dc7c2f | 275 | kstat_named_t arcstat_mutex_miss; |
e49f1e20 WA |
276 | /* |
277 | * Number of buffers skipped because they have I/O in progress, are | |
278 | * indrect prefetch buffers that have not lived long enough, or are | |
279 | * not from the spa we're trying to evict from. | |
280 | */ | |
34dc7c2f | 281 | kstat_named_t arcstat_evict_skip; |
ca0bf58d PS |
282 | /* |
283 | * Number of times arc_evict_state() was unable to evict enough | |
284 | * buffers to reach its target amount. | |
285 | */ | |
286 | kstat_named_t arcstat_evict_not_enough; | |
428870ff BB |
287 | kstat_named_t arcstat_evict_l2_cached; |
288 | kstat_named_t arcstat_evict_l2_eligible; | |
289 | kstat_named_t arcstat_evict_l2_ineligible; | |
ca0bf58d | 290 | kstat_named_t arcstat_evict_l2_skip; |
34dc7c2f BB |
291 | kstat_named_t arcstat_hash_elements; |
292 | kstat_named_t arcstat_hash_elements_max; | |
293 | kstat_named_t arcstat_hash_collisions; | |
294 | kstat_named_t arcstat_hash_chains; | |
295 | kstat_named_t arcstat_hash_chain_max; | |
296 | kstat_named_t arcstat_p; | |
297 | kstat_named_t arcstat_c; | |
298 | kstat_named_t arcstat_c_min; | |
299 | kstat_named_t arcstat_c_max; | |
300 | kstat_named_t arcstat_size; | |
301 | kstat_named_t arcstat_hdr_size; | |
d164b209 | 302 | kstat_named_t arcstat_data_size; |
cc7f677c | 303 | kstat_named_t arcstat_meta_size; |
d164b209 | 304 | kstat_named_t arcstat_other_size; |
13be560d BB |
305 | kstat_named_t arcstat_anon_size; |
306 | kstat_named_t arcstat_anon_evict_data; | |
307 | kstat_named_t arcstat_anon_evict_metadata; | |
308 | kstat_named_t arcstat_mru_size; | |
309 | kstat_named_t arcstat_mru_evict_data; | |
310 | kstat_named_t arcstat_mru_evict_metadata; | |
311 | kstat_named_t arcstat_mru_ghost_size; | |
312 | kstat_named_t arcstat_mru_ghost_evict_data; | |
313 | kstat_named_t arcstat_mru_ghost_evict_metadata; | |
314 | kstat_named_t arcstat_mfu_size; | |
315 | kstat_named_t arcstat_mfu_evict_data; | |
316 | kstat_named_t arcstat_mfu_evict_metadata; | |
317 | kstat_named_t arcstat_mfu_ghost_size; | |
318 | kstat_named_t arcstat_mfu_ghost_evict_data; | |
319 | kstat_named_t arcstat_mfu_ghost_evict_metadata; | |
34dc7c2f BB |
320 | kstat_named_t arcstat_l2_hits; |
321 | kstat_named_t arcstat_l2_misses; | |
322 | kstat_named_t arcstat_l2_feeds; | |
323 | kstat_named_t arcstat_l2_rw_clash; | |
d164b209 BB |
324 | kstat_named_t arcstat_l2_read_bytes; |
325 | kstat_named_t arcstat_l2_write_bytes; | |
34dc7c2f BB |
326 | kstat_named_t arcstat_l2_writes_sent; |
327 | kstat_named_t arcstat_l2_writes_done; | |
328 | kstat_named_t arcstat_l2_writes_error; | |
ca0bf58d | 329 | kstat_named_t arcstat_l2_writes_lock_retry; |
34dc7c2f BB |
330 | kstat_named_t arcstat_l2_evict_lock_retry; |
331 | kstat_named_t arcstat_l2_evict_reading; | |
b9541d6b | 332 | kstat_named_t arcstat_l2_evict_l1cached; |
34dc7c2f | 333 | kstat_named_t arcstat_l2_free_on_write; |
ca0bf58d | 334 | kstat_named_t arcstat_l2_cdata_free_on_write; |
34dc7c2f BB |
335 | kstat_named_t arcstat_l2_abort_lowmem; |
336 | kstat_named_t arcstat_l2_cksum_bad; | |
337 | kstat_named_t arcstat_l2_io_error; | |
338 | kstat_named_t arcstat_l2_size; | |
3a17a7a9 | 339 | kstat_named_t arcstat_l2_asize; |
34dc7c2f | 340 | kstat_named_t arcstat_l2_hdr_size; |
3a17a7a9 SK |
341 | kstat_named_t arcstat_l2_compress_successes; |
342 | kstat_named_t arcstat_l2_compress_zeros; | |
343 | kstat_named_t arcstat_l2_compress_failures; | |
34dc7c2f | 344 | kstat_named_t arcstat_memory_throttle_count; |
1eb5bfa3 GW |
345 | kstat_named_t arcstat_duplicate_buffers; |
346 | kstat_named_t arcstat_duplicate_buffers_size; | |
347 | kstat_named_t arcstat_duplicate_reads; | |
7cb67b45 BB |
348 | kstat_named_t arcstat_memory_direct_count; |
349 | kstat_named_t arcstat_memory_indirect_count; | |
1834f2d8 BB |
350 | kstat_named_t arcstat_no_grow; |
351 | kstat_named_t arcstat_tempreserve; | |
352 | kstat_named_t arcstat_loaned_bytes; | |
ab26409d | 353 | kstat_named_t arcstat_prune; |
1834f2d8 BB |
354 | kstat_named_t arcstat_meta_used; |
355 | kstat_named_t arcstat_meta_limit; | |
356 | kstat_named_t arcstat_meta_max; | |
ca0bf58d | 357 | kstat_named_t arcstat_meta_min; |
34dc7c2f BB |
358 | } arc_stats_t; |
359 | ||
360 | static arc_stats_t arc_stats = { | |
361 | { "hits", KSTAT_DATA_UINT64 }, | |
362 | { "misses", KSTAT_DATA_UINT64 }, | |
363 | { "demand_data_hits", KSTAT_DATA_UINT64 }, | |
364 | { "demand_data_misses", KSTAT_DATA_UINT64 }, | |
365 | { "demand_metadata_hits", KSTAT_DATA_UINT64 }, | |
366 | { "demand_metadata_misses", KSTAT_DATA_UINT64 }, | |
367 | { "prefetch_data_hits", KSTAT_DATA_UINT64 }, | |
368 | { "prefetch_data_misses", KSTAT_DATA_UINT64 }, | |
369 | { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, | |
370 | { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, | |
371 | { "mru_hits", KSTAT_DATA_UINT64 }, | |
372 | { "mru_ghost_hits", KSTAT_DATA_UINT64 }, | |
373 | { "mfu_hits", KSTAT_DATA_UINT64 }, | |
374 | { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, | |
375 | { "deleted", KSTAT_DATA_UINT64 }, | |
34dc7c2f BB |
376 | { "mutex_miss", KSTAT_DATA_UINT64 }, |
377 | { "evict_skip", KSTAT_DATA_UINT64 }, | |
ca0bf58d | 378 | { "evict_not_enough", KSTAT_DATA_UINT64 }, |
428870ff BB |
379 | { "evict_l2_cached", KSTAT_DATA_UINT64 }, |
380 | { "evict_l2_eligible", KSTAT_DATA_UINT64 }, | |
381 | { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, | |
ca0bf58d | 382 | { "evict_l2_skip", KSTAT_DATA_UINT64 }, |
34dc7c2f BB |
383 | { "hash_elements", KSTAT_DATA_UINT64 }, |
384 | { "hash_elements_max", KSTAT_DATA_UINT64 }, | |
385 | { "hash_collisions", KSTAT_DATA_UINT64 }, | |
386 | { "hash_chains", KSTAT_DATA_UINT64 }, | |
387 | { "hash_chain_max", KSTAT_DATA_UINT64 }, | |
388 | { "p", KSTAT_DATA_UINT64 }, | |
389 | { "c", KSTAT_DATA_UINT64 }, | |
390 | { "c_min", KSTAT_DATA_UINT64 }, | |
391 | { "c_max", KSTAT_DATA_UINT64 }, | |
392 | { "size", KSTAT_DATA_UINT64 }, | |
393 | { "hdr_size", KSTAT_DATA_UINT64 }, | |
d164b209 | 394 | { "data_size", KSTAT_DATA_UINT64 }, |
cc7f677c | 395 | { "meta_size", KSTAT_DATA_UINT64 }, |
d164b209 | 396 | { "other_size", KSTAT_DATA_UINT64 }, |
13be560d BB |
397 | { "anon_size", KSTAT_DATA_UINT64 }, |
398 | { "anon_evict_data", KSTAT_DATA_UINT64 }, | |
399 | { "anon_evict_metadata", KSTAT_DATA_UINT64 }, | |
400 | { "mru_size", KSTAT_DATA_UINT64 }, | |
401 | { "mru_evict_data", KSTAT_DATA_UINT64 }, | |
402 | { "mru_evict_metadata", KSTAT_DATA_UINT64 }, | |
403 | { "mru_ghost_size", KSTAT_DATA_UINT64 }, | |
404 | { "mru_ghost_evict_data", KSTAT_DATA_UINT64 }, | |
405 | { "mru_ghost_evict_metadata", KSTAT_DATA_UINT64 }, | |
406 | { "mfu_size", KSTAT_DATA_UINT64 }, | |
407 | { "mfu_evict_data", KSTAT_DATA_UINT64 }, | |
408 | { "mfu_evict_metadata", KSTAT_DATA_UINT64 }, | |
409 | { "mfu_ghost_size", KSTAT_DATA_UINT64 }, | |
410 | { "mfu_ghost_evict_data", KSTAT_DATA_UINT64 }, | |
411 | { "mfu_ghost_evict_metadata", KSTAT_DATA_UINT64 }, | |
34dc7c2f BB |
412 | { "l2_hits", KSTAT_DATA_UINT64 }, |
413 | { "l2_misses", KSTAT_DATA_UINT64 }, | |
414 | { "l2_feeds", KSTAT_DATA_UINT64 }, | |
415 | { "l2_rw_clash", KSTAT_DATA_UINT64 }, | |
d164b209 BB |
416 | { "l2_read_bytes", KSTAT_DATA_UINT64 }, |
417 | { "l2_write_bytes", KSTAT_DATA_UINT64 }, | |
34dc7c2f BB |
418 | { "l2_writes_sent", KSTAT_DATA_UINT64 }, |
419 | { "l2_writes_done", KSTAT_DATA_UINT64 }, | |
420 | { "l2_writes_error", KSTAT_DATA_UINT64 }, | |
ca0bf58d | 421 | { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, |
34dc7c2f BB |
422 | { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, |
423 | { "l2_evict_reading", KSTAT_DATA_UINT64 }, | |
b9541d6b | 424 | { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, |
34dc7c2f | 425 | { "l2_free_on_write", KSTAT_DATA_UINT64 }, |
ca0bf58d | 426 | { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, |
34dc7c2f BB |
427 | { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, |
428 | { "l2_cksum_bad", KSTAT_DATA_UINT64 }, | |
429 | { "l2_io_error", KSTAT_DATA_UINT64 }, | |
430 | { "l2_size", KSTAT_DATA_UINT64 }, | |
3a17a7a9 | 431 | { "l2_asize", KSTAT_DATA_UINT64 }, |
34dc7c2f | 432 | { "l2_hdr_size", KSTAT_DATA_UINT64 }, |
3a17a7a9 SK |
433 | { "l2_compress_successes", KSTAT_DATA_UINT64 }, |
434 | { "l2_compress_zeros", KSTAT_DATA_UINT64 }, | |
435 | { "l2_compress_failures", KSTAT_DATA_UINT64 }, | |
1834f2d8 | 436 | { "memory_throttle_count", KSTAT_DATA_UINT64 }, |
1eb5bfa3 GW |
437 | { "duplicate_buffers", KSTAT_DATA_UINT64 }, |
438 | { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, | |
439 | { "duplicate_reads", KSTAT_DATA_UINT64 }, | |
7cb67b45 BB |
440 | { "memory_direct_count", KSTAT_DATA_UINT64 }, |
441 | { "memory_indirect_count", KSTAT_DATA_UINT64 }, | |
1834f2d8 BB |
442 | { "arc_no_grow", KSTAT_DATA_UINT64 }, |
443 | { "arc_tempreserve", KSTAT_DATA_UINT64 }, | |
444 | { "arc_loaned_bytes", KSTAT_DATA_UINT64 }, | |
ab26409d | 445 | { "arc_prune", KSTAT_DATA_UINT64 }, |
1834f2d8 BB |
446 | { "arc_meta_used", KSTAT_DATA_UINT64 }, |
447 | { "arc_meta_limit", KSTAT_DATA_UINT64 }, | |
448 | { "arc_meta_max", KSTAT_DATA_UINT64 }, | |
ca0bf58d | 449 | { "arc_meta_min", KSTAT_DATA_UINT64 }, |
34dc7c2f BB |
450 | }; |
451 | ||
452 | #define ARCSTAT(stat) (arc_stats.stat.value.ui64) | |
453 | ||
454 | #define ARCSTAT_INCR(stat, val) \ | |
d3cc8b15 | 455 | atomic_add_64(&arc_stats.stat.value.ui64, (val)) |
34dc7c2f | 456 | |
428870ff | 457 | #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) |
34dc7c2f BB |
458 | #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) |
459 | ||
460 | #define ARCSTAT_MAX(stat, val) { \ | |
461 | uint64_t m; \ | |
462 | while ((val) > (m = arc_stats.stat.value.ui64) && \ | |
463 | (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ | |
464 | continue; \ | |
465 | } | |
466 | ||
467 | #define ARCSTAT_MAXSTAT(stat) \ | |
468 | ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) | |
469 | ||
470 | /* | |
471 | * We define a macro to allow ARC hits/misses to be easily broken down by | |
472 | * two separate conditions, giving a total of four different subtypes for | |
473 | * each of hits and misses (so eight statistics total). | |
474 | */ | |
475 | #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ | |
476 | if (cond1) { \ | |
477 | if (cond2) { \ | |
478 | ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ | |
479 | } else { \ | |
480 | ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ | |
481 | } \ | |
482 | } else { \ | |
483 | if (cond2) { \ | |
484 | ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ | |
485 | } else { \ | |
486 | ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ | |
487 | } \ | |
488 | } | |
489 | ||
490 | kstat_t *arc_ksp; | |
428870ff | 491 | static arc_state_t *arc_anon; |
34dc7c2f BB |
492 | static arc_state_t *arc_mru; |
493 | static arc_state_t *arc_mru_ghost; | |
494 | static arc_state_t *arc_mfu; | |
495 | static arc_state_t *arc_mfu_ghost; | |
496 | static arc_state_t *arc_l2c_only; | |
497 | ||
498 | /* | |
499 | * There are several ARC variables that are critical to export as kstats -- | |
500 | * but we don't want to have to grovel around in the kstat whenever we wish to | |
501 | * manipulate them. For these variables, we therefore define them to be in | |
502 | * terms of the statistic variable. This assures that we are not introducing | |
503 | * the possibility of inconsistency by having shadow copies of the variables, | |
504 | * while still allowing the code to be readable. | |
505 | */ | |
506 | #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ | |
507 | #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ | |
508 | #define arc_c ARCSTAT(arcstat_c) /* target size of cache */ | |
509 | #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ | |
510 | #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ | |
1834f2d8 BB |
511 | #define arc_no_grow ARCSTAT(arcstat_no_grow) |
512 | #define arc_tempreserve ARCSTAT(arcstat_tempreserve) | |
513 | #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes) | |
23c0a133 | 514 | #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ |
ca0bf58d | 515 | #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ |
23c0a133 GW |
516 | #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ |
517 | #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ | |
34dc7c2f | 518 | |
3a17a7a9 SK |
519 | #define L2ARC_IS_VALID_COMPRESS(_c_) \ |
520 | ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) | |
521 | ||
ab26409d BB |
522 | static list_t arc_prune_list; |
523 | static kmutex_t arc_prune_mtx; | |
f6046738 | 524 | static taskq_t *arc_prune_taskq; |
34dc7c2f | 525 | static arc_buf_t *arc_eviction_list; |
34dc7c2f | 526 | static arc_buf_hdr_t arc_eviction_hdr; |
428870ff | 527 | |
34dc7c2f BB |
528 | #define GHOST_STATE(state) \ |
529 | ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ | |
530 | (state) == arc_l2c_only) | |
531 | ||
2a432414 GW |
532 | #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) |
533 | #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) | |
534 | #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) | |
535 | #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) | |
536 | #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) | |
537 | #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) | |
b9541d6b | 538 | |
2a432414 | 539 | #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) |
b9541d6b | 540 | #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) |
2a432414 | 541 | #define HDR_L2_READING(hdr) \ |
b9541d6b CW |
542 | (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ |
543 | ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) | |
2a432414 GW |
544 | #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) |
545 | #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) | |
546 | #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) | |
34dc7c2f | 547 | |
b9541d6b CW |
548 | #define HDR_ISTYPE_METADATA(hdr) \ |
549 | ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) | |
550 | #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) | |
551 | ||
552 | #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) | |
553 | #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) | |
554 | ||
555 | /* For storing compression mode in b_flags */ | |
556 | #define HDR_COMPRESS_OFFSET 24 | |
557 | #define HDR_COMPRESS_NBITS 7 | |
558 | ||
559 | #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET(hdr->b_flags, \ | |
560 | HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS)) | |
561 | #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \ | |
562 | HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp)) | |
563 | ||
34dc7c2f BB |
564 | /* |
565 | * Other sizes | |
566 | */ | |
567 | ||
b9541d6b CW |
568 | #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) |
569 | #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) | |
34dc7c2f BB |
570 | |
571 | /* | |
572 | * Hash table routines | |
573 | */ | |
574 | ||
00b46022 BB |
575 | #define HT_LOCK_ALIGN 64 |
576 | #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN))) | |
34dc7c2f BB |
577 | |
578 | struct ht_lock { | |
579 | kmutex_t ht_lock; | |
580 | #ifdef _KERNEL | |
00b46022 | 581 | unsigned char pad[HT_LOCK_PAD]; |
34dc7c2f BB |
582 | #endif |
583 | }; | |
584 | ||
b31d8ea7 | 585 | #define BUF_LOCKS 8192 |
34dc7c2f BB |
586 | typedef struct buf_hash_table { |
587 | uint64_t ht_mask; | |
588 | arc_buf_hdr_t **ht_table; | |
589 | struct ht_lock ht_locks[BUF_LOCKS]; | |
590 | } buf_hash_table_t; | |
591 | ||
592 | static buf_hash_table_t buf_hash_table; | |
593 | ||
594 | #define BUF_HASH_INDEX(spa, dva, birth) \ | |
595 | (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) | |
596 | #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) | |
597 | #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) | |
428870ff BB |
598 | #define HDR_LOCK(hdr) \ |
599 | (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) | |
34dc7c2f BB |
600 | |
601 | uint64_t zfs_crc64_table[256]; | |
602 | ||
603 | /* | |
604 | * Level 2 ARC | |
605 | */ | |
606 | ||
607 | #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ | |
3a17a7a9 SK |
608 | #define L2ARC_HEADROOM 2 /* num of writes */ |
609 | /* | |
610 | * If we discover during ARC scan any buffers to be compressed, we boost | |
611 | * our headroom for the next scanning cycle by this percentage multiple. | |
612 | */ | |
613 | #define L2ARC_HEADROOM_BOOST 200 | |
d164b209 BB |
614 | #define L2ARC_FEED_SECS 1 /* caching interval secs */ |
615 | #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ | |
34dc7c2f | 616 | |
d962d5da PS |
617 | /* |
618 | * Used to distinguish headers that are being process by | |
619 | * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk | |
620 | * address. This can happen when the header is added to the l2arc's list | |
621 | * of buffers to write in the first stage of l2arc_write_buffers(), but | |
622 | * has not yet been written out which happens in the second stage of | |
623 | * l2arc_write_buffers(). | |
624 | */ | |
625 | #define L2ARC_ADDR_UNSET ((uint64_t)(-1)) | |
626 | ||
34dc7c2f BB |
627 | #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) |
628 | #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) | |
629 | ||
d3cc8b15 | 630 | /* L2ARC Performance Tunables */ |
abd8610c BB |
631 | unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */ |
632 | unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */ | |
633 | unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */ | |
3a17a7a9 | 634 | unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; |
abd8610c BB |
635 | unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ |
636 | unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */ | |
637 | int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ | |
3a17a7a9 | 638 | int l2arc_nocompress = B_FALSE; /* don't compress bufs */ |
abd8610c | 639 | int l2arc_feed_again = B_TRUE; /* turbo warmup */ |
c93504f0 | 640 | int l2arc_norw = B_FALSE; /* no reads during writes */ |
34dc7c2f BB |
641 | |
642 | /* | |
643 | * L2ARC Internals | |
644 | */ | |
34dc7c2f BB |
645 | static list_t L2ARC_dev_list; /* device list */ |
646 | static list_t *l2arc_dev_list; /* device list pointer */ | |
647 | static kmutex_t l2arc_dev_mtx; /* device list mutex */ | |
648 | static l2arc_dev_t *l2arc_dev_last; /* last device used */ | |
34dc7c2f BB |
649 | static list_t L2ARC_free_on_write; /* free after write buf list */ |
650 | static list_t *l2arc_free_on_write; /* free after write list ptr */ | |
651 | static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ | |
652 | static uint64_t l2arc_ndev; /* number of devices */ | |
653 | ||
654 | typedef struct l2arc_read_callback { | |
3a17a7a9 SK |
655 | arc_buf_t *l2rcb_buf; /* read buffer */ |
656 | spa_t *l2rcb_spa; /* spa */ | |
657 | blkptr_t l2rcb_bp; /* original blkptr */ | |
5dbd68a3 | 658 | zbookmark_phys_t l2rcb_zb; /* original bookmark */ |
3a17a7a9 SK |
659 | int l2rcb_flags; /* original flags */ |
660 | enum zio_compress l2rcb_compress; /* applied compress */ | |
34dc7c2f BB |
661 | } l2arc_read_callback_t; |
662 | ||
34dc7c2f BB |
663 | typedef struct l2arc_data_free { |
664 | /* protected by l2arc_free_on_write_mtx */ | |
665 | void *l2df_data; | |
666 | size_t l2df_size; | |
667 | void (*l2df_func)(void *, size_t); | |
668 | list_node_t l2df_list_node; | |
669 | } l2arc_data_free_t; | |
670 | ||
671 | static kmutex_t l2arc_feed_thr_lock; | |
672 | static kcondvar_t l2arc_feed_thr_cv; | |
673 | static uint8_t l2arc_thread_exit; | |
674 | ||
2a432414 GW |
675 | static void arc_get_data_buf(arc_buf_t *); |
676 | static void arc_access(arc_buf_hdr_t *, kmutex_t *); | |
ca0bf58d | 677 | static boolean_t arc_is_overflowing(void); |
2a432414 | 678 | static void arc_buf_watch(arc_buf_t *); |
ca67b33a | 679 | static void arc_tuning_update(void); |
2a432414 | 680 | |
b9541d6b CW |
681 | static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); |
682 | static uint32_t arc_bufc_to_flags(arc_buf_contents_t); | |
683 | ||
2a432414 GW |
684 | static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); |
685 | static void l2arc_read_done(zio_t *); | |
34dc7c2f | 686 | |
b9541d6b | 687 | static boolean_t l2arc_compress_buf(arc_buf_hdr_t *); |
2a432414 GW |
688 | static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); |
689 | static void l2arc_release_cdata_buf(arc_buf_hdr_t *); | |
3a17a7a9 | 690 | |
34dc7c2f | 691 | static uint64_t |
d164b209 | 692 | buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) |
34dc7c2f | 693 | { |
34dc7c2f BB |
694 | uint8_t *vdva = (uint8_t *)dva; |
695 | uint64_t crc = -1ULL; | |
696 | int i; | |
697 | ||
698 | ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); | |
699 | ||
700 | for (i = 0; i < sizeof (dva_t); i++) | |
701 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; | |
702 | ||
d164b209 | 703 | crc ^= (spa>>8) ^ birth; |
34dc7c2f BB |
704 | |
705 | return (crc); | |
706 | } | |
707 | ||
708 | #define BUF_EMPTY(buf) \ | |
709 | ((buf)->b_dva.dva_word[0] == 0 && \ | |
b9541d6b | 710 | (buf)->b_dva.dva_word[1] == 0) |
34dc7c2f BB |
711 | |
712 | #define BUF_EQUAL(spa, dva, birth, buf) \ | |
713 | ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ | |
714 | ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ | |
715 | ((buf)->b_birth == birth) && ((buf)->b_spa == spa) | |
716 | ||
428870ff BB |
717 | static void |
718 | buf_discard_identity(arc_buf_hdr_t *hdr) | |
719 | { | |
720 | hdr->b_dva.dva_word[0] = 0; | |
721 | hdr->b_dva.dva_word[1] = 0; | |
722 | hdr->b_birth = 0; | |
428870ff BB |
723 | } |
724 | ||
34dc7c2f | 725 | static arc_buf_hdr_t * |
9b67f605 | 726 | buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) |
34dc7c2f | 727 | { |
9b67f605 MA |
728 | const dva_t *dva = BP_IDENTITY(bp); |
729 | uint64_t birth = BP_PHYSICAL_BIRTH(bp); | |
34dc7c2f BB |
730 | uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); |
731 | kmutex_t *hash_lock = BUF_HASH_LOCK(idx); | |
2a432414 | 732 | arc_buf_hdr_t *hdr; |
34dc7c2f BB |
733 | |
734 | mutex_enter(hash_lock); | |
2a432414 GW |
735 | for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; |
736 | hdr = hdr->b_hash_next) { | |
737 | if (BUF_EQUAL(spa, dva, birth, hdr)) { | |
34dc7c2f | 738 | *lockp = hash_lock; |
2a432414 | 739 | return (hdr); |
34dc7c2f BB |
740 | } |
741 | } | |
742 | mutex_exit(hash_lock); | |
743 | *lockp = NULL; | |
744 | return (NULL); | |
745 | } | |
746 | ||
747 | /* | |
748 | * Insert an entry into the hash table. If there is already an element | |
749 | * equal to elem in the hash table, then the already existing element | |
750 | * will be returned and the new element will not be inserted. | |
751 | * Otherwise returns NULL. | |
b9541d6b | 752 | * If lockp == NULL, the caller is assumed to already hold the hash lock. |
34dc7c2f BB |
753 | */ |
754 | static arc_buf_hdr_t * | |
2a432414 | 755 | buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) |
34dc7c2f | 756 | { |
2a432414 | 757 | uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); |
34dc7c2f | 758 | kmutex_t *hash_lock = BUF_HASH_LOCK(idx); |
2a432414 | 759 | arc_buf_hdr_t *fhdr; |
34dc7c2f BB |
760 | uint32_t i; |
761 | ||
2a432414 GW |
762 | ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); |
763 | ASSERT(hdr->b_birth != 0); | |
764 | ASSERT(!HDR_IN_HASH_TABLE(hdr)); | |
b9541d6b CW |
765 | |
766 | if (lockp != NULL) { | |
767 | *lockp = hash_lock; | |
768 | mutex_enter(hash_lock); | |
769 | } else { | |
770 | ASSERT(MUTEX_HELD(hash_lock)); | |
771 | } | |
772 | ||
2a432414 GW |
773 | for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; |
774 | fhdr = fhdr->b_hash_next, i++) { | |
775 | if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) | |
776 | return (fhdr); | |
34dc7c2f BB |
777 | } |
778 | ||
2a432414 GW |
779 | hdr->b_hash_next = buf_hash_table.ht_table[idx]; |
780 | buf_hash_table.ht_table[idx] = hdr; | |
781 | hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; | |
34dc7c2f BB |
782 | |
783 | /* collect some hash table performance data */ | |
784 | if (i > 0) { | |
785 | ARCSTAT_BUMP(arcstat_hash_collisions); | |
786 | if (i == 1) | |
787 | ARCSTAT_BUMP(arcstat_hash_chains); | |
788 | ||
789 | ARCSTAT_MAX(arcstat_hash_chain_max, i); | |
790 | } | |
791 | ||
792 | ARCSTAT_BUMP(arcstat_hash_elements); | |
793 | ARCSTAT_MAXSTAT(arcstat_hash_elements); | |
794 | ||
795 | return (NULL); | |
796 | } | |
797 | ||
798 | static void | |
2a432414 | 799 | buf_hash_remove(arc_buf_hdr_t *hdr) |
34dc7c2f | 800 | { |
2a432414 GW |
801 | arc_buf_hdr_t *fhdr, **hdrp; |
802 | uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); | |
34dc7c2f BB |
803 | |
804 | ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); | |
2a432414 | 805 | ASSERT(HDR_IN_HASH_TABLE(hdr)); |
34dc7c2f | 806 | |
2a432414 GW |
807 | hdrp = &buf_hash_table.ht_table[idx]; |
808 | while ((fhdr = *hdrp) != hdr) { | |
809 | ASSERT(fhdr != NULL); | |
810 | hdrp = &fhdr->b_hash_next; | |
34dc7c2f | 811 | } |
2a432414 GW |
812 | *hdrp = hdr->b_hash_next; |
813 | hdr->b_hash_next = NULL; | |
814 | hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; | |
34dc7c2f BB |
815 | |
816 | /* collect some hash table performance data */ | |
817 | ARCSTAT_BUMPDOWN(arcstat_hash_elements); | |
818 | ||
819 | if (buf_hash_table.ht_table[idx] && | |
820 | buf_hash_table.ht_table[idx]->b_hash_next == NULL) | |
821 | ARCSTAT_BUMPDOWN(arcstat_hash_chains); | |
822 | } | |
823 | ||
824 | /* | |
825 | * Global data structures and functions for the buf kmem cache. | |
826 | */ | |
b9541d6b CW |
827 | static kmem_cache_t *hdr_full_cache; |
828 | static kmem_cache_t *hdr_l2only_cache; | |
34dc7c2f BB |
829 | static kmem_cache_t *buf_cache; |
830 | ||
831 | static void | |
832 | buf_fini(void) | |
833 | { | |
834 | int i; | |
835 | ||
00b46022 | 836 | #if defined(_KERNEL) && defined(HAVE_SPL) |
d1d7e268 MK |
837 | /* |
838 | * Large allocations which do not require contiguous pages | |
839 | * should be using vmem_free() in the linux kernel\ | |
840 | */ | |
00b46022 BB |
841 | vmem_free(buf_hash_table.ht_table, |
842 | (buf_hash_table.ht_mask + 1) * sizeof (void *)); | |
843 | #else | |
34dc7c2f BB |
844 | kmem_free(buf_hash_table.ht_table, |
845 | (buf_hash_table.ht_mask + 1) * sizeof (void *)); | |
00b46022 | 846 | #endif |
34dc7c2f BB |
847 | for (i = 0; i < BUF_LOCKS; i++) |
848 | mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); | |
b9541d6b CW |
849 | kmem_cache_destroy(hdr_full_cache); |
850 | kmem_cache_destroy(hdr_l2only_cache); | |
34dc7c2f BB |
851 | kmem_cache_destroy(buf_cache); |
852 | } | |
853 | ||
854 | /* | |
855 | * Constructor callback - called when the cache is empty | |
856 | * and a new buf is requested. | |
857 | */ | |
858 | /* ARGSUSED */ | |
859 | static int | |
b9541d6b CW |
860 | hdr_full_cons(void *vbuf, void *unused, int kmflag) |
861 | { | |
862 | arc_buf_hdr_t *hdr = vbuf; | |
863 | ||
864 | bzero(hdr, HDR_FULL_SIZE); | |
865 | cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); | |
866 | refcount_create(&hdr->b_l1hdr.b_refcnt); | |
867 | mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); | |
868 | list_link_init(&hdr->b_l1hdr.b_arc_node); | |
869 | list_link_init(&hdr->b_l2hdr.b_l2node); | |
ca0bf58d | 870 | multilist_link_init(&hdr->b_l1hdr.b_arc_node); |
b9541d6b CW |
871 | arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); |
872 | ||
873 | return (0); | |
874 | } | |
875 | ||
876 | /* ARGSUSED */ | |
877 | static int | |
878 | hdr_l2only_cons(void *vbuf, void *unused, int kmflag) | |
34dc7c2f | 879 | { |
2a432414 GW |
880 | arc_buf_hdr_t *hdr = vbuf; |
881 | ||
b9541d6b CW |
882 | bzero(hdr, HDR_L2ONLY_SIZE); |
883 | arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); | |
34dc7c2f | 884 | |
34dc7c2f BB |
885 | return (0); |
886 | } | |
887 | ||
b128c09f BB |
888 | /* ARGSUSED */ |
889 | static int | |
890 | buf_cons(void *vbuf, void *unused, int kmflag) | |
891 | { | |
892 | arc_buf_t *buf = vbuf; | |
893 | ||
894 | bzero(buf, sizeof (arc_buf_t)); | |
428870ff | 895 | mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); |
d164b209 BB |
896 | arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); |
897 | ||
b128c09f BB |
898 | return (0); |
899 | } | |
900 | ||
34dc7c2f BB |
901 | /* |
902 | * Destructor callback - called when a cached buf is | |
903 | * no longer required. | |
904 | */ | |
905 | /* ARGSUSED */ | |
906 | static void | |
b9541d6b | 907 | hdr_full_dest(void *vbuf, void *unused) |
34dc7c2f | 908 | { |
2a432414 | 909 | arc_buf_hdr_t *hdr = vbuf; |
34dc7c2f | 910 | |
2a432414 | 911 | ASSERT(BUF_EMPTY(hdr)); |
b9541d6b CW |
912 | cv_destroy(&hdr->b_l1hdr.b_cv); |
913 | refcount_destroy(&hdr->b_l1hdr.b_refcnt); | |
914 | mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); | |
ca0bf58d | 915 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); |
b9541d6b CW |
916 | arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); |
917 | } | |
918 | ||
919 | /* ARGSUSED */ | |
920 | static void | |
921 | hdr_l2only_dest(void *vbuf, void *unused) | |
922 | { | |
923 | ASSERTV(arc_buf_hdr_t *hdr = vbuf); | |
924 | ||
925 | ASSERT(BUF_EMPTY(hdr)); | |
926 | arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); | |
34dc7c2f BB |
927 | } |
928 | ||
b128c09f BB |
929 | /* ARGSUSED */ |
930 | static void | |
931 | buf_dest(void *vbuf, void *unused) | |
932 | { | |
933 | arc_buf_t *buf = vbuf; | |
934 | ||
428870ff | 935 | mutex_destroy(&buf->b_evict_lock); |
d164b209 | 936 | arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); |
b128c09f BB |
937 | } |
938 | ||
34dc7c2f BB |
939 | static void |
940 | buf_init(void) | |
941 | { | |
942 | uint64_t *ct; | |
943 | uint64_t hsize = 1ULL << 12; | |
944 | int i, j; | |
945 | ||
946 | /* | |
947 | * The hash table is big enough to fill all of physical memory | |
49ddb315 MA |
948 | * with an average block size of zfs_arc_average_blocksize (default 8K). |
949 | * By default, the table will take up | |
950 | * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). | |
34dc7c2f | 951 | */ |
49ddb315 | 952 | while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) |
34dc7c2f BB |
953 | hsize <<= 1; |
954 | retry: | |
955 | buf_hash_table.ht_mask = hsize - 1; | |
00b46022 | 956 | #if defined(_KERNEL) && defined(HAVE_SPL) |
d1d7e268 MK |
957 | /* |
958 | * Large allocations which do not require contiguous pages | |
959 | * should be using vmem_alloc() in the linux kernel | |
960 | */ | |
00b46022 BB |
961 | buf_hash_table.ht_table = |
962 | vmem_zalloc(hsize * sizeof (void*), KM_SLEEP); | |
963 | #else | |
34dc7c2f BB |
964 | buf_hash_table.ht_table = |
965 | kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); | |
00b46022 | 966 | #endif |
34dc7c2f BB |
967 | if (buf_hash_table.ht_table == NULL) { |
968 | ASSERT(hsize > (1ULL << 8)); | |
969 | hsize >>= 1; | |
970 | goto retry; | |
971 | } | |
972 | ||
b9541d6b CW |
973 | hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, |
974 | 0, hdr_full_cons, hdr_full_dest, NULL, NULL, NULL, 0); | |
975 | hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", | |
976 | HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, NULL, | |
977 | NULL, NULL, 0); | |
34dc7c2f | 978 | buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), |
b128c09f | 979 | 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); |
34dc7c2f BB |
980 | |
981 | for (i = 0; i < 256; i++) | |
982 | for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) | |
983 | *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); | |
984 | ||
985 | for (i = 0; i < BUF_LOCKS; i++) { | |
986 | mutex_init(&buf_hash_table.ht_locks[i].ht_lock, | |
40d06e3c | 987 | NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f BB |
988 | } |
989 | } | |
990 | ||
b9541d6b CW |
991 | /* |
992 | * Transition between the two allocation states for the arc_buf_hdr struct. | |
993 | * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without | |
994 | * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller | |
995 | * version is used when a cache buffer is only in the L2ARC in order to reduce | |
996 | * memory usage. | |
997 | */ | |
998 | static arc_buf_hdr_t * | |
999 | arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) | |
1000 | { | |
1001 | arc_buf_hdr_t *nhdr; | |
1002 | l2arc_dev_t *dev; | |
1003 | ||
1004 | ASSERT(HDR_HAS_L2HDR(hdr)); | |
1005 | ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || | |
1006 | (old == hdr_l2only_cache && new == hdr_full_cache)); | |
1007 | ||
1008 | dev = hdr->b_l2hdr.b_dev; | |
1009 | nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); | |
1010 | ||
1011 | ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); | |
1012 | buf_hash_remove(hdr); | |
1013 | ||
1014 | bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); | |
d962d5da | 1015 | |
b9541d6b CW |
1016 | if (new == hdr_full_cache) { |
1017 | nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; | |
1018 | /* | |
1019 | * arc_access and arc_change_state need to be aware that a | |
1020 | * header has just come out of L2ARC, so we set its state to | |
1021 | * l2c_only even though it's about to change. | |
1022 | */ | |
1023 | nhdr->b_l1hdr.b_state = arc_l2c_only; | |
ca0bf58d PS |
1024 | |
1025 | /* Verify previous threads set to NULL before freeing */ | |
1026 | ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL); | |
b9541d6b CW |
1027 | } else { |
1028 | ASSERT(hdr->b_l1hdr.b_buf == NULL); | |
1029 | ASSERT0(hdr->b_l1hdr.b_datacnt); | |
ca0bf58d PS |
1030 | |
1031 | /* | |
1032 | * If we've reached here, We must have been called from | |
1033 | * arc_evict_hdr(), as such we should have already been | |
1034 | * removed from any ghost list we were previously on | |
1035 | * (which protects us from racing with arc_evict_state), | |
1036 | * thus no locking is needed during this check. | |
1037 | */ | |
1038 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); | |
1039 | ||
b9541d6b | 1040 | /* |
ca0bf58d PS |
1041 | * A buffer must not be moved into the arc_l2c_only |
1042 | * state if it's not finished being written out to the | |
1043 | * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field | |
1044 | * might try to be accessed, even though it was removed. | |
b9541d6b | 1045 | */ |
ca0bf58d PS |
1046 | VERIFY(!HDR_L2_WRITING(hdr)); |
1047 | VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); | |
1048 | ||
b9541d6b CW |
1049 | nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; |
1050 | } | |
1051 | /* | |
1052 | * The header has been reallocated so we need to re-insert it into any | |
1053 | * lists it was on. | |
1054 | */ | |
1055 | (void) buf_hash_insert(nhdr, NULL); | |
1056 | ||
1057 | ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); | |
1058 | ||
1059 | mutex_enter(&dev->l2ad_mtx); | |
1060 | ||
1061 | /* | |
1062 | * We must place the realloc'ed header back into the list at | |
1063 | * the same spot. Otherwise, if it's placed earlier in the list, | |
1064 | * l2arc_write_buffers() could find it during the function's | |
1065 | * write phase, and try to write it out to the l2arc. | |
1066 | */ | |
1067 | list_insert_after(&dev->l2ad_buflist, hdr, nhdr); | |
1068 | list_remove(&dev->l2ad_buflist, hdr); | |
1069 | ||
1070 | mutex_exit(&dev->l2ad_mtx); | |
1071 | ||
d962d5da PS |
1072 | /* |
1073 | * Since we're using the pointer address as the tag when | |
1074 | * incrementing and decrementing the l2ad_alloc refcount, we | |
1075 | * must remove the old pointer (that we're about to destroy) and | |
1076 | * add the new pointer to the refcount. Otherwise we'd remove | |
1077 | * the wrong pointer address when calling arc_hdr_destroy() later. | |
1078 | */ | |
1079 | ||
1080 | (void) refcount_remove_many(&dev->l2ad_alloc, | |
1081 | hdr->b_l2hdr.b_asize, hdr); | |
1082 | ||
1083 | (void) refcount_add_many(&dev->l2ad_alloc, | |
1084 | nhdr->b_l2hdr.b_asize, nhdr); | |
1085 | ||
b9541d6b CW |
1086 | buf_discard_identity(hdr); |
1087 | hdr->b_freeze_cksum = NULL; | |
1088 | kmem_cache_free(old, hdr); | |
1089 | ||
1090 | return (nhdr); | |
1091 | } | |
1092 | ||
1093 | ||
34dc7c2f BB |
1094 | #define ARC_MINTIME (hz>>4) /* 62 ms */ |
1095 | ||
1096 | static void | |
1097 | arc_cksum_verify(arc_buf_t *buf) | |
1098 | { | |
1099 | zio_cksum_t zc; | |
1100 | ||
1101 | if (!(zfs_flags & ZFS_DEBUG_MODIFY)) | |
1102 | return; | |
1103 | ||
b9541d6b CW |
1104 | mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); |
1105 | if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { | |
1106 | mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); | |
34dc7c2f BB |
1107 | return; |
1108 | } | |
1109 | fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); | |
1110 | if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) | |
1111 | panic("buffer modified while frozen!"); | |
b9541d6b | 1112 | mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); |
34dc7c2f BB |
1113 | } |
1114 | ||
1115 | static int | |
1116 | arc_cksum_equal(arc_buf_t *buf) | |
1117 | { | |
1118 | zio_cksum_t zc; | |
1119 | int equal; | |
1120 | ||
b9541d6b | 1121 | mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); |
34dc7c2f BB |
1122 | fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); |
1123 | equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); | |
b9541d6b | 1124 | mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); |
34dc7c2f BB |
1125 | |
1126 | return (equal); | |
1127 | } | |
1128 | ||
1129 | static void | |
1130 | arc_cksum_compute(arc_buf_t *buf, boolean_t force) | |
1131 | { | |
1132 | if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) | |
1133 | return; | |
1134 | ||
b9541d6b | 1135 | mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); |
34dc7c2f | 1136 | if (buf->b_hdr->b_freeze_cksum != NULL) { |
b9541d6b | 1137 | mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); |
34dc7c2f BB |
1138 | return; |
1139 | } | |
409dc1a5 | 1140 | buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), |
79c76d5b | 1141 | KM_SLEEP); |
34dc7c2f BB |
1142 | fletcher_2_native(buf->b_data, buf->b_hdr->b_size, |
1143 | buf->b_hdr->b_freeze_cksum); | |
b9541d6b | 1144 | mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); |
498877ba MA |
1145 | arc_buf_watch(buf); |
1146 | } | |
1147 | ||
1148 | #ifndef _KERNEL | |
1149 | void | |
1150 | arc_buf_sigsegv(int sig, siginfo_t *si, void *unused) | |
1151 | { | |
1152 | panic("Got SIGSEGV at address: 0x%lx\n", (long) si->si_addr); | |
1153 | } | |
1154 | #endif | |
1155 | ||
1156 | /* ARGSUSED */ | |
1157 | static void | |
1158 | arc_buf_unwatch(arc_buf_t *buf) | |
1159 | { | |
1160 | #ifndef _KERNEL | |
1161 | if (arc_watch) { | |
1162 | ASSERT0(mprotect(buf->b_data, buf->b_hdr->b_size, | |
1163 | PROT_READ | PROT_WRITE)); | |
1164 | } | |
1165 | #endif | |
1166 | } | |
1167 | ||
1168 | /* ARGSUSED */ | |
1169 | static void | |
1170 | arc_buf_watch(arc_buf_t *buf) | |
1171 | { | |
1172 | #ifndef _KERNEL | |
1173 | if (arc_watch) | |
1174 | ASSERT0(mprotect(buf->b_data, buf->b_hdr->b_size, PROT_READ)); | |
1175 | #endif | |
34dc7c2f BB |
1176 | } |
1177 | ||
b9541d6b CW |
1178 | static arc_buf_contents_t |
1179 | arc_buf_type(arc_buf_hdr_t *hdr) | |
1180 | { | |
1181 | if (HDR_ISTYPE_METADATA(hdr)) { | |
1182 | return (ARC_BUFC_METADATA); | |
1183 | } else { | |
1184 | return (ARC_BUFC_DATA); | |
1185 | } | |
1186 | } | |
1187 | ||
1188 | static uint32_t | |
1189 | arc_bufc_to_flags(arc_buf_contents_t type) | |
1190 | { | |
1191 | switch (type) { | |
1192 | case ARC_BUFC_DATA: | |
1193 | /* metadata field is 0 if buffer contains normal data */ | |
1194 | return (0); | |
1195 | case ARC_BUFC_METADATA: | |
1196 | return (ARC_FLAG_BUFC_METADATA); | |
1197 | default: | |
1198 | break; | |
1199 | } | |
1200 | panic("undefined ARC buffer type!"); | |
1201 | return ((uint32_t)-1); | |
1202 | } | |
1203 | ||
34dc7c2f BB |
1204 | void |
1205 | arc_buf_thaw(arc_buf_t *buf) | |
1206 | { | |
1207 | if (zfs_flags & ZFS_DEBUG_MODIFY) { | |
b9541d6b | 1208 | if (buf->b_hdr->b_l1hdr.b_state != arc_anon) |
34dc7c2f | 1209 | panic("modifying non-anon buffer!"); |
b9541d6b | 1210 | if (HDR_IO_IN_PROGRESS(buf->b_hdr)) |
34dc7c2f BB |
1211 | panic("modifying buffer while i/o in progress!"); |
1212 | arc_cksum_verify(buf); | |
1213 | } | |
1214 | ||
b9541d6b | 1215 | mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); |
34dc7c2f BB |
1216 | if (buf->b_hdr->b_freeze_cksum != NULL) { |
1217 | kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); | |
1218 | buf->b_hdr->b_freeze_cksum = NULL; | |
1219 | } | |
428870ff | 1220 | |
b9541d6b | 1221 | mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); |
498877ba MA |
1222 | |
1223 | arc_buf_unwatch(buf); | |
34dc7c2f BB |
1224 | } |
1225 | ||
1226 | void | |
1227 | arc_buf_freeze(arc_buf_t *buf) | |
1228 | { | |
428870ff BB |
1229 | kmutex_t *hash_lock; |
1230 | ||
34dc7c2f BB |
1231 | if (!(zfs_flags & ZFS_DEBUG_MODIFY)) |
1232 | return; | |
1233 | ||
428870ff BB |
1234 | hash_lock = HDR_LOCK(buf->b_hdr); |
1235 | mutex_enter(hash_lock); | |
1236 | ||
34dc7c2f | 1237 | ASSERT(buf->b_hdr->b_freeze_cksum != NULL || |
b9541d6b | 1238 | buf->b_hdr->b_l1hdr.b_state == arc_anon); |
34dc7c2f | 1239 | arc_cksum_compute(buf, B_FALSE); |
428870ff | 1240 | mutex_exit(hash_lock); |
498877ba | 1241 | |
34dc7c2f BB |
1242 | } |
1243 | ||
1244 | static void | |
2a432414 | 1245 | add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) |
34dc7c2f | 1246 | { |
b9541d6b CW |
1247 | arc_state_t *state; |
1248 | ||
1249 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
34dc7c2f BB |
1250 | ASSERT(MUTEX_HELD(hash_lock)); |
1251 | ||
b9541d6b CW |
1252 | state = hdr->b_l1hdr.b_state; |
1253 | ||
1254 | if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && | |
1255 | (state != arc_anon)) { | |
1256 | /* We don't use the L2-only state list. */ | |
1257 | if (state != arc_l2c_only) { | |
ca0bf58d | 1258 | arc_buf_contents_t type = arc_buf_type(hdr); |
b9541d6b | 1259 | uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; |
ca0bf58d PS |
1260 | multilist_t *list = &state->arcs_list[type]; |
1261 | uint64_t *size = &state->arcs_lsize[type]; | |
1262 | ||
1263 | multilist_remove(list, hdr); | |
b9541d6b | 1264 | |
b9541d6b CW |
1265 | if (GHOST_STATE(state)) { |
1266 | ASSERT0(hdr->b_l1hdr.b_datacnt); | |
1267 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); | |
1268 | delta = hdr->b_size; | |
1269 | } | |
1270 | ASSERT(delta > 0); | |
1271 | ASSERT3U(*size, >=, delta); | |
1272 | atomic_add_64(size, -delta); | |
34dc7c2f | 1273 | } |
b128c09f | 1274 | /* remove the prefetch flag if we get a reference */ |
b9541d6b | 1275 | hdr->b_flags &= ~ARC_FLAG_PREFETCH; |
34dc7c2f BB |
1276 | } |
1277 | } | |
1278 | ||
1279 | static int | |
2a432414 | 1280 | remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) |
34dc7c2f BB |
1281 | { |
1282 | int cnt; | |
b9541d6b | 1283 | arc_state_t *state = hdr->b_l1hdr.b_state; |
34dc7c2f | 1284 | |
b9541d6b | 1285 | ASSERT(HDR_HAS_L1HDR(hdr)); |
34dc7c2f BB |
1286 | ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); |
1287 | ASSERT(!GHOST_STATE(state)); | |
1288 | ||
b9541d6b CW |
1289 | /* |
1290 | * arc_l2c_only counts as a ghost state so we don't need to explicitly | |
1291 | * check to prevent usage of the arc_l2c_only list. | |
1292 | */ | |
1293 | if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && | |
34dc7c2f | 1294 | (state != arc_anon)) { |
ca0bf58d PS |
1295 | arc_buf_contents_t type = arc_buf_type(hdr); |
1296 | multilist_t *list = &state->arcs_list[type]; | |
1297 | uint64_t *size = &state->arcs_lsize[type]; | |
1298 | ||
1299 | multilist_insert(list, hdr); | |
34dc7c2f | 1300 | |
b9541d6b CW |
1301 | ASSERT(hdr->b_l1hdr.b_datacnt > 0); |
1302 | atomic_add_64(size, hdr->b_size * | |
1303 | hdr->b_l1hdr.b_datacnt); | |
34dc7c2f BB |
1304 | } |
1305 | return (cnt); | |
1306 | } | |
1307 | ||
e0b0ca98 BB |
1308 | /* |
1309 | * Returns detailed information about a specific arc buffer. When the | |
1310 | * state_index argument is set the function will calculate the arc header | |
1311 | * list position for its arc state. Since this requires a linear traversal | |
1312 | * callers are strongly encourage not to do this. However, it can be helpful | |
1313 | * for targeted analysis so the functionality is provided. | |
1314 | */ | |
1315 | void | |
1316 | arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index) | |
1317 | { | |
1318 | arc_buf_hdr_t *hdr = ab->b_hdr; | |
b9541d6b CW |
1319 | l1arc_buf_hdr_t *l1hdr = NULL; |
1320 | l2arc_buf_hdr_t *l2hdr = NULL; | |
1321 | arc_state_t *state = NULL; | |
1322 | ||
1323 | if (HDR_HAS_L1HDR(hdr)) { | |
1324 | l1hdr = &hdr->b_l1hdr; | |
1325 | state = l1hdr->b_state; | |
1326 | } | |
1327 | if (HDR_HAS_L2HDR(hdr)) | |
1328 | l2hdr = &hdr->b_l2hdr; | |
e0b0ca98 | 1329 | |
d1d7e268 | 1330 | memset(abi, 0, sizeof (arc_buf_info_t)); |
e0b0ca98 | 1331 | abi->abi_flags = hdr->b_flags; |
b9541d6b CW |
1332 | |
1333 | if (l1hdr) { | |
1334 | abi->abi_datacnt = l1hdr->b_datacnt; | |
1335 | abi->abi_access = l1hdr->b_arc_access; | |
1336 | abi->abi_mru_hits = l1hdr->b_mru_hits; | |
1337 | abi->abi_mru_ghost_hits = l1hdr->b_mru_ghost_hits; | |
1338 | abi->abi_mfu_hits = l1hdr->b_mfu_hits; | |
1339 | abi->abi_mfu_ghost_hits = l1hdr->b_mfu_ghost_hits; | |
1340 | abi->abi_holds = refcount_count(&l1hdr->b_refcnt); | |
1341 | } | |
1342 | ||
1343 | if (l2hdr) { | |
1344 | abi->abi_l2arc_dattr = l2hdr->b_daddr; | |
1345 | abi->abi_l2arc_asize = l2hdr->b_asize; | |
1346 | abi->abi_l2arc_compress = HDR_GET_COMPRESS(hdr); | |
1347 | abi->abi_l2arc_hits = l2hdr->b_hits; | |
1348 | } | |
1349 | ||
e0b0ca98 | 1350 | abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON; |
b9541d6b | 1351 | abi->abi_state_contents = arc_buf_type(hdr); |
e0b0ca98 | 1352 | abi->abi_size = hdr->b_size; |
e0b0ca98 BB |
1353 | } |
1354 | ||
34dc7c2f | 1355 | /* |
ca0bf58d | 1356 | * Move the supplied buffer to the indicated state. The hash lock |
34dc7c2f BB |
1357 | * for the buffer must be held by the caller. |
1358 | */ | |
1359 | static void | |
2a432414 GW |
1360 | arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, |
1361 | kmutex_t *hash_lock) | |
34dc7c2f | 1362 | { |
b9541d6b CW |
1363 | arc_state_t *old_state; |
1364 | int64_t refcnt; | |
1365 | uint32_t datacnt; | |
34dc7c2f | 1366 | uint64_t from_delta, to_delta; |
b9541d6b CW |
1367 | arc_buf_contents_t buftype = arc_buf_type(hdr); |
1368 | ||
1369 | /* | |
1370 | * We almost always have an L1 hdr here, since we call arc_hdr_realloc() | |
1371 | * in arc_read() when bringing a buffer out of the L2ARC. However, the | |
1372 | * L1 hdr doesn't always exist when we change state to arc_anon before | |
1373 | * destroying a header, in which case reallocating to add the L1 hdr is | |
1374 | * pointless. | |
1375 | */ | |
1376 | if (HDR_HAS_L1HDR(hdr)) { | |
1377 | old_state = hdr->b_l1hdr.b_state; | |
1378 | refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); | |
1379 | datacnt = hdr->b_l1hdr.b_datacnt; | |
1380 | } else { | |
1381 | old_state = arc_l2c_only; | |
1382 | refcnt = 0; | |
1383 | datacnt = 0; | |
1384 | } | |
34dc7c2f BB |
1385 | |
1386 | ASSERT(MUTEX_HELD(hash_lock)); | |
e8b96c60 | 1387 | ASSERT3P(new_state, !=, old_state); |
b9541d6b CW |
1388 | ASSERT(refcnt == 0 || datacnt > 0); |
1389 | ASSERT(!GHOST_STATE(new_state) || datacnt == 0); | |
1390 | ASSERT(old_state != arc_anon || datacnt <= 1); | |
34dc7c2f | 1391 | |
b9541d6b | 1392 | from_delta = to_delta = datacnt * hdr->b_size; |
34dc7c2f BB |
1393 | |
1394 | /* | |
1395 | * If this buffer is evictable, transfer it from the | |
1396 | * old state list to the new state list. | |
1397 | */ | |
1398 | if (refcnt == 0) { | |
b9541d6b | 1399 | if (old_state != arc_anon && old_state != arc_l2c_only) { |
b9541d6b | 1400 | uint64_t *size = &old_state->arcs_lsize[buftype]; |
34dc7c2f | 1401 | |
b9541d6b | 1402 | ASSERT(HDR_HAS_L1HDR(hdr)); |
ca0bf58d | 1403 | multilist_remove(&old_state->arcs_list[buftype], hdr); |
34dc7c2f BB |
1404 | |
1405 | /* | |
1406 | * If prefetching out of the ghost cache, | |
428870ff | 1407 | * we will have a non-zero datacnt. |
34dc7c2f | 1408 | */ |
b9541d6b | 1409 | if (GHOST_STATE(old_state) && datacnt == 0) { |
34dc7c2f | 1410 | /* ghost elements have a ghost size */ |
b9541d6b | 1411 | ASSERT(hdr->b_l1hdr.b_buf == NULL); |
2a432414 | 1412 | from_delta = hdr->b_size; |
34dc7c2f BB |
1413 | } |
1414 | ASSERT3U(*size, >=, from_delta); | |
1415 | atomic_add_64(size, -from_delta); | |
34dc7c2f | 1416 | } |
b9541d6b | 1417 | if (new_state != arc_anon && new_state != arc_l2c_only) { |
b9541d6b | 1418 | uint64_t *size = &new_state->arcs_lsize[buftype]; |
34dc7c2f | 1419 | |
b9541d6b CW |
1420 | /* |
1421 | * An L1 header always exists here, since if we're | |
1422 | * moving to some L1-cached state (i.e. not l2c_only or | |
1423 | * anonymous), we realloc the header to add an L1hdr | |
1424 | * beforehand. | |
1425 | */ | |
1426 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
ca0bf58d | 1427 | multilist_insert(&new_state->arcs_list[buftype], hdr); |
34dc7c2f BB |
1428 | |
1429 | /* ghost elements have a ghost size */ | |
1430 | if (GHOST_STATE(new_state)) { | |
b9541d6b CW |
1431 | ASSERT0(datacnt); |
1432 | ASSERT(hdr->b_l1hdr.b_buf == NULL); | |
2a432414 | 1433 | to_delta = hdr->b_size; |
34dc7c2f BB |
1434 | } |
1435 | atomic_add_64(size, to_delta); | |
34dc7c2f BB |
1436 | } |
1437 | } | |
1438 | ||
2a432414 GW |
1439 | ASSERT(!BUF_EMPTY(hdr)); |
1440 | if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) | |
1441 | buf_hash_remove(hdr); | |
34dc7c2f | 1442 | |
b9541d6b CW |
1443 | /* adjust state sizes (ignore arc_l2c_only) */ |
1444 | if (to_delta && new_state != arc_l2c_only) | |
34dc7c2f | 1445 | atomic_add_64(&new_state->arcs_size, to_delta); |
b9541d6b | 1446 | if (from_delta && old_state != arc_l2c_only) { |
34dc7c2f BB |
1447 | ASSERT3U(old_state->arcs_size, >=, from_delta); |
1448 | atomic_add_64(&old_state->arcs_size, -from_delta); | |
1449 | } | |
b9541d6b CW |
1450 | if (HDR_HAS_L1HDR(hdr)) |
1451 | hdr->b_l1hdr.b_state = new_state; | |
34dc7c2f | 1452 | |
b9541d6b CW |
1453 | /* |
1454 | * L2 headers should never be on the L2 state list since they don't | |
1455 | * have L1 headers allocated. | |
1456 | */ | |
ca0bf58d PS |
1457 | ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && |
1458 | multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); | |
34dc7c2f BB |
1459 | } |
1460 | ||
1461 | void | |
d164b209 | 1462 | arc_space_consume(uint64_t space, arc_space_type_t type) |
34dc7c2f | 1463 | { |
d164b209 BB |
1464 | ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); |
1465 | ||
1466 | switch (type) { | |
e75c13c3 BB |
1467 | default: |
1468 | break; | |
d164b209 BB |
1469 | case ARC_SPACE_DATA: |
1470 | ARCSTAT_INCR(arcstat_data_size, space); | |
1471 | break; | |
cc7f677c PS |
1472 | case ARC_SPACE_META: |
1473 | ARCSTAT_INCR(arcstat_meta_size, space); | |
1474 | break; | |
d164b209 BB |
1475 | case ARC_SPACE_OTHER: |
1476 | ARCSTAT_INCR(arcstat_other_size, space); | |
1477 | break; | |
1478 | case ARC_SPACE_HDRS: | |
1479 | ARCSTAT_INCR(arcstat_hdr_size, space); | |
1480 | break; | |
1481 | case ARC_SPACE_L2HDRS: | |
1482 | ARCSTAT_INCR(arcstat_l2_hdr_size, space); | |
1483 | break; | |
1484 | } | |
1485 | ||
596a8935 | 1486 | if (type != ARC_SPACE_DATA) { |
cc7f677c | 1487 | ARCSTAT_INCR(arcstat_meta_used, space); |
596a8935 BB |
1488 | if (arc_meta_max < arc_meta_used) |
1489 | arc_meta_max = arc_meta_used; | |
1490 | } | |
cc7f677c | 1491 | |
34dc7c2f BB |
1492 | atomic_add_64(&arc_size, space); |
1493 | } | |
1494 | ||
1495 | void | |
d164b209 | 1496 | arc_space_return(uint64_t space, arc_space_type_t type) |
34dc7c2f | 1497 | { |
d164b209 BB |
1498 | ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); |
1499 | ||
1500 | switch (type) { | |
e75c13c3 BB |
1501 | default: |
1502 | break; | |
d164b209 BB |
1503 | case ARC_SPACE_DATA: |
1504 | ARCSTAT_INCR(arcstat_data_size, -space); | |
1505 | break; | |
cc7f677c PS |
1506 | case ARC_SPACE_META: |
1507 | ARCSTAT_INCR(arcstat_meta_size, -space); | |
1508 | break; | |
d164b209 BB |
1509 | case ARC_SPACE_OTHER: |
1510 | ARCSTAT_INCR(arcstat_other_size, -space); | |
1511 | break; | |
1512 | case ARC_SPACE_HDRS: | |
1513 | ARCSTAT_INCR(arcstat_hdr_size, -space); | |
1514 | break; | |
1515 | case ARC_SPACE_L2HDRS: | |
1516 | ARCSTAT_INCR(arcstat_l2_hdr_size, -space); | |
1517 | break; | |
1518 | } | |
1519 | ||
cc7f677c PS |
1520 | if (type != ARC_SPACE_DATA) { |
1521 | ASSERT(arc_meta_used >= space); | |
cc7f677c PS |
1522 | ARCSTAT_INCR(arcstat_meta_used, -space); |
1523 | } | |
1524 | ||
34dc7c2f BB |
1525 | ASSERT(arc_size >= space); |
1526 | atomic_add_64(&arc_size, -space); | |
1527 | } | |
1528 | ||
34dc7c2f | 1529 | arc_buf_t * |
5f6d0b6f | 1530 | arc_buf_alloc(spa_t *spa, uint64_t size, void *tag, arc_buf_contents_t type) |
34dc7c2f BB |
1531 | { |
1532 | arc_buf_hdr_t *hdr; | |
1533 | arc_buf_t *buf; | |
1534 | ||
f1512ee6 | 1535 | VERIFY3U(size, <=, spa_maxblocksize(spa)); |
b9541d6b | 1536 | hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); |
34dc7c2f | 1537 | ASSERT(BUF_EMPTY(hdr)); |
b9541d6b | 1538 | ASSERT3P(hdr->b_freeze_cksum, ==, NULL); |
34dc7c2f | 1539 | hdr->b_size = size; |
3541dc6d | 1540 | hdr->b_spa = spa_load_guid(spa); |
b9541d6b CW |
1541 | hdr->b_l1hdr.b_mru_hits = 0; |
1542 | hdr->b_l1hdr.b_mru_ghost_hits = 0; | |
1543 | hdr->b_l1hdr.b_mfu_hits = 0; | |
1544 | hdr->b_l1hdr.b_mfu_ghost_hits = 0; | |
1545 | hdr->b_l1hdr.b_l2_hits = 0; | |
1546 | ||
34dc7c2f BB |
1547 | buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); |
1548 | buf->b_hdr = hdr; | |
1549 | buf->b_data = NULL; | |
1550 | buf->b_efunc = NULL; | |
1551 | buf->b_private = NULL; | |
1552 | buf->b_next = NULL; | |
b9541d6b CW |
1553 | |
1554 | hdr->b_flags = arc_bufc_to_flags(type); | |
1555 | hdr->b_flags |= ARC_FLAG_HAS_L1HDR; | |
1556 | ||
1557 | hdr->b_l1hdr.b_buf = buf; | |
1558 | hdr->b_l1hdr.b_state = arc_anon; | |
1559 | hdr->b_l1hdr.b_arc_access = 0; | |
1560 | hdr->b_l1hdr.b_datacnt = 1; | |
ca0bf58d | 1561 | hdr->b_l1hdr.b_tmp_cdata = NULL; |
b9541d6b | 1562 | |
34dc7c2f | 1563 | arc_get_data_buf(buf); |
b9541d6b CW |
1564 | |
1565 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); | |
1566 | (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); | |
34dc7c2f BB |
1567 | |
1568 | return (buf); | |
1569 | } | |
1570 | ||
9babb374 BB |
1571 | static char *arc_onloan_tag = "onloan"; |
1572 | ||
1573 | /* | |
1574 | * Loan out an anonymous arc buffer. Loaned buffers are not counted as in | |
1575 | * flight data by arc_tempreserve_space() until they are "returned". Loaned | |
1576 | * buffers must be returned to the arc before they can be used by the DMU or | |
1577 | * freed. | |
1578 | */ | |
1579 | arc_buf_t * | |
5f6d0b6f | 1580 | arc_loan_buf(spa_t *spa, uint64_t size) |
9babb374 BB |
1581 | { |
1582 | arc_buf_t *buf; | |
1583 | ||
1584 | buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); | |
1585 | ||
1586 | atomic_add_64(&arc_loaned_bytes, size); | |
1587 | return (buf); | |
1588 | } | |
1589 | ||
1590 | /* | |
1591 | * Return a loaned arc buffer to the arc. | |
1592 | */ | |
1593 | void | |
1594 | arc_return_buf(arc_buf_t *buf, void *tag) | |
1595 | { | |
1596 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
1597 | ||
9babb374 | 1598 | ASSERT(buf->b_data != NULL); |
b9541d6b CW |
1599 | ASSERT(HDR_HAS_L1HDR(hdr)); |
1600 | (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); | |
1601 | (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); | |
9babb374 BB |
1602 | |
1603 | atomic_add_64(&arc_loaned_bytes, -hdr->b_size); | |
1604 | } | |
1605 | ||
428870ff BB |
1606 | /* Detach an arc_buf from a dbuf (tag) */ |
1607 | void | |
1608 | arc_loan_inuse_buf(arc_buf_t *buf, void *tag) | |
1609 | { | |
b9541d6b | 1610 | arc_buf_hdr_t *hdr = buf->b_hdr; |
428870ff BB |
1611 | |
1612 | ASSERT(buf->b_data != NULL); | |
b9541d6b CW |
1613 | ASSERT(HDR_HAS_L1HDR(hdr)); |
1614 | (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); | |
1615 | (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); | |
428870ff BB |
1616 | buf->b_efunc = NULL; |
1617 | buf->b_private = NULL; | |
1618 | ||
1619 | atomic_add_64(&arc_loaned_bytes, hdr->b_size); | |
1620 | } | |
1621 | ||
34dc7c2f BB |
1622 | static arc_buf_t * |
1623 | arc_buf_clone(arc_buf_t *from) | |
1624 | { | |
1625 | arc_buf_t *buf; | |
1626 | arc_buf_hdr_t *hdr = from->b_hdr; | |
1627 | uint64_t size = hdr->b_size; | |
1628 | ||
b9541d6b CW |
1629 | ASSERT(HDR_HAS_L1HDR(hdr)); |
1630 | ASSERT(hdr->b_l1hdr.b_state != arc_anon); | |
428870ff | 1631 | |
34dc7c2f BB |
1632 | buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); |
1633 | buf->b_hdr = hdr; | |
1634 | buf->b_data = NULL; | |
1635 | buf->b_efunc = NULL; | |
1636 | buf->b_private = NULL; | |
b9541d6b CW |
1637 | buf->b_next = hdr->b_l1hdr.b_buf; |
1638 | hdr->b_l1hdr.b_buf = buf; | |
34dc7c2f BB |
1639 | arc_get_data_buf(buf); |
1640 | bcopy(from->b_data, buf->b_data, size); | |
1eb5bfa3 GW |
1641 | |
1642 | /* | |
1643 | * This buffer already exists in the arc so create a duplicate | |
1644 | * copy for the caller. If the buffer is associated with user data | |
1645 | * then track the size and number of duplicates. These stats will be | |
1646 | * updated as duplicate buffers are created and destroyed. | |
1647 | */ | |
b9541d6b | 1648 | if (HDR_ISTYPE_DATA(hdr)) { |
1eb5bfa3 GW |
1649 | ARCSTAT_BUMP(arcstat_duplicate_buffers); |
1650 | ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); | |
1651 | } | |
b9541d6b | 1652 | hdr->b_l1hdr.b_datacnt += 1; |
34dc7c2f BB |
1653 | return (buf); |
1654 | } | |
1655 | ||
1656 | void | |
1657 | arc_buf_add_ref(arc_buf_t *buf, void* tag) | |
1658 | { | |
1659 | arc_buf_hdr_t *hdr; | |
1660 | kmutex_t *hash_lock; | |
1661 | ||
1662 | /* | |
b128c09f BB |
1663 | * Check to see if this buffer is evicted. Callers |
1664 | * must verify b_data != NULL to know if the add_ref | |
1665 | * was successful. | |
34dc7c2f | 1666 | */ |
428870ff | 1667 | mutex_enter(&buf->b_evict_lock); |
b128c09f | 1668 | if (buf->b_data == NULL) { |
428870ff | 1669 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
1670 | return; |
1671 | } | |
428870ff | 1672 | hash_lock = HDR_LOCK(buf->b_hdr); |
34dc7c2f | 1673 | mutex_enter(hash_lock); |
428870ff | 1674 | hdr = buf->b_hdr; |
b9541d6b | 1675 | ASSERT(HDR_HAS_L1HDR(hdr)); |
428870ff BB |
1676 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); |
1677 | mutex_exit(&buf->b_evict_lock); | |
34dc7c2f | 1678 | |
b9541d6b CW |
1679 | ASSERT(hdr->b_l1hdr.b_state == arc_mru || |
1680 | hdr->b_l1hdr.b_state == arc_mfu); | |
1681 | ||
34dc7c2f | 1682 | add_reference(hdr, hash_lock, tag); |
d164b209 | 1683 | DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); |
34dc7c2f BB |
1684 | arc_access(hdr, hash_lock); |
1685 | mutex_exit(hash_lock); | |
1686 | ARCSTAT_BUMP(arcstat_hits); | |
b9541d6b CW |
1687 | ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), |
1688 | demand, prefetch, !HDR_ISTYPE_METADATA(hdr), | |
34dc7c2f BB |
1689 | data, metadata, hits); |
1690 | } | |
1691 | ||
ca0bf58d PS |
1692 | static void |
1693 | arc_buf_free_on_write(void *data, size_t size, | |
1694 | void (*free_func)(void *, size_t)) | |
1695 | { | |
1696 | l2arc_data_free_t *df; | |
1697 | ||
1698 | df = kmem_alloc(sizeof (*df), KM_SLEEP); | |
1699 | df->l2df_data = data; | |
1700 | df->l2df_size = size; | |
1701 | df->l2df_func = free_func; | |
1702 | mutex_enter(&l2arc_free_on_write_mtx); | |
1703 | list_insert_head(l2arc_free_on_write, df); | |
1704 | mutex_exit(&l2arc_free_on_write_mtx); | |
1705 | } | |
1706 | ||
34dc7c2f BB |
1707 | /* |
1708 | * Free the arc data buffer. If it is an l2arc write in progress, | |
1709 | * the buffer is placed on l2arc_free_on_write to be freed later. | |
1710 | */ | |
1711 | static void | |
498877ba | 1712 | arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) |
34dc7c2f | 1713 | { |
498877ba MA |
1714 | arc_buf_hdr_t *hdr = buf->b_hdr; |
1715 | ||
34dc7c2f | 1716 | if (HDR_L2_WRITING(hdr)) { |
ca0bf58d | 1717 | arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); |
34dc7c2f BB |
1718 | ARCSTAT_BUMP(arcstat_l2_free_on_write); |
1719 | } else { | |
498877ba | 1720 | free_func(buf->b_data, hdr->b_size); |
34dc7c2f BB |
1721 | } |
1722 | } | |
1723 | ||
ca0bf58d PS |
1724 | static void |
1725 | arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) | |
1726 | { | |
1727 | ASSERT(HDR_HAS_L2HDR(hdr)); | |
1728 | ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); | |
1729 | ||
1730 | /* | |
1731 | * The b_tmp_cdata field is linked off of the b_l1hdr, so if | |
1732 | * that doesn't exist, the header is in the arc_l2c_only state, | |
1733 | * and there isn't anything to free (it's already been freed). | |
1734 | */ | |
1735 | if (!HDR_HAS_L1HDR(hdr)) | |
1736 | return; | |
1737 | ||
1738 | /* | |
1739 | * The header isn't being written to the l2arc device, thus it | |
1740 | * shouldn't have a b_tmp_cdata to free. | |
1741 | */ | |
1742 | if (!HDR_L2_WRITING(hdr)) { | |
1743 | ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); | |
1744 | return; | |
1745 | } | |
1746 | ||
1747 | /* | |
1748 | * The header does not have compression enabled. This can be due | |
1749 | * to the buffer not being compressible, or because we're | |
1750 | * freeing the buffer before the second phase of | |
1751 | * l2arc_write_buffer() has started (which does the compression | |
1752 | * step). In either case, b_tmp_cdata does not point to a | |
1753 | * separately compressed buffer, so there's nothing to free (it | |
1754 | * points to the same buffer as the arc_buf_t's b_data field). | |
1755 | */ | |
1756 | if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) { | |
1757 | hdr->b_l1hdr.b_tmp_cdata = NULL; | |
1758 | return; | |
1759 | } | |
1760 | ||
1761 | /* | |
1762 | * There's nothing to free since the buffer was all zero's and | |
1763 | * compressed to a zero length buffer. | |
1764 | */ | |
1765 | if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_EMPTY) { | |
1766 | ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); | |
1767 | return; | |
1768 | } | |
1769 | ||
1770 | ASSERT(L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr))); | |
1771 | ||
1772 | arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, | |
1773 | hdr->b_size, zio_data_buf_free); | |
1774 | ||
1775 | ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); | |
1776 | hdr->b_l1hdr.b_tmp_cdata = NULL; | |
1777 | } | |
1778 | ||
bd089c54 MA |
1779 | /* |
1780 | * Free up buf->b_data and if 'remove' is set, then pull the | |
1781 | * arc_buf_t off of the the arc_buf_hdr_t's list and free it. | |
1782 | */ | |
34dc7c2f | 1783 | static void |
ca0bf58d | 1784 | arc_buf_destroy(arc_buf_t *buf, boolean_t remove) |
34dc7c2f BB |
1785 | { |
1786 | arc_buf_t **bufp; | |
1787 | ||
1788 | /* free up data associated with the buf */ | |
b9541d6b CW |
1789 | if (buf->b_data != NULL) { |
1790 | arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; | |
34dc7c2f | 1791 | uint64_t size = buf->b_hdr->b_size; |
b9541d6b | 1792 | arc_buf_contents_t type = arc_buf_type(buf->b_hdr); |
34dc7c2f BB |
1793 | |
1794 | arc_cksum_verify(buf); | |
498877ba | 1795 | arc_buf_unwatch(buf); |
428870ff | 1796 | |
ca0bf58d PS |
1797 | if (type == ARC_BUFC_METADATA) { |
1798 | arc_buf_data_free(buf, zio_buf_free); | |
1799 | arc_space_return(size, ARC_SPACE_META); | |
1800 | } else { | |
1801 | ASSERT(type == ARC_BUFC_DATA); | |
1802 | arc_buf_data_free(buf, zio_data_buf_free); | |
1803 | arc_space_return(size, ARC_SPACE_DATA); | |
34dc7c2f | 1804 | } |
ca0bf58d PS |
1805 | |
1806 | /* protected by hash lock, if in the hash table */ | |
1807 | if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { | |
34dc7c2f BB |
1808 | uint64_t *cnt = &state->arcs_lsize[type]; |
1809 | ||
b9541d6b CW |
1810 | ASSERT(refcount_is_zero( |
1811 | &buf->b_hdr->b_l1hdr.b_refcnt)); | |
1812 | ASSERT(state != arc_anon && state != arc_l2c_only); | |
34dc7c2f BB |
1813 | |
1814 | ASSERT3U(*cnt, >=, size); | |
1815 | atomic_add_64(cnt, -size); | |
1816 | } | |
1817 | ASSERT3U(state->arcs_size, >=, size); | |
1818 | atomic_add_64(&state->arcs_size, -size); | |
1819 | buf->b_data = NULL; | |
1eb5bfa3 GW |
1820 | |
1821 | /* | |
1822 | * If we're destroying a duplicate buffer make sure | |
1823 | * that the appropriate statistics are updated. | |
1824 | */ | |
b9541d6b CW |
1825 | if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && |
1826 | HDR_ISTYPE_DATA(buf->b_hdr)) { | |
1eb5bfa3 GW |
1827 | ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); |
1828 | ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); | |
1829 | } | |
b9541d6b CW |
1830 | ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); |
1831 | buf->b_hdr->b_l1hdr.b_datacnt -= 1; | |
34dc7c2f BB |
1832 | } |
1833 | ||
1834 | /* only remove the buf if requested */ | |
bd089c54 | 1835 | if (!remove) |
34dc7c2f BB |
1836 | return; |
1837 | ||
1838 | /* remove the buf from the hdr list */ | |
b9541d6b CW |
1839 | for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; |
1840 | bufp = &(*bufp)->b_next) | |
34dc7c2f BB |
1841 | continue; |
1842 | *bufp = buf->b_next; | |
428870ff | 1843 | buf->b_next = NULL; |
34dc7c2f BB |
1844 | |
1845 | ASSERT(buf->b_efunc == NULL); | |
1846 | ||
1847 | /* clean up the buf */ | |
1848 | buf->b_hdr = NULL; | |
1849 | kmem_cache_free(buf_cache, buf); | |
1850 | } | |
1851 | ||
d962d5da PS |
1852 | static void |
1853 | arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) | |
1854 | { | |
1855 | l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; | |
1856 | l2arc_dev_t *dev = l2hdr->b_dev; | |
1857 | ||
1858 | ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); | |
1859 | ASSERT(HDR_HAS_L2HDR(hdr)); | |
1860 | ||
1861 | list_remove(&dev->l2ad_buflist, hdr); | |
1862 | ||
1863 | arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); | |
1864 | ||
1865 | /* | |
1866 | * We don't want to leak the b_tmp_cdata buffer that was | |
1867 | * allocated in l2arc_write_buffers() | |
1868 | */ | |
1869 | arc_buf_l2_cdata_free(hdr); | |
1870 | ||
1871 | /* | |
1872 | * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then | |
1873 | * this header is being processed by l2arc_write_buffers() (i.e. | |
1874 | * it's in the first stage of l2arc_write_buffers()). | |
1875 | * Re-affirming that truth here, just to serve as a reminder. If | |
1876 | * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or | |
1877 | * may not have its HDR_L2_WRITING flag set. (the write may have | |
1878 | * completed, in which case HDR_L2_WRITING will be false and the | |
1879 | * b_daddr field will point to the address of the buffer on disk). | |
1880 | */ | |
1881 | IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr)); | |
1882 | ||
1883 | /* | |
1884 | * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with | |
1885 | * l2arc_write_buffers(). Since we've just removed this header | |
1886 | * from the l2arc buffer list, this header will never reach the | |
1887 | * second stage of l2arc_write_buffers(), which increments the | |
1888 | * accounting stats for this header. Thus, we must be careful | |
1889 | * not to decrement them for this header either. | |
1890 | */ | |
1891 | if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) { | |
1892 | ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); | |
1893 | ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); | |
1894 | ||
1895 | vdev_space_update(dev->l2ad_vdev, | |
1896 | -l2hdr->b_asize, 0, 0); | |
1897 | ||
1898 | (void) refcount_remove_many(&dev->l2ad_alloc, | |
1899 | l2hdr->b_asize, hdr); | |
1900 | } | |
1901 | ||
1902 | hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; | |
1903 | } | |
1904 | ||
34dc7c2f BB |
1905 | static void |
1906 | arc_hdr_destroy(arc_buf_hdr_t *hdr) | |
1907 | { | |
b9541d6b CW |
1908 | if (HDR_HAS_L1HDR(hdr)) { |
1909 | ASSERT(hdr->b_l1hdr.b_buf == NULL || | |
1910 | hdr->b_l1hdr.b_datacnt > 0); | |
1911 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); | |
1912 | ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); | |
1913 | } | |
34dc7c2f | 1914 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); |
b9541d6b CW |
1915 | ASSERT(!HDR_IN_HASH_TABLE(hdr)); |
1916 | ||
1917 | if (HDR_HAS_L2HDR(hdr)) { | |
d962d5da PS |
1918 | l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; |
1919 | boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); | |
428870ff | 1920 | |
d962d5da PS |
1921 | if (!buflist_held) |
1922 | mutex_enter(&dev->l2ad_mtx); | |
b9541d6b | 1923 | |
ca0bf58d | 1924 | /* |
d962d5da PS |
1925 | * Even though we checked this conditional above, we |
1926 | * need to check this again now that we have the | |
1927 | * l2ad_mtx. This is because we could be racing with | |
1928 | * another thread calling l2arc_evict() which might have | |
1929 | * destroyed this header's L2 portion as we were waiting | |
1930 | * to acquire the l2ad_mtx. If that happens, we don't | |
1931 | * want to re-destroy the header's L2 portion. | |
ca0bf58d | 1932 | */ |
d962d5da PS |
1933 | if (HDR_HAS_L2HDR(hdr)) |
1934 | arc_hdr_l2hdr_destroy(hdr); | |
428870ff BB |
1935 | |
1936 | if (!buflist_held) | |
d962d5da | 1937 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f BB |
1938 | } |
1939 | ||
b9541d6b | 1940 | if (!BUF_EMPTY(hdr)) |
428870ff | 1941 | buf_discard_identity(hdr); |
b9541d6b | 1942 | |
34dc7c2f BB |
1943 | if (hdr->b_freeze_cksum != NULL) { |
1944 | kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); | |
1945 | hdr->b_freeze_cksum = NULL; | |
1946 | } | |
1947 | ||
b9541d6b CW |
1948 | if (HDR_HAS_L1HDR(hdr)) { |
1949 | while (hdr->b_l1hdr.b_buf) { | |
1950 | arc_buf_t *buf = hdr->b_l1hdr.b_buf; | |
1951 | ||
1952 | if (buf->b_efunc != NULL) { | |
ca0bf58d | 1953 | mutex_enter(&arc_user_evicts_lock); |
b9541d6b CW |
1954 | mutex_enter(&buf->b_evict_lock); |
1955 | ASSERT(buf->b_hdr != NULL); | |
ca0bf58d | 1956 | arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE); |
b9541d6b CW |
1957 | hdr->b_l1hdr.b_buf = buf->b_next; |
1958 | buf->b_hdr = &arc_eviction_hdr; | |
1959 | buf->b_next = arc_eviction_list; | |
1960 | arc_eviction_list = buf; | |
1961 | mutex_exit(&buf->b_evict_lock); | |
ca0bf58d PS |
1962 | cv_signal(&arc_user_evicts_cv); |
1963 | mutex_exit(&arc_user_evicts_lock); | |
b9541d6b | 1964 | } else { |
ca0bf58d | 1965 | arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE); |
b9541d6b CW |
1966 | } |
1967 | } | |
1968 | } | |
1969 | ||
34dc7c2f | 1970 | ASSERT3P(hdr->b_hash_next, ==, NULL); |
b9541d6b | 1971 | if (HDR_HAS_L1HDR(hdr)) { |
ca0bf58d | 1972 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); |
b9541d6b CW |
1973 | ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); |
1974 | kmem_cache_free(hdr_full_cache, hdr); | |
1975 | } else { | |
1976 | kmem_cache_free(hdr_l2only_cache, hdr); | |
1977 | } | |
34dc7c2f BB |
1978 | } |
1979 | ||
1980 | void | |
1981 | arc_buf_free(arc_buf_t *buf, void *tag) | |
1982 | { | |
1983 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
b9541d6b | 1984 | int hashed = hdr->b_l1hdr.b_state != arc_anon; |
34dc7c2f BB |
1985 | |
1986 | ASSERT(buf->b_efunc == NULL); | |
1987 | ASSERT(buf->b_data != NULL); | |
1988 | ||
1989 | if (hashed) { | |
1990 | kmutex_t *hash_lock = HDR_LOCK(hdr); | |
1991 | ||
1992 | mutex_enter(hash_lock); | |
428870ff BB |
1993 | hdr = buf->b_hdr; |
1994 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
1995 | ||
34dc7c2f | 1996 | (void) remove_reference(hdr, hash_lock, tag); |
b9541d6b | 1997 | if (hdr->b_l1hdr.b_datacnt > 1) { |
ca0bf58d | 1998 | arc_buf_destroy(buf, TRUE); |
428870ff | 1999 | } else { |
b9541d6b | 2000 | ASSERT(buf == hdr->b_l1hdr.b_buf); |
428870ff | 2001 | ASSERT(buf->b_efunc == NULL); |
2a432414 | 2002 | hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; |
428870ff | 2003 | } |
34dc7c2f BB |
2004 | mutex_exit(hash_lock); |
2005 | } else if (HDR_IO_IN_PROGRESS(hdr)) { | |
2006 | int destroy_hdr; | |
2007 | /* | |
2008 | * We are in the middle of an async write. Don't destroy | |
2009 | * this buffer unless the write completes before we finish | |
2010 | * decrementing the reference count. | |
2011 | */ | |
ca0bf58d | 2012 | mutex_enter(&arc_user_evicts_lock); |
34dc7c2f | 2013 | (void) remove_reference(hdr, NULL, tag); |
b9541d6b | 2014 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
34dc7c2f | 2015 | destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); |
ca0bf58d | 2016 | mutex_exit(&arc_user_evicts_lock); |
34dc7c2f BB |
2017 | if (destroy_hdr) |
2018 | arc_hdr_destroy(hdr); | |
2019 | } else { | |
428870ff | 2020 | if (remove_reference(hdr, NULL, tag) > 0) |
ca0bf58d | 2021 | arc_buf_destroy(buf, TRUE); |
428870ff | 2022 | else |
34dc7c2f | 2023 | arc_hdr_destroy(hdr); |
34dc7c2f BB |
2024 | } |
2025 | } | |
2026 | ||
13fe0198 | 2027 | boolean_t |
34dc7c2f BB |
2028 | arc_buf_remove_ref(arc_buf_t *buf, void* tag) |
2029 | { | |
2030 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
b4f7f105 | 2031 | kmutex_t *hash_lock = NULL; |
13fe0198 | 2032 | boolean_t no_callback = (buf->b_efunc == NULL); |
34dc7c2f | 2033 | |
b9541d6b CW |
2034 | if (hdr->b_l1hdr.b_state == arc_anon) { |
2035 | ASSERT(hdr->b_l1hdr.b_datacnt == 1); | |
34dc7c2f BB |
2036 | arc_buf_free(buf, tag); |
2037 | return (no_callback); | |
2038 | } | |
2039 | ||
b4f7f105 | 2040 | hash_lock = HDR_LOCK(hdr); |
34dc7c2f | 2041 | mutex_enter(hash_lock); |
428870ff | 2042 | hdr = buf->b_hdr; |
b9541d6b | 2043 | ASSERT(hdr->b_l1hdr.b_datacnt > 0); |
428870ff | 2044 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); |
b9541d6b | 2045 | ASSERT(hdr->b_l1hdr.b_state != arc_anon); |
34dc7c2f BB |
2046 | ASSERT(buf->b_data != NULL); |
2047 | ||
2048 | (void) remove_reference(hdr, hash_lock, tag); | |
b9541d6b | 2049 | if (hdr->b_l1hdr.b_datacnt > 1) { |
34dc7c2f | 2050 | if (no_callback) |
ca0bf58d | 2051 | arc_buf_destroy(buf, TRUE); |
34dc7c2f | 2052 | } else if (no_callback) { |
b9541d6b | 2053 | ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); |
428870ff | 2054 | ASSERT(buf->b_efunc == NULL); |
2a432414 | 2055 | hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; |
34dc7c2f | 2056 | } |
b9541d6b CW |
2057 | ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || |
2058 | refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); | |
34dc7c2f BB |
2059 | mutex_exit(hash_lock); |
2060 | return (no_callback); | |
2061 | } | |
2062 | ||
5f6d0b6f | 2063 | uint64_t |
34dc7c2f BB |
2064 | arc_buf_size(arc_buf_t *buf) |
2065 | { | |
2066 | return (buf->b_hdr->b_size); | |
2067 | } | |
2068 | ||
1eb5bfa3 GW |
2069 | /* |
2070 | * Called from the DMU to determine if the current buffer should be | |
2071 | * evicted. In order to ensure proper locking, the eviction must be initiated | |
2072 | * from the DMU. Return true if the buffer is associated with user data and | |
2073 | * duplicate buffers still exist. | |
2074 | */ | |
2075 | boolean_t | |
2076 | arc_buf_eviction_needed(arc_buf_t *buf) | |
2077 | { | |
2078 | arc_buf_hdr_t *hdr; | |
2079 | boolean_t evict_needed = B_FALSE; | |
2080 | ||
2081 | if (zfs_disable_dup_eviction) | |
2082 | return (B_FALSE); | |
2083 | ||
2084 | mutex_enter(&buf->b_evict_lock); | |
2085 | hdr = buf->b_hdr; | |
2086 | if (hdr == NULL) { | |
2087 | /* | |
2088 | * We are in arc_do_user_evicts(); let that function | |
2089 | * perform the eviction. | |
2090 | */ | |
2091 | ASSERT(buf->b_data == NULL); | |
2092 | mutex_exit(&buf->b_evict_lock); | |
2093 | return (B_FALSE); | |
2094 | } else if (buf->b_data == NULL) { | |
2095 | /* | |
2096 | * We have already been added to the arc eviction list; | |
2097 | * recommend eviction. | |
2098 | */ | |
2099 | ASSERT3P(hdr, ==, &arc_eviction_hdr); | |
2100 | mutex_exit(&buf->b_evict_lock); | |
2101 | return (B_TRUE); | |
2102 | } | |
2103 | ||
b9541d6b | 2104 | if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) |
1eb5bfa3 GW |
2105 | evict_needed = B_TRUE; |
2106 | ||
2107 | mutex_exit(&buf->b_evict_lock); | |
2108 | return (evict_needed); | |
2109 | } | |
2110 | ||
34dc7c2f | 2111 | /* |
ca0bf58d PS |
2112 | * Evict the arc_buf_hdr that is provided as a parameter. The resultant |
2113 | * state of the header is dependent on its state prior to entering this | |
2114 | * function. The following transitions are possible: | |
34dc7c2f | 2115 | * |
ca0bf58d PS |
2116 | * - arc_mru -> arc_mru_ghost |
2117 | * - arc_mfu -> arc_mfu_ghost | |
2118 | * - arc_mru_ghost -> arc_l2c_only | |
2119 | * - arc_mru_ghost -> deleted | |
2120 | * - arc_mfu_ghost -> arc_l2c_only | |
2121 | * - arc_mfu_ghost -> deleted | |
34dc7c2f | 2122 | */ |
ca0bf58d PS |
2123 | static int64_t |
2124 | arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) | |
34dc7c2f | 2125 | { |
ca0bf58d PS |
2126 | arc_state_t *evicted_state, *state; |
2127 | int64_t bytes_evicted = 0; | |
34dc7c2f | 2128 | |
ca0bf58d PS |
2129 | ASSERT(MUTEX_HELD(hash_lock)); |
2130 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
e8b96c60 | 2131 | |
ca0bf58d PS |
2132 | state = hdr->b_l1hdr.b_state; |
2133 | if (GHOST_STATE(state)) { | |
2134 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
2135 | ASSERT(hdr->b_l1hdr.b_buf == NULL); | |
e8b96c60 MA |
2136 | |
2137 | /* | |
ca0bf58d PS |
2138 | * l2arc_write_buffers() relies on a header's L1 portion |
2139 | * (i.e. its b_tmp_cdata field) during its write phase. | |
2140 | * Thus, we cannot push a header onto the arc_l2c_only | |
2141 | * state (removing its L1 piece) until the header is | |
2142 | * done being written to the l2arc. | |
e8b96c60 | 2143 | */ |
ca0bf58d PS |
2144 | if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { |
2145 | ARCSTAT_BUMP(arcstat_evict_l2_skip); | |
2146 | return (bytes_evicted); | |
e8b96c60 MA |
2147 | } |
2148 | ||
ca0bf58d PS |
2149 | ARCSTAT_BUMP(arcstat_deleted); |
2150 | bytes_evicted += hdr->b_size; | |
428870ff | 2151 | |
ca0bf58d | 2152 | DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); |
428870ff | 2153 | |
ca0bf58d PS |
2154 | if (HDR_HAS_L2HDR(hdr)) { |
2155 | /* | |
2156 | * This buffer is cached on the 2nd Level ARC; | |
2157 | * don't destroy the header. | |
2158 | */ | |
2159 | arc_change_state(arc_l2c_only, hdr, hash_lock); | |
2160 | /* | |
2161 | * dropping from L1+L2 cached to L2-only, | |
2162 | * realloc to remove the L1 header. | |
2163 | */ | |
2164 | hdr = arc_hdr_realloc(hdr, hdr_full_cache, | |
2165 | hdr_l2only_cache); | |
34dc7c2f | 2166 | } else { |
ca0bf58d PS |
2167 | arc_change_state(arc_anon, hdr, hash_lock); |
2168 | arc_hdr_destroy(hdr); | |
34dc7c2f | 2169 | } |
ca0bf58d | 2170 | return (bytes_evicted); |
34dc7c2f BB |
2171 | } |
2172 | ||
ca0bf58d PS |
2173 | ASSERT(state == arc_mru || state == arc_mfu); |
2174 | evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; | |
34dc7c2f | 2175 | |
ca0bf58d PS |
2176 | /* prefetch buffers have a minimum lifespan */ |
2177 | if (HDR_IO_IN_PROGRESS(hdr) || | |
2178 | ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && | |
2179 | ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < | |
2180 | arc_min_prefetch_lifespan)) { | |
2181 | ARCSTAT_BUMP(arcstat_evict_skip); | |
2182 | return (bytes_evicted); | |
da8ccd0e PS |
2183 | } |
2184 | ||
ca0bf58d PS |
2185 | ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); |
2186 | ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); | |
2187 | while (hdr->b_l1hdr.b_buf) { | |
2188 | arc_buf_t *buf = hdr->b_l1hdr.b_buf; | |
2189 | if (!mutex_tryenter(&buf->b_evict_lock)) { | |
2190 | ARCSTAT_BUMP(arcstat_mutex_miss); | |
2191 | break; | |
2192 | } | |
2193 | if (buf->b_data != NULL) | |
2194 | bytes_evicted += hdr->b_size; | |
2195 | if (buf->b_efunc != NULL) { | |
2196 | mutex_enter(&arc_user_evicts_lock); | |
2197 | arc_buf_destroy(buf, FALSE); | |
2198 | hdr->b_l1hdr.b_buf = buf->b_next; | |
2199 | buf->b_hdr = &arc_eviction_hdr; | |
2200 | buf->b_next = arc_eviction_list; | |
2201 | arc_eviction_list = buf; | |
2202 | cv_signal(&arc_user_evicts_cv); | |
2203 | mutex_exit(&arc_user_evicts_lock); | |
2204 | mutex_exit(&buf->b_evict_lock); | |
2205 | } else { | |
2206 | mutex_exit(&buf->b_evict_lock); | |
2207 | arc_buf_destroy(buf, TRUE); | |
2208 | } | |
2209 | } | |
34dc7c2f | 2210 | |
ca0bf58d PS |
2211 | if (HDR_HAS_L2HDR(hdr)) { |
2212 | ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size); | |
2213 | } else { | |
2214 | if (l2arc_write_eligible(hdr->b_spa, hdr)) | |
2215 | ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size); | |
2216 | else | |
2217 | ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size); | |
2218 | } | |
34dc7c2f | 2219 | |
ca0bf58d PS |
2220 | if (hdr->b_l1hdr.b_datacnt == 0) { |
2221 | arc_change_state(evicted_state, hdr, hash_lock); | |
2222 | ASSERT(HDR_IN_HASH_TABLE(hdr)); | |
2223 | hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; | |
2224 | hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; | |
2225 | DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); | |
2226 | } | |
34dc7c2f | 2227 | |
ca0bf58d | 2228 | return (bytes_evicted); |
34dc7c2f BB |
2229 | } |
2230 | ||
ca0bf58d PS |
2231 | static uint64_t |
2232 | arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, | |
2233 | uint64_t spa, int64_t bytes) | |
34dc7c2f | 2234 | { |
ca0bf58d PS |
2235 | multilist_sublist_t *mls; |
2236 | uint64_t bytes_evicted = 0; | |
2237 | arc_buf_hdr_t *hdr; | |
34dc7c2f | 2238 | kmutex_t *hash_lock; |
ca0bf58d | 2239 | int evict_count = 0; |
34dc7c2f | 2240 | |
ca0bf58d PS |
2241 | ASSERT3P(marker, !=, NULL); |
2242 | ASSERTV(if (bytes < 0) ASSERT(bytes == ARC_EVICT_ALL)); | |
2243 | ||
2244 | mls = multilist_sublist_lock(ml, idx); | |
572e2857 | 2245 | |
ca0bf58d PS |
2246 | for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; |
2247 | hdr = multilist_sublist_prev(mls, marker)) { | |
2248 | if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || | |
2249 | (evict_count >= zfs_arc_evict_batch_limit)) | |
2250 | break; | |
2251 | ||
2252 | /* | |
2253 | * To keep our iteration location, move the marker | |
2254 | * forward. Since we're not holding hdr's hash lock, we | |
2255 | * must be very careful and not remove 'hdr' from the | |
2256 | * sublist. Otherwise, other consumers might mistake the | |
2257 | * 'hdr' as not being on a sublist when they call the | |
2258 | * multilist_link_active() function (they all rely on | |
2259 | * the hash lock protecting concurrent insertions and | |
2260 | * removals). multilist_sublist_move_forward() was | |
2261 | * specifically implemented to ensure this is the case | |
2262 | * (only 'marker' will be removed and re-inserted). | |
2263 | */ | |
2264 | multilist_sublist_move_forward(mls, marker); | |
2265 | ||
2266 | /* | |
2267 | * The only case where the b_spa field should ever be | |
2268 | * zero, is the marker headers inserted by | |
2269 | * arc_evict_state(). It's possible for multiple threads | |
2270 | * to be calling arc_evict_state() concurrently (e.g. | |
2271 | * dsl_pool_close() and zio_inject_fault()), so we must | |
2272 | * skip any markers we see from these other threads. | |
2273 | */ | |
2a432414 | 2274 | if (hdr->b_spa == 0) |
572e2857 BB |
2275 | continue; |
2276 | ||
ca0bf58d PS |
2277 | /* we're only interested in evicting buffers of a certain spa */ |
2278 | if (spa != 0 && hdr->b_spa != spa) { | |
2279 | ARCSTAT_BUMP(arcstat_evict_skip); | |
428870ff | 2280 | continue; |
ca0bf58d PS |
2281 | } |
2282 | ||
2283 | hash_lock = HDR_LOCK(hdr); | |
e8b96c60 MA |
2284 | |
2285 | /* | |
ca0bf58d PS |
2286 | * We aren't calling this function from any code path |
2287 | * that would already be holding a hash lock, so we're | |
2288 | * asserting on this assumption to be defensive in case | |
2289 | * this ever changes. Without this check, it would be | |
2290 | * possible to incorrectly increment arcstat_mutex_miss | |
2291 | * below (e.g. if the code changed such that we called | |
2292 | * this function with a hash lock held). | |
e8b96c60 | 2293 | */ |
ca0bf58d PS |
2294 | ASSERT(!MUTEX_HELD(hash_lock)); |
2295 | ||
34dc7c2f | 2296 | if (mutex_tryenter(hash_lock)) { |
ca0bf58d PS |
2297 | uint64_t evicted = arc_evict_hdr(hdr, hash_lock); |
2298 | mutex_exit(hash_lock); | |
34dc7c2f | 2299 | |
ca0bf58d | 2300 | bytes_evicted += evicted; |
34dc7c2f | 2301 | |
572e2857 | 2302 | /* |
ca0bf58d PS |
2303 | * If evicted is zero, arc_evict_hdr() must have |
2304 | * decided to skip this header, don't increment | |
2305 | * evict_count in this case. | |
572e2857 | 2306 | */ |
ca0bf58d PS |
2307 | if (evicted != 0) |
2308 | evict_count++; | |
2309 | ||
2310 | /* | |
2311 | * If arc_size isn't overflowing, signal any | |
2312 | * threads that might happen to be waiting. | |
2313 | * | |
2314 | * For each header evicted, we wake up a single | |
2315 | * thread. If we used cv_broadcast, we could | |
2316 | * wake up "too many" threads causing arc_size | |
2317 | * to significantly overflow arc_c; since | |
2318 | * arc_get_data_buf() doesn't check for overflow | |
2319 | * when it's woken up (it doesn't because it's | |
2320 | * possible for the ARC to be overflowing while | |
2321 | * full of un-evictable buffers, and the | |
2322 | * function should proceed in this case). | |
2323 | * | |
2324 | * If threads are left sleeping, due to not | |
2325 | * using cv_broadcast, they will be woken up | |
2326 | * just before arc_reclaim_thread() sleeps. | |
2327 | */ | |
2328 | mutex_enter(&arc_reclaim_lock); | |
2329 | if (!arc_is_overflowing()) | |
2330 | cv_signal(&arc_reclaim_waiters_cv); | |
2331 | mutex_exit(&arc_reclaim_lock); | |
e8b96c60 | 2332 | } else { |
ca0bf58d | 2333 | ARCSTAT_BUMP(arcstat_mutex_miss); |
e8b96c60 | 2334 | } |
34dc7c2f | 2335 | } |
34dc7c2f | 2336 | |
ca0bf58d | 2337 | multilist_sublist_unlock(mls); |
34dc7c2f | 2338 | |
ca0bf58d | 2339 | return (bytes_evicted); |
34dc7c2f BB |
2340 | } |
2341 | ||
ca0bf58d PS |
2342 | /* |
2343 | * Evict buffers from the given arc state, until we've removed the | |
2344 | * specified number of bytes. Move the removed buffers to the | |
2345 | * appropriate evict state. | |
2346 | * | |
2347 | * This function makes a "best effort". It skips over any buffers | |
2348 | * it can't get a hash_lock on, and so, may not catch all candidates. | |
2349 | * It may also return without evicting as much space as requested. | |
2350 | * | |
2351 | * If bytes is specified using the special value ARC_EVICT_ALL, this | |
2352 | * will evict all available (i.e. unlocked and evictable) buffers from | |
2353 | * the given arc state; which is used by arc_flush(). | |
2354 | */ | |
2355 | static uint64_t | |
2356 | arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, | |
2357 | arc_buf_contents_t type) | |
34dc7c2f | 2358 | { |
ca0bf58d PS |
2359 | uint64_t total_evicted = 0; |
2360 | multilist_t *ml = &state->arcs_list[type]; | |
2361 | int num_sublists; | |
2362 | arc_buf_hdr_t **markers; | |
2363 | int i; | |
2364 | ||
2365 | ASSERTV(if (bytes < 0) ASSERT(bytes == ARC_EVICT_ALL)); | |
2366 | ||
2367 | num_sublists = multilist_get_num_sublists(ml); | |
d164b209 BB |
2368 | |
2369 | /* | |
ca0bf58d PS |
2370 | * If we've tried to evict from each sublist, made some |
2371 | * progress, but still have not hit the target number of bytes | |
2372 | * to evict, we want to keep trying. The markers allow us to | |
2373 | * pick up where we left off for each individual sublist, rather | |
2374 | * than starting from the tail each time. | |
d164b209 | 2375 | */ |
ca0bf58d PS |
2376 | markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); |
2377 | for (i = 0; i < num_sublists; i++) { | |
2378 | multilist_sublist_t *mls; | |
34dc7c2f | 2379 | |
ca0bf58d PS |
2380 | markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); |
2381 | ||
2382 | /* | |
2383 | * A b_spa of 0 is used to indicate that this header is | |
2384 | * a marker. This fact is used in arc_adjust_type() and | |
2385 | * arc_evict_state_impl(). | |
2386 | */ | |
2387 | markers[i]->b_spa = 0; | |
34dc7c2f | 2388 | |
ca0bf58d PS |
2389 | mls = multilist_sublist_lock(ml, i); |
2390 | multilist_sublist_insert_tail(mls, markers[i]); | |
2391 | multilist_sublist_unlock(mls); | |
34dc7c2f BB |
2392 | } |
2393 | ||
d164b209 | 2394 | /* |
ca0bf58d PS |
2395 | * While we haven't hit our target number of bytes to evict, or |
2396 | * we're evicting all available buffers. | |
d164b209 | 2397 | */ |
ca0bf58d PS |
2398 | while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { |
2399 | /* | |
2400 | * Start eviction using a randomly selected sublist, | |
2401 | * this is to try and evenly balance eviction across all | |
2402 | * sublists. Always starting at the same sublist | |
2403 | * (e.g. index 0) would cause evictions to favor certain | |
2404 | * sublists over others. | |
2405 | */ | |
2406 | int sublist_idx = multilist_get_random_index(ml); | |
2407 | uint64_t scan_evicted = 0; | |
34dc7c2f | 2408 | |
ca0bf58d PS |
2409 | for (i = 0; i < num_sublists; i++) { |
2410 | uint64_t bytes_remaining; | |
2411 | uint64_t bytes_evicted; | |
d164b209 | 2412 | |
ca0bf58d PS |
2413 | if (bytes == ARC_EVICT_ALL) |
2414 | bytes_remaining = ARC_EVICT_ALL; | |
2415 | else if (total_evicted < bytes) | |
2416 | bytes_remaining = bytes - total_evicted; | |
2417 | else | |
2418 | break; | |
34dc7c2f | 2419 | |
ca0bf58d PS |
2420 | bytes_evicted = arc_evict_state_impl(ml, sublist_idx, |
2421 | markers[sublist_idx], spa, bytes_remaining); | |
2422 | ||
2423 | scan_evicted += bytes_evicted; | |
2424 | total_evicted += bytes_evicted; | |
2425 | ||
2426 | /* we've reached the end, wrap to the beginning */ | |
2427 | if (++sublist_idx >= num_sublists) | |
2428 | sublist_idx = 0; | |
2429 | } | |
2430 | ||
2431 | /* | |
2432 | * If we didn't evict anything during this scan, we have | |
2433 | * no reason to believe we'll evict more during another | |
2434 | * scan, so break the loop. | |
2435 | */ | |
2436 | if (scan_evicted == 0) { | |
2437 | /* This isn't possible, let's make that obvious */ | |
2438 | ASSERT3S(bytes, !=, 0); | |
34dc7c2f | 2439 | |
ca0bf58d PS |
2440 | /* |
2441 | * When bytes is ARC_EVICT_ALL, the only way to | |
2442 | * break the loop is when scan_evicted is zero. | |
2443 | * In that case, we actually have evicted enough, | |
2444 | * so we don't want to increment the kstat. | |
2445 | */ | |
2446 | if (bytes != ARC_EVICT_ALL) { | |
2447 | ASSERT3S(total_evicted, <, bytes); | |
2448 | ARCSTAT_BUMP(arcstat_evict_not_enough); | |
2449 | } | |
d164b209 | 2450 | |
ca0bf58d PS |
2451 | break; |
2452 | } | |
d164b209 | 2453 | } |
34dc7c2f | 2454 | |
ca0bf58d PS |
2455 | for (i = 0; i < num_sublists; i++) { |
2456 | multilist_sublist_t *mls = multilist_sublist_lock(ml, i); | |
2457 | multilist_sublist_remove(mls, markers[i]); | |
2458 | multilist_sublist_unlock(mls); | |
34dc7c2f | 2459 | |
ca0bf58d | 2460 | kmem_cache_free(hdr_full_cache, markers[i]); |
34dc7c2f | 2461 | } |
ca0bf58d PS |
2462 | kmem_free(markers, sizeof (*markers) * num_sublists); |
2463 | ||
2464 | return (total_evicted); | |
2465 | } | |
2466 | ||
2467 | /* | |
2468 | * Flush all "evictable" data of the given type from the arc state | |
2469 | * specified. This will not evict any "active" buffers (i.e. referenced). | |
2470 | * | |
2471 | * When 'retry' is set to FALSE, the function will make a single pass | |
2472 | * over the state and evict any buffers that it can. Since it doesn't | |
2473 | * continually retry the eviction, it might end up leaving some buffers | |
2474 | * in the ARC due to lock misses. | |
2475 | * | |
2476 | * When 'retry' is set to TRUE, the function will continually retry the | |
2477 | * eviction until *all* evictable buffers have been removed from the | |
2478 | * state. As a result, if concurrent insertions into the state are | |
2479 | * allowed (e.g. if the ARC isn't shutting down), this function might | |
2480 | * wind up in an infinite loop, continually trying to evict buffers. | |
2481 | */ | |
2482 | static uint64_t | |
2483 | arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, | |
2484 | boolean_t retry) | |
2485 | { | |
2486 | uint64_t evicted = 0; | |
2487 | ||
2488 | while (state->arcs_lsize[type] != 0) { | |
2489 | evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); | |
2490 | ||
2491 | if (!retry) | |
2492 | break; | |
2493 | } | |
2494 | ||
2495 | return (evicted); | |
34dc7c2f BB |
2496 | } |
2497 | ||
ab26409d | 2498 | /* |
f6046738 BB |
2499 | * Helper function for arc_prune() it is responsible for safely handling |
2500 | * the execution of a registered arc_prune_func_t. | |
ab26409d BB |
2501 | */ |
2502 | static void | |
f6046738 | 2503 | arc_prune_task(void *ptr) |
ab26409d | 2504 | { |
f6046738 BB |
2505 | arc_prune_t *ap = (arc_prune_t *)ptr; |
2506 | arc_prune_func_t *func = ap->p_pfunc; | |
ab26409d | 2507 | |
f6046738 BB |
2508 | if (func != NULL) |
2509 | func(ap->p_adjust, ap->p_private); | |
ab26409d | 2510 | |
f6046738 BB |
2511 | /* Callback unregistered concurrently with execution */ |
2512 | if (refcount_remove(&ap->p_refcnt, func) == 0) { | |
2513 | ASSERT(!list_link_active(&ap->p_node)); | |
2514 | refcount_destroy(&ap->p_refcnt); | |
2515 | kmem_free(ap, sizeof (*ap)); | |
2516 | } | |
2517 | } | |
ab26409d | 2518 | |
f6046738 BB |
2519 | /* |
2520 | * Notify registered consumers they must drop holds on a portion of the ARC | |
2521 | * buffered they reference. This provides a mechanism to ensure the ARC can | |
2522 | * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers. This | |
2523 | * is analogous to dnlc_reduce_cache() but more generic. | |
2524 | * | |
2525 | * This operation is performed asyncronously so it may be safely called | |
ca67b33a | 2526 | * in the context of the arc_reclaim_thread(). A reference is taken here |
f6046738 BB |
2527 | * for each registered arc_prune_t and the arc_prune_task() is responsible |
2528 | * for releasing it once the registered arc_prune_func_t has completed. | |
2529 | */ | |
2530 | static void | |
2531 | arc_prune_async(int64_t adjust) | |
2532 | { | |
2533 | arc_prune_t *ap; | |
ab26409d | 2534 | |
f6046738 BB |
2535 | mutex_enter(&arc_prune_mtx); |
2536 | for (ap = list_head(&arc_prune_list); ap != NULL; | |
2537 | ap = list_next(&arc_prune_list, ap)) { | |
ab26409d | 2538 | |
f6046738 BB |
2539 | if (refcount_count(&ap->p_refcnt) >= 2) |
2540 | continue; | |
ab26409d | 2541 | |
f6046738 BB |
2542 | refcount_add(&ap->p_refcnt, ap->p_pfunc); |
2543 | ap->p_adjust = adjust; | |
2544 | taskq_dispatch(arc_prune_taskq, arc_prune_task, ap, TQ_SLEEP); | |
2545 | ARCSTAT_BUMP(arcstat_prune); | |
ab26409d | 2546 | } |
ab26409d BB |
2547 | mutex_exit(&arc_prune_mtx); |
2548 | } | |
2549 | ||
f6046738 BB |
2550 | static void |
2551 | arc_prune(int64_t adjust) | |
2552 | { | |
2553 | arc_prune_async(adjust); | |
2554 | taskq_wait_outstanding(arc_prune_taskq, 0); | |
2555 | } | |
2556 | ||
ca0bf58d PS |
2557 | /* |
2558 | * Evict the specified number of bytes from the state specified, | |
2559 | * restricting eviction to the spa and type given. This function | |
2560 | * prevents us from trying to evict more from a state's list than | |
2561 | * is "evictable", and to skip evicting altogether when passed a | |
2562 | * negative value for "bytes". In contrast, arc_evict_state() will | |
2563 | * evict everything it can, when passed a negative value for "bytes". | |
2564 | */ | |
2565 | static uint64_t | |
2566 | arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, | |
2567 | arc_buf_contents_t type) | |
2568 | { | |
2569 | int64_t delta; | |
2570 | ||
2571 | if (bytes > 0 && state->arcs_lsize[type] > 0) { | |
2572 | delta = MIN(state->arcs_lsize[type], bytes); | |
2573 | return (arc_evict_state(state, spa, delta, type)); | |
2574 | } | |
2575 | ||
2576 | return (0); | |
2577 | } | |
2578 | ||
2579 | /* | |
2580 | * The goal of this function is to evict enough meta data buffers from the | |
2581 | * ARC in order to enforce the arc_meta_limit. Achieving this is slightly | |
2582 | * more complicated than it appears because it is common for data buffers | |
2583 | * to have holds on meta data buffers. In addition, dnode meta data buffers | |
2584 | * will be held by the dnodes in the block preventing them from being freed. | |
2585 | * This means we can't simply traverse the ARC and expect to always find | |
2586 | * enough unheld meta data buffer to release. | |
2587 | * | |
2588 | * Therefore, this function has been updated to make alternating passes | |
2589 | * over the ARC releasing data buffers and then newly unheld meta data | |
2590 | * buffers. This ensures forward progress is maintained and arc_meta_used | |
2591 | * will decrease. Normally this is sufficient, but if required the ARC | |
2592 | * will call the registered prune callbacks causing dentry and inodes to | |
2593 | * be dropped from the VFS cache. This will make dnode meta data buffers | |
2594 | * available for reclaim. | |
2595 | */ | |
2596 | static uint64_t | |
f6046738 | 2597 | arc_adjust_meta_balanced(void) |
ca0bf58d PS |
2598 | { |
2599 | int64_t adjustmnt, delta, prune = 0; | |
2600 | uint64_t total_evicted = 0; | |
2601 | arc_buf_contents_t type = ARC_BUFC_DATA; | |
ca67b33a | 2602 | int restarts = MAX(zfs_arc_meta_adjust_restarts, 0); |
ca0bf58d PS |
2603 | |
2604 | restart: | |
2605 | /* | |
2606 | * This slightly differs than the way we evict from the mru in | |
2607 | * arc_adjust because we don't have a "target" value (i.e. no | |
2608 | * "meta" arc_p). As a result, I think we can completely | |
2609 | * cannibalize the metadata in the MRU before we evict the | |
2610 | * metadata from the MFU. I think we probably need to implement a | |
2611 | * "metadata arc_p" value to do this properly. | |
2612 | */ | |
2613 | adjustmnt = arc_meta_used - arc_meta_limit; | |
2614 | ||
2615 | if (adjustmnt > 0 && arc_mru->arcs_lsize[type] > 0) { | |
2616 | delta = MIN(arc_mru->arcs_lsize[type], adjustmnt); | |
2617 | total_evicted += arc_adjust_impl(arc_mru, 0, delta, type); | |
2618 | adjustmnt -= delta; | |
2619 | } | |
2620 | ||
2621 | /* | |
2622 | * We can't afford to recalculate adjustmnt here. If we do, | |
2623 | * new metadata buffers can sneak into the MRU or ANON lists, | |
2624 | * thus penalize the MFU metadata. Although the fudge factor is | |
2625 | * small, it has been empirically shown to be significant for | |
2626 | * certain workloads (e.g. creating many empty directories). As | |
2627 | * such, we use the original calculation for adjustmnt, and | |
2628 | * simply decrement the amount of data evicted from the MRU. | |
2629 | */ | |
2630 | ||
2631 | if (adjustmnt > 0 && arc_mfu->arcs_lsize[type] > 0) { | |
2632 | delta = MIN(arc_mfu->arcs_lsize[type], adjustmnt); | |
2633 | total_evicted += arc_adjust_impl(arc_mfu, 0, delta, type); | |
2634 | } | |
2635 | ||
2636 | adjustmnt = arc_meta_used - arc_meta_limit; | |
2637 | ||
2638 | if (adjustmnt > 0 && arc_mru_ghost->arcs_lsize[type] > 0) { | |
2639 | delta = MIN(adjustmnt, | |
2640 | arc_mru_ghost->arcs_lsize[type]); | |
2641 | total_evicted += arc_adjust_impl(arc_mru_ghost, 0, delta, type); | |
2642 | adjustmnt -= delta; | |
2643 | } | |
2644 | ||
2645 | if (adjustmnt > 0 && arc_mfu_ghost->arcs_lsize[type] > 0) { | |
2646 | delta = MIN(adjustmnt, | |
2647 | arc_mfu_ghost->arcs_lsize[type]); | |
2648 | total_evicted += arc_adjust_impl(arc_mfu_ghost, 0, delta, type); | |
2649 | } | |
2650 | ||
2651 | /* | |
2652 | * If after attempting to make the requested adjustment to the ARC | |
2653 | * the meta limit is still being exceeded then request that the | |
2654 | * higher layers drop some cached objects which have holds on ARC | |
2655 | * meta buffers. Requests to the upper layers will be made with | |
2656 | * increasingly large scan sizes until the ARC is below the limit. | |
2657 | */ | |
2658 | if (arc_meta_used > arc_meta_limit) { | |
2659 | if (type == ARC_BUFC_DATA) { | |
2660 | type = ARC_BUFC_METADATA; | |
2661 | } else { | |
2662 | type = ARC_BUFC_DATA; | |
2663 | ||
2664 | if (zfs_arc_meta_prune) { | |
2665 | prune += zfs_arc_meta_prune; | |
f6046738 | 2666 | arc_prune_async(prune); |
ca0bf58d PS |
2667 | } |
2668 | } | |
2669 | ||
2670 | if (restarts > 0) { | |
2671 | restarts--; | |
2672 | goto restart; | |
2673 | } | |
2674 | } | |
2675 | return (total_evicted); | |
2676 | } | |
2677 | ||
f6046738 BB |
2678 | /* |
2679 | * Evict metadata buffers from the cache, such that arc_meta_used is | |
2680 | * capped by the arc_meta_limit tunable. | |
2681 | */ | |
2682 | static uint64_t | |
2683 | arc_adjust_meta_only(void) | |
2684 | { | |
2685 | uint64_t total_evicted = 0; | |
2686 | int64_t target; | |
2687 | ||
2688 | /* | |
2689 | * If we're over the meta limit, we want to evict enough | |
2690 | * metadata to get back under the meta limit. We don't want to | |
2691 | * evict so much that we drop the MRU below arc_p, though. If | |
2692 | * we're over the meta limit more than we're over arc_p, we | |
2693 | * evict some from the MRU here, and some from the MFU below. | |
2694 | */ | |
2695 | target = MIN((int64_t)(arc_meta_used - arc_meta_limit), | |
2696 | (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size - arc_p)); | |
2697 | ||
2698 | total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); | |
2699 | ||
2700 | /* | |
2701 | * Similar to the above, we want to evict enough bytes to get us | |
2702 | * below the meta limit, but not so much as to drop us below the | |
2703 | * space alloted to the MFU (which is defined as arc_c - arc_p). | |
2704 | */ | |
2705 | target = MIN((int64_t)(arc_meta_used - arc_meta_limit), | |
2706 | (int64_t)(arc_mfu->arcs_size - (arc_c - arc_p))); | |
2707 | ||
2708 | total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); | |
2709 | ||
2710 | return (total_evicted); | |
2711 | } | |
2712 | ||
2713 | static uint64_t | |
2714 | arc_adjust_meta(void) | |
2715 | { | |
2716 | if (zfs_arc_meta_strategy == ARC_STRATEGY_META_ONLY) | |
2717 | return (arc_adjust_meta_only()); | |
2718 | else | |
2719 | return (arc_adjust_meta_balanced()); | |
2720 | } | |
2721 | ||
ca0bf58d PS |
2722 | /* |
2723 | * Return the type of the oldest buffer in the given arc state | |
2724 | * | |
2725 | * This function will select a random sublist of type ARC_BUFC_DATA and | |
2726 | * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist | |
2727 | * is compared, and the type which contains the "older" buffer will be | |
2728 | * returned. | |
2729 | */ | |
2730 | static arc_buf_contents_t | |
2731 | arc_adjust_type(arc_state_t *state) | |
2732 | { | |
2733 | multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; | |
2734 | multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; | |
2735 | int data_idx = multilist_get_random_index(data_ml); | |
2736 | int meta_idx = multilist_get_random_index(meta_ml); | |
2737 | multilist_sublist_t *data_mls; | |
2738 | multilist_sublist_t *meta_mls; | |
2739 | arc_buf_contents_t type; | |
2740 | arc_buf_hdr_t *data_hdr; | |
2741 | arc_buf_hdr_t *meta_hdr; | |
2742 | ||
2743 | /* | |
2744 | * We keep the sublist lock until we're finished, to prevent | |
2745 | * the headers from being destroyed via arc_evict_state(). | |
2746 | */ | |
2747 | data_mls = multilist_sublist_lock(data_ml, data_idx); | |
2748 | meta_mls = multilist_sublist_lock(meta_ml, meta_idx); | |
2749 | ||
2750 | /* | |
2751 | * These two loops are to ensure we skip any markers that | |
2752 | * might be at the tail of the lists due to arc_evict_state(). | |
2753 | */ | |
2754 | ||
2755 | for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; | |
2756 | data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { | |
2757 | if (data_hdr->b_spa != 0) | |
2758 | break; | |
2759 | } | |
2760 | ||
2761 | for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; | |
2762 | meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { | |
2763 | if (meta_hdr->b_spa != 0) | |
2764 | break; | |
2765 | } | |
2766 | ||
2767 | if (data_hdr == NULL && meta_hdr == NULL) { | |
2768 | type = ARC_BUFC_DATA; | |
2769 | } else if (data_hdr == NULL) { | |
2770 | ASSERT3P(meta_hdr, !=, NULL); | |
2771 | type = ARC_BUFC_METADATA; | |
2772 | } else if (meta_hdr == NULL) { | |
2773 | ASSERT3P(data_hdr, !=, NULL); | |
2774 | type = ARC_BUFC_DATA; | |
2775 | } else { | |
2776 | ASSERT3P(data_hdr, !=, NULL); | |
2777 | ASSERT3P(meta_hdr, !=, NULL); | |
2778 | ||
2779 | /* The headers can't be on the sublist without an L1 header */ | |
2780 | ASSERT(HDR_HAS_L1HDR(data_hdr)); | |
2781 | ASSERT(HDR_HAS_L1HDR(meta_hdr)); | |
2782 | ||
2783 | if (data_hdr->b_l1hdr.b_arc_access < | |
2784 | meta_hdr->b_l1hdr.b_arc_access) { | |
2785 | type = ARC_BUFC_DATA; | |
2786 | } else { | |
2787 | type = ARC_BUFC_METADATA; | |
2788 | } | |
2789 | } | |
2790 | ||
2791 | multilist_sublist_unlock(meta_mls); | |
2792 | multilist_sublist_unlock(data_mls); | |
2793 | ||
2794 | return (type); | |
2795 | } | |
2796 | ||
2797 | /* | |
2798 | * Evict buffers from the cache, such that arc_size is capped by arc_c. | |
2799 | */ | |
2800 | static uint64_t | |
2801 | arc_adjust(void) | |
2802 | { | |
2803 | uint64_t total_evicted = 0; | |
2804 | uint64_t bytes; | |
2805 | int64_t target; | |
2806 | ||
2807 | /* | |
2808 | * If we're over arc_meta_limit, we want to correct that before | |
2809 | * potentially evicting data buffers below. | |
2810 | */ | |
2811 | total_evicted += arc_adjust_meta(); | |
2812 | ||
2813 | /* | |
2814 | * Adjust MRU size | |
2815 | * | |
2816 | * If we're over the target cache size, we want to evict enough | |
2817 | * from the list to get back to our target size. We don't want | |
2818 | * to evict too much from the MRU, such that it drops below | |
2819 | * arc_p. So, if we're over our target cache size more than | |
2820 | * the MRU is over arc_p, we'll evict enough to get back to | |
2821 | * arc_p here, and then evict more from the MFU below. | |
2822 | */ | |
2823 | target = MIN((int64_t)(arc_size - arc_c), | |
2824 | (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - | |
2825 | arc_p)); | |
2826 | ||
2827 | /* | |
2828 | * If we're below arc_meta_min, always prefer to evict data. | |
2829 | * Otherwise, try to satisfy the requested number of bytes to | |
2830 | * evict from the type which contains older buffers; in an | |
2831 | * effort to keep newer buffers in the cache regardless of their | |
2832 | * type. If we cannot satisfy the number of bytes from this | |
2833 | * type, spill over into the next type. | |
2834 | */ | |
2835 | if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && | |
2836 | arc_meta_used > arc_meta_min) { | |
2837 | bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); | |
2838 | total_evicted += bytes; | |
2839 | ||
2840 | /* | |
2841 | * If we couldn't evict our target number of bytes from | |
2842 | * metadata, we try to get the rest from data. | |
2843 | */ | |
2844 | target -= bytes; | |
2845 | ||
2846 | total_evicted += | |
2847 | arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); | |
2848 | } else { | |
2849 | bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); | |
2850 | total_evicted += bytes; | |
2851 | ||
2852 | /* | |
2853 | * If we couldn't evict our target number of bytes from | |
2854 | * data, we try to get the rest from metadata. | |
2855 | */ | |
2856 | target -= bytes; | |
2857 | ||
2858 | total_evicted += | |
2859 | arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); | |
2860 | } | |
2861 | ||
2862 | /* | |
2863 | * Adjust MFU size | |
2864 | * | |
2865 | * Now that we've tried to evict enough from the MRU to get its | |
2866 | * size back to arc_p, if we're still above the target cache | |
2867 | * size, we evict the rest from the MFU. | |
2868 | */ | |
2869 | target = arc_size - arc_c; | |
2870 | ||
a7b10a93 | 2871 | if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && |
ca0bf58d PS |
2872 | arc_meta_used > arc_meta_min) { |
2873 | bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); | |
2874 | total_evicted += bytes; | |
2875 | ||
2876 | /* | |
2877 | * If we couldn't evict our target number of bytes from | |
2878 | * metadata, we try to get the rest from data. | |
2879 | */ | |
2880 | target -= bytes; | |
2881 | ||
2882 | total_evicted += | |
2883 | arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); | |
2884 | } else { | |
2885 | bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); | |
2886 | total_evicted += bytes; | |
2887 | ||
2888 | /* | |
2889 | * If we couldn't evict our target number of bytes from | |
2890 | * data, we try to get the rest from data. | |
2891 | */ | |
2892 | target -= bytes; | |
2893 | ||
2894 | total_evicted += | |
2895 | arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); | |
2896 | } | |
2897 | ||
2898 | /* | |
2899 | * Adjust ghost lists | |
2900 | * | |
2901 | * In addition to the above, the ARC also defines target values | |
2902 | * for the ghost lists. The sum of the mru list and mru ghost | |
2903 | * list should never exceed the target size of the cache, and | |
2904 | * the sum of the mru list, mfu list, mru ghost list, and mfu | |
2905 | * ghost list should never exceed twice the target size of the | |
2906 | * cache. The following logic enforces these limits on the ghost | |
2907 | * caches, and evicts from them as needed. | |
2908 | */ | |
2909 | target = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; | |
2910 | ||
2911 | bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); | |
2912 | total_evicted += bytes; | |
2913 | ||
2914 | target -= bytes; | |
2915 | ||
2916 | total_evicted += | |
2917 | arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); | |
2918 | ||
2919 | /* | |
2920 | * We assume the sum of the mru list and mfu list is less than | |
2921 | * or equal to arc_c (we enforced this above), which means we | |
2922 | * can use the simpler of the two equations below: | |
2923 | * | |
2924 | * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c | |
2925 | * mru ghost + mfu ghost <= arc_c | |
2926 | */ | |
2927 | target = arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; | |
2928 | ||
2929 | bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); | |
2930 | total_evicted += bytes; | |
2931 | ||
2932 | target -= bytes; | |
2933 | ||
2934 | total_evicted += | |
2935 | arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); | |
2936 | ||
2937 | return (total_evicted); | |
2938 | } | |
2939 | ||
34dc7c2f BB |
2940 | static void |
2941 | arc_do_user_evicts(void) | |
2942 | { | |
ca0bf58d | 2943 | mutex_enter(&arc_user_evicts_lock); |
34dc7c2f BB |
2944 | while (arc_eviction_list != NULL) { |
2945 | arc_buf_t *buf = arc_eviction_list; | |
2946 | arc_eviction_list = buf->b_next; | |
428870ff | 2947 | mutex_enter(&buf->b_evict_lock); |
34dc7c2f | 2948 | buf->b_hdr = NULL; |
428870ff | 2949 | mutex_exit(&buf->b_evict_lock); |
ca0bf58d | 2950 | mutex_exit(&arc_user_evicts_lock); |
34dc7c2f BB |
2951 | |
2952 | if (buf->b_efunc != NULL) | |
bd089c54 | 2953 | VERIFY0(buf->b_efunc(buf->b_private)); |
34dc7c2f BB |
2954 | |
2955 | buf->b_efunc = NULL; | |
2956 | buf->b_private = NULL; | |
2957 | kmem_cache_free(buf_cache, buf); | |
ca0bf58d | 2958 | mutex_enter(&arc_user_evicts_lock); |
34dc7c2f | 2959 | } |
ca0bf58d | 2960 | mutex_exit(&arc_user_evicts_lock); |
34dc7c2f BB |
2961 | } |
2962 | ||
ca0bf58d PS |
2963 | void |
2964 | arc_flush(spa_t *spa, boolean_t retry) | |
ab26409d | 2965 | { |
ca0bf58d | 2966 | uint64_t guid = 0; |
94520ca4 | 2967 | |
bc888666 | 2968 | /* |
ca0bf58d PS |
2969 | * If retry is TRUE, a spa must not be specified since we have |
2970 | * no good way to determine if all of a spa's buffers have been | |
2971 | * evicted from an arc state. | |
bc888666 | 2972 | */ |
ca0bf58d | 2973 | ASSERT(!retry || spa == 0); |
d164b209 | 2974 | |
b9541d6b | 2975 | if (spa != NULL) |
3541dc6d | 2976 | guid = spa_load_guid(spa); |
d164b209 | 2977 | |
ca0bf58d PS |
2978 | (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); |
2979 | (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); | |
2980 | ||
2981 | (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); | |
2982 | (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); | |
2983 | ||
2984 | (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); | |
2985 | (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); | |
34dc7c2f | 2986 | |
ca0bf58d PS |
2987 | (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); |
2988 | (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); | |
34dc7c2f | 2989 | |
34dc7c2f | 2990 | arc_do_user_evicts(); |
34dc7c2f BB |
2991 | ASSERT(spa || arc_eviction_list == NULL); |
2992 | } | |
2993 | ||
34dc7c2f | 2994 | void |
ca67b33a | 2995 | arc_shrink(int64_t to_free) |
34dc7c2f BB |
2996 | { |
2997 | if (arc_c > arc_c_min) { | |
302f753f | 2998 | |
34dc7c2f BB |
2999 | if (arc_c > arc_c_min + to_free) |
3000 | atomic_add_64(&arc_c, -to_free); | |
3001 | else | |
3002 | arc_c = arc_c_min; | |
3003 | ||
ca67b33a | 3004 | atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); |
34dc7c2f BB |
3005 | if (arc_c > arc_size) |
3006 | arc_c = MAX(arc_size, arc_c_min); | |
3007 | if (arc_p > arc_c) | |
3008 | arc_p = (arc_c >> 1); | |
3009 | ASSERT(arc_c >= arc_c_min); | |
3010 | ASSERT((int64_t)arc_p >= 0); | |
3011 | } | |
3012 | ||
3013 | if (arc_size > arc_c) | |
ca0bf58d | 3014 | (void) arc_adjust(); |
34dc7c2f BB |
3015 | } |
3016 | ||
ca67b33a MA |
3017 | typedef enum free_memory_reason_t { |
3018 | FMR_UNKNOWN, | |
3019 | FMR_NEEDFREE, | |
3020 | FMR_LOTSFREE, | |
3021 | FMR_SWAPFS_MINFREE, | |
3022 | FMR_PAGES_PP_MAXIMUM, | |
3023 | FMR_HEAP_ARENA, | |
3024 | FMR_ZIO_ARENA, | |
3025 | } free_memory_reason_t; | |
3026 | ||
3027 | int64_t last_free_memory; | |
3028 | free_memory_reason_t last_free_reason; | |
3029 | ||
3030 | #ifdef _KERNEL | |
3031 | #ifdef __linux__ | |
3032 | /* | |
3033 | * expiration time for arc_no_grow set by direct memory reclaim. | |
3034 | */ | |
3035 | static clock_t arc_grow_time = 0; | |
3036 | #else | |
3037 | /* | |
3038 | * Additional reserve of pages for pp_reserve. | |
3039 | */ | |
3040 | int64_t arc_pages_pp_reserve = 64; | |
3041 | ||
3042 | /* | |
3043 | * Additional reserve of pages for swapfs. | |
3044 | */ | |
3045 | int64_t arc_swapfs_reserve = 64; | |
3046 | #endif | |
3047 | #endif /* _KERNEL */ | |
3048 | ||
3049 | /* | |
3050 | * Return the amount of memory that can be consumed before reclaim will be | |
3051 | * needed. Positive if there is sufficient free memory, negative indicates | |
3052 | * the amount of memory that needs to be freed up. | |
3053 | */ | |
3054 | static int64_t | |
3055 | arc_available_memory(void) | |
3056 | { | |
3057 | int64_t lowest = INT64_MAX; | |
3058 | free_memory_reason_t r = FMR_UNKNOWN; | |
3059 | ||
3060 | #ifdef _KERNEL | |
3061 | #ifdef __linux__ | |
3062 | /* | |
3063 | * Under Linux we are not allowed to directly interrogate the global | |
3064 | * memory state. Instead rely on observing that direct reclaim has | |
3065 | * recently occurred therefore the system must be low on memory. The | |
3066 | * exact values returned are not critical but should be small. | |
3067 | */ | |
3068 | if (ddi_time_after_eq(ddi_get_lbolt(), arc_grow_time)) | |
3069 | lowest = PAGE_SIZE; | |
3070 | else | |
3071 | lowest = -PAGE_SIZE; | |
3072 | #else | |
3073 | int64_t n; | |
3074 | ||
3075 | /* | |
3076 | * Platforms like illumos have greater visibility in to the memory | |
3077 | * subsystem and can return a more detailed analysis of memory. | |
3078 | */ | |
3079 | if (needfree > 0) { | |
3080 | n = PAGESIZE * (-needfree); | |
3081 | if (n < lowest) { | |
3082 | lowest = n; | |
3083 | r = FMR_NEEDFREE; | |
3084 | } | |
3085 | } | |
3086 | ||
3087 | /* | |
3088 | * check that we're out of range of the pageout scanner. It starts to | |
3089 | * schedule paging if freemem is less than lotsfree and needfree. | |
3090 | * lotsfree is the high-water mark for pageout, and needfree is the | |
3091 | * number of needed free pages. We add extra pages here to make sure | |
3092 | * the scanner doesn't start up while we're freeing memory. | |
3093 | */ | |
3094 | n = PAGESIZE * (freemem - lotsfree - needfree - desfree); | |
3095 | if (n < lowest) { | |
3096 | lowest = n; | |
3097 | r = FMR_LOTSFREE; | |
3098 | } | |
3099 | ||
3100 | /* | |
3101 | * check to make sure that swapfs has enough space so that anon | |
3102 | * reservations can still succeed. anon_resvmem() checks that the | |
3103 | * availrmem is greater than swapfs_minfree, and the number of reserved | |
3104 | * swap pages. We also add a bit of extra here just to prevent | |
3105 | * circumstances from getting really dire. | |
3106 | */ | |
3107 | n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - | |
3108 | desfree - arc_swapfs_reserve); | |
3109 | if (n < lowest) { | |
3110 | lowest = n; | |
3111 | r = FMR_SWAPFS_MINFREE; | |
3112 | } | |
3113 | ||
3114 | ||
3115 | /* | |
3116 | * Check that we have enough availrmem that memory locking (e.g., via | |
3117 | * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum | |
3118 | * stores the number of pages that cannot be locked; when availrmem | |
3119 | * drops below pages_pp_maximum, page locking mechanisms such as | |
3120 | * page_pp_lock() will fail.) | |
3121 | */ | |
3122 | n = PAGESIZE * (availrmem - pages_pp_maximum - | |
3123 | arc_pages_pp_reserve); | |
3124 | if (n < lowest) { | |
3125 | lowest = n; | |
3126 | r = FMR_PAGES_PP_MAXIMUM; | |
3127 | } | |
3128 | ||
3129 | #if defined(__i386) | |
3130 | /* | |
3131 | * If we're on an i386 platform, it's possible that we'll exhaust the | |
3132 | * kernel heap space before we ever run out of available physical | |
3133 | * memory. Most checks of the size of the heap_area compare against | |
3134 | * tune.t_minarmem, which is the minimum available real memory that we | |
3135 | * can have in the system. However, this is generally fixed at 25 pages | |
3136 | * which is so low that it's useless. In this comparison, we seek to | |
3137 | * calculate the total heap-size, and reclaim if more than 3/4ths of the | |
3138 | * heap is allocated. (Or, in the calculation, if less than 1/4th is | |
3139 | * free) | |
3140 | */ | |
3141 | n = vmem_size(heap_arena, VMEM_FREE) - | |
3142 | (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); | |
3143 | if (n < lowest) { | |
3144 | lowest = n; | |
3145 | r = FMR_HEAP_ARENA; | |
3146 | } | |
3147 | #endif | |
3148 | ||
3149 | /* | |
3150 | * If zio data pages are being allocated out of a separate heap segment, | |
3151 | * then enforce that the size of available vmem for this arena remains | |
3152 | * above about 1/16th free. | |
3153 | * | |
3154 | * Note: The 1/16th arena free requirement was put in place | |
3155 | * to aggressively evict memory from the arc in order to avoid | |
3156 | * memory fragmentation issues. | |
3157 | */ | |
3158 | if (zio_arena != NULL) { | |
3159 | n = vmem_size(zio_arena, VMEM_FREE) - | |
3160 | (vmem_size(zio_arena, VMEM_ALLOC) >> 4); | |
3161 | if (n < lowest) { | |
3162 | lowest = n; | |
3163 | r = FMR_ZIO_ARENA; | |
3164 | } | |
3165 | } | |
3166 | #endif /* __linux__ */ | |
3167 | #else | |
3168 | /* Every 100 calls, free a small amount */ | |
3169 | if (spa_get_random(100) == 0) | |
3170 | lowest = -1024; | |
3171 | #endif | |
3172 | ||
3173 | last_free_memory = lowest; | |
3174 | last_free_reason = r; | |
3175 | ||
3176 | return (lowest); | |
3177 | } | |
3178 | ||
3179 | /* | |
3180 | * Determine if the system is under memory pressure and is asking | |
3181 | * to reclaim memory. A return value of TRUE indicates that the system | |
3182 | * is under memory pressure and that the arc should adjust accordingly. | |
3183 | */ | |
3184 | static boolean_t | |
3185 | arc_reclaim_needed(void) | |
3186 | { | |
3187 | return (arc_available_memory() < 0); | |
3188 | } | |
3189 | ||
34dc7c2f | 3190 | static void |
ca67b33a | 3191 | arc_kmem_reap_now(void) |
34dc7c2f BB |
3192 | { |
3193 | size_t i; | |
3194 | kmem_cache_t *prev_cache = NULL; | |
3195 | kmem_cache_t *prev_data_cache = NULL; | |
3196 | extern kmem_cache_t *zio_buf_cache[]; | |
3197 | extern kmem_cache_t *zio_data_buf_cache[]; | |
669dedb3 | 3198 | extern kmem_cache_t *range_seg_cache; |
34dc7c2f | 3199 | |
f6046738 BB |
3200 | if ((arc_meta_used >= arc_meta_limit) && zfs_arc_meta_prune) { |
3201 | /* | |
3202 | * We are exceeding our meta-data cache limit. | |
3203 | * Prune some entries to release holds on meta-data. | |
3204 | */ | |
3205 | arc_prune(zfs_arc_meta_prune); | |
3206 | } | |
3207 | ||
34dc7c2f BB |
3208 | for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { |
3209 | if (zio_buf_cache[i] != prev_cache) { | |
3210 | prev_cache = zio_buf_cache[i]; | |
3211 | kmem_cache_reap_now(zio_buf_cache[i]); | |
3212 | } | |
3213 | if (zio_data_buf_cache[i] != prev_data_cache) { | |
3214 | prev_data_cache = zio_data_buf_cache[i]; | |
3215 | kmem_cache_reap_now(zio_data_buf_cache[i]); | |
3216 | } | |
3217 | } | |
ca0bf58d | 3218 | kmem_cache_reap_now(buf_cache); |
b9541d6b CW |
3219 | kmem_cache_reap_now(hdr_full_cache); |
3220 | kmem_cache_reap_now(hdr_l2only_cache); | |
669dedb3 | 3221 | kmem_cache_reap_now(range_seg_cache); |
ca67b33a MA |
3222 | |
3223 | if (zio_arena != NULL) { | |
3224 | /* | |
3225 | * Ask the vmem arena to reclaim unused memory from its | |
3226 | * quantum caches. | |
3227 | */ | |
3228 | vmem_qcache_reap(zio_arena); | |
3229 | } | |
34dc7c2f BB |
3230 | } |
3231 | ||
302f753f | 3232 | /* |
ca0bf58d PS |
3233 | * Threads can block in arc_get_data_buf() waiting for this thread to evict |
3234 | * enough data and signal them to proceed. When this happens, the threads in | |
3235 | * arc_get_data_buf() are sleeping while holding the hash lock for their | |
3236 | * particular arc header. Thus, we must be careful to never sleep on a | |
3237 | * hash lock in this thread. This is to prevent the following deadlock: | |
3238 | * | |
3239 | * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", | |
3240 | * waiting for the reclaim thread to signal it. | |
3241 | * | |
3242 | * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, | |
3243 | * fails, and goes to sleep forever. | |
3244 | * | |
3245 | * This possible deadlock is avoided by always acquiring a hash lock | |
3246 | * using mutex_tryenter() from arc_reclaim_thread(). | |
302f753f | 3247 | */ |
34dc7c2f | 3248 | static void |
ca67b33a | 3249 | arc_reclaim_thread(void) |
34dc7c2f | 3250 | { |
ca67b33a MA |
3251 | fstrans_cookie_t cookie = spl_fstrans_mark(); |
3252 | clock_t growtime = 0; | |
34dc7c2f BB |
3253 | callb_cpr_t cpr; |
3254 | ||
ca0bf58d | 3255 | CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); |
34dc7c2f | 3256 | |
ca0bf58d | 3257 | mutex_enter(&arc_reclaim_lock); |
ca67b33a MA |
3258 | while (!arc_reclaim_thread_exit) { |
3259 | int64_t to_free; | |
3260 | int64_t free_memory = arc_available_memory(); | |
3261 | uint64_t evicted = 0; | |
302f753f | 3262 | |
ca67b33a | 3263 | arc_tuning_update(); |
34dc7c2f | 3264 | |
ca67b33a | 3265 | mutex_exit(&arc_reclaim_lock); |
34dc7c2f | 3266 | |
ca67b33a | 3267 | if (free_memory < 0) { |
34dc7c2f | 3268 | |
ca67b33a | 3269 | arc_no_grow = B_TRUE; |
b128c09f | 3270 | arc_warm = B_TRUE; |
34dc7c2f | 3271 | |
ca67b33a MA |
3272 | /* |
3273 | * Wait at least zfs_grow_retry (default 5) seconds | |
3274 | * before considering growing. | |
3275 | */ | |
3276 | growtime = ddi_get_lbolt() + (arc_grow_retry * hz); | |
6a8f9b6b | 3277 | |
ca67b33a | 3278 | arc_kmem_reap_now(); |
34dc7c2f | 3279 | |
ca67b33a MA |
3280 | /* |
3281 | * If we are still low on memory, shrink the ARC | |
3282 | * so that we have arc_shrink_min free space. | |
3283 | */ | |
3284 | free_memory = arc_available_memory(); | |
34dc7c2f | 3285 | |
ca67b33a MA |
3286 | to_free = (arc_c >> arc_shrink_shift) - free_memory; |
3287 | if (to_free > 0) { | |
3288 | #ifdef _KERNEL | |
3289 | to_free = MAX(to_free, ptob(needfree)); | |
3290 | #endif | |
3291 | arc_shrink(to_free); | |
3292 | } | |
3293 | } else if (free_memory < arc_c >> arc_no_grow_shift) { | |
3294 | arc_no_grow = B_TRUE; | |
3295 | } else if (ddi_get_lbolt() >= growtime) { | |
3296 | arc_no_grow = B_FALSE; | |
3297 | } | |
bce45ec9 | 3298 | |
ca67b33a | 3299 | evicted = arc_adjust(); |
bce45ec9 | 3300 | |
ca67b33a | 3301 | mutex_enter(&arc_reclaim_lock); |
bce45ec9 | 3302 | |
ca67b33a MA |
3303 | /* |
3304 | * If evicted is zero, we couldn't evict anything via | |
3305 | * arc_adjust(). This could be due to hash lock | |
3306 | * collisions, but more likely due to the majority of | |
3307 | * arc buffers being unevictable. Therefore, even if | |
3308 | * arc_size is above arc_c, another pass is unlikely to | |
3309 | * be helpful and could potentially cause us to enter an | |
3310 | * infinite loop. | |
3311 | */ | |
3312 | if (arc_size <= arc_c || evicted == 0) { | |
3313 | /* | |
3314 | * We're either no longer overflowing, or we | |
3315 | * can't evict anything more, so we should wake | |
3316 | * up any threads before we go to sleep. | |
3317 | */ | |
3318 | cv_broadcast(&arc_reclaim_waiters_cv); | |
bce45ec9 | 3319 | |
ca67b33a MA |
3320 | /* |
3321 | * Block until signaled, or after one second (we | |
3322 | * might need to perform arc_kmem_reap_now() | |
3323 | * even if we aren't being signalled) | |
3324 | */ | |
3325 | CALLB_CPR_SAFE_BEGIN(&cpr); | |
3326 | (void) cv_timedwait_sig(&arc_reclaim_thread_cv, | |
3327 | &arc_reclaim_lock, ddi_get_lbolt() + hz); | |
3328 | CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); | |
3329 | } | |
ca0bf58d | 3330 | } |
bce45ec9 | 3331 | |
ca67b33a | 3332 | arc_reclaim_thread_exit = FALSE; |
ca0bf58d PS |
3333 | cv_broadcast(&arc_reclaim_thread_cv); |
3334 | CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ | |
3335 | spl_fstrans_unmark(cookie); | |
3336 | thread_exit(); | |
3337 | } | |
3338 | ||
3339 | static void | |
3340 | arc_user_evicts_thread(void) | |
3341 | { | |
ca67b33a | 3342 | fstrans_cookie_t cookie = spl_fstrans_mark(); |
ca0bf58d | 3343 | callb_cpr_t cpr; |
bce45ec9 | 3344 | |
ca0bf58d | 3345 | CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG); |
bce45ec9 | 3346 | |
ca0bf58d PS |
3347 | mutex_enter(&arc_user_evicts_lock); |
3348 | while (!arc_user_evicts_thread_exit) { | |
3349 | mutex_exit(&arc_user_evicts_lock); | |
3350 | ||
3351 | arc_do_user_evicts(); | |
3352 | ||
3353 | /* | |
3354 | * This is necessary in order for the mdb ::arc dcmd to | |
3355 | * show up to date information. Since the ::arc command | |
3356 | * does not call the kstat's update function, without | |
3357 | * this call, the command may show stale stats for the | |
3358 | * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even | |
3359 | * with this change, the data might be up to 1 second | |
3360 | * out of date; but that should suffice. The arc_state_t | |
3361 | * structures can be queried directly if more accurate | |
3362 | * information is needed. | |
3363 | */ | |
3364 | if (arc_ksp != NULL) | |
3365 | arc_ksp->ks_update(arc_ksp, KSTAT_READ); | |
3366 | ||
3367 | mutex_enter(&arc_user_evicts_lock); | |
3368 | ||
3369 | /* | |
3370 | * Block until signaled, or after one second (we need to | |
3371 | * call the arc's kstat update function regularly). | |
3372 | */ | |
3373 | CALLB_CPR_SAFE_BEGIN(&cpr); | |
b64ccd6c | 3374 | (void) cv_timedwait_sig(&arc_user_evicts_cv, |
ca0bf58d PS |
3375 | &arc_user_evicts_lock, ddi_get_lbolt() + hz); |
3376 | CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock); | |
34dc7c2f BB |
3377 | } |
3378 | ||
ca0bf58d PS |
3379 | arc_user_evicts_thread_exit = FALSE; |
3380 | cv_broadcast(&arc_user_evicts_cv); | |
3381 | CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */ | |
40d06e3c | 3382 | spl_fstrans_unmark(cookie); |
34dc7c2f BB |
3383 | thread_exit(); |
3384 | } | |
3385 | ||
7cb67b45 BB |
3386 | #ifdef _KERNEL |
3387 | /* | |
302f753f BB |
3388 | * Determine the amount of memory eligible for eviction contained in the |
3389 | * ARC. All clean data reported by the ghost lists can always be safely | |
3390 | * evicted. Due to arc_c_min, the same does not hold for all clean data | |
3391 | * contained by the regular mru and mfu lists. | |
3392 | * | |
3393 | * In the case of the regular mru and mfu lists, we need to report as | |
3394 | * much clean data as possible, such that evicting that same reported | |
3395 | * data will not bring arc_size below arc_c_min. Thus, in certain | |
3396 | * circumstances, the total amount of clean data in the mru and mfu | |
3397 | * lists might not actually be evictable. | |
3398 | * | |
3399 | * The following two distinct cases are accounted for: | |
3400 | * | |
3401 | * 1. The sum of the amount of dirty data contained by both the mru and | |
3402 | * mfu lists, plus the ARC's other accounting (e.g. the anon list), | |
3403 | * is greater than or equal to arc_c_min. | |
3404 | * (i.e. amount of dirty data >= arc_c_min) | |
3405 | * | |
3406 | * This is the easy case; all clean data contained by the mru and mfu | |
3407 | * lists is evictable. Evicting all clean data can only drop arc_size | |
3408 | * to the amount of dirty data, which is greater than arc_c_min. | |
3409 | * | |
3410 | * 2. The sum of the amount of dirty data contained by both the mru and | |
3411 | * mfu lists, plus the ARC's other accounting (e.g. the anon list), | |
3412 | * is less than arc_c_min. | |
3413 | * (i.e. arc_c_min > amount of dirty data) | |
3414 | * | |
3415 | * 2.1. arc_size is greater than or equal arc_c_min. | |
3416 | * (i.e. arc_size >= arc_c_min > amount of dirty data) | |
3417 | * | |
3418 | * In this case, not all clean data from the regular mru and mfu | |
3419 | * lists is actually evictable; we must leave enough clean data | |
3420 | * to keep arc_size above arc_c_min. Thus, the maximum amount of | |
3421 | * evictable data from the two lists combined, is exactly the | |
3422 | * difference between arc_size and arc_c_min. | |
3423 | * | |
3424 | * 2.2. arc_size is less than arc_c_min | |
3425 | * (i.e. arc_c_min > arc_size > amount of dirty data) | |
3426 | * | |
3427 | * In this case, none of the data contained in the mru and mfu | |
3428 | * lists is evictable, even if it's clean. Since arc_size is | |
3429 | * already below arc_c_min, evicting any more would only | |
3430 | * increase this negative difference. | |
7cb67b45 | 3431 | */ |
302f753f BB |
3432 | static uint64_t |
3433 | arc_evictable_memory(void) { | |
3434 | uint64_t arc_clean = | |
3435 | arc_mru->arcs_lsize[ARC_BUFC_DATA] + | |
3436 | arc_mru->arcs_lsize[ARC_BUFC_METADATA] + | |
3437 | arc_mfu->arcs_lsize[ARC_BUFC_DATA] + | |
3438 | arc_mfu->arcs_lsize[ARC_BUFC_METADATA]; | |
3439 | uint64_t ghost_clean = | |
3440 | arc_mru_ghost->arcs_lsize[ARC_BUFC_DATA] + | |
3441 | arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] + | |
3442 | arc_mfu_ghost->arcs_lsize[ARC_BUFC_DATA] + | |
3443 | arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA]; | |
3444 | uint64_t arc_dirty = MAX((int64_t)arc_size - (int64_t)arc_clean, 0); | |
3445 | ||
3446 | if (arc_dirty >= arc_c_min) | |
3447 | return (ghost_clean + arc_clean); | |
3448 | ||
3449 | return (ghost_clean + MAX((int64_t)arc_size - (int64_t)arc_c_min, 0)); | |
3450 | } | |
3451 | ||
ed6e9cc2 TC |
3452 | /* |
3453 | * If sc->nr_to_scan is zero, the caller is requesting a query of the | |
3454 | * number of objects which can potentially be freed. If it is nonzero, | |
3455 | * the request is to free that many objects. | |
3456 | * | |
3457 | * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks | |
3458 | * in struct shrinker and also require the shrinker to return the number | |
3459 | * of objects freed. | |
3460 | * | |
3461 | * Older kernels require the shrinker to return the number of freeable | |
3462 | * objects following the freeing of nr_to_free. | |
3463 | */ | |
3464 | static spl_shrinker_t | |
7e7baeca | 3465 | __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc) |
7cb67b45 | 3466 | { |
ed6e9cc2 | 3467 | int64_t pages; |
7cb67b45 | 3468 | |
302f753f BB |
3469 | /* The arc is considered warm once reclaim has occurred */ |
3470 | if (unlikely(arc_warm == B_FALSE)) | |
3471 | arc_warm = B_TRUE; | |
7cb67b45 | 3472 | |
302f753f | 3473 | /* Return the potential number of reclaimable pages */ |
ed6e9cc2 | 3474 | pages = btop((int64_t)arc_evictable_memory()); |
302f753f BB |
3475 | if (sc->nr_to_scan == 0) |
3476 | return (pages); | |
3fd70ee6 BB |
3477 | |
3478 | /* Not allowed to perform filesystem reclaim */ | |
7e7baeca | 3479 | if (!(sc->gfp_mask & __GFP_FS)) |
ed6e9cc2 | 3480 | return (SHRINK_STOP); |
3fd70ee6 | 3481 | |
7cb67b45 | 3482 | /* Reclaim in progress */ |
ca0bf58d | 3483 | if (mutex_tryenter(&arc_reclaim_lock) == 0) |
ed6e9cc2 | 3484 | return (SHRINK_STOP); |
7cb67b45 | 3485 | |
ca0bf58d PS |
3486 | mutex_exit(&arc_reclaim_lock); |
3487 | ||
302f753f BB |
3488 | /* |
3489 | * Evict the requested number of pages by shrinking arc_c the | |
3490 | * requested amount. If there is nothing left to evict just | |
3491 | * reap whatever we can from the various arc slabs. | |
3492 | */ | |
3493 | if (pages > 0) { | |
ca67b33a MA |
3494 | arc_shrink(ptob(sc->nr_to_scan)); |
3495 | arc_kmem_reap_now(); | |
ed6e9cc2 TC |
3496 | #ifdef HAVE_SPLIT_SHRINKER_CALLBACK |
3497 | pages = MAX(pages - btop(arc_evictable_memory()), 0); | |
3498 | #else | |
1e3cb67b | 3499 | pages = btop(arc_evictable_memory()); |
ed6e9cc2 | 3500 | #endif |
302f753f | 3501 | } else { |
ca67b33a | 3502 | arc_kmem_reap_now(); |
ed6e9cc2 | 3503 | pages = SHRINK_STOP; |
302f753f BB |
3504 | } |
3505 | ||
ca0bf58d PS |
3506 | /* |
3507 | * We've reaped what we can, wake up threads. | |
3508 | */ | |
3509 | cv_broadcast(&arc_reclaim_waiters_cv); | |
3510 | ||
302f753f BB |
3511 | /* |
3512 | * When direct reclaim is observed it usually indicates a rapid | |
3513 | * increase in memory pressure. This occurs because the kswapd | |
3514 | * threads were unable to asynchronously keep enough free memory | |
3515 | * available. In this case set arc_no_grow to briefly pause arc | |
3516 | * growth to avoid compounding the memory pressure. | |
3517 | */ | |
7cb67b45 | 3518 | if (current_is_kswapd()) { |
302f753f | 3519 | ARCSTAT_BUMP(arcstat_memory_indirect_count); |
7cb67b45 | 3520 | } else { |
302f753f | 3521 | arc_no_grow = B_TRUE; |
bce45ec9 | 3522 | arc_grow_time = ddi_get_lbolt() + (zfs_arc_grow_retry * hz); |
302f753f | 3523 | ARCSTAT_BUMP(arcstat_memory_direct_count); |
7cb67b45 BB |
3524 | } |
3525 | ||
1e3cb67b | 3526 | return (pages); |
7cb67b45 | 3527 | } |
7e7baeca | 3528 | SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func); |
7cb67b45 BB |
3529 | |
3530 | SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS); | |
3531 | #endif /* _KERNEL */ | |
3532 | ||
34dc7c2f BB |
3533 | /* |
3534 | * Adapt arc info given the number of bytes we are trying to add and | |
3535 | * the state that we are comming from. This function is only called | |
3536 | * when we are adding new content to the cache. | |
3537 | */ | |
3538 | static void | |
3539 | arc_adapt(int bytes, arc_state_t *state) | |
3540 | { | |
3541 | int mult; | |
3542 | ||
3543 | if (state == arc_l2c_only) | |
3544 | return; | |
3545 | ||
3546 | ASSERT(bytes > 0); | |
3547 | /* | |
3548 | * Adapt the target size of the MRU list: | |
3549 | * - if we just hit in the MRU ghost list, then increase | |
3550 | * the target size of the MRU list. | |
3551 | * - if we just hit in the MFU ghost list, then increase | |
3552 | * the target size of the MFU list by decreasing the | |
3553 | * target size of the MRU list. | |
3554 | */ | |
3555 | if (state == arc_mru_ghost) { | |
3556 | mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? | |
3557 | 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); | |
62422785 PS |
3558 | |
3559 | if (!zfs_arc_p_dampener_disable) | |
3560 | mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ | |
34dc7c2f | 3561 | |
f521ce1b | 3562 | arc_p = MIN(arc_c, arc_p + bytes * mult); |
34dc7c2f | 3563 | } else if (state == arc_mfu_ghost) { |
d164b209 BB |
3564 | uint64_t delta; |
3565 | ||
34dc7c2f BB |
3566 | mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? |
3567 | 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); | |
62422785 PS |
3568 | |
3569 | if (!zfs_arc_p_dampener_disable) | |
3570 | mult = MIN(mult, 10); | |
34dc7c2f | 3571 | |
d164b209 | 3572 | delta = MIN(bytes * mult, arc_p); |
f521ce1b | 3573 | arc_p = MAX(0, arc_p - delta); |
34dc7c2f BB |
3574 | } |
3575 | ASSERT((int64_t)arc_p >= 0); | |
3576 | ||
ca67b33a MA |
3577 | if (arc_reclaim_needed()) { |
3578 | cv_signal(&arc_reclaim_thread_cv); | |
3579 | return; | |
3580 | } | |
3581 | ||
34dc7c2f BB |
3582 | if (arc_no_grow) |
3583 | return; | |
3584 | ||
3585 | if (arc_c >= arc_c_max) | |
3586 | return; | |
3587 | ||
3588 | /* | |
3589 | * If we're within (2 * maxblocksize) bytes of the target | |
3590 | * cache size, increment the target cache size | |
3591 | */ | |
121b3cae TC |
3592 | VERIFY3U(arc_c, >=, 2ULL << SPA_MAXBLOCKSHIFT); |
3593 | if (arc_size >= arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { | |
34dc7c2f BB |
3594 | atomic_add_64(&arc_c, (int64_t)bytes); |
3595 | if (arc_c > arc_c_max) | |
3596 | arc_c = arc_c_max; | |
3597 | else if (state == arc_anon) | |
3598 | atomic_add_64(&arc_p, (int64_t)bytes); | |
3599 | if (arc_p > arc_c) | |
3600 | arc_p = arc_c; | |
3601 | } | |
3602 | ASSERT((int64_t)arc_p >= 0); | |
3603 | } | |
3604 | ||
3605 | /* | |
ca0bf58d PS |
3606 | * Check if arc_size has grown past our upper threshold, determined by |
3607 | * zfs_arc_overflow_shift. | |
34dc7c2f | 3608 | */ |
ca0bf58d PS |
3609 | static boolean_t |
3610 | arc_is_overflowing(void) | |
34dc7c2f | 3611 | { |
ca0bf58d PS |
3612 | /* Always allow at least one block of overflow */ |
3613 | uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, | |
3614 | arc_c >> zfs_arc_overflow_shift); | |
34dc7c2f | 3615 | |
ca0bf58d | 3616 | return (arc_size >= arc_c + overflow); |
34dc7c2f BB |
3617 | } |
3618 | ||
3619 | /* | |
ca0bf58d PS |
3620 | * The buffer, supplied as the first argument, needs a data block. If we |
3621 | * are hitting the hard limit for the cache size, we must sleep, waiting | |
3622 | * for the eviction thread to catch up. If we're past the target size | |
3623 | * but below the hard limit, we'll only signal the reclaim thread and | |
3624 | * continue on. | |
34dc7c2f BB |
3625 | */ |
3626 | static void | |
3627 | arc_get_data_buf(arc_buf_t *buf) | |
3628 | { | |
b9541d6b | 3629 | arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; |
34dc7c2f | 3630 | uint64_t size = buf->b_hdr->b_size; |
b9541d6b | 3631 | arc_buf_contents_t type = arc_buf_type(buf->b_hdr); |
34dc7c2f BB |
3632 | |
3633 | arc_adapt(size, state); | |
3634 | ||
3635 | /* | |
ca0bf58d PS |
3636 | * If arc_size is currently overflowing, and has grown past our |
3637 | * upper limit, we must be adding data faster than the evict | |
3638 | * thread can evict. Thus, to ensure we don't compound the | |
3639 | * problem by adding more data and forcing arc_size to grow even | |
3640 | * further past it's target size, we halt and wait for the | |
3641 | * eviction thread to catch up. | |
3642 | * | |
3643 | * It's also possible that the reclaim thread is unable to evict | |
3644 | * enough buffers to get arc_size below the overflow limit (e.g. | |
3645 | * due to buffers being un-evictable, or hash lock collisions). | |
3646 | * In this case, we want to proceed regardless if we're | |
3647 | * overflowing; thus we don't use a while loop here. | |
34dc7c2f | 3648 | */ |
ca0bf58d PS |
3649 | if (arc_is_overflowing()) { |
3650 | mutex_enter(&arc_reclaim_lock); | |
3651 | ||
3652 | /* | |
3653 | * Now that we've acquired the lock, we may no longer be | |
3654 | * over the overflow limit, lets check. | |
3655 | * | |
3656 | * We're ignoring the case of spurious wake ups. If that | |
3657 | * were to happen, it'd let this thread consume an ARC | |
3658 | * buffer before it should have (i.e. before we're under | |
3659 | * the overflow limit and were signalled by the reclaim | |
3660 | * thread). As long as that is a rare occurrence, it | |
3661 | * shouldn't cause any harm. | |
3662 | */ | |
3663 | if (arc_is_overflowing()) { | |
3664 | cv_signal(&arc_reclaim_thread_cv); | |
3665 | cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); | |
34dc7c2f | 3666 | } |
34dc7c2f | 3667 | |
ca0bf58d | 3668 | mutex_exit(&arc_reclaim_lock); |
34dc7c2f | 3669 | } |
ab26409d | 3670 | |
da8ccd0e | 3671 | if (type == ARC_BUFC_METADATA) { |
ca0bf58d PS |
3672 | buf->b_data = zio_buf_alloc(size); |
3673 | arc_space_consume(size, ARC_SPACE_META); | |
3674 | } else { | |
3675 | ASSERT(type == ARC_BUFC_DATA); | |
3676 | buf->b_data = zio_data_buf_alloc(size); | |
3677 | arc_space_consume(size, ARC_SPACE_DATA); | |
da8ccd0e PS |
3678 | } |
3679 | ||
34dc7c2f BB |
3680 | /* |
3681 | * Update the state size. Note that ghost states have a | |
3682 | * "ghost size" and so don't need to be updated. | |
3683 | */ | |
b9541d6b | 3684 | if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { |
34dc7c2f BB |
3685 | arc_buf_hdr_t *hdr = buf->b_hdr; |
3686 | ||
b9541d6b | 3687 | atomic_add_64(&hdr->b_l1hdr.b_state->arcs_size, size); |
ca0bf58d PS |
3688 | |
3689 | /* | |
3690 | * If this is reached via arc_read, the link is | |
3691 | * protected by the hash lock. If reached via | |
3692 | * arc_buf_alloc, the header should not be accessed by | |
3693 | * any other thread. And, if reached via arc_read_done, | |
3694 | * the hash lock will protect it if it's found in the | |
3695 | * hash table; otherwise no other thread should be | |
3696 | * trying to [add|remove]_reference it. | |
3697 | */ | |
3698 | if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { | |
b9541d6b CW |
3699 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
3700 | atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], | |
3701 | size); | |
34dc7c2f BB |
3702 | } |
3703 | /* | |
3704 | * If we are growing the cache, and we are adding anonymous | |
3705 | * data, and we have outgrown arc_p, update arc_p | |
3706 | */ | |
ca0bf58d | 3707 | if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && |
34dc7c2f BB |
3708 | arc_anon->arcs_size + arc_mru->arcs_size > arc_p) |
3709 | arc_p = MIN(arc_c, arc_p + size); | |
3710 | } | |
3711 | } | |
3712 | ||
3713 | /* | |
3714 | * This routine is called whenever a buffer is accessed. | |
3715 | * NOTE: the hash lock is dropped in this function. | |
3716 | */ | |
3717 | static void | |
2a432414 | 3718 | arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) |
34dc7c2f | 3719 | { |
428870ff BB |
3720 | clock_t now; |
3721 | ||
34dc7c2f | 3722 | ASSERT(MUTEX_HELD(hash_lock)); |
b9541d6b | 3723 | ASSERT(HDR_HAS_L1HDR(hdr)); |
34dc7c2f | 3724 | |
b9541d6b | 3725 | if (hdr->b_l1hdr.b_state == arc_anon) { |
34dc7c2f BB |
3726 | /* |
3727 | * This buffer is not in the cache, and does not | |
3728 | * appear in our "ghost" list. Add the new buffer | |
3729 | * to the MRU state. | |
3730 | */ | |
3731 | ||
b9541d6b CW |
3732 | ASSERT0(hdr->b_l1hdr.b_arc_access); |
3733 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); | |
2a432414 GW |
3734 | DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); |
3735 | arc_change_state(arc_mru, hdr, hash_lock); | |
34dc7c2f | 3736 | |
b9541d6b | 3737 | } else if (hdr->b_l1hdr.b_state == arc_mru) { |
428870ff BB |
3738 | now = ddi_get_lbolt(); |
3739 | ||
34dc7c2f BB |
3740 | /* |
3741 | * If this buffer is here because of a prefetch, then either: | |
3742 | * - clear the flag if this is a "referencing" read | |
3743 | * (any subsequent access will bump this into the MFU state). | |
3744 | * or | |
3745 | * - move the buffer to the head of the list if this is | |
3746 | * another prefetch (to make it less likely to be evicted). | |
3747 | */ | |
b9541d6b CW |
3748 | if (HDR_PREFETCH(hdr)) { |
3749 | if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { | |
ca0bf58d PS |
3750 | /* link protected by hash lock */ |
3751 | ASSERT(multilist_link_active( | |
b9541d6b | 3752 | &hdr->b_l1hdr.b_arc_node)); |
34dc7c2f | 3753 | } else { |
2a432414 | 3754 | hdr->b_flags &= ~ARC_FLAG_PREFETCH; |
b9541d6b | 3755 | atomic_inc_32(&hdr->b_l1hdr.b_mru_hits); |
34dc7c2f BB |
3756 | ARCSTAT_BUMP(arcstat_mru_hits); |
3757 | } | |
b9541d6b | 3758 | hdr->b_l1hdr.b_arc_access = now; |
34dc7c2f BB |
3759 | return; |
3760 | } | |
3761 | ||
3762 | /* | |
3763 | * This buffer has been "accessed" only once so far, | |
3764 | * but it is still in the cache. Move it to the MFU | |
3765 | * state. | |
3766 | */ | |
b9541d6b CW |
3767 | if (ddi_time_after(now, hdr->b_l1hdr.b_arc_access + |
3768 | ARC_MINTIME)) { | |
34dc7c2f BB |
3769 | /* |
3770 | * More than 125ms have passed since we | |
3771 | * instantiated this buffer. Move it to the | |
3772 | * most frequently used state. | |
3773 | */ | |
b9541d6b | 3774 | hdr->b_l1hdr.b_arc_access = now; |
2a432414 GW |
3775 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); |
3776 | arc_change_state(arc_mfu, hdr, hash_lock); | |
34dc7c2f | 3777 | } |
b9541d6b | 3778 | atomic_inc_32(&hdr->b_l1hdr.b_mru_hits); |
34dc7c2f | 3779 | ARCSTAT_BUMP(arcstat_mru_hits); |
b9541d6b | 3780 | } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { |
34dc7c2f BB |
3781 | arc_state_t *new_state; |
3782 | /* | |
3783 | * This buffer has been "accessed" recently, but | |
3784 | * was evicted from the cache. Move it to the | |
3785 | * MFU state. | |
3786 | */ | |
3787 | ||
b9541d6b | 3788 | if (HDR_PREFETCH(hdr)) { |
34dc7c2f | 3789 | new_state = arc_mru; |
b9541d6b | 3790 | if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) |
2a432414 GW |
3791 | hdr->b_flags &= ~ARC_FLAG_PREFETCH; |
3792 | DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); | |
34dc7c2f BB |
3793 | } else { |
3794 | new_state = arc_mfu; | |
2a432414 | 3795 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); |
34dc7c2f BB |
3796 | } |
3797 | ||
b9541d6b | 3798 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); |
2a432414 | 3799 | arc_change_state(new_state, hdr, hash_lock); |
34dc7c2f | 3800 | |
b9541d6b | 3801 | atomic_inc_32(&hdr->b_l1hdr.b_mru_ghost_hits); |
34dc7c2f | 3802 | ARCSTAT_BUMP(arcstat_mru_ghost_hits); |
b9541d6b | 3803 | } else if (hdr->b_l1hdr.b_state == arc_mfu) { |
34dc7c2f BB |
3804 | /* |
3805 | * This buffer has been accessed more than once and is | |
3806 | * still in the cache. Keep it in the MFU state. | |
3807 | * | |
3808 | * NOTE: an add_reference() that occurred when we did | |
3809 | * the arc_read() will have kicked this off the list. | |
3810 | * If it was a prefetch, we will explicitly move it to | |
3811 | * the head of the list now. | |
3812 | */ | |
b9541d6b CW |
3813 | if ((HDR_PREFETCH(hdr)) != 0) { |
3814 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); | |
ca0bf58d PS |
3815 | /* link protected by hash_lock */ |
3816 | ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); | |
34dc7c2f | 3817 | } |
b9541d6b | 3818 | atomic_inc_32(&hdr->b_l1hdr.b_mfu_hits); |
34dc7c2f | 3819 | ARCSTAT_BUMP(arcstat_mfu_hits); |
b9541d6b CW |
3820 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); |
3821 | } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { | |
34dc7c2f BB |
3822 | arc_state_t *new_state = arc_mfu; |
3823 | /* | |
3824 | * This buffer has been accessed more than once but has | |
3825 | * been evicted from the cache. Move it back to the | |
3826 | * MFU state. | |
3827 | */ | |
3828 | ||
b9541d6b | 3829 | if (HDR_PREFETCH(hdr)) { |
34dc7c2f BB |
3830 | /* |
3831 | * This is a prefetch access... | |
3832 | * move this block back to the MRU state. | |
3833 | */ | |
b9541d6b | 3834 | ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); |
34dc7c2f BB |
3835 | new_state = arc_mru; |
3836 | } | |
3837 | ||
b9541d6b | 3838 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); |
2a432414 GW |
3839 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); |
3840 | arc_change_state(new_state, hdr, hash_lock); | |
34dc7c2f | 3841 | |
b9541d6b | 3842 | atomic_inc_32(&hdr->b_l1hdr.b_mfu_ghost_hits); |
34dc7c2f | 3843 | ARCSTAT_BUMP(arcstat_mfu_ghost_hits); |
b9541d6b | 3844 | } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { |
34dc7c2f BB |
3845 | /* |
3846 | * This buffer is on the 2nd Level ARC. | |
3847 | */ | |
3848 | ||
b9541d6b | 3849 | hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); |
2a432414 GW |
3850 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); |
3851 | arc_change_state(arc_mfu, hdr, hash_lock); | |
34dc7c2f | 3852 | } else { |
b9541d6b CW |
3853 | cmn_err(CE_PANIC, "invalid arc state 0x%p", |
3854 | hdr->b_l1hdr.b_state); | |
34dc7c2f BB |
3855 | } |
3856 | } | |
3857 | ||
3858 | /* a generic arc_done_func_t which you can use */ | |
3859 | /* ARGSUSED */ | |
3860 | void | |
3861 | arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) | |
3862 | { | |
428870ff BB |
3863 | if (zio == NULL || zio->io_error == 0) |
3864 | bcopy(buf->b_data, arg, buf->b_hdr->b_size); | |
13fe0198 | 3865 | VERIFY(arc_buf_remove_ref(buf, arg)); |
34dc7c2f BB |
3866 | } |
3867 | ||
3868 | /* a generic arc_done_func_t */ | |
3869 | void | |
3870 | arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) | |
3871 | { | |
3872 | arc_buf_t **bufp = arg; | |
3873 | if (zio && zio->io_error) { | |
13fe0198 | 3874 | VERIFY(arc_buf_remove_ref(buf, arg)); |
34dc7c2f BB |
3875 | *bufp = NULL; |
3876 | } else { | |
3877 | *bufp = buf; | |
428870ff | 3878 | ASSERT(buf->b_data); |
34dc7c2f BB |
3879 | } |
3880 | } | |
3881 | ||
3882 | static void | |
3883 | arc_read_done(zio_t *zio) | |
3884 | { | |
9b67f605 | 3885 | arc_buf_hdr_t *hdr; |
34dc7c2f BB |
3886 | arc_buf_t *buf; |
3887 | arc_buf_t *abuf; /* buffer we're assigning to callback */ | |
9b67f605 | 3888 | kmutex_t *hash_lock = NULL; |
34dc7c2f BB |
3889 | arc_callback_t *callback_list, *acb; |
3890 | int freeable = FALSE; | |
3891 | ||
3892 | buf = zio->io_private; | |
3893 | hdr = buf->b_hdr; | |
3894 | ||
3895 | /* | |
3896 | * The hdr was inserted into hash-table and removed from lists | |
3897 | * prior to starting I/O. We should find this header, since | |
3898 | * it's in the hash table, and it should be legit since it's | |
3899 | * not possible to evict it during the I/O. The only possible | |
3900 | * reason for it not to be found is if we were freed during the | |
3901 | * read. | |
3902 | */ | |
9b67f605 MA |
3903 | if (HDR_IN_HASH_TABLE(hdr)) { |
3904 | arc_buf_hdr_t *found; | |
3905 | ||
3906 | ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); | |
3907 | ASSERT3U(hdr->b_dva.dva_word[0], ==, | |
3908 | BP_IDENTITY(zio->io_bp)->dva_word[0]); | |
3909 | ASSERT3U(hdr->b_dva.dva_word[1], ==, | |
3910 | BP_IDENTITY(zio->io_bp)->dva_word[1]); | |
3911 | ||
3912 | found = buf_hash_find(hdr->b_spa, zio->io_bp, | |
3913 | &hash_lock); | |
3914 | ||
3915 | ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && | |
3916 | hash_lock == NULL) || | |
3917 | (found == hdr && | |
3918 | DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || | |
3919 | (found == hdr && HDR_L2_READING(hdr))); | |
3920 | } | |
34dc7c2f | 3921 | |
2a432414 | 3922 | hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; |
b9541d6b | 3923 | if (l2arc_noprefetch && HDR_PREFETCH(hdr)) |
2a432414 | 3924 | hdr->b_flags &= ~ARC_FLAG_L2CACHE; |
34dc7c2f BB |
3925 | |
3926 | /* byteswap if necessary */ | |
b9541d6b | 3927 | callback_list = hdr->b_l1hdr.b_acb; |
34dc7c2f | 3928 | ASSERT(callback_list != NULL); |
428870ff | 3929 | if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { |
9ae529ec CS |
3930 | dmu_object_byteswap_t bswap = |
3931 | DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); | |
b01615d5 RY |
3932 | if (BP_GET_LEVEL(zio->io_bp) > 0) |
3933 | byteswap_uint64_array(buf->b_data, hdr->b_size); | |
3934 | else | |
3935 | dmu_ot_byteswap[bswap].ob_func(buf->b_data, hdr->b_size); | |
b128c09f | 3936 | } |
34dc7c2f BB |
3937 | |
3938 | arc_cksum_compute(buf, B_FALSE); | |
498877ba | 3939 | arc_buf_watch(buf); |
34dc7c2f | 3940 | |
b9541d6b CW |
3941 | if (hash_lock && zio->io_error == 0 && |
3942 | hdr->b_l1hdr.b_state == arc_anon) { | |
428870ff BB |
3943 | /* |
3944 | * Only call arc_access on anonymous buffers. This is because | |
3945 | * if we've issued an I/O for an evicted buffer, we've already | |
3946 | * called arc_access (to prevent any simultaneous readers from | |
3947 | * getting confused). | |
3948 | */ | |
3949 | arc_access(hdr, hash_lock); | |
3950 | } | |
3951 | ||
34dc7c2f BB |
3952 | /* create copies of the data buffer for the callers */ |
3953 | abuf = buf; | |
3954 | for (acb = callback_list; acb; acb = acb->acb_next) { | |
3955 | if (acb->acb_done) { | |
1eb5bfa3 GW |
3956 | if (abuf == NULL) { |
3957 | ARCSTAT_BUMP(arcstat_duplicate_reads); | |
34dc7c2f | 3958 | abuf = arc_buf_clone(buf); |
1eb5bfa3 | 3959 | } |
34dc7c2f BB |
3960 | acb->acb_buf = abuf; |
3961 | abuf = NULL; | |
3962 | } | |
3963 | } | |
b9541d6b | 3964 | hdr->b_l1hdr.b_acb = NULL; |
2a432414 | 3965 | hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; |
34dc7c2f | 3966 | ASSERT(!HDR_BUF_AVAILABLE(hdr)); |
428870ff BB |
3967 | if (abuf == buf) { |
3968 | ASSERT(buf->b_efunc == NULL); | |
b9541d6b | 3969 | ASSERT(hdr->b_l1hdr.b_datacnt == 1); |
2a432414 | 3970 | hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; |
428870ff | 3971 | } |
34dc7c2f | 3972 | |
b9541d6b CW |
3973 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || |
3974 | callback_list != NULL); | |
34dc7c2f BB |
3975 | |
3976 | if (zio->io_error != 0) { | |
2a432414 | 3977 | hdr->b_flags |= ARC_FLAG_IO_ERROR; |
b9541d6b | 3978 | if (hdr->b_l1hdr.b_state != arc_anon) |
34dc7c2f BB |
3979 | arc_change_state(arc_anon, hdr, hash_lock); |
3980 | if (HDR_IN_HASH_TABLE(hdr)) | |
3981 | buf_hash_remove(hdr); | |
b9541d6b | 3982 | freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); |
34dc7c2f BB |
3983 | } |
3984 | ||
3985 | /* | |
3986 | * Broadcast before we drop the hash_lock to avoid the possibility | |
3987 | * that the hdr (and hence the cv) might be freed before we get to | |
3988 | * the cv_broadcast(). | |
3989 | */ | |
b9541d6b | 3990 | cv_broadcast(&hdr->b_l1hdr.b_cv); |
34dc7c2f | 3991 | |
b9541d6b | 3992 | if (hash_lock != NULL) { |
34dc7c2f BB |
3993 | mutex_exit(hash_lock); |
3994 | } else { | |
3995 | /* | |
3996 | * This block was freed while we waited for the read to | |
3997 | * complete. It has been removed from the hash table and | |
3998 | * moved to the anonymous state (so that it won't show up | |
3999 | * in the cache). | |
4000 | */ | |
b9541d6b CW |
4001 | ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); |
4002 | freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); | |
34dc7c2f BB |
4003 | } |
4004 | ||
4005 | /* execute each callback and free its structure */ | |
4006 | while ((acb = callback_list) != NULL) { | |
4007 | if (acb->acb_done) | |
4008 | acb->acb_done(zio, acb->acb_buf, acb->acb_private); | |
4009 | ||
4010 | if (acb->acb_zio_dummy != NULL) { | |
4011 | acb->acb_zio_dummy->io_error = zio->io_error; | |
4012 | zio_nowait(acb->acb_zio_dummy); | |
4013 | } | |
4014 | ||
4015 | callback_list = acb->acb_next; | |
4016 | kmem_free(acb, sizeof (arc_callback_t)); | |
4017 | } | |
4018 | ||
4019 | if (freeable) | |
4020 | arc_hdr_destroy(hdr); | |
4021 | } | |
4022 | ||
4023 | /* | |
5c839890 | 4024 | * "Read" the block at the specified DVA (in bp) via the |
34dc7c2f BB |
4025 | * cache. If the block is found in the cache, invoke the provided |
4026 | * callback immediately and return. Note that the `zio' parameter | |
4027 | * in the callback will be NULL in this case, since no IO was | |
4028 | * required. If the block is not in the cache pass the read request | |
4029 | * on to the spa with a substitute callback function, so that the | |
4030 | * requested block will be added to the cache. | |
4031 | * | |
4032 | * If a read request arrives for a block that has a read in-progress, | |
4033 | * either wait for the in-progress read to complete (and return the | |
4034 | * results); or, if this is a read with a "done" func, add a record | |
4035 | * to the read to invoke the "done" func when the read completes, | |
4036 | * and return; or just return. | |
4037 | * | |
4038 | * arc_read_done() will invoke all the requested "done" functions | |
4039 | * for readers of this block. | |
4040 | */ | |
4041 | int | |
294f6806 | 4042 | arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, |
2a432414 GW |
4043 | void *private, zio_priority_t priority, int zio_flags, |
4044 | arc_flags_t *arc_flags, const zbookmark_phys_t *zb) | |
34dc7c2f | 4045 | { |
9b67f605 | 4046 | arc_buf_hdr_t *hdr = NULL; |
d4ed6673 | 4047 | arc_buf_t *buf = NULL; |
9b67f605 | 4048 | kmutex_t *hash_lock = NULL; |
34dc7c2f | 4049 | zio_t *rzio; |
3541dc6d | 4050 | uint64_t guid = spa_load_guid(spa); |
1421c891 | 4051 | int rc = 0; |
34dc7c2f | 4052 | |
9b67f605 MA |
4053 | ASSERT(!BP_IS_EMBEDDED(bp) || |
4054 | BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); | |
4055 | ||
34dc7c2f | 4056 | top: |
9b67f605 MA |
4057 | if (!BP_IS_EMBEDDED(bp)) { |
4058 | /* | |
4059 | * Embedded BP's have no DVA and require no I/O to "read". | |
4060 | * Create an anonymous arc buf to back it. | |
4061 | */ | |
4062 | hdr = buf_hash_find(guid, bp, &hash_lock); | |
4063 | } | |
4064 | ||
b9541d6b | 4065 | if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { |
34dc7c2f | 4066 | |
2a432414 | 4067 | *arc_flags |= ARC_FLAG_CACHED; |
34dc7c2f BB |
4068 | |
4069 | if (HDR_IO_IN_PROGRESS(hdr)) { | |
4070 | ||
2a432414 | 4071 | if (*arc_flags & ARC_FLAG_WAIT) { |
b9541d6b | 4072 | cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); |
34dc7c2f BB |
4073 | mutex_exit(hash_lock); |
4074 | goto top; | |
4075 | } | |
2a432414 | 4076 | ASSERT(*arc_flags & ARC_FLAG_NOWAIT); |
34dc7c2f BB |
4077 | |
4078 | if (done) { | |
4079 | arc_callback_t *acb = NULL; | |
4080 | ||
4081 | acb = kmem_zalloc(sizeof (arc_callback_t), | |
79c76d5b | 4082 | KM_SLEEP); |
34dc7c2f BB |
4083 | acb->acb_done = done; |
4084 | acb->acb_private = private; | |
34dc7c2f BB |
4085 | if (pio != NULL) |
4086 | acb->acb_zio_dummy = zio_null(pio, | |
d164b209 | 4087 | spa, NULL, NULL, NULL, zio_flags); |
34dc7c2f BB |
4088 | |
4089 | ASSERT(acb->acb_done != NULL); | |
b9541d6b CW |
4090 | acb->acb_next = hdr->b_l1hdr.b_acb; |
4091 | hdr->b_l1hdr.b_acb = acb; | |
34dc7c2f BB |
4092 | add_reference(hdr, hash_lock, private); |
4093 | mutex_exit(hash_lock); | |
1421c891 | 4094 | goto out; |
34dc7c2f BB |
4095 | } |
4096 | mutex_exit(hash_lock); | |
1421c891 | 4097 | goto out; |
34dc7c2f BB |
4098 | } |
4099 | ||
b9541d6b CW |
4100 | ASSERT(hdr->b_l1hdr.b_state == arc_mru || |
4101 | hdr->b_l1hdr.b_state == arc_mfu); | |
34dc7c2f BB |
4102 | |
4103 | if (done) { | |
4104 | add_reference(hdr, hash_lock, private); | |
4105 | /* | |
4106 | * If this block is already in use, create a new | |
4107 | * copy of the data so that we will be guaranteed | |
4108 | * that arc_release() will always succeed. | |
4109 | */ | |
b9541d6b | 4110 | buf = hdr->b_l1hdr.b_buf; |
34dc7c2f BB |
4111 | ASSERT(buf); |
4112 | ASSERT(buf->b_data); | |
4113 | if (HDR_BUF_AVAILABLE(hdr)) { | |
4114 | ASSERT(buf->b_efunc == NULL); | |
2a432414 | 4115 | hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; |
34dc7c2f BB |
4116 | } else { |
4117 | buf = arc_buf_clone(buf); | |
4118 | } | |
428870ff | 4119 | |
2a432414 | 4120 | } else if (*arc_flags & ARC_FLAG_PREFETCH && |
b9541d6b | 4121 | refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { |
2a432414 | 4122 | hdr->b_flags |= ARC_FLAG_PREFETCH; |
34dc7c2f BB |
4123 | } |
4124 | DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); | |
4125 | arc_access(hdr, hash_lock); | |
2a432414 GW |
4126 | if (*arc_flags & ARC_FLAG_L2CACHE) |
4127 | hdr->b_flags |= ARC_FLAG_L2CACHE; | |
4128 | if (*arc_flags & ARC_FLAG_L2COMPRESS) | |
4129 | hdr->b_flags |= ARC_FLAG_L2COMPRESS; | |
34dc7c2f BB |
4130 | mutex_exit(hash_lock); |
4131 | ARCSTAT_BUMP(arcstat_hits); | |
b9541d6b CW |
4132 | ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), |
4133 | demand, prefetch, !HDR_ISTYPE_METADATA(hdr), | |
34dc7c2f BB |
4134 | data, metadata, hits); |
4135 | ||
4136 | if (done) | |
4137 | done(NULL, buf, private); | |
4138 | } else { | |
4139 | uint64_t size = BP_GET_LSIZE(bp); | |
9b67f605 | 4140 | arc_callback_t *acb; |
b128c09f | 4141 | vdev_t *vd = NULL; |
a117a6d6 | 4142 | uint64_t addr = 0; |
d164b209 | 4143 | boolean_t devw = B_FALSE; |
0ed212dc | 4144 | enum zio_compress b_compress = ZIO_COMPRESS_OFF; |
b9541d6b | 4145 | int32_t b_asize = 0; |
34dc7c2f | 4146 | |
5f6d0b6f BB |
4147 | /* |
4148 | * Gracefully handle a damaged logical block size as a | |
4149 | * checksum error by passing a dummy zio to the done callback. | |
4150 | */ | |
f1512ee6 | 4151 | if (size > spa_maxblocksize(spa)) { |
5f6d0b6f BB |
4152 | if (done) { |
4153 | rzio = zio_null(pio, spa, NULL, | |
4154 | NULL, NULL, zio_flags); | |
4155 | rzio->io_error = ECKSUM; | |
4156 | done(rzio, buf, private); | |
4157 | zio_nowait(rzio); | |
4158 | } | |
4159 | rc = ECKSUM; | |
4160 | goto out; | |
4161 | } | |
4162 | ||
34dc7c2f BB |
4163 | if (hdr == NULL) { |
4164 | /* this block is not in the cache */ | |
9b67f605 | 4165 | arc_buf_hdr_t *exists = NULL; |
34dc7c2f BB |
4166 | arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); |
4167 | buf = arc_buf_alloc(spa, size, private, type); | |
4168 | hdr = buf->b_hdr; | |
9b67f605 MA |
4169 | if (!BP_IS_EMBEDDED(bp)) { |
4170 | hdr->b_dva = *BP_IDENTITY(bp); | |
4171 | hdr->b_birth = BP_PHYSICAL_BIRTH(bp); | |
9b67f605 MA |
4172 | exists = buf_hash_insert(hdr, &hash_lock); |
4173 | } | |
4174 | if (exists != NULL) { | |
34dc7c2f BB |
4175 | /* somebody beat us to the hash insert */ |
4176 | mutex_exit(hash_lock); | |
428870ff | 4177 | buf_discard_identity(hdr); |
34dc7c2f BB |
4178 | (void) arc_buf_remove_ref(buf, private); |
4179 | goto top; /* restart the IO request */ | |
4180 | } | |
2a432414 | 4181 | |
34dc7c2f | 4182 | /* if this is a prefetch, we don't have a reference */ |
2a432414 | 4183 | if (*arc_flags & ARC_FLAG_PREFETCH) { |
34dc7c2f BB |
4184 | (void) remove_reference(hdr, hash_lock, |
4185 | private); | |
2a432414 | 4186 | hdr->b_flags |= ARC_FLAG_PREFETCH; |
34dc7c2f | 4187 | } |
2a432414 GW |
4188 | if (*arc_flags & ARC_FLAG_L2CACHE) |
4189 | hdr->b_flags |= ARC_FLAG_L2CACHE; | |
4190 | if (*arc_flags & ARC_FLAG_L2COMPRESS) | |
4191 | hdr->b_flags |= ARC_FLAG_L2COMPRESS; | |
34dc7c2f | 4192 | if (BP_GET_LEVEL(bp) > 0) |
2a432414 | 4193 | hdr->b_flags |= ARC_FLAG_INDIRECT; |
34dc7c2f | 4194 | } else { |
b9541d6b CW |
4195 | /* |
4196 | * This block is in the ghost cache. If it was L2-only | |
4197 | * (and thus didn't have an L1 hdr), we realloc the | |
4198 | * header to add an L1 hdr. | |
4199 | */ | |
4200 | if (!HDR_HAS_L1HDR(hdr)) { | |
4201 | hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, | |
4202 | hdr_full_cache); | |
4203 | } | |
4204 | ||
4205 | ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); | |
34dc7c2f | 4206 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); |
b9541d6b | 4207 | ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
ca0bf58d | 4208 | ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); |
34dc7c2f BB |
4209 | |
4210 | /* if this is a prefetch, we don't have a reference */ | |
2a432414 GW |
4211 | if (*arc_flags & ARC_FLAG_PREFETCH) |
4212 | hdr->b_flags |= ARC_FLAG_PREFETCH; | |
34dc7c2f BB |
4213 | else |
4214 | add_reference(hdr, hash_lock, private); | |
2a432414 GW |
4215 | if (*arc_flags & ARC_FLAG_L2CACHE) |
4216 | hdr->b_flags |= ARC_FLAG_L2CACHE; | |
4217 | if (*arc_flags & ARC_FLAG_L2COMPRESS) | |
4218 | hdr->b_flags |= ARC_FLAG_L2COMPRESS; | |
34dc7c2f BB |
4219 | buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); |
4220 | buf->b_hdr = hdr; | |
4221 | buf->b_data = NULL; | |
4222 | buf->b_efunc = NULL; | |
4223 | buf->b_private = NULL; | |
4224 | buf->b_next = NULL; | |
b9541d6b CW |
4225 | hdr->b_l1hdr.b_buf = buf; |
4226 | ASSERT0(hdr->b_l1hdr.b_datacnt); | |
4227 | hdr->b_l1hdr.b_datacnt = 1; | |
428870ff BB |
4228 | arc_get_data_buf(buf); |
4229 | arc_access(hdr, hash_lock); | |
34dc7c2f BB |
4230 | } |
4231 | ||
b9541d6b | 4232 | ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); |
428870ff | 4233 | |
79c76d5b | 4234 | acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); |
34dc7c2f BB |
4235 | acb->acb_done = done; |
4236 | acb->acb_private = private; | |
34dc7c2f | 4237 | |
b9541d6b CW |
4238 | ASSERT(hdr->b_l1hdr.b_acb == NULL); |
4239 | hdr->b_l1hdr.b_acb = acb; | |
2a432414 | 4240 | hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; |
34dc7c2f | 4241 | |
b9541d6b CW |
4242 | if (HDR_HAS_L2HDR(hdr) && |
4243 | (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { | |
4244 | devw = hdr->b_l2hdr.b_dev->l2ad_writing; | |
4245 | addr = hdr->b_l2hdr.b_daddr; | |
4246 | b_compress = HDR_GET_COMPRESS(hdr); | |
4247 | b_asize = hdr->b_l2hdr.b_asize; | |
b128c09f BB |
4248 | /* |
4249 | * Lock out device removal. | |
4250 | */ | |
4251 | if (vdev_is_dead(vd) || | |
4252 | !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) | |
4253 | vd = NULL; | |
4254 | } | |
4255 | ||
9b67f605 MA |
4256 | if (hash_lock != NULL) |
4257 | mutex_exit(hash_lock); | |
b128c09f | 4258 | |
e49f1e20 WA |
4259 | /* |
4260 | * At this point, we have a level 1 cache miss. Try again in | |
4261 | * L2ARC if possible. | |
4262 | */ | |
34dc7c2f | 4263 | ASSERT3U(hdr->b_size, ==, size); |
428870ff | 4264 | DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, |
5dbd68a3 | 4265 | uint64_t, size, zbookmark_phys_t *, zb); |
34dc7c2f | 4266 | ARCSTAT_BUMP(arcstat_misses); |
b9541d6b CW |
4267 | ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), |
4268 | demand, prefetch, !HDR_ISTYPE_METADATA(hdr), | |
34dc7c2f BB |
4269 | data, metadata, misses); |
4270 | ||
d164b209 | 4271 | if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { |
34dc7c2f BB |
4272 | /* |
4273 | * Read from the L2ARC if the following are true: | |
b128c09f BB |
4274 | * 1. The L2ARC vdev was previously cached. |
4275 | * 2. This buffer still has L2ARC metadata. | |
4276 | * 3. This buffer isn't currently writing to the L2ARC. | |
4277 | * 4. The L2ARC entry wasn't evicted, which may | |
4278 | * also have invalidated the vdev. | |
d164b209 | 4279 | * 5. This isn't prefetch and l2arc_noprefetch is set. |
34dc7c2f | 4280 | */ |
b9541d6b | 4281 | if (HDR_HAS_L2HDR(hdr) && |
d164b209 BB |
4282 | !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && |
4283 | !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { | |
34dc7c2f BB |
4284 | l2arc_read_callback_t *cb; |
4285 | ||
4286 | DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); | |
4287 | ARCSTAT_BUMP(arcstat_l2_hits); | |
b9541d6b | 4288 | atomic_inc_32(&hdr->b_l2hdr.b_hits); |
34dc7c2f | 4289 | |
34dc7c2f | 4290 | cb = kmem_zalloc(sizeof (l2arc_read_callback_t), |
79c76d5b | 4291 | KM_SLEEP); |
34dc7c2f BB |
4292 | cb->l2rcb_buf = buf; |
4293 | cb->l2rcb_spa = spa; | |
4294 | cb->l2rcb_bp = *bp; | |
4295 | cb->l2rcb_zb = *zb; | |
b128c09f | 4296 | cb->l2rcb_flags = zio_flags; |
0ed212dc | 4297 | cb->l2rcb_compress = b_compress; |
34dc7c2f | 4298 | |
a117a6d6 GW |
4299 | ASSERT(addr >= VDEV_LABEL_START_SIZE && |
4300 | addr + size < vd->vdev_psize - | |
4301 | VDEV_LABEL_END_SIZE); | |
4302 | ||
34dc7c2f | 4303 | /* |
b128c09f BB |
4304 | * l2arc read. The SCL_L2ARC lock will be |
4305 | * released by l2arc_read_done(). | |
3a17a7a9 SK |
4306 | * Issue a null zio if the underlying buffer |
4307 | * was squashed to zero size by compression. | |
34dc7c2f | 4308 | */ |
0ed212dc | 4309 | if (b_compress == ZIO_COMPRESS_EMPTY) { |
3a17a7a9 SK |
4310 | rzio = zio_null(pio, spa, vd, |
4311 | l2arc_read_done, cb, | |
4312 | zio_flags | ZIO_FLAG_DONT_CACHE | | |
4313 | ZIO_FLAG_CANFAIL | | |
4314 | ZIO_FLAG_DONT_PROPAGATE | | |
4315 | ZIO_FLAG_DONT_RETRY); | |
4316 | } else { | |
4317 | rzio = zio_read_phys(pio, vd, addr, | |
0ed212dc BP |
4318 | b_asize, buf->b_data, |
4319 | ZIO_CHECKSUM_OFF, | |
3a17a7a9 SK |
4320 | l2arc_read_done, cb, priority, |
4321 | zio_flags | ZIO_FLAG_DONT_CACHE | | |
4322 | ZIO_FLAG_CANFAIL | | |
4323 | ZIO_FLAG_DONT_PROPAGATE | | |
4324 | ZIO_FLAG_DONT_RETRY, B_FALSE); | |
4325 | } | |
34dc7c2f BB |
4326 | DTRACE_PROBE2(l2arc__read, vdev_t *, vd, |
4327 | zio_t *, rzio); | |
0ed212dc | 4328 | ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); |
34dc7c2f | 4329 | |
2a432414 | 4330 | if (*arc_flags & ARC_FLAG_NOWAIT) { |
b128c09f | 4331 | zio_nowait(rzio); |
1421c891 | 4332 | goto out; |
b128c09f | 4333 | } |
34dc7c2f | 4334 | |
2a432414 | 4335 | ASSERT(*arc_flags & ARC_FLAG_WAIT); |
b128c09f | 4336 | if (zio_wait(rzio) == 0) |
1421c891 | 4337 | goto out; |
b128c09f BB |
4338 | |
4339 | /* l2arc read error; goto zio_read() */ | |
34dc7c2f BB |
4340 | } else { |
4341 | DTRACE_PROBE1(l2arc__miss, | |
4342 | arc_buf_hdr_t *, hdr); | |
4343 | ARCSTAT_BUMP(arcstat_l2_misses); | |
4344 | if (HDR_L2_WRITING(hdr)) | |
4345 | ARCSTAT_BUMP(arcstat_l2_rw_clash); | |
b128c09f | 4346 | spa_config_exit(spa, SCL_L2ARC, vd); |
34dc7c2f | 4347 | } |
d164b209 BB |
4348 | } else { |
4349 | if (vd != NULL) | |
4350 | spa_config_exit(spa, SCL_L2ARC, vd); | |
4351 | if (l2arc_ndev != 0) { | |
4352 | DTRACE_PROBE1(l2arc__miss, | |
4353 | arc_buf_hdr_t *, hdr); | |
4354 | ARCSTAT_BUMP(arcstat_l2_misses); | |
4355 | } | |
34dc7c2f | 4356 | } |
34dc7c2f BB |
4357 | |
4358 | rzio = zio_read(pio, spa, bp, buf->b_data, size, | |
b128c09f | 4359 | arc_read_done, buf, priority, zio_flags, zb); |
34dc7c2f | 4360 | |
2a432414 | 4361 | if (*arc_flags & ARC_FLAG_WAIT) { |
1421c891 PS |
4362 | rc = zio_wait(rzio); |
4363 | goto out; | |
4364 | } | |
34dc7c2f | 4365 | |
2a432414 | 4366 | ASSERT(*arc_flags & ARC_FLAG_NOWAIT); |
34dc7c2f BB |
4367 | zio_nowait(rzio); |
4368 | } | |
1421c891 PS |
4369 | |
4370 | out: | |
4371 | spa_read_history_add(spa, zb, *arc_flags); | |
4372 | return (rc); | |
34dc7c2f BB |
4373 | } |
4374 | ||
ab26409d BB |
4375 | arc_prune_t * |
4376 | arc_add_prune_callback(arc_prune_func_t *func, void *private) | |
4377 | { | |
4378 | arc_prune_t *p; | |
4379 | ||
d1d7e268 | 4380 | p = kmem_alloc(sizeof (*p), KM_SLEEP); |
ab26409d BB |
4381 | p->p_pfunc = func; |
4382 | p->p_private = private; | |
4383 | list_link_init(&p->p_node); | |
4384 | refcount_create(&p->p_refcnt); | |
4385 | ||
4386 | mutex_enter(&arc_prune_mtx); | |
4387 | refcount_add(&p->p_refcnt, &arc_prune_list); | |
4388 | list_insert_head(&arc_prune_list, p); | |
4389 | mutex_exit(&arc_prune_mtx); | |
4390 | ||
4391 | return (p); | |
4392 | } | |
4393 | ||
4394 | void | |
4395 | arc_remove_prune_callback(arc_prune_t *p) | |
4396 | { | |
4397 | mutex_enter(&arc_prune_mtx); | |
4398 | list_remove(&arc_prune_list, p); | |
4399 | if (refcount_remove(&p->p_refcnt, &arc_prune_list) == 0) { | |
4400 | refcount_destroy(&p->p_refcnt); | |
4401 | kmem_free(p, sizeof (*p)); | |
4402 | } | |
4403 | mutex_exit(&arc_prune_mtx); | |
4404 | } | |
4405 | ||
34dc7c2f BB |
4406 | void |
4407 | arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) | |
4408 | { | |
4409 | ASSERT(buf->b_hdr != NULL); | |
b9541d6b CW |
4410 | ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); |
4411 | ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || | |
4412 | func == NULL); | |
428870ff BB |
4413 | ASSERT(buf->b_efunc == NULL); |
4414 | ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); | |
4415 | ||
34dc7c2f BB |
4416 | buf->b_efunc = func; |
4417 | buf->b_private = private; | |
4418 | } | |
4419 | ||
df4474f9 MA |
4420 | /* |
4421 | * Notify the arc that a block was freed, and thus will never be used again. | |
4422 | */ | |
4423 | void | |
4424 | arc_freed(spa_t *spa, const blkptr_t *bp) | |
4425 | { | |
4426 | arc_buf_hdr_t *hdr; | |
4427 | kmutex_t *hash_lock; | |
4428 | uint64_t guid = spa_load_guid(spa); | |
4429 | ||
9b67f605 MA |
4430 | ASSERT(!BP_IS_EMBEDDED(bp)); |
4431 | ||
4432 | hdr = buf_hash_find(guid, bp, &hash_lock); | |
df4474f9 MA |
4433 | if (hdr == NULL) |
4434 | return; | |
4435 | if (HDR_BUF_AVAILABLE(hdr)) { | |
b9541d6b | 4436 | arc_buf_t *buf = hdr->b_l1hdr.b_buf; |
df4474f9 | 4437 | add_reference(hdr, hash_lock, FTAG); |
2a432414 | 4438 | hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; |
df4474f9 MA |
4439 | mutex_exit(hash_lock); |
4440 | ||
4441 | arc_release(buf, FTAG); | |
4442 | (void) arc_buf_remove_ref(buf, FTAG); | |
4443 | } else { | |
4444 | mutex_exit(hash_lock); | |
4445 | } | |
4446 | ||
4447 | } | |
4448 | ||
34dc7c2f | 4449 | /* |
bd089c54 MA |
4450 | * Clear the user eviction callback set by arc_set_callback(), first calling |
4451 | * it if it exists. Because the presence of a callback keeps an arc_buf cached | |
4452 | * clearing the callback may result in the arc_buf being destroyed. However, | |
4453 | * it will not result in the *last* arc_buf being destroyed, hence the data | |
4454 | * will remain cached in the ARC. We make a copy of the arc buffer here so | |
4455 | * that we can process the callback without holding any locks. | |
4456 | * | |
4457 | * It's possible that the callback is already in the process of being cleared | |
4458 | * by another thread. In this case we can not clear the callback. | |
4459 | * | |
4460 | * Returns B_TRUE if the callback was successfully called and cleared. | |
34dc7c2f | 4461 | */ |
bd089c54 MA |
4462 | boolean_t |
4463 | arc_clear_callback(arc_buf_t *buf) | |
34dc7c2f BB |
4464 | { |
4465 | arc_buf_hdr_t *hdr; | |
4466 | kmutex_t *hash_lock; | |
bd089c54 MA |
4467 | arc_evict_func_t *efunc = buf->b_efunc; |
4468 | void *private = buf->b_private; | |
34dc7c2f | 4469 | |
428870ff | 4470 | mutex_enter(&buf->b_evict_lock); |
34dc7c2f BB |
4471 | hdr = buf->b_hdr; |
4472 | if (hdr == NULL) { | |
4473 | /* | |
4474 | * We are in arc_do_user_evicts(). | |
4475 | */ | |
4476 | ASSERT(buf->b_data == NULL); | |
428870ff | 4477 | mutex_exit(&buf->b_evict_lock); |
bd089c54 | 4478 | return (B_FALSE); |
b128c09f | 4479 | } else if (buf->b_data == NULL) { |
34dc7c2f | 4480 | /* |
b128c09f BB |
4481 | * We are on the eviction list; process this buffer now |
4482 | * but let arc_do_user_evicts() do the reaping. | |
34dc7c2f | 4483 | */ |
b128c09f | 4484 | buf->b_efunc = NULL; |
428870ff | 4485 | mutex_exit(&buf->b_evict_lock); |
bd089c54 MA |
4486 | VERIFY0(efunc(private)); |
4487 | return (B_TRUE); | |
34dc7c2f | 4488 | } |
b128c09f BB |
4489 | hash_lock = HDR_LOCK(hdr); |
4490 | mutex_enter(hash_lock); | |
428870ff BB |
4491 | hdr = buf->b_hdr; |
4492 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
34dc7c2f | 4493 | |
b9541d6b CW |
4494 | ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, |
4495 | hdr->b_l1hdr.b_datacnt); | |
4496 | ASSERT(hdr->b_l1hdr.b_state == arc_mru || | |
4497 | hdr->b_l1hdr.b_state == arc_mfu); | |
34dc7c2f | 4498 | |
bd089c54 MA |
4499 | buf->b_efunc = NULL; |
4500 | buf->b_private = NULL; | |
34dc7c2f | 4501 | |
b9541d6b | 4502 | if (hdr->b_l1hdr.b_datacnt > 1) { |
bd089c54 | 4503 | mutex_exit(&buf->b_evict_lock); |
ca0bf58d | 4504 | arc_buf_destroy(buf, TRUE); |
bd089c54 | 4505 | } else { |
b9541d6b | 4506 | ASSERT(buf == hdr->b_l1hdr.b_buf); |
2a432414 | 4507 | hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; |
bd089c54 | 4508 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f | 4509 | } |
34dc7c2f | 4510 | |
bd089c54 MA |
4511 | mutex_exit(hash_lock); |
4512 | VERIFY0(efunc(private)); | |
4513 | return (B_TRUE); | |
34dc7c2f BB |
4514 | } |
4515 | ||
4516 | /* | |
e49f1e20 WA |
4517 | * Release this buffer from the cache, making it an anonymous buffer. This |
4518 | * must be done after a read and prior to modifying the buffer contents. | |
34dc7c2f | 4519 | * If the buffer has more than one reference, we must make |
b128c09f | 4520 | * a new hdr for the buffer. |
34dc7c2f BB |
4521 | */ |
4522 | void | |
4523 | arc_release(arc_buf_t *buf, void *tag) | |
4524 | { | |
b9541d6b CW |
4525 | kmutex_t *hash_lock; |
4526 | arc_state_t *state; | |
4527 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
34dc7c2f | 4528 | |
428870ff | 4529 | /* |
ca0bf58d | 4530 | * It would be nice to assert that if its DMU metadata (level > |
428870ff BB |
4531 | * 0 || it's the dnode file), then it must be syncing context. |
4532 | * But we don't know that information at this level. | |
4533 | */ | |
4534 | ||
4535 | mutex_enter(&buf->b_evict_lock); | |
b128c09f | 4536 | |
ca0bf58d PS |
4537 | ASSERT(HDR_HAS_L1HDR(hdr)); |
4538 | ||
b9541d6b CW |
4539 | /* |
4540 | * We don't grab the hash lock prior to this check, because if | |
4541 | * the buffer's header is in the arc_anon state, it won't be | |
4542 | * linked into the hash table. | |
4543 | */ | |
4544 | if (hdr->b_l1hdr.b_state == arc_anon) { | |
4545 | mutex_exit(&buf->b_evict_lock); | |
4546 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
4547 | ASSERT(!HDR_IN_HASH_TABLE(hdr)); | |
4548 | ASSERT(!HDR_HAS_L2HDR(hdr)); | |
4549 | ASSERT(BUF_EMPTY(hdr)); | |
34dc7c2f | 4550 | |
b9541d6b CW |
4551 | ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); |
4552 | ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); | |
4553 | ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); | |
4554 | ||
4555 | ASSERT3P(buf->b_efunc, ==, NULL); | |
4556 | ASSERT3P(buf->b_private, ==, NULL); | |
4557 | ||
4558 | hdr->b_l1hdr.b_arc_access = 0; | |
4559 | arc_buf_thaw(buf); | |
4560 | ||
4561 | return; | |
34dc7c2f BB |
4562 | } |
4563 | ||
b9541d6b CW |
4564 | hash_lock = HDR_LOCK(hdr); |
4565 | mutex_enter(hash_lock); | |
4566 | ||
4567 | /* | |
4568 | * This assignment is only valid as long as the hash_lock is | |
4569 | * held, we must be careful not to reference state or the | |
4570 | * b_state field after dropping the lock. | |
4571 | */ | |
4572 | state = hdr->b_l1hdr.b_state; | |
4573 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
4574 | ASSERT3P(state, !=, arc_anon); | |
4575 | ||
4576 | /* this buffer is not on any list */ | |
4577 | ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); | |
4578 | ||
4579 | if (HDR_HAS_L2HDR(hdr)) { | |
b9541d6b | 4580 | mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); |
ca0bf58d PS |
4581 | |
4582 | /* | |
d962d5da PS |
4583 | * We have to recheck this conditional again now that |
4584 | * we're holding the l2ad_mtx to prevent a race with | |
4585 | * another thread which might be concurrently calling | |
4586 | * l2arc_evict(). In that case, l2arc_evict() might have | |
4587 | * destroyed the header's L2 portion as we were waiting | |
4588 | * to acquire the l2ad_mtx. | |
ca0bf58d | 4589 | */ |
d962d5da PS |
4590 | if (HDR_HAS_L2HDR(hdr)) |
4591 | arc_hdr_l2hdr_destroy(hdr); | |
ca0bf58d | 4592 | |
b9541d6b | 4593 | mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); |
b128c09f BB |
4594 | } |
4595 | ||
34dc7c2f BB |
4596 | /* |
4597 | * Do we have more than one buf? | |
4598 | */ | |
b9541d6b | 4599 | if (hdr->b_l1hdr.b_datacnt > 1) { |
34dc7c2f BB |
4600 | arc_buf_hdr_t *nhdr; |
4601 | arc_buf_t **bufp; | |
4602 | uint64_t blksz = hdr->b_size; | |
d164b209 | 4603 | uint64_t spa = hdr->b_spa; |
b9541d6b | 4604 | arc_buf_contents_t type = arc_buf_type(hdr); |
34dc7c2f BB |
4605 | uint32_t flags = hdr->b_flags; |
4606 | ||
b9541d6b | 4607 | ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); |
34dc7c2f | 4608 | /* |
428870ff BB |
4609 | * Pull the data off of this hdr and attach it to |
4610 | * a new anonymous hdr. | |
34dc7c2f BB |
4611 | */ |
4612 | (void) remove_reference(hdr, hash_lock, tag); | |
b9541d6b | 4613 | bufp = &hdr->b_l1hdr.b_buf; |
34dc7c2f BB |
4614 | while (*bufp != buf) |
4615 | bufp = &(*bufp)->b_next; | |
428870ff | 4616 | *bufp = buf->b_next; |
34dc7c2f BB |
4617 | buf->b_next = NULL; |
4618 | ||
b9541d6b CW |
4619 | ASSERT3P(state, !=, arc_l2c_only); |
4620 | ASSERT3U(state->arcs_size, >=, hdr->b_size); | |
4621 | atomic_add_64(&state->arcs_size, -hdr->b_size); | |
4622 | if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { | |
4623 | uint64_t *size; | |
4624 | ||
4625 | ASSERT3P(state, !=, arc_l2c_only); | |
4626 | size = &state->arcs_lsize[type]; | |
34dc7c2f BB |
4627 | ASSERT3U(*size, >=, hdr->b_size); |
4628 | atomic_add_64(size, -hdr->b_size); | |
4629 | } | |
1eb5bfa3 GW |
4630 | |
4631 | /* | |
4632 | * We're releasing a duplicate user data buffer, update | |
4633 | * our statistics accordingly. | |
4634 | */ | |
b9541d6b | 4635 | if (HDR_ISTYPE_DATA(hdr)) { |
1eb5bfa3 GW |
4636 | ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); |
4637 | ARCSTAT_INCR(arcstat_duplicate_buffers_size, | |
4638 | -hdr->b_size); | |
4639 | } | |
b9541d6b | 4640 | hdr->b_l1hdr.b_datacnt -= 1; |
34dc7c2f | 4641 | arc_cksum_verify(buf); |
498877ba | 4642 | arc_buf_unwatch(buf); |
34dc7c2f BB |
4643 | |
4644 | mutex_exit(hash_lock); | |
4645 | ||
b9541d6b | 4646 | nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); |
34dc7c2f BB |
4647 | nhdr->b_size = blksz; |
4648 | nhdr->b_spa = spa; | |
b9541d6b CW |
4649 | |
4650 | nhdr->b_l1hdr.b_mru_hits = 0; | |
4651 | nhdr->b_l1hdr.b_mru_ghost_hits = 0; | |
4652 | nhdr->b_l1hdr.b_mfu_hits = 0; | |
4653 | nhdr->b_l1hdr.b_mfu_ghost_hits = 0; | |
4654 | nhdr->b_l1hdr.b_l2_hits = 0; | |
2a432414 | 4655 | nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; |
b9541d6b CW |
4656 | nhdr->b_flags |= arc_bufc_to_flags(type); |
4657 | nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; | |
4658 | ||
4659 | nhdr->b_l1hdr.b_buf = buf; | |
4660 | nhdr->b_l1hdr.b_datacnt = 1; | |
4661 | nhdr->b_l1hdr.b_state = arc_anon; | |
4662 | nhdr->b_l1hdr.b_arc_access = 0; | |
ca0bf58d | 4663 | nhdr->b_l1hdr.b_tmp_cdata = NULL; |
34dc7c2f | 4664 | nhdr->b_freeze_cksum = NULL; |
b9541d6b CW |
4665 | |
4666 | (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); | |
34dc7c2f | 4667 | buf->b_hdr = nhdr; |
428870ff | 4668 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
4669 | atomic_add_64(&arc_anon->arcs_size, blksz); |
4670 | } else { | |
428870ff | 4671 | mutex_exit(&buf->b_evict_lock); |
b9541d6b | 4672 | ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); |
ca0bf58d PS |
4673 | /* protected by hash lock, or hdr is on arc_anon */ |
4674 | ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); | |
34dc7c2f | 4675 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); |
b9541d6b CW |
4676 | hdr->b_l1hdr.b_mru_hits = 0; |
4677 | hdr->b_l1hdr.b_mru_ghost_hits = 0; | |
4678 | hdr->b_l1hdr.b_mfu_hits = 0; | |
4679 | hdr->b_l1hdr.b_mfu_ghost_hits = 0; | |
4680 | hdr->b_l1hdr.b_l2_hits = 0; | |
4681 | arc_change_state(arc_anon, hdr, hash_lock); | |
4682 | hdr->b_l1hdr.b_arc_access = 0; | |
4683 | mutex_exit(hash_lock); | |
34dc7c2f | 4684 | |
428870ff | 4685 | buf_discard_identity(hdr); |
34dc7c2f BB |
4686 | arc_buf_thaw(buf); |
4687 | } | |
4688 | buf->b_efunc = NULL; | |
4689 | buf->b_private = NULL; | |
34dc7c2f BB |
4690 | } |
4691 | ||
4692 | int | |
4693 | arc_released(arc_buf_t *buf) | |
4694 | { | |
b128c09f BB |
4695 | int released; |
4696 | ||
428870ff | 4697 | mutex_enter(&buf->b_evict_lock); |
b9541d6b CW |
4698 | released = (buf->b_data != NULL && |
4699 | buf->b_hdr->b_l1hdr.b_state == arc_anon); | |
428870ff | 4700 | mutex_exit(&buf->b_evict_lock); |
b128c09f | 4701 | return (released); |
34dc7c2f BB |
4702 | } |
4703 | ||
34dc7c2f BB |
4704 | #ifdef ZFS_DEBUG |
4705 | int | |
4706 | arc_referenced(arc_buf_t *buf) | |
4707 | { | |
b128c09f BB |
4708 | int referenced; |
4709 | ||
428870ff | 4710 | mutex_enter(&buf->b_evict_lock); |
b9541d6b | 4711 | referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); |
428870ff | 4712 | mutex_exit(&buf->b_evict_lock); |
b128c09f | 4713 | return (referenced); |
34dc7c2f BB |
4714 | } |
4715 | #endif | |
4716 | ||
4717 | static void | |
4718 | arc_write_ready(zio_t *zio) | |
4719 | { | |
4720 | arc_write_callback_t *callback = zio->io_private; | |
4721 | arc_buf_t *buf = callback->awcb_buf; | |
4722 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
4723 | ||
b9541d6b CW |
4724 | ASSERT(HDR_HAS_L1HDR(hdr)); |
4725 | ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); | |
4726 | ASSERT(hdr->b_l1hdr.b_datacnt > 0); | |
b128c09f BB |
4727 | callback->awcb_ready(zio, buf, callback->awcb_private); |
4728 | ||
34dc7c2f BB |
4729 | /* |
4730 | * If the IO is already in progress, then this is a re-write | |
b128c09f BB |
4731 | * attempt, so we need to thaw and re-compute the cksum. |
4732 | * It is the responsibility of the callback to handle the | |
4733 | * accounting for any re-write attempt. | |
34dc7c2f BB |
4734 | */ |
4735 | if (HDR_IO_IN_PROGRESS(hdr)) { | |
b9541d6b | 4736 | mutex_enter(&hdr->b_l1hdr.b_freeze_lock); |
34dc7c2f BB |
4737 | if (hdr->b_freeze_cksum != NULL) { |
4738 | kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); | |
4739 | hdr->b_freeze_cksum = NULL; | |
4740 | } | |
b9541d6b | 4741 | mutex_exit(&hdr->b_l1hdr.b_freeze_lock); |
34dc7c2f BB |
4742 | } |
4743 | arc_cksum_compute(buf, B_FALSE); | |
2a432414 | 4744 | hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; |
34dc7c2f BB |
4745 | } |
4746 | ||
e8b96c60 MA |
4747 | /* |
4748 | * The SPA calls this callback for each physical write that happens on behalf | |
4749 | * of a logical write. See the comment in dbuf_write_physdone() for details. | |
4750 | */ | |
4751 | static void | |
4752 | arc_write_physdone(zio_t *zio) | |
4753 | { | |
4754 | arc_write_callback_t *cb = zio->io_private; | |
4755 | if (cb->awcb_physdone != NULL) | |
4756 | cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); | |
4757 | } | |
4758 | ||
34dc7c2f BB |
4759 | static void |
4760 | arc_write_done(zio_t *zio) | |
4761 | { | |
4762 | arc_write_callback_t *callback = zio->io_private; | |
4763 | arc_buf_t *buf = callback->awcb_buf; | |
4764 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
4765 | ||
b9541d6b | 4766 | ASSERT(hdr->b_l1hdr.b_acb == NULL); |
428870ff BB |
4767 | |
4768 | if (zio->io_error == 0) { | |
9b67f605 | 4769 | if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { |
b0bc7a84 MG |
4770 | buf_discard_identity(hdr); |
4771 | } else { | |
4772 | hdr->b_dva = *BP_IDENTITY(zio->io_bp); | |
4773 | hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); | |
b0bc7a84 | 4774 | } |
428870ff BB |
4775 | } else { |
4776 | ASSERT(BUF_EMPTY(hdr)); | |
4777 | } | |
34dc7c2f | 4778 | |
34dc7c2f | 4779 | /* |
9b67f605 MA |
4780 | * If the block to be written was all-zero or compressed enough to be |
4781 | * embedded in the BP, no write was performed so there will be no | |
4782 | * dva/birth/checksum. The buffer must therefore remain anonymous | |
4783 | * (and uncached). | |
34dc7c2f BB |
4784 | */ |
4785 | if (!BUF_EMPTY(hdr)) { | |
4786 | arc_buf_hdr_t *exists; | |
4787 | kmutex_t *hash_lock; | |
4788 | ||
428870ff BB |
4789 | ASSERT(zio->io_error == 0); |
4790 | ||
34dc7c2f BB |
4791 | arc_cksum_verify(buf); |
4792 | ||
4793 | exists = buf_hash_insert(hdr, &hash_lock); | |
b9541d6b | 4794 | if (exists != NULL) { |
34dc7c2f BB |
4795 | /* |
4796 | * This can only happen if we overwrite for | |
4797 | * sync-to-convergence, because we remove | |
4798 | * buffers from the hash table when we arc_free(). | |
4799 | */ | |
428870ff BB |
4800 | if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { |
4801 | if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) | |
4802 | panic("bad overwrite, hdr=%p exists=%p", | |
4803 | (void *)hdr, (void *)exists); | |
b9541d6b CW |
4804 | ASSERT(refcount_is_zero( |
4805 | &exists->b_l1hdr.b_refcnt)); | |
428870ff BB |
4806 | arc_change_state(arc_anon, exists, hash_lock); |
4807 | mutex_exit(hash_lock); | |
4808 | arc_hdr_destroy(exists); | |
4809 | exists = buf_hash_insert(hdr, &hash_lock); | |
4810 | ASSERT3P(exists, ==, NULL); | |
03c6040b GW |
4811 | } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { |
4812 | /* nopwrite */ | |
4813 | ASSERT(zio->io_prop.zp_nopwrite); | |
4814 | if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) | |
4815 | panic("bad nopwrite, hdr=%p exists=%p", | |
4816 | (void *)hdr, (void *)exists); | |
428870ff BB |
4817 | } else { |
4818 | /* Dedup */ | |
b9541d6b CW |
4819 | ASSERT(hdr->b_l1hdr.b_datacnt == 1); |
4820 | ASSERT(hdr->b_l1hdr.b_state == arc_anon); | |
428870ff BB |
4821 | ASSERT(BP_GET_DEDUP(zio->io_bp)); |
4822 | ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); | |
4823 | } | |
34dc7c2f | 4824 | } |
2a432414 | 4825 | hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; |
b128c09f | 4826 | /* if it's not anon, we are doing a scrub */ |
b9541d6b | 4827 | if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) |
b128c09f | 4828 | arc_access(hdr, hash_lock); |
34dc7c2f | 4829 | mutex_exit(hash_lock); |
34dc7c2f | 4830 | } else { |
2a432414 | 4831 | hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; |
34dc7c2f BB |
4832 | } |
4833 | ||
b9541d6b | 4834 | ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); |
428870ff | 4835 | callback->awcb_done(zio, buf, callback->awcb_private); |
34dc7c2f BB |
4836 | |
4837 | kmem_free(callback, sizeof (arc_write_callback_t)); | |
4838 | } | |
4839 | ||
4840 | zio_t * | |
428870ff | 4841 | arc_write(zio_t *pio, spa_t *spa, uint64_t txg, |
3a17a7a9 | 4842 | blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, |
e8b96c60 MA |
4843 | const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, |
4844 | arc_done_func_t *done, void *private, zio_priority_t priority, | |
5dbd68a3 | 4845 | int zio_flags, const zbookmark_phys_t *zb) |
34dc7c2f BB |
4846 | { |
4847 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
4848 | arc_write_callback_t *callback; | |
b128c09f | 4849 | zio_t *zio; |
34dc7c2f | 4850 | |
b128c09f | 4851 | ASSERT(ready != NULL); |
428870ff | 4852 | ASSERT(done != NULL); |
34dc7c2f | 4853 | ASSERT(!HDR_IO_ERROR(hdr)); |
b9541d6b CW |
4854 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); |
4855 | ASSERT(hdr->b_l1hdr.b_acb == NULL); | |
4856 | ASSERT(hdr->b_l1hdr.b_datacnt > 0); | |
b128c09f | 4857 | if (l2arc) |
2a432414 | 4858 | hdr->b_flags |= ARC_FLAG_L2CACHE; |
3a17a7a9 | 4859 | if (l2arc_compress) |
2a432414 | 4860 | hdr->b_flags |= ARC_FLAG_L2COMPRESS; |
79c76d5b | 4861 | callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); |
34dc7c2f | 4862 | callback->awcb_ready = ready; |
e8b96c60 | 4863 | callback->awcb_physdone = physdone; |
34dc7c2f BB |
4864 | callback->awcb_done = done; |
4865 | callback->awcb_private = private; | |
4866 | callback->awcb_buf = buf; | |
b128c09f | 4867 | |
428870ff | 4868 | zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, |
e8b96c60 MA |
4869 | arc_write_ready, arc_write_physdone, arc_write_done, callback, |
4870 | priority, zio_flags, zb); | |
34dc7c2f BB |
4871 | |
4872 | return (zio); | |
4873 | } | |
4874 | ||
34dc7c2f | 4875 | static int |
e8b96c60 | 4876 | arc_memory_throttle(uint64_t reserve, uint64_t txg) |
34dc7c2f BB |
4877 | { |
4878 | #ifdef _KERNEL | |
0c5493d4 BB |
4879 | if (zfs_arc_memory_throttle_disable) |
4880 | return (0); | |
4881 | ||
ca67b33a MA |
4882 | if (freemem > physmem * arc_lotsfree_percent / 100) |
4883 | return (0); | |
4884 | ||
4885 | if (arc_reclaim_needed()) { | |
4886 | /* memory is low, delay before restarting */ | |
34dc7c2f | 4887 | ARCSTAT_INCR(arcstat_memory_throttle_count, 1); |
570827e1 | 4888 | DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim); |
2e528b49 | 4889 | return (SET_ERROR(EAGAIN)); |
34dc7c2f | 4890 | } |
34dc7c2f BB |
4891 | #endif |
4892 | return (0); | |
4893 | } | |
4894 | ||
4895 | void | |
4896 | arc_tempreserve_clear(uint64_t reserve) | |
4897 | { | |
4898 | atomic_add_64(&arc_tempreserve, -reserve); | |
4899 | ASSERT((int64_t)arc_tempreserve >= 0); | |
4900 | } | |
4901 | ||
4902 | int | |
4903 | arc_tempreserve_space(uint64_t reserve, uint64_t txg) | |
4904 | { | |
4905 | int error; | |
9babb374 | 4906 | uint64_t anon_size; |
34dc7c2f | 4907 | |
34dc7c2f BB |
4908 | if (reserve > arc_c/4 && !arc_no_grow) |
4909 | arc_c = MIN(arc_c_max, reserve * 4); | |
12f9a6a3 BB |
4910 | |
4911 | /* | |
4912 | * Throttle when the calculated memory footprint for the TXG | |
4913 | * exceeds the target ARC size. | |
4914 | */ | |
570827e1 BB |
4915 | if (reserve > arc_c) { |
4916 | DMU_TX_STAT_BUMP(dmu_tx_memory_reserve); | |
12f9a6a3 | 4917 | return (SET_ERROR(ERESTART)); |
570827e1 | 4918 | } |
34dc7c2f | 4919 | |
9babb374 BB |
4920 | /* |
4921 | * Don't count loaned bufs as in flight dirty data to prevent long | |
4922 | * network delays from blocking transactions that are ready to be | |
4923 | * assigned to a txg. | |
4924 | */ | |
4925 | anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); | |
4926 | ||
34dc7c2f BB |
4927 | /* |
4928 | * Writes will, almost always, require additional memory allocations | |
d3cc8b15 | 4929 | * in order to compress/encrypt/etc the data. We therefore need to |
34dc7c2f BB |
4930 | * make sure that there is sufficient available memory for this. |
4931 | */ | |
e8b96c60 MA |
4932 | error = arc_memory_throttle(reserve, txg); |
4933 | if (error != 0) | |
34dc7c2f BB |
4934 | return (error); |
4935 | ||
4936 | /* | |
4937 | * Throttle writes when the amount of dirty data in the cache | |
4938 | * gets too large. We try to keep the cache less than half full | |
4939 | * of dirty blocks so that our sync times don't grow too large. | |
4940 | * Note: if two requests come in concurrently, we might let them | |
4941 | * both succeed, when one of them should fail. Not a huge deal. | |
4942 | */ | |
9babb374 BB |
4943 | |
4944 | if (reserve + arc_tempreserve + anon_size > arc_c / 2 && | |
4945 | anon_size > arc_c / 4) { | |
34dc7c2f BB |
4946 | dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " |
4947 | "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", | |
4948 | arc_tempreserve>>10, | |
4949 | arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, | |
4950 | arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, | |
4951 | reserve>>10, arc_c>>10); | |
570827e1 | 4952 | DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle); |
2e528b49 | 4953 | return (SET_ERROR(ERESTART)); |
34dc7c2f BB |
4954 | } |
4955 | atomic_add_64(&arc_tempreserve, reserve); | |
4956 | return (0); | |
4957 | } | |
4958 | ||
13be560d BB |
4959 | static void |
4960 | arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, | |
4961 | kstat_named_t *evict_data, kstat_named_t *evict_metadata) | |
4962 | { | |
4963 | size->value.ui64 = state->arcs_size; | |
4964 | evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; | |
4965 | evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; | |
4966 | } | |
4967 | ||
4968 | static int | |
4969 | arc_kstat_update(kstat_t *ksp, int rw) | |
4970 | { | |
4971 | arc_stats_t *as = ksp->ks_data; | |
4972 | ||
4973 | if (rw == KSTAT_WRITE) { | |
2e528b49 | 4974 | return (SET_ERROR(EACCES)); |
13be560d BB |
4975 | } else { |
4976 | arc_kstat_update_state(arc_anon, | |
4977 | &as->arcstat_anon_size, | |
4978 | &as->arcstat_anon_evict_data, | |
4979 | &as->arcstat_anon_evict_metadata); | |
4980 | arc_kstat_update_state(arc_mru, | |
4981 | &as->arcstat_mru_size, | |
4982 | &as->arcstat_mru_evict_data, | |
4983 | &as->arcstat_mru_evict_metadata); | |
4984 | arc_kstat_update_state(arc_mru_ghost, | |
4985 | &as->arcstat_mru_ghost_size, | |
4986 | &as->arcstat_mru_ghost_evict_data, | |
4987 | &as->arcstat_mru_ghost_evict_metadata); | |
4988 | arc_kstat_update_state(arc_mfu, | |
4989 | &as->arcstat_mfu_size, | |
4990 | &as->arcstat_mfu_evict_data, | |
4991 | &as->arcstat_mfu_evict_metadata); | |
fc41c640 | 4992 | arc_kstat_update_state(arc_mfu_ghost, |
13be560d BB |
4993 | &as->arcstat_mfu_ghost_size, |
4994 | &as->arcstat_mfu_ghost_evict_data, | |
4995 | &as->arcstat_mfu_ghost_evict_metadata); | |
4996 | } | |
4997 | ||
4998 | return (0); | |
4999 | } | |
5000 | ||
ca0bf58d PS |
5001 | /* |
5002 | * This function *must* return indices evenly distributed between all | |
5003 | * sublists of the multilist. This is needed due to how the ARC eviction | |
5004 | * code is laid out; arc_evict_state() assumes ARC buffers are evenly | |
5005 | * distributed between all sublists and uses this assumption when | |
5006 | * deciding which sublist to evict from and how much to evict from it. | |
5007 | */ | |
5008 | unsigned int | |
5009 | arc_state_multilist_index_func(multilist_t *ml, void *obj) | |
5010 | { | |
5011 | arc_buf_hdr_t *hdr = obj; | |
5012 | ||
5013 | /* | |
5014 | * We rely on b_dva to generate evenly distributed index | |
5015 | * numbers using buf_hash below. So, as an added precaution, | |
5016 | * let's make sure we never add empty buffers to the arc lists. | |
5017 | */ | |
5018 | ASSERT(!BUF_EMPTY(hdr)); | |
5019 | ||
5020 | /* | |
5021 | * The assumption here, is the hash value for a given | |
5022 | * arc_buf_hdr_t will remain constant throughout its lifetime | |
5023 | * (i.e. its b_spa, b_dva, and b_birth fields don't change). | |
5024 | * Thus, we don't need to store the header's sublist index | |
5025 | * on insertion, as this index can be recalculated on removal. | |
5026 | * | |
5027 | * Also, the low order bits of the hash value are thought to be | |
5028 | * distributed evenly. Otherwise, in the case that the multilist | |
5029 | * has a power of two number of sublists, each sublists' usage | |
5030 | * would not be evenly distributed. | |
5031 | */ | |
5032 | return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % | |
5033 | multilist_get_num_sublists(ml)); | |
5034 | } | |
5035 | ||
ca67b33a MA |
5036 | /* |
5037 | * Called during module initialization and periodically thereafter to | |
5038 | * apply reasonable changes to the exposed performance tunings. Non-zero | |
5039 | * zfs_* values which differ from the currently set values will be applied. | |
5040 | */ | |
5041 | static void | |
5042 | arc_tuning_update(void) | |
5043 | { | |
5044 | /* Valid range: 64M - <all physical memory> */ | |
5045 | if ((zfs_arc_max) && (zfs_arc_max != arc_c_max) && | |
5046 | (zfs_arc_max > 64 << 20) && (zfs_arc_max < ptob(physmem)) && | |
5047 | (zfs_arc_max > arc_c_min)) { | |
5048 | arc_c_max = zfs_arc_max; | |
5049 | arc_c = arc_c_max; | |
5050 | arc_p = (arc_c >> 1); | |
5051 | arc_meta_limit = MIN(arc_meta_limit, arc_c_max); | |
5052 | } | |
5053 | ||
5054 | /* Valid range: 32M - <arc_c_max> */ | |
5055 | if ((zfs_arc_min) && (zfs_arc_min != arc_c_min) && | |
5056 | (zfs_arc_min >= 2ULL << SPA_MAXBLOCKSHIFT) && | |
5057 | (zfs_arc_min <= arc_c_max)) { | |
5058 | arc_c_min = zfs_arc_min; | |
5059 | arc_c = MAX(arc_c, arc_c_min); | |
5060 | } | |
5061 | ||
5062 | /* Valid range: 16M - <arc_c_max> */ | |
5063 | if ((zfs_arc_meta_min) && (zfs_arc_meta_min != arc_meta_min) && | |
5064 | (zfs_arc_meta_min >= 1ULL << SPA_MAXBLOCKSHIFT) && | |
5065 | (zfs_arc_meta_min <= arc_c_max)) { | |
5066 | arc_meta_min = zfs_arc_meta_min; | |
5067 | arc_meta_limit = MAX(arc_meta_limit, arc_meta_min); | |
5068 | } | |
5069 | ||
5070 | /* Valid range: <arc_meta_min> - <arc_c_max> */ | |
5071 | if ((zfs_arc_meta_limit) && (zfs_arc_meta_limit != arc_meta_limit) && | |
5072 | (zfs_arc_meta_limit >= zfs_arc_meta_min) && | |
5073 | (zfs_arc_meta_limit <= arc_c_max)) | |
5074 | arc_meta_limit = zfs_arc_meta_limit; | |
5075 | ||
5076 | /* Valid range: 1 - N */ | |
5077 | if (zfs_arc_grow_retry) | |
5078 | arc_grow_retry = zfs_arc_grow_retry; | |
5079 | ||
5080 | /* Valid range: 1 - N */ | |
5081 | if (zfs_arc_shrink_shift) { | |
5082 | arc_shrink_shift = zfs_arc_shrink_shift; | |
5083 | arc_no_grow_shift = MIN(arc_no_grow_shift, arc_shrink_shift -1); | |
5084 | } | |
5085 | ||
5086 | /* Valid range: 1 - N ticks */ | |
5087 | if (zfs_arc_min_prefetch_lifespan) | |
5088 | arc_min_prefetch_lifespan = zfs_arc_min_prefetch_lifespan; | |
5089 | } | |
5090 | ||
34dc7c2f BB |
5091 | void |
5092 | arc_init(void) | |
5093 | { | |
ca67b33a MA |
5094 | /* |
5095 | * allmem is "all memory that we could possibly use". | |
5096 | */ | |
5097 | #ifdef _KERNEL | |
5098 | uint64_t allmem = ptob(physmem); | |
5099 | #else | |
5100 | uint64_t allmem = (physmem * PAGESIZE) / 2; | |
5101 | #endif | |
5102 | ||
ca0bf58d PS |
5103 | mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); |
5104 | cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); | |
5105 | cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); | |
5106 | ||
5107 | mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL); | |
5108 | cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL); | |
34dc7c2f BB |
5109 | |
5110 | /* Convert seconds to clock ticks */ | |
ca67b33a | 5111 | arc_min_prefetch_lifespan = 1 * hz; |
34dc7c2f BB |
5112 | |
5113 | /* Start out with 1/8 of all memory */ | |
ca67b33a | 5114 | arc_c = allmem / 8; |
34dc7c2f BB |
5115 | |
5116 | #ifdef _KERNEL | |
5117 | /* | |
5118 | * On architectures where the physical memory can be larger | |
5119 | * than the addressable space (intel in 32-bit mode), we may | |
5120 | * need to limit the cache to 1/8 of VM size. | |
5121 | */ | |
5122 | arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); | |
ca67b33a | 5123 | |
7cb67b45 BB |
5124 | /* |
5125 | * Register a shrinker to support synchronous (direct) memory | |
5126 | * reclaim from the arc. This is done to prevent kswapd from | |
5127 | * swapping out pages when it is preferable to shrink the arc. | |
5128 | */ | |
5129 | spl_register_shrinker(&arc_shrinker); | |
34dc7c2f BB |
5130 | #endif |
5131 | ||
ca67b33a | 5132 | /* Set min cache to allow safe operation of arc_adapt() */ |
121b3cae | 5133 | arc_c_min = 2ULL << SPA_MAXBLOCKSHIFT; |
ca67b33a MA |
5134 | /* Set max to 1/2 of all memory */ |
5135 | arc_c_max = allmem / 2; | |
34dc7c2f BB |
5136 | |
5137 | arc_c = arc_c_max; | |
5138 | arc_p = (arc_c >> 1); | |
5139 | ||
ca67b33a MA |
5140 | /* Set min to 1/2 of arc_c_min */ |
5141 | arc_meta_min = 1ULL << SPA_MAXBLOCKSHIFT; | |
5142 | /* Initialize maximum observed usage to zero */ | |
1834f2d8 | 5143 | arc_meta_max = 0; |
ca67b33a MA |
5144 | /* Set limit to 3/4 of arc_c_max with a floor of arc_meta_min */ |
5145 | arc_meta_limit = MAX((3 * arc_c_max) / 4, arc_meta_min); | |
34dc7c2f | 5146 | |
ca67b33a MA |
5147 | /* Apply user specified tunings */ |
5148 | arc_tuning_update(); | |
c52fca13 | 5149 | |
ca0bf58d | 5150 | if (zfs_arc_num_sublists_per_state < 1) |
ca67b33a | 5151 | zfs_arc_num_sublists_per_state = MAX(boot_ncpus, 1); |
ca0bf58d | 5152 | |
34dc7c2f BB |
5153 | /* if kmem_flags are set, lets try to use less memory */ |
5154 | if (kmem_debugging()) | |
5155 | arc_c = arc_c / 2; | |
5156 | if (arc_c < arc_c_min) | |
5157 | arc_c = arc_c_min; | |
5158 | ||
5159 | arc_anon = &ARC_anon; | |
5160 | arc_mru = &ARC_mru; | |
5161 | arc_mru_ghost = &ARC_mru_ghost; | |
5162 | arc_mfu = &ARC_mfu; | |
5163 | arc_mfu_ghost = &ARC_mfu_ghost; | |
5164 | arc_l2c_only = &ARC_l2c_only; | |
5165 | arc_size = 0; | |
5166 | ||
ca0bf58d | 5167 | multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], |
b9541d6b | 5168 | sizeof (arc_buf_hdr_t), |
ca0bf58d PS |
5169 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
5170 | zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); | |
5171 | multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], | |
b9541d6b | 5172 | sizeof (arc_buf_hdr_t), |
ca0bf58d PS |
5173 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
5174 | zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); | |
5175 | multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], | |
b9541d6b | 5176 | sizeof (arc_buf_hdr_t), |
ca0bf58d PS |
5177 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
5178 | zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); | |
5179 | multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], | |
b9541d6b | 5180 | sizeof (arc_buf_hdr_t), |
ca0bf58d PS |
5181 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
5182 | zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); | |
5183 | multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], | |
b9541d6b | 5184 | sizeof (arc_buf_hdr_t), |
ca0bf58d PS |
5185 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
5186 | zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); | |
5187 | multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], | |
b9541d6b | 5188 | sizeof (arc_buf_hdr_t), |
ca0bf58d PS |
5189 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
5190 | zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); | |
5191 | multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], | |
b9541d6b | 5192 | sizeof (arc_buf_hdr_t), |
ca0bf58d PS |
5193 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
5194 | zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); | |
5195 | multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], | |
b9541d6b | 5196 | sizeof (arc_buf_hdr_t), |
ca0bf58d PS |
5197 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
5198 | zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); | |
5199 | multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], | |
b9541d6b | 5200 | sizeof (arc_buf_hdr_t), |
ca0bf58d PS |
5201 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
5202 | zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); | |
5203 | multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], | |
b9541d6b | 5204 | sizeof (arc_buf_hdr_t), |
ca0bf58d PS |
5205 | offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), |
5206 | zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); | |
34dc7c2f | 5207 | |
e0b0ca98 BB |
5208 | arc_anon->arcs_state = ARC_STATE_ANON; |
5209 | arc_mru->arcs_state = ARC_STATE_MRU; | |
5210 | arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST; | |
5211 | arc_mfu->arcs_state = ARC_STATE_MFU; | |
5212 | arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST; | |
5213 | arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY; | |
5214 | ||
34dc7c2f BB |
5215 | buf_init(); |
5216 | ||
ca0bf58d PS |
5217 | arc_reclaim_thread_exit = FALSE; |
5218 | arc_user_evicts_thread_exit = FALSE; | |
ab26409d BB |
5219 | list_create(&arc_prune_list, sizeof (arc_prune_t), |
5220 | offsetof(arc_prune_t, p_node)); | |
34dc7c2f | 5221 | arc_eviction_list = NULL; |
ab26409d | 5222 | mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f BB |
5223 | bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); |
5224 | ||
f6046738 | 5225 | arc_prune_taskq = taskq_create("arc_prune", max_ncpus, minclsyspri, |
aa9af22c | 5226 | max_ncpus, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC); |
f6046738 | 5227 | |
34dc7c2f BB |
5228 | arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, |
5229 | sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); | |
5230 | ||
5231 | if (arc_ksp != NULL) { | |
5232 | arc_ksp->ks_data = &arc_stats; | |
13be560d | 5233 | arc_ksp->ks_update = arc_kstat_update; |
34dc7c2f BB |
5234 | kstat_install(arc_ksp); |
5235 | } | |
5236 | ||
ca67b33a | 5237 | (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, |
34dc7c2f BB |
5238 | TS_RUN, minclsyspri); |
5239 | ||
ca0bf58d PS |
5240 | (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0, |
5241 | TS_RUN, minclsyspri); | |
5242 | ||
34dc7c2f | 5243 | arc_dead = FALSE; |
b128c09f | 5244 | arc_warm = B_FALSE; |
34dc7c2f | 5245 | |
e8b96c60 MA |
5246 | /* |
5247 | * Calculate maximum amount of dirty data per pool. | |
5248 | * | |
5249 | * If it has been set by a module parameter, take that. | |
5250 | * Otherwise, use a percentage of physical memory defined by | |
5251 | * zfs_dirty_data_max_percent (default 10%) with a cap at | |
5252 | * zfs_dirty_data_max_max (default 25% of physical memory). | |
5253 | */ | |
5254 | if (zfs_dirty_data_max_max == 0) | |
5255 | zfs_dirty_data_max_max = physmem * PAGESIZE * | |
5256 | zfs_dirty_data_max_max_percent / 100; | |
5257 | ||
5258 | if (zfs_dirty_data_max == 0) { | |
5259 | zfs_dirty_data_max = physmem * PAGESIZE * | |
5260 | zfs_dirty_data_max_percent / 100; | |
5261 | zfs_dirty_data_max = MIN(zfs_dirty_data_max, | |
5262 | zfs_dirty_data_max_max); | |
5263 | } | |
34dc7c2f BB |
5264 | } |
5265 | ||
5266 | void | |
5267 | arc_fini(void) | |
5268 | { | |
ab26409d BB |
5269 | arc_prune_t *p; |
5270 | ||
7cb67b45 BB |
5271 | #ifdef _KERNEL |
5272 | spl_unregister_shrinker(&arc_shrinker); | |
5273 | #endif /* _KERNEL */ | |
5274 | ||
ca0bf58d PS |
5275 | mutex_enter(&arc_reclaim_lock); |
5276 | arc_reclaim_thread_exit = TRUE; | |
5277 | /* | |
5278 | * The reclaim thread will set arc_reclaim_thread_exit back to | |
5279 | * FALSE when it is finished exiting; we're waiting for that. | |
5280 | */ | |
5281 | while (arc_reclaim_thread_exit) { | |
5282 | cv_signal(&arc_reclaim_thread_cv); | |
5283 | cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); | |
5284 | } | |
5285 | mutex_exit(&arc_reclaim_lock); | |
5286 | ||
5287 | mutex_enter(&arc_user_evicts_lock); | |
5288 | arc_user_evicts_thread_exit = TRUE; | |
5289 | /* | |
5290 | * The user evicts thread will set arc_user_evicts_thread_exit | |
5291 | * to FALSE when it is finished exiting; we're waiting for that. | |
5292 | */ | |
5293 | while (arc_user_evicts_thread_exit) { | |
5294 | cv_signal(&arc_user_evicts_cv); | |
5295 | cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock); | |
5296 | } | |
5297 | mutex_exit(&arc_user_evicts_lock); | |
34dc7c2f | 5298 | |
ca0bf58d PS |
5299 | /* Use TRUE to ensure *all* buffers are evicted */ |
5300 | arc_flush(NULL, TRUE); | |
34dc7c2f BB |
5301 | |
5302 | arc_dead = TRUE; | |
5303 | ||
5304 | if (arc_ksp != NULL) { | |
5305 | kstat_delete(arc_ksp); | |
5306 | arc_ksp = NULL; | |
5307 | } | |
5308 | ||
f6046738 BB |
5309 | taskq_wait(arc_prune_taskq); |
5310 | taskq_destroy(arc_prune_taskq); | |
5311 | ||
ab26409d BB |
5312 | mutex_enter(&arc_prune_mtx); |
5313 | while ((p = list_head(&arc_prune_list)) != NULL) { | |
5314 | list_remove(&arc_prune_list, p); | |
5315 | refcount_remove(&p->p_refcnt, &arc_prune_list); | |
5316 | refcount_destroy(&p->p_refcnt); | |
5317 | kmem_free(p, sizeof (*p)); | |
5318 | } | |
5319 | mutex_exit(&arc_prune_mtx); | |
5320 | ||
5321 | list_destroy(&arc_prune_list); | |
5322 | mutex_destroy(&arc_prune_mtx); | |
ca0bf58d PS |
5323 | mutex_destroy(&arc_reclaim_lock); |
5324 | cv_destroy(&arc_reclaim_thread_cv); | |
5325 | cv_destroy(&arc_reclaim_waiters_cv); | |
5326 | ||
5327 | mutex_destroy(&arc_user_evicts_lock); | |
5328 | cv_destroy(&arc_user_evicts_cv); | |
5329 | ||
5330 | multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); | |
5331 | multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); | |
5332 | multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); | |
5333 | multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); | |
5334 | multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); | |
5335 | multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); | |
5336 | multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); | |
5337 | multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); | |
5338 | multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]); | |
5339 | multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]); | |
34dc7c2f BB |
5340 | |
5341 | buf_fini(); | |
9babb374 | 5342 | |
b9541d6b | 5343 | ASSERT0(arc_loaned_bytes); |
34dc7c2f BB |
5344 | } |
5345 | ||
5346 | /* | |
5347 | * Level 2 ARC | |
5348 | * | |
5349 | * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. | |
5350 | * It uses dedicated storage devices to hold cached data, which are populated | |
5351 | * using large infrequent writes. The main role of this cache is to boost | |
5352 | * the performance of random read workloads. The intended L2ARC devices | |
5353 | * include short-stroked disks, solid state disks, and other media with | |
5354 | * substantially faster read latency than disk. | |
5355 | * | |
5356 | * +-----------------------+ | |
5357 | * | ARC | | |
5358 | * +-----------------------+ | |
5359 | * | ^ ^ | |
5360 | * | | | | |
5361 | * l2arc_feed_thread() arc_read() | |
5362 | * | | | | |
5363 | * | l2arc read | | |
5364 | * V | | | |
5365 | * +---------------+ | | |
5366 | * | L2ARC | | | |
5367 | * +---------------+ | | |
5368 | * | ^ | | |
5369 | * l2arc_write() | | | |
5370 | * | | | | |
5371 | * V | | | |
5372 | * +-------+ +-------+ | |
5373 | * | vdev | | vdev | | |
5374 | * | cache | | cache | | |
5375 | * +-------+ +-------+ | |
5376 | * +=========+ .-----. | |
5377 | * : L2ARC : |-_____-| | |
5378 | * : devices : | Disks | | |
5379 | * +=========+ `-_____-' | |
5380 | * | |
5381 | * Read requests are satisfied from the following sources, in order: | |
5382 | * | |
5383 | * 1) ARC | |
5384 | * 2) vdev cache of L2ARC devices | |
5385 | * 3) L2ARC devices | |
5386 | * 4) vdev cache of disks | |
5387 | * 5) disks | |
5388 | * | |
5389 | * Some L2ARC device types exhibit extremely slow write performance. | |
5390 | * To accommodate for this there are some significant differences between | |
5391 | * the L2ARC and traditional cache design: | |
5392 | * | |
5393 | * 1. There is no eviction path from the ARC to the L2ARC. Evictions from | |
5394 | * the ARC behave as usual, freeing buffers and placing headers on ghost | |
5395 | * lists. The ARC does not send buffers to the L2ARC during eviction as | |
5396 | * this would add inflated write latencies for all ARC memory pressure. | |
5397 | * | |
5398 | * 2. The L2ARC attempts to cache data from the ARC before it is evicted. | |
5399 | * It does this by periodically scanning buffers from the eviction-end of | |
5400 | * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are | |
3a17a7a9 SK |
5401 | * not already there. It scans until a headroom of buffers is satisfied, |
5402 | * which itself is a buffer for ARC eviction. If a compressible buffer is | |
5403 | * found during scanning and selected for writing to an L2ARC device, we | |
5404 | * temporarily boost scanning headroom during the next scan cycle to make | |
5405 | * sure we adapt to compression effects (which might significantly reduce | |
5406 | * the data volume we write to L2ARC). The thread that does this is | |
34dc7c2f BB |
5407 | * l2arc_feed_thread(), illustrated below; example sizes are included to |
5408 | * provide a better sense of ratio than this diagram: | |
5409 | * | |
5410 | * head --> tail | |
5411 | * +---------------------+----------+ | |
5412 | * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC | |
5413 | * +---------------------+----------+ | o L2ARC eligible | |
5414 | * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer | |
5415 | * +---------------------+----------+ | | |
5416 | * 15.9 Gbytes ^ 32 Mbytes | | |
5417 | * headroom | | |
5418 | * l2arc_feed_thread() | |
5419 | * | | |
5420 | * l2arc write hand <--[oooo]--' | |
5421 | * | 8 Mbyte | |
5422 | * | write max | |
5423 | * V | |
5424 | * +==============================+ | |
5425 | * L2ARC dev |####|#|###|###| |####| ... | | |
5426 | * +==============================+ | |
5427 | * 32 Gbytes | |
5428 | * | |
5429 | * 3. If an ARC buffer is copied to the L2ARC but then hit instead of | |
5430 | * evicted, then the L2ARC has cached a buffer much sooner than it probably | |
5431 | * needed to, potentially wasting L2ARC device bandwidth and storage. It is | |
5432 | * safe to say that this is an uncommon case, since buffers at the end of | |
5433 | * the ARC lists have moved there due to inactivity. | |
5434 | * | |
5435 | * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, | |
5436 | * then the L2ARC simply misses copying some buffers. This serves as a | |
5437 | * pressure valve to prevent heavy read workloads from both stalling the ARC | |
5438 | * with waits and clogging the L2ARC with writes. This also helps prevent | |
5439 | * the potential for the L2ARC to churn if it attempts to cache content too | |
5440 | * quickly, such as during backups of the entire pool. | |
5441 | * | |
b128c09f BB |
5442 | * 5. After system boot and before the ARC has filled main memory, there are |
5443 | * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru | |
5444 | * lists can remain mostly static. Instead of searching from tail of these | |
5445 | * lists as pictured, the l2arc_feed_thread() will search from the list heads | |
5446 | * for eligible buffers, greatly increasing its chance of finding them. | |
5447 | * | |
5448 | * The L2ARC device write speed is also boosted during this time so that | |
5449 | * the L2ARC warms up faster. Since there have been no ARC evictions yet, | |
5450 | * there are no L2ARC reads, and no fear of degrading read performance | |
5451 | * through increased writes. | |
5452 | * | |
5453 | * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that | |
34dc7c2f BB |
5454 | * the vdev queue can aggregate them into larger and fewer writes. Each |
5455 | * device is written to in a rotor fashion, sweeping writes through | |
5456 | * available space then repeating. | |
5457 | * | |
b128c09f | 5458 | * 7. The L2ARC does not store dirty content. It never needs to flush |
34dc7c2f BB |
5459 | * write buffers back to disk based storage. |
5460 | * | |
b128c09f | 5461 | * 8. If an ARC buffer is written (and dirtied) which also exists in the |
34dc7c2f BB |
5462 | * L2ARC, the now stale L2ARC buffer is immediately dropped. |
5463 | * | |
5464 | * The performance of the L2ARC can be tweaked by a number of tunables, which | |
5465 | * may be necessary for different workloads: | |
5466 | * | |
5467 | * l2arc_write_max max write bytes per interval | |
b128c09f | 5468 | * l2arc_write_boost extra write bytes during device warmup |
34dc7c2f | 5469 | * l2arc_noprefetch skip caching prefetched buffers |
3a17a7a9 | 5470 | * l2arc_nocompress skip compressing buffers |
34dc7c2f | 5471 | * l2arc_headroom number of max device writes to precache |
3a17a7a9 SK |
5472 | * l2arc_headroom_boost when we find compressed buffers during ARC |
5473 | * scanning, we multiply headroom by this | |
5474 | * percentage factor for the next scan cycle, | |
5475 | * since more compressed buffers are likely to | |
5476 | * be present | |
34dc7c2f BB |
5477 | * l2arc_feed_secs seconds between L2ARC writing |
5478 | * | |
5479 | * Tunables may be removed or added as future performance improvements are | |
5480 | * integrated, and also may become zpool properties. | |
d164b209 BB |
5481 | * |
5482 | * There are three key functions that control how the L2ARC warms up: | |
5483 | * | |
5484 | * l2arc_write_eligible() check if a buffer is eligible to cache | |
5485 | * l2arc_write_size() calculate how much to write | |
5486 | * l2arc_write_interval() calculate sleep delay between writes | |
5487 | * | |
5488 | * These three functions determine what to write, how much, and how quickly | |
5489 | * to send writes. | |
34dc7c2f BB |
5490 | */ |
5491 | ||
d164b209 | 5492 | static boolean_t |
2a432414 | 5493 | l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) |
d164b209 BB |
5494 | { |
5495 | /* | |
5496 | * A buffer is *not* eligible for the L2ARC if it: | |
5497 | * 1. belongs to a different spa. | |
428870ff BB |
5498 | * 2. is already cached on the L2ARC. |
5499 | * 3. has an I/O in progress (it may be an incomplete read). | |
5500 | * 4. is flagged not eligible (zfs property). | |
d164b209 | 5501 | */ |
b9541d6b | 5502 | if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) || |
2a432414 | 5503 | HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr)) |
d164b209 BB |
5504 | return (B_FALSE); |
5505 | ||
5506 | return (B_TRUE); | |
5507 | } | |
5508 | ||
5509 | static uint64_t | |
3a17a7a9 | 5510 | l2arc_write_size(void) |
d164b209 BB |
5511 | { |
5512 | uint64_t size; | |
5513 | ||
3a17a7a9 SK |
5514 | /* |
5515 | * Make sure our globals have meaningful values in case the user | |
5516 | * altered them. | |
5517 | */ | |
5518 | size = l2arc_write_max; | |
5519 | if (size == 0) { | |
5520 | cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " | |
5521 | "be greater than zero, resetting it to the default (%d)", | |
5522 | L2ARC_WRITE_SIZE); | |
5523 | size = l2arc_write_max = L2ARC_WRITE_SIZE; | |
5524 | } | |
d164b209 BB |
5525 | |
5526 | if (arc_warm == B_FALSE) | |
3a17a7a9 | 5527 | size += l2arc_write_boost; |
d164b209 BB |
5528 | |
5529 | return (size); | |
5530 | ||
5531 | } | |
5532 | ||
5533 | static clock_t | |
5534 | l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) | |
5535 | { | |
428870ff | 5536 | clock_t interval, next, now; |
d164b209 BB |
5537 | |
5538 | /* | |
5539 | * If the ARC lists are busy, increase our write rate; if the | |
5540 | * lists are stale, idle back. This is achieved by checking | |
5541 | * how much we previously wrote - if it was more than half of | |
5542 | * what we wanted, schedule the next write much sooner. | |
5543 | */ | |
5544 | if (l2arc_feed_again && wrote > (wanted / 2)) | |
5545 | interval = (hz * l2arc_feed_min_ms) / 1000; | |
5546 | else | |
5547 | interval = hz * l2arc_feed_secs; | |
5548 | ||
428870ff BB |
5549 | now = ddi_get_lbolt(); |
5550 | next = MAX(now, MIN(now + interval, began + interval)); | |
d164b209 BB |
5551 | |
5552 | return (next); | |
5553 | } | |
5554 | ||
34dc7c2f BB |
5555 | /* |
5556 | * Cycle through L2ARC devices. This is how L2ARC load balances. | |
b128c09f | 5557 | * If a device is returned, this also returns holding the spa config lock. |
34dc7c2f BB |
5558 | */ |
5559 | static l2arc_dev_t * | |
5560 | l2arc_dev_get_next(void) | |
5561 | { | |
b128c09f | 5562 | l2arc_dev_t *first, *next = NULL; |
34dc7c2f | 5563 | |
b128c09f BB |
5564 | /* |
5565 | * Lock out the removal of spas (spa_namespace_lock), then removal | |
5566 | * of cache devices (l2arc_dev_mtx). Once a device has been selected, | |
5567 | * both locks will be dropped and a spa config lock held instead. | |
5568 | */ | |
5569 | mutex_enter(&spa_namespace_lock); | |
5570 | mutex_enter(&l2arc_dev_mtx); | |
5571 | ||
5572 | /* if there are no vdevs, there is nothing to do */ | |
5573 | if (l2arc_ndev == 0) | |
5574 | goto out; | |
5575 | ||
5576 | first = NULL; | |
5577 | next = l2arc_dev_last; | |
5578 | do { | |
5579 | /* loop around the list looking for a non-faulted vdev */ | |
5580 | if (next == NULL) { | |
34dc7c2f | 5581 | next = list_head(l2arc_dev_list); |
b128c09f BB |
5582 | } else { |
5583 | next = list_next(l2arc_dev_list, next); | |
5584 | if (next == NULL) | |
5585 | next = list_head(l2arc_dev_list); | |
5586 | } | |
5587 | ||
5588 | /* if we have come back to the start, bail out */ | |
5589 | if (first == NULL) | |
5590 | first = next; | |
5591 | else if (next == first) | |
5592 | break; | |
5593 | ||
5594 | } while (vdev_is_dead(next->l2ad_vdev)); | |
5595 | ||
5596 | /* if we were unable to find any usable vdevs, return NULL */ | |
5597 | if (vdev_is_dead(next->l2ad_vdev)) | |
5598 | next = NULL; | |
34dc7c2f BB |
5599 | |
5600 | l2arc_dev_last = next; | |
5601 | ||
b128c09f BB |
5602 | out: |
5603 | mutex_exit(&l2arc_dev_mtx); | |
5604 | ||
5605 | /* | |
5606 | * Grab the config lock to prevent the 'next' device from being | |
5607 | * removed while we are writing to it. | |
5608 | */ | |
5609 | if (next != NULL) | |
5610 | spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); | |
5611 | mutex_exit(&spa_namespace_lock); | |
5612 | ||
34dc7c2f BB |
5613 | return (next); |
5614 | } | |
5615 | ||
b128c09f BB |
5616 | /* |
5617 | * Free buffers that were tagged for destruction. | |
5618 | */ | |
5619 | static void | |
0bc8fd78 | 5620 | l2arc_do_free_on_write(void) |
b128c09f BB |
5621 | { |
5622 | list_t *buflist; | |
5623 | l2arc_data_free_t *df, *df_prev; | |
5624 | ||
5625 | mutex_enter(&l2arc_free_on_write_mtx); | |
5626 | buflist = l2arc_free_on_write; | |
5627 | ||
5628 | for (df = list_tail(buflist); df; df = df_prev) { | |
5629 | df_prev = list_prev(buflist, df); | |
5630 | ASSERT(df->l2df_data != NULL); | |
5631 | ASSERT(df->l2df_func != NULL); | |
5632 | df->l2df_func(df->l2df_data, df->l2df_size); | |
5633 | list_remove(buflist, df); | |
5634 | kmem_free(df, sizeof (l2arc_data_free_t)); | |
5635 | } | |
5636 | ||
5637 | mutex_exit(&l2arc_free_on_write_mtx); | |
5638 | } | |
5639 | ||
34dc7c2f BB |
5640 | /* |
5641 | * A write to a cache device has completed. Update all headers to allow | |
5642 | * reads from these buffers to begin. | |
5643 | */ | |
5644 | static void | |
5645 | l2arc_write_done(zio_t *zio) | |
5646 | { | |
5647 | l2arc_write_callback_t *cb; | |
5648 | l2arc_dev_t *dev; | |
5649 | list_t *buflist; | |
2a432414 | 5650 | arc_buf_hdr_t *head, *hdr, *hdr_prev; |
34dc7c2f | 5651 | kmutex_t *hash_lock; |
3bec585e | 5652 | int64_t bytes_dropped = 0; |
34dc7c2f BB |
5653 | |
5654 | cb = zio->io_private; | |
5655 | ASSERT(cb != NULL); | |
5656 | dev = cb->l2wcb_dev; | |
5657 | ASSERT(dev != NULL); | |
5658 | head = cb->l2wcb_head; | |
5659 | ASSERT(head != NULL); | |
b9541d6b | 5660 | buflist = &dev->l2ad_buflist; |
34dc7c2f BB |
5661 | ASSERT(buflist != NULL); |
5662 | DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, | |
5663 | l2arc_write_callback_t *, cb); | |
5664 | ||
5665 | if (zio->io_error != 0) | |
5666 | ARCSTAT_BUMP(arcstat_l2_writes_error); | |
5667 | ||
34dc7c2f BB |
5668 | /* |
5669 | * All writes completed, or an error was hit. | |
5670 | */ | |
ca0bf58d PS |
5671 | top: |
5672 | mutex_enter(&dev->l2ad_mtx); | |
2a432414 GW |
5673 | for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { |
5674 | hdr_prev = list_prev(buflist, hdr); | |
34dc7c2f | 5675 | |
2a432414 | 5676 | hash_lock = HDR_LOCK(hdr); |
ca0bf58d PS |
5677 | |
5678 | /* | |
5679 | * We cannot use mutex_enter or else we can deadlock | |
5680 | * with l2arc_write_buffers (due to swapping the order | |
5681 | * the hash lock and l2ad_mtx are taken). | |
5682 | */ | |
34dc7c2f BB |
5683 | if (!mutex_tryenter(hash_lock)) { |
5684 | /* | |
ca0bf58d PS |
5685 | * Missed the hash lock. We must retry so we |
5686 | * don't leave the ARC_FLAG_L2_WRITING bit set. | |
34dc7c2f | 5687 | */ |
ca0bf58d PS |
5688 | ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); |
5689 | ||
5690 | /* | |
5691 | * We don't want to rescan the headers we've | |
5692 | * already marked as having been written out, so | |
5693 | * we reinsert the head node so we can pick up | |
5694 | * where we left off. | |
5695 | */ | |
5696 | list_remove(buflist, head); | |
5697 | list_insert_after(buflist, hdr, head); | |
5698 | ||
5699 | mutex_exit(&dev->l2ad_mtx); | |
5700 | ||
5701 | /* | |
5702 | * We wait for the hash lock to become available | |
5703 | * to try and prevent busy waiting, and increase | |
5704 | * the chance we'll be able to acquire the lock | |
5705 | * the next time around. | |
5706 | */ | |
5707 | mutex_enter(hash_lock); | |
5708 | mutex_exit(hash_lock); | |
5709 | goto top; | |
34dc7c2f BB |
5710 | } |
5711 | ||
b9541d6b | 5712 | /* |
ca0bf58d PS |
5713 | * We could not have been moved into the arc_l2c_only |
5714 | * state while in-flight due to our ARC_FLAG_L2_WRITING | |
5715 | * bit being set. Let's just ensure that's being enforced. | |
5716 | */ | |
5717 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
5718 | ||
5719 | /* | |
5720 | * We may have allocated a buffer for L2ARC compression, | |
5721 | * we must release it to avoid leaking this data. | |
b9541d6b | 5722 | */ |
ca0bf58d | 5723 | l2arc_release_cdata_buf(hdr); |
b9541d6b | 5724 | |
34dc7c2f BB |
5725 | if (zio->io_error != 0) { |
5726 | /* | |
b128c09f | 5727 | * Error - drop L2ARC entry. |
34dc7c2f | 5728 | */ |
2a432414 | 5729 | list_remove(buflist, hdr); |
b9541d6b CW |
5730 | hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; |
5731 | ||
5732 | ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); | |
2a432414 | 5733 | ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); |
d962d5da PS |
5734 | |
5735 | bytes_dropped += hdr->b_l2hdr.b_asize; | |
5736 | (void) refcount_remove_many(&dev->l2ad_alloc, | |
5737 | hdr->b_l2hdr.b_asize, hdr); | |
34dc7c2f BB |
5738 | } |
5739 | ||
5740 | /* | |
ca0bf58d PS |
5741 | * Allow ARC to begin reads and ghost list evictions to |
5742 | * this L2ARC entry. | |
34dc7c2f | 5743 | */ |
2a432414 | 5744 | hdr->b_flags &= ~ARC_FLAG_L2_WRITING; |
34dc7c2f BB |
5745 | |
5746 | mutex_exit(hash_lock); | |
5747 | } | |
5748 | ||
5749 | atomic_inc_64(&l2arc_writes_done); | |
5750 | list_remove(buflist, head); | |
b9541d6b CW |
5751 | ASSERT(!HDR_HAS_L1HDR(head)); |
5752 | kmem_cache_free(hdr_l2only_cache, head); | |
5753 | mutex_exit(&dev->l2ad_mtx); | |
34dc7c2f | 5754 | |
3bec585e SK |
5755 | vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); |
5756 | ||
b128c09f | 5757 | l2arc_do_free_on_write(); |
34dc7c2f BB |
5758 | |
5759 | kmem_free(cb, sizeof (l2arc_write_callback_t)); | |
5760 | } | |
5761 | ||
5762 | /* | |
5763 | * A read to a cache device completed. Validate buffer contents before | |
5764 | * handing over to the regular ARC routines. | |
5765 | */ | |
5766 | static void | |
5767 | l2arc_read_done(zio_t *zio) | |
5768 | { | |
5769 | l2arc_read_callback_t *cb; | |
5770 | arc_buf_hdr_t *hdr; | |
5771 | arc_buf_t *buf; | |
34dc7c2f | 5772 | kmutex_t *hash_lock; |
b128c09f BB |
5773 | int equal; |
5774 | ||
5775 | ASSERT(zio->io_vd != NULL); | |
5776 | ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); | |
5777 | ||
5778 | spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); | |
34dc7c2f BB |
5779 | |
5780 | cb = zio->io_private; | |
5781 | ASSERT(cb != NULL); | |
5782 | buf = cb->l2rcb_buf; | |
5783 | ASSERT(buf != NULL); | |
34dc7c2f | 5784 | |
428870ff | 5785 | hash_lock = HDR_LOCK(buf->b_hdr); |
34dc7c2f | 5786 | mutex_enter(hash_lock); |
428870ff BB |
5787 | hdr = buf->b_hdr; |
5788 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
34dc7c2f | 5789 | |
3a17a7a9 SK |
5790 | /* |
5791 | * If the buffer was compressed, decompress it first. | |
5792 | */ | |
5793 | if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) | |
5794 | l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); | |
5795 | ASSERT(zio->io_data != NULL); | |
5796 | ||
34dc7c2f BB |
5797 | /* |
5798 | * Check this survived the L2ARC journey. | |
5799 | */ | |
5800 | equal = arc_cksum_equal(buf); | |
5801 | if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { | |
5802 | mutex_exit(hash_lock); | |
5803 | zio->io_private = buf; | |
b128c09f BB |
5804 | zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ |
5805 | zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ | |
34dc7c2f BB |
5806 | arc_read_done(zio); |
5807 | } else { | |
5808 | mutex_exit(hash_lock); | |
5809 | /* | |
5810 | * Buffer didn't survive caching. Increment stats and | |
5811 | * reissue to the original storage device. | |
5812 | */ | |
b128c09f | 5813 | if (zio->io_error != 0) { |
34dc7c2f | 5814 | ARCSTAT_BUMP(arcstat_l2_io_error); |
b128c09f | 5815 | } else { |
2e528b49 | 5816 | zio->io_error = SET_ERROR(EIO); |
b128c09f | 5817 | } |
34dc7c2f BB |
5818 | if (!equal) |
5819 | ARCSTAT_BUMP(arcstat_l2_cksum_bad); | |
5820 | ||
34dc7c2f | 5821 | /* |
b128c09f BB |
5822 | * If there's no waiter, issue an async i/o to the primary |
5823 | * storage now. If there *is* a waiter, the caller must | |
5824 | * issue the i/o in a context where it's OK to block. | |
34dc7c2f | 5825 | */ |
d164b209 BB |
5826 | if (zio->io_waiter == NULL) { |
5827 | zio_t *pio = zio_unique_parent(zio); | |
5828 | ||
5829 | ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); | |
5830 | ||
5831 | zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, | |
b128c09f BB |
5832 | buf->b_data, zio->io_size, arc_read_done, buf, |
5833 | zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); | |
d164b209 | 5834 | } |
34dc7c2f BB |
5835 | } |
5836 | ||
5837 | kmem_free(cb, sizeof (l2arc_read_callback_t)); | |
5838 | } | |
5839 | ||
5840 | /* | |
5841 | * This is the list priority from which the L2ARC will search for pages to | |
5842 | * cache. This is used within loops (0..3) to cycle through lists in the | |
5843 | * desired order. This order can have a significant effect on cache | |
5844 | * performance. | |
5845 | * | |
5846 | * Currently the metadata lists are hit first, MFU then MRU, followed by | |
5847 | * the data lists. This function returns a locked list, and also returns | |
5848 | * the lock pointer. | |
5849 | */ | |
ca0bf58d PS |
5850 | static multilist_sublist_t * |
5851 | l2arc_sublist_lock(int list_num) | |
34dc7c2f | 5852 | { |
ca0bf58d PS |
5853 | multilist_t *ml = NULL; |
5854 | unsigned int idx; | |
34dc7c2f BB |
5855 | |
5856 | ASSERT(list_num >= 0 && list_num <= 3); | |
5857 | ||
5858 | switch (list_num) { | |
5859 | case 0: | |
ca0bf58d | 5860 | ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; |
34dc7c2f BB |
5861 | break; |
5862 | case 1: | |
ca0bf58d | 5863 | ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; |
34dc7c2f BB |
5864 | break; |
5865 | case 2: | |
ca0bf58d | 5866 | ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; |
34dc7c2f BB |
5867 | break; |
5868 | case 3: | |
ca0bf58d | 5869 | ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; |
34dc7c2f BB |
5870 | break; |
5871 | } | |
5872 | ||
ca0bf58d PS |
5873 | /* |
5874 | * Return a randomly-selected sublist. This is acceptable | |
5875 | * because the caller feeds only a little bit of data for each | |
5876 | * call (8MB). Subsequent calls will result in different | |
5877 | * sublists being selected. | |
5878 | */ | |
5879 | idx = multilist_get_random_index(ml); | |
5880 | return (multilist_sublist_lock(ml, idx)); | |
34dc7c2f BB |
5881 | } |
5882 | ||
5883 | /* | |
5884 | * Evict buffers from the device write hand to the distance specified in | |
5885 | * bytes. This distance may span populated buffers, it may span nothing. | |
5886 | * This is clearing a region on the L2ARC device ready for writing. | |
5887 | * If the 'all' boolean is set, every buffer is evicted. | |
5888 | */ | |
5889 | static void | |
5890 | l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) | |
5891 | { | |
5892 | list_t *buflist; | |
2a432414 | 5893 | arc_buf_hdr_t *hdr, *hdr_prev; |
34dc7c2f BB |
5894 | kmutex_t *hash_lock; |
5895 | uint64_t taddr; | |
5896 | ||
b9541d6b | 5897 | buflist = &dev->l2ad_buflist; |
34dc7c2f BB |
5898 | |
5899 | if (!all && dev->l2ad_first) { | |
5900 | /* | |
5901 | * This is the first sweep through the device. There is | |
5902 | * nothing to evict. | |
5903 | */ | |
5904 | return; | |
5905 | } | |
5906 | ||
b128c09f | 5907 | if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { |
34dc7c2f BB |
5908 | /* |
5909 | * When nearing the end of the device, evict to the end | |
5910 | * before the device write hand jumps to the start. | |
5911 | */ | |
5912 | taddr = dev->l2ad_end; | |
5913 | } else { | |
5914 | taddr = dev->l2ad_hand + distance; | |
5915 | } | |
5916 | DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, | |
5917 | uint64_t, taddr, boolean_t, all); | |
5918 | ||
5919 | top: | |
b9541d6b | 5920 | mutex_enter(&dev->l2ad_mtx); |
2a432414 GW |
5921 | for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { |
5922 | hdr_prev = list_prev(buflist, hdr); | |
34dc7c2f | 5923 | |
2a432414 | 5924 | hash_lock = HDR_LOCK(hdr); |
ca0bf58d PS |
5925 | |
5926 | /* | |
5927 | * We cannot use mutex_enter or else we can deadlock | |
5928 | * with l2arc_write_buffers (due to swapping the order | |
5929 | * the hash lock and l2ad_mtx are taken). | |
5930 | */ | |
34dc7c2f BB |
5931 | if (!mutex_tryenter(hash_lock)) { |
5932 | /* | |
5933 | * Missed the hash lock. Retry. | |
5934 | */ | |
5935 | ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); | |
b9541d6b | 5936 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f BB |
5937 | mutex_enter(hash_lock); |
5938 | mutex_exit(hash_lock); | |
5939 | goto top; | |
5940 | } | |
5941 | ||
2a432414 | 5942 | if (HDR_L2_WRITE_HEAD(hdr)) { |
34dc7c2f BB |
5943 | /* |
5944 | * We hit a write head node. Leave it for | |
5945 | * l2arc_write_done(). | |
5946 | */ | |
2a432414 | 5947 | list_remove(buflist, hdr); |
34dc7c2f BB |
5948 | mutex_exit(hash_lock); |
5949 | continue; | |
5950 | } | |
5951 | ||
b9541d6b CW |
5952 | if (!all && HDR_HAS_L2HDR(hdr) && |
5953 | (hdr->b_l2hdr.b_daddr > taddr || | |
5954 | hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { | |
34dc7c2f BB |
5955 | /* |
5956 | * We've evicted to the target address, | |
5957 | * or the end of the device. | |
5958 | */ | |
5959 | mutex_exit(hash_lock); | |
5960 | break; | |
5961 | } | |
5962 | ||
b9541d6b CW |
5963 | ASSERT(HDR_HAS_L2HDR(hdr)); |
5964 | if (!HDR_HAS_L1HDR(hdr)) { | |
2a432414 | 5965 | ASSERT(!HDR_L2_READING(hdr)); |
34dc7c2f BB |
5966 | /* |
5967 | * This doesn't exist in the ARC. Destroy. | |
5968 | * arc_hdr_destroy() will call list_remove() | |
5969 | * and decrement arcstat_l2_size. | |
5970 | */ | |
2a432414 GW |
5971 | arc_change_state(arc_anon, hdr, hash_lock); |
5972 | arc_hdr_destroy(hdr); | |
34dc7c2f | 5973 | } else { |
b9541d6b CW |
5974 | ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); |
5975 | ARCSTAT_BUMP(arcstat_l2_evict_l1cached); | |
b128c09f BB |
5976 | /* |
5977 | * Invalidate issued or about to be issued | |
5978 | * reads, since we may be about to write | |
5979 | * over this location. | |
5980 | */ | |
2a432414 | 5981 | if (HDR_L2_READING(hdr)) { |
b128c09f | 5982 | ARCSTAT_BUMP(arcstat_l2_evict_reading); |
2a432414 | 5983 | hdr->b_flags |= ARC_FLAG_L2_EVICTED; |
b128c09f BB |
5984 | } |
5985 | ||
ca0bf58d PS |
5986 | /* Ensure this header has finished being written */ |
5987 | ASSERT(!HDR_L2_WRITING(hdr)); | |
5988 | ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); | |
d962d5da PS |
5989 | |
5990 | arc_hdr_l2hdr_destroy(hdr); | |
34dc7c2f BB |
5991 | } |
5992 | mutex_exit(hash_lock); | |
5993 | } | |
b9541d6b | 5994 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f BB |
5995 | } |
5996 | ||
5997 | /* | |
5998 | * Find and write ARC buffers to the L2ARC device. | |
5999 | * | |
2a432414 | 6000 | * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid |
34dc7c2f | 6001 | * for reading until they have completed writing. |
3a17a7a9 SK |
6002 | * The headroom_boost is an in-out parameter used to maintain headroom boost |
6003 | * state between calls to this function. | |
6004 | * | |
6005 | * Returns the number of bytes actually written (which may be smaller than | |
6006 | * the delta by which the device hand has changed due to alignment). | |
34dc7c2f | 6007 | */ |
d164b209 | 6008 | static uint64_t |
3a17a7a9 SK |
6009 | l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, |
6010 | boolean_t *headroom_boost) | |
34dc7c2f | 6011 | { |
2a432414 | 6012 | arc_buf_hdr_t *hdr, *hdr_prev, *head; |
ef56b078 AG |
6013 | uint64_t write_asize, write_sz, headroom, buf_compress_minsz, |
6014 | stats_size; | |
34dc7c2f | 6015 | void *buf_data; |
3a17a7a9 | 6016 | boolean_t full; |
34dc7c2f BB |
6017 | l2arc_write_callback_t *cb; |
6018 | zio_t *pio, *wzio; | |
3541dc6d | 6019 | uint64_t guid = spa_load_guid(spa); |
d6320ddb | 6020 | int try; |
3a17a7a9 | 6021 | const boolean_t do_headroom_boost = *headroom_boost; |
34dc7c2f | 6022 | |
34dc7c2f BB |
6023 | ASSERT(dev->l2ad_vdev != NULL); |
6024 | ||
3a17a7a9 SK |
6025 | /* Lower the flag now, we might want to raise it again later. */ |
6026 | *headroom_boost = B_FALSE; | |
6027 | ||
34dc7c2f | 6028 | pio = NULL; |
ef56b078 | 6029 | write_sz = write_asize = 0; |
34dc7c2f | 6030 | full = B_FALSE; |
b9541d6b | 6031 | head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); |
2a432414 | 6032 | head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; |
b9541d6b | 6033 | head->b_flags |= ARC_FLAG_HAS_L2HDR; |
34dc7c2f | 6034 | |
3a17a7a9 SK |
6035 | /* |
6036 | * We will want to try to compress buffers that are at least 2x the | |
6037 | * device sector size. | |
6038 | */ | |
6039 | buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; | |
6040 | ||
34dc7c2f BB |
6041 | /* |
6042 | * Copy buffers for L2ARC writing. | |
6043 | */ | |
d6320ddb | 6044 | for (try = 0; try <= 3; try++) { |
ca0bf58d | 6045 | multilist_sublist_t *mls = l2arc_sublist_lock(try); |
3a17a7a9 SK |
6046 | uint64_t passed_sz = 0; |
6047 | ||
b128c09f BB |
6048 | /* |
6049 | * L2ARC fast warmup. | |
6050 | * | |
6051 | * Until the ARC is warm and starts to evict, read from the | |
6052 | * head of the ARC lists rather than the tail. | |
6053 | */ | |
b128c09f | 6054 | if (arc_warm == B_FALSE) |
ca0bf58d | 6055 | hdr = multilist_sublist_head(mls); |
b128c09f | 6056 | else |
ca0bf58d | 6057 | hdr = multilist_sublist_tail(mls); |
b128c09f | 6058 | |
3a17a7a9 SK |
6059 | headroom = target_sz * l2arc_headroom; |
6060 | if (do_headroom_boost) | |
6061 | headroom = (headroom * l2arc_headroom_boost) / 100; | |
6062 | ||
2a432414 | 6063 | for (; hdr; hdr = hdr_prev) { |
3a17a7a9 SK |
6064 | kmutex_t *hash_lock; |
6065 | uint64_t buf_sz; | |
ef56b078 | 6066 | uint64_t buf_a_sz; |
3a17a7a9 | 6067 | |
b128c09f | 6068 | if (arc_warm == B_FALSE) |
ca0bf58d | 6069 | hdr_prev = multilist_sublist_next(mls, hdr); |
b128c09f | 6070 | else |
ca0bf58d | 6071 | hdr_prev = multilist_sublist_prev(mls, hdr); |
34dc7c2f | 6072 | |
2a432414 | 6073 | hash_lock = HDR_LOCK(hdr); |
3a17a7a9 | 6074 | if (!mutex_tryenter(hash_lock)) { |
34dc7c2f BB |
6075 | /* |
6076 | * Skip this buffer rather than waiting. | |
6077 | */ | |
6078 | continue; | |
6079 | } | |
6080 | ||
2a432414 | 6081 | passed_sz += hdr->b_size; |
34dc7c2f BB |
6082 | if (passed_sz > headroom) { |
6083 | /* | |
6084 | * Searched too far. | |
6085 | */ | |
6086 | mutex_exit(hash_lock); | |
6087 | break; | |
6088 | } | |
6089 | ||
2a432414 | 6090 | if (!l2arc_write_eligible(guid, hdr)) { |
34dc7c2f BB |
6091 | mutex_exit(hash_lock); |
6092 | continue; | |
6093 | } | |
6094 | ||
ef56b078 AG |
6095 | /* |
6096 | * Assume that the buffer is not going to be compressed | |
6097 | * and could take more space on disk because of a larger | |
6098 | * disk block size. | |
6099 | */ | |
6100 | buf_sz = hdr->b_size; | |
6101 | buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); | |
6102 | ||
6103 | if ((write_asize + buf_a_sz) > target_sz) { | |
34dc7c2f BB |
6104 | full = B_TRUE; |
6105 | mutex_exit(hash_lock); | |
6106 | break; | |
6107 | } | |
6108 | ||
34dc7c2f BB |
6109 | if (pio == NULL) { |
6110 | /* | |
6111 | * Insert a dummy header on the buflist so | |
6112 | * l2arc_write_done() can find where the | |
6113 | * write buffers begin without searching. | |
6114 | */ | |
ca0bf58d | 6115 | mutex_enter(&dev->l2ad_mtx); |
b9541d6b | 6116 | list_insert_head(&dev->l2ad_buflist, head); |
ca0bf58d | 6117 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f | 6118 | |
409dc1a5 | 6119 | cb = kmem_alloc(sizeof (l2arc_write_callback_t), |
79c76d5b | 6120 | KM_SLEEP); |
34dc7c2f BB |
6121 | cb->l2wcb_dev = dev; |
6122 | cb->l2wcb_head = head; | |
6123 | pio = zio_root(spa, l2arc_write_done, cb, | |
6124 | ZIO_FLAG_CANFAIL); | |
6125 | } | |
6126 | ||
6127 | /* | |
6128 | * Create and add a new L2ARC header. | |
6129 | */ | |
b9541d6b CW |
6130 | hdr->b_l2hdr.b_dev = dev; |
6131 | arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); | |
2a432414 | 6132 | hdr->b_flags |= ARC_FLAG_L2_WRITING; |
3a17a7a9 SK |
6133 | /* |
6134 | * Temporarily stash the data buffer in b_tmp_cdata. | |
6135 | * The subsequent write step will pick it up from | |
b9541d6b | 6136 | * there. This is because can't access b_l1hdr.b_buf |
3a17a7a9 SK |
6137 | * without holding the hash_lock, which we in turn |
6138 | * can't access without holding the ARC list locks | |
6139 | * (which we want to avoid during compression/writing) | |
6140 | */ | |
b9541d6b CW |
6141 | HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); |
6142 | hdr->b_l2hdr.b_asize = hdr->b_size; | |
6143 | hdr->b_l2hdr.b_hits = 0; | |
6144 | hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; | |
3a17a7a9 | 6145 | |
d962d5da PS |
6146 | /* |
6147 | * Explicitly set the b_daddr field to a known | |
6148 | * value which means "invalid address". This | |
6149 | * enables us to differentiate which stage of | |
6150 | * l2arc_write_buffers() the particular header | |
6151 | * is in (e.g. this loop, or the one below). | |
6152 | * ARC_FLAG_L2_WRITING is not enough to make | |
6153 | * this distinction, and we need to know in | |
6154 | * order to do proper l2arc vdev accounting in | |
6155 | * arc_release() and arc_hdr_destroy(). | |
6156 | * | |
6157 | * Note, we can't use a new flag to distinguish | |
6158 | * the two stages because we don't hold the | |
6159 | * header's hash_lock below, in the second stage | |
6160 | * of this function. Thus, we can't simply | |
6161 | * change the b_flags field to denote that the | |
6162 | * IO has been sent. We can change the b_daddr | |
6163 | * field of the L2 portion, though, since we'll | |
6164 | * be holding the l2ad_mtx; which is why we're | |
6165 | * using it to denote the header's state change. | |
6166 | */ | |
6167 | hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET; | |
b9541d6b | 6168 | hdr->b_flags |= ARC_FLAG_HAS_L2HDR; |
3a17a7a9 | 6169 | |
ca0bf58d | 6170 | mutex_enter(&dev->l2ad_mtx); |
b9541d6b | 6171 | list_insert_head(&dev->l2ad_buflist, hdr); |
ca0bf58d | 6172 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f BB |
6173 | |
6174 | /* | |
6175 | * Compute and store the buffer cksum before | |
6176 | * writing. On debug the cksum is verified first. | |
6177 | */ | |
b9541d6b CW |
6178 | arc_cksum_verify(hdr->b_l1hdr.b_buf); |
6179 | arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); | |
34dc7c2f BB |
6180 | |
6181 | mutex_exit(hash_lock); | |
6182 | ||
3a17a7a9 | 6183 | write_sz += buf_sz; |
ef56b078 | 6184 | write_asize += buf_a_sz; |
3a17a7a9 SK |
6185 | } |
6186 | ||
ca0bf58d | 6187 | multilist_sublist_unlock(mls); |
3a17a7a9 SK |
6188 | |
6189 | if (full == B_TRUE) | |
6190 | break; | |
6191 | } | |
6192 | ||
6193 | /* No buffers selected for writing? */ | |
6194 | if (pio == NULL) { | |
6195 | ASSERT0(write_sz); | |
b9541d6b CW |
6196 | ASSERT(!HDR_HAS_L1HDR(head)); |
6197 | kmem_cache_free(hdr_l2only_cache, head); | |
3a17a7a9 SK |
6198 | return (0); |
6199 | } | |
6200 | ||
ca0bf58d PS |
6201 | mutex_enter(&dev->l2ad_mtx); |
6202 | ||
ef56b078 AG |
6203 | /* |
6204 | * Note that elsewhere in this file arcstat_l2_asize | |
6205 | * and the used space on l2ad_vdev are updated using b_asize, | |
6206 | * which is not necessarily rounded up to the device block size. | |
6207 | * Too keep accounting consistent we do the same here as well: | |
6208 | * stats_size accumulates the sum of b_asize of the written buffers, | |
6209 | * while write_asize accumulates the sum of b_asize rounded up | |
6210 | * to the device block size. | |
6211 | * The latter sum is used only to validate the corectness of the code. | |
6212 | */ | |
6213 | stats_size = 0; | |
6214 | write_asize = 0; | |
6215 | ||
3a17a7a9 SK |
6216 | /* |
6217 | * Now start writing the buffers. We're starting at the write head | |
6218 | * and work backwards, retracing the course of the buffer selector | |
6219 | * loop above. | |
6220 | */ | |
b9541d6b CW |
6221 | for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; |
6222 | hdr = list_prev(&dev->l2ad_buflist, hdr)) { | |
3a17a7a9 SK |
6223 | uint64_t buf_sz; |
6224 | ||
ca0bf58d PS |
6225 | /* |
6226 | * We rely on the L1 portion of the header below, so | |
6227 | * it's invalid for this header to have been evicted out | |
6228 | * of the ghost cache, prior to being written out. The | |
6229 | * ARC_FLAG_L2_WRITING bit ensures this won't happen. | |
6230 | */ | |
6231 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
6232 | ||
3a17a7a9 SK |
6233 | /* |
6234 | * We shouldn't need to lock the buffer here, since we flagged | |
2a432414 GW |
6235 | * it as ARC_FLAG_L2_WRITING in the previous step, but we must |
6236 | * take care to only access its L2 cache parameters. In | |
b9541d6b | 6237 | * particular, hdr->l1hdr.b_buf may be invalid by now due to |
2a432414 | 6238 | * ARC eviction. |
3a17a7a9 | 6239 | */ |
b9541d6b | 6240 | hdr->b_l2hdr.b_daddr = dev->l2ad_hand; |
3a17a7a9 | 6241 | |
b9541d6b CW |
6242 | if ((!l2arc_nocompress && HDR_L2COMPRESS(hdr)) && |
6243 | hdr->b_l2hdr.b_asize >= buf_compress_minsz) { | |
6244 | if (l2arc_compress_buf(hdr)) { | |
3a17a7a9 SK |
6245 | /* |
6246 | * If compression succeeded, enable headroom | |
6247 | * boost on the next scan cycle. | |
6248 | */ | |
6249 | *headroom_boost = B_TRUE; | |
6250 | } | |
6251 | } | |
6252 | ||
6253 | /* | |
6254 | * Pick up the buffer data we had previously stashed away | |
6255 | * (and now potentially also compressed). | |
6256 | */ | |
b9541d6b CW |
6257 | buf_data = hdr->b_l1hdr.b_tmp_cdata; |
6258 | buf_sz = hdr->b_l2hdr.b_asize; | |
3a17a7a9 | 6259 | |
d962d5da PS |
6260 | /* |
6261 | * We need to do this regardless if buf_sz is zero or | |
6262 | * not, otherwise, when this l2hdr is evicted we'll | |
6263 | * remove a reference that was never added. | |
6264 | */ | |
6265 | (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr); | |
6266 | ||
3a17a7a9 SK |
6267 | /* Compression may have squashed the buffer to zero length. */ |
6268 | if (buf_sz != 0) { | |
ef56b078 | 6269 | uint64_t buf_a_sz; |
3a17a7a9 | 6270 | |
34dc7c2f BB |
6271 | wzio = zio_write_phys(pio, dev->l2ad_vdev, |
6272 | dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, | |
6273 | NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, | |
6274 | ZIO_FLAG_CANFAIL, B_FALSE); | |
6275 | ||
6276 | DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, | |
6277 | zio_t *, wzio); | |
6278 | (void) zio_nowait(wzio); | |
6279 | ||
ef56b078 | 6280 | stats_size += buf_sz; |
d962d5da | 6281 | |
b128c09f BB |
6282 | /* |
6283 | * Keep the clock hand suitably device-aligned. | |
6284 | */ | |
ef56b078 AG |
6285 | buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); |
6286 | write_asize += buf_a_sz; | |
6287 | dev->l2ad_hand += buf_a_sz; | |
34dc7c2f | 6288 | } |
34dc7c2f | 6289 | } |
34dc7c2f | 6290 | |
b9541d6b | 6291 | mutex_exit(&dev->l2ad_mtx); |
34dc7c2f | 6292 | |
3a17a7a9 | 6293 | ASSERT3U(write_asize, <=, target_sz); |
34dc7c2f | 6294 | ARCSTAT_BUMP(arcstat_l2_writes_sent); |
3a17a7a9 | 6295 | ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); |
34dc7c2f | 6296 | ARCSTAT_INCR(arcstat_l2_size, write_sz); |
ef56b078 AG |
6297 | ARCSTAT_INCR(arcstat_l2_asize, stats_size); |
6298 | vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0); | |
34dc7c2f BB |
6299 | |
6300 | /* | |
6301 | * Bump device hand to the device start if it is approaching the end. | |
6302 | * l2arc_evict() will already have evicted ahead for this case. | |
6303 | */ | |
b128c09f | 6304 | if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { |
34dc7c2f | 6305 | dev->l2ad_hand = dev->l2ad_start; |
34dc7c2f BB |
6306 | dev->l2ad_first = B_FALSE; |
6307 | } | |
6308 | ||
d164b209 | 6309 | dev->l2ad_writing = B_TRUE; |
34dc7c2f | 6310 | (void) zio_wait(pio); |
d164b209 BB |
6311 | dev->l2ad_writing = B_FALSE; |
6312 | ||
3a17a7a9 SK |
6313 | return (write_asize); |
6314 | } | |
6315 | ||
6316 | /* | |
6317 | * Compresses an L2ARC buffer. | |
b9541d6b | 6318 | * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its |
3a17a7a9 SK |
6319 | * size in l2hdr->b_asize. This routine tries to compress the data and |
6320 | * depending on the compression result there are three possible outcomes: | |
6321 | * *) The buffer was incompressible. The original l2hdr contents were left | |
6322 | * untouched and are ready for writing to an L2 device. | |
6323 | * *) The buffer was all-zeros, so there is no need to write it to an L2 | |
6324 | * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is | |
6325 | * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. | |
6326 | * *) Compression succeeded and b_tmp_cdata was replaced with a temporary | |
6327 | * data buffer which holds the compressed data to be written, and b_asize | |
6328 | * tells us how much data there is. b_compress is set to the appropriate | |
6329 | * compression algorithm. Once writing is done, invoke | |
6330 | * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. | |
6331 | * | |
6332 | * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the | |
6333 | * buffer was incompressible). | |
6334 | */ | |
6335 | static boolean_t | |
b9541d6b | 6336 | l2arc_compress_buf(arc_buf_hdr_t *hdr) |
3a17a7a9 SK |
6337 | { |
6338 | void *cdata; | |
9b67f605 | 6339 | size_t csize, len, rounded; |
b9541d6b | 6340 | l2arc_buf_hdr_t *l2hdr; |
3a17a7a9 | 6341 | |
b9541d6b CW |
6342 | ASSERT(HDR_HAS_L2HDR(hdr)); |
6343 | ||
6344 | l2hdr = &hdr->b_l2hdr; | |
6345 | ||
6346 | ASSERT(HDR_HAS_L1HDR(hdr)); | |
6347 | ASSERT(HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF); | |
6348 | ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); | |
3a17a7a9 SK |
6349 | |
6350 | len = l2hdr->b_asize; | |
6351 | cdata = zio_data_buf_alloc(len); | |
b9541d6b CW |
6352 | ASSERT3P(cdata, !=, NULL); |
6353 | csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata, | |
3a17a7a9 SK |
6354 | cdata, l2hdr->b_asize); |
6355 | ||
9b67f605 MA |
6356 | rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); |
6357 | if (rounded > csize) { | |
6358 | bzero((char *)cdata + csize, rounded - csize); | |
6359 | csize = rounded; | |
6360 | } | |
6361 | ||
3a17a7a9 SK |
6362 | if (csize == 0) { |
6363 | /* zero block, indicate that there's nothing to write */ | |
6364 | zio_data_buf_free(cdata, len); | |
b9541d6b | 6365 | HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_EMPTY); |
3a17a7a9 | 6366 | l2hdr->b_asize = 0; |
b9541d6b | 6367 | hdr->b_l1hdr.b_tmp_cdata = NULL; |
3a17a7a9 SK |
6368 | ARCSTAT_BUMP(arcstat_l2_compress_zeros); |
6369 | return (B_TRUE); | |
6370 | } else if (csize > 0 && csize < len) { | |
6371 | /* | |
6372 | * Compression succeeded, we'll keep the cdata around for | |
6373 | * writing and release it afterwards. | |
6374 | */ | |
b9541d6b | 6375 | HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_LZ4); |
3a17a7a9 | 6376 | l2hdr->b_asize = csize; |
b9541d6b | 6377 | hdr->b_l1hdr.b_tmp_cdata = cdata; |
3a17a7a9 SK |
6378 | ARCSTAT_BUMP(arcstat_l2_compress_successes); |
6379 | return (B_TRUE); | |
6380 | } else { | |
6381 | /* | |
6382 | * Compression failed, release the compressed buffer. | |
6383 | * l2hdr will be left unmodified. | |
6384 | */ | |
6385 | zio_data_buf_free(cdata, len); | |
6386 | ARCSTAT_BUMP(arcstat_l2_compress_failures); | |
6387 | return (B_FALSE); | |
6388 | } | |
6389 | } | |
6390 | ||
6391 | /* | |
6392 | * Decompresses a zio read back from an l2arc device. On success, the | |
6393 | * underlying zio's io_data buffer is overwritten by the uncompressed | |
6394 | * version. On decompression error (corrupt compressed stream), the | |
6395 | * zio->io_error value is set to signal an I/O error. | |
6396 | * | |
6397 | * Please note that the compressed data stream is not checksummed, so | |
6398 | * if the underlying device is experiencing data corruption, we may feed | |
6399 | * corrupt data to the decompressor, so the decompressor needs to be | |
6400 | * able to handle this situation (LZ4 does). | |
6401 | */ | |
6402 | static void | |
6403 | l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) | |
6404 | { | |
6405 | uint64_t csize; | |
6406 | void *cdata; | |
6407 | ||
6408 | ASSERT(L2ARC_IS_VALID_COMPRESS(c)); | |
6409 | ||
6410 | if (zio->io_error != 0) { | |
6411 | /* | |
6412 | * An io error has occured, just restore the original io | |
6413 | * size in preparation for a main pool read. | |
6414 | */ | |
6415 | zio->io_orig_size = zio->io_size = hdr->b_size; | |
6416 | return; | |
6417 | } | |
6418 | ||
6419 | if (c == ZIO_COMPRESS_EMPTY) { | |
6420 | /* | |
6421 | * An empty buffer results in a null zio, which means we | |
6422 | * need to fill its io_data after we're done restoring the | |
6423 | * buffer's contents. | |
6424 | */ | |
b9541d6b CW |
6425 | ASSERT(hdr->b_l1hdr.b_buf != NULL); |
6426 | bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); | |
6427 | zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; | |
3a17a7a9 SK |
6428 | } else { |
6429 | ASSERT(zio->io_data != NULL); | |
6430 | /* | |
6431 | * We copy the compressed data from the start of the arc buffer | |
6432 | * (the zio_read will have pulled in only what we need, the | |
6433 | * rest is garbage which we will overwrite at decompression) | |
6434 | * and then decompress back to the ARC data buffer. This way we | |
6435 | * can minimize copying by simply decompressing back over the | |
6436 | * original compressed data (rather than decompressing to an | |
6437 | * aux buffer and then copying back the uncompressed buffer, | |
6438 | * which is likely to be much larger). | |
6439 | */ | |
6440 | csize = zio->io_size; | |
6441 | cdata = zio_data_buf_alloc(csize); | |
6442 | bcopy(zio->io_data, cdata, csize); | |
6443 | if (zio_decompress_data(c, cdata, zio->io_data, csize, | |
6444 | hdr->b_size) != 0) | |
2e528b49 | 6445 | zio->io_error = SET_ERROR(EIO); |
3a17a7a9 SK |
6446 | zio_data_buf_free(cdata, csize); |
6447 | } | |
6448 | ||
6449 | /* Restore the expected uncompressed IO size. */ | |
6450 | zio->io_orig_size = zio->io_size = hdr->b_size; | |
6451 | } | |
6452 | ||
6453 | /* | |
6454 | * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. | |
6455 | * This buffer serves as a temporary holder of compressed data while | |
6456 | * the buffer entry is being written to an l2arc device. Once that is | |
6457 | * done, we can dispose of it. | |
6458 | */ | |
6459 | static void | |
2a432414 | 6460 | l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) |
3a17a7a9 | 6461 | { |
ca0bf58d PS |
6462 | enum zio_compress comp = HDR_GET_COMPRESS(hdr); |
6463 | ||
b9541d6b | 6464 | ASSERT(HDR_HAS_L1HDR(hdr)); |
ca0bf58d PS |
6465 | ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp)); |
6466 | ||
6467 | if (comp == ZIO_COMPRESS_OFF) { | |
6468 | /* | |
6469 | * In this case, b_tmp_cdata points to the same buffer | |
6470 | * as the arc_buf_t's b_data field. We don't want to | |
6471 | * free it, since the arc_buf_t will handle that. | |
6472 | */ | |
6473 | hdr->b_l1hdr.b_tmp_cdata = NULL; | |
6474 | } else if (comp == ZIO_COMPRESS_EMPTY) { | |
6475 | /* | |
6476 | * In this case, b_tmp_cdata was compressed to an empty | |
6477 | * buffer, thus there's nothing to free and b_tmp_cdata | |
6478 | * should have been set to NULL in l2arc_write_buffers(). | |
6479 | */ | |
6480 | ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); | |
6481 | } else { | |
3a17a7a9 SK |
6482 | /* |
6483 | * If the data was compressed, then we've allocated a | |
6484 | * temporary buffer for it, so now we need to release it. | |
6485 | */ | |
b9541d6b CW |
6486 | ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); |
6487 | zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, | |
6488 | hdr->b_size); | |
ca0bf58d | 6489 | hdr->b_l1hdr.b_tmp_cdata = NULL; |
3a17a7a9 | 6490 | } |
ca0bf58d | 6491 | |
34dc7c2f BB |
6492 | } |
6493 | ||
6494 | /* | |
6495 | * This thread feeds the L2ARC at regular intervals. This is the beating | |
6496 | * heart of the L2ARC. | |
6497 | */ | |
6498 | static void | |
6499 | l2arc_feed_thread(void) | |
6500 | { | |
6501 | callb_cpr_t cpr; | |
6502 | l2arc_dev_t *dev; | |
6503 | spa_t *spa; | |
d164b209 | 6504 | uint64_t size, wrote; |
428870ff | 6505 | clock_t begin, next = ddi_get_lbolt(); |
3a17a7a9 | 6506 | boolean_t headroom_boost = B_FALSE; |
40d06e3c | 6507 | fstrans_cookie_t cookie; |
34dc7c2f BB |
6508 | |
6509 | CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); | |
6510 | ||
6511 | mutex_enter(&l2arc_feed_thr_lock); | |
6512 | ||
40d06e3c | 6513 | cookie = spl_fstrans_mark(); |
34dc7c2f | 6514 | while (l2arc_thread_exit == 0) { |
34dc7c2f | 6515 | CALLB_CPR_SAFE_BEGIN(&cpr); |
b64ccd6c | 6516 | (void) cv_timedwait_sig(&l2arc_feed_thr_cv, |
5b63b3eb | 6517 | &l2arc_feed_thr_lock, next); |
34dc7c2f | 6518 | CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); |
428870ff | 6519 | next = ddi_get_lbolt() + hz; |
34dc7c2f BB |
6520 | |
6521 | /* | |
b128c09f | 6522 | * Quick check for L2ARC devices. |
34dc7c2f BB |
6523 | */ |
6524 | mutex_enter(&l2arc_dev_mtx); | |
6525 | if (l2arc_ndev == 0) { | |
6526 | mutex_exit(&l2arc_dev_mtx); | |
6527 | continue; | |
6528 | } | |
b128c09f | 6529 | mutex_exit(&l2arc_dev_mtx); |
428870ff | 6530 | begin = ddi_get_lbolt(); |
34dc7c2f BB |
6531 | |
6532 | /* | |
b128c09f BB |
6533 | * This selects the next l2arc device to write to, and in |
6534 | * doing so the next spa to feed from: dev->l2ad_spa. This | |
6535 | * will return NULL if there are now no l2arc devices or if | |
6536 | * they are all faulted. | |
6537 | * | |
6538 | * If a device is returned, its spa's config lock is also | |
6539 | * held to prevent device removal. l2arc_dev_get_next() | |
6540 | * will grab and release l2arc_dev_mtx. | |
34dc7c2f | 6541 | */ |
b128c09f | 6542 | if ((dev = l2arc_dev_get_next()) == NULL) |
34dc7c2f | 6543 | continue; |
b128c09f BB |
6544 | |
6545 | spa = dev->l2ad_spa; | |
6546 | ASSERT(spa != NULL); | |
34dc7c2f | 6547 | |
572e2857 BB |
6548 | /* |
6549 | * If the pool is read-only then force the feed thread to | |
6550 | * sleep a little longer. | |
6551 | */ | |
6552 | if (!spa_writeable(spa)) { | |
6553 | next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; | |
6554 | spa_config_exit(spa, SCL_L2ARC, dev); | |
6555 | continue; | |
6556 | } | |
6557 | ||
34dc7c2f | 6558 | /* |
b128c09f | 6559 | * Avoid contributing to memory pressure. |
34dc7c2f | 6560 | */ |
ca67b33a | 6561 | if (arc_reclaim_needed()) { |
b128c09f BB |
6562 | ARCSTAT_BUMP(arcstat_l2_abort_lowmem); |
6563 | spa_config_exit(spa, SCL_L2ARC, dev); | |
34dc7c2f BB |
6564 | continue; |
6565 | } | |
b128c09f | 6566 | |
34dc7c2f BB |
6567 | ARCSTAT_BUMP(arcstat_l2_feeds); |
6568 | ||
3a17a7a9 | 6569 | size = l2arc_write_size(); |
b128c09f | 6570 | |
34dc7c2f BB |
6571 | /* |
6572 | * Evict L2ARC buffers that will be overwritten. | |
6573 | */ | |
b128c09f | 6574 | l2arc_evict(dev, size, B_FALSE); |
34dc7c2f BB |
6575 | |
6576 | /* | |
6577 | * Write ARC buffers. | |
6578 | */ | |
3a17a7a9 | 6579 | wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); |
d164b209 BB |
6580 | |
6581 | /* | |
6582 | * Calculate interval between writes. | |
6583 | */ | |
6584 | next = l2arc_write_interval(begin, size, wrote); | |
b128c09f | 6585 | spa_config_exit(spa, SCL_L2ARC, dev); |
34dc7c2f | 6586 | } |
40d06e3c | 6587 | spl_fstrans_unmark(cookie); |
34dc7c2f BB |
6588 | |
6589 | l2arc_thread_exit = 0; | |
6590 | cv_broadcast(&l2arc_feed_thr_cv); | |
6591 | CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ | |
6592 | thread_exit(); | |
6593 | } | |
6594 | ||
b128c09f BB |
6595 | boolean_t |
6596 | l2arc_vdev_present(vdev_t *vd) | |
6597 | { | |
6598 | l2arc_dev_t *dev; | |
6599 | ||
6600 | mutex_enter(&l2arc_dev_mtx); | |
6601 | for (dev = list_head(l2arc_dev_list); dev != NULL; | |
6602 | dev = list_next(l2arc_dev_list, dev)) { | |
6603 | if (dev->l2ad_vdev == vd) | |
6604 | break; | |
6605 | } | |
6606 | mutex_exit(&l2arc_dev_mtx); | |
6607 | ||
6608 | return (dev != NULL); | |
6609 | } | |
6610 | ||
34dc7c2f BB |
6611 | /* |
6612 | * Add a vdev for use by the L2ARC. By this point the spa has already | |
6613 | * validated the vdev and opened it. | |
6614 | */ | |
6615 | void | |
9babb374 | 6616 | l2arc_add_vdev(spa_t *spa, vdev_t *vd) |
34dc7c2f BB |
6617 | { |
6618 | l2arc_dev_t *adddev; | |
6619 | ||
b128c09f BB |
6620 | ASSERT(!l2arc_vdev_present(vd)); |
6621 | ||
34dc7c2f BB |
6622 | /* |
6623 | * Create a new l2arc device entry. | |
6624 | */ | |
6625 | adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); | |
6626 | adddev->l2ad_spa = spa; | |
6627 | adddev->l2ad_vdev = vd; | |
9babb374 BB |
6628 | adddev->l2ad_start = VDEV_LABEL_START_SIZE; |
6629 | adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); | |
34dc7c2f | 6630 | adddev->l2ad_hand = adddev->l2ad_start; |
34dc7c2f | 6631 | adddev->l2ad_first = B_TRUE; |
d164b209 | 6632 | adddev->l2ad_writing = B_FALSE; |
98f72a53 | 6633 | list_link_init(&adddev->l2ad_node); |
34dc7c2f | 6634 | |
b9541d6b | 6635 | mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f BB |
6636 | /* |
6637 | * This is a list of all ARC buffers that are still valid on the | |
6638 | * device. | |
6639 | */ | |
b9541d6b CW |
6640 | list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), |
6641 | offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); | |
34dc7c2f | 6642 | |
428870ff | 6643 | vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); |
d962d5da | 6644 | refcount_create(&adddev->l2ad_alloc); |
34dc7c2f BB |
6645 | |
6646 | /* | |
6647 | * Add device to global list | |
6648 | */ | |
6649 | mutex_enter(&l2arc_dev_mtx); | |
6650 | list_insert_head(l2arc_dev_list, adddev); | |
6651 | atomic_inc_64(&l2arc_ndev); | |
6652 | mutex_exit(&l2arc_dev_mtx); | |
6653 | } | |
6654 | ||
6655 | /* | |
6656 | * Remove a vdev from the L2ARC. | |
6657 | */ | |
6658 | void | |
6659 | l2arc_remove_vdev(vdev_t *vd) | |
6660 | { | |
6661 | l2arc_dev_t *dev, *nextdev, *remdev = NULL; | |
6662 | ||
34dc7c2f BB |
6663 | /* |
6664 | * Find the device by vdev | |
6665 | */ | |
6666 | mutex_enter(&l2arc_dev_mtx); | |
6667 | for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { | |
6668 | nextdev = list_next(l2arc_dev_list, dev); | |
6669 | if (vd == dev->l2ad_vdev) { | |
6670 | remdev = dev; | |
6671 | break; | |
6672 | } | |
6673 | } | |
6674 | ASSERT(remdev != NULL); | |
6675 | ||
6676 | /* | |
6677 | * Remove device from global list | |
6678 | */ | |
6679 | list_remove(l2arc_dev_list, remdev); | |
6680 | l2arc_dev_last = NULL; /* may have been invalidated */ | |
b128c09f BB |
6681 | atomic_dec_64(&l2arc_ndev); |
6682 | mutex_exit(&l2arc_dev_mtx); | |
34dc7c2f BB |
6683 | |
6684 | /* | |
6685 | * Clear all buflists and ARC references. L2ARC device flush. | |
6686 | */ | |
6687 | l2arc_evict(remdev, 0, B_TRUE); | |
b9541d6b CW |
6688 | list_destroy(&remdev->l2ad_buflist); |
6689 | mutex_destroy(&remdev->l2ad_mtx); | |
d962d5da | 6690 | refcount_destroy(&remdev->l2ad_alloc); |
34dc7c2f | 6691 | kmem_free(remdev, sizeof (l2arc_dev_t)); |
34dc7c2f BB |
6692 | } |
6693 | ||
6694 | void | |
b128c09f | 6695 | l2arc_init(void) |
34dc7c2f BB |
6696 | { |
6697 | l2arc_thread_exit = 0; | |
6698 | l2arc_ndev = 0; | |
6699 | l2arc_writes_sent = 0; | |
6700 | l2arc_writes_done = 0; | |
6701 | ||
6702 | mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); | |
6703 | cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); | |
6704 | mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); | |
34dc7c2f BB |
6705 | mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); |
6706 | ||
6707 | l2arc_dev_list = &L2ARC_dev_list; | |
6708 | l2arc_free_on_write = &L2ARC_free_on_write; | |
6709 | list_create(l2arc_dev_list, sizeof (l2arc_dev_t), | |
6710 | offsetof(l2arc_dev_t, l2ad_node)); | |
6711 | list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), | |
6712 | offsetof(l2arc_data_free_t, l2df_list_node)); | |
34dc7c2f BB |
6713 | } |
6714 | ||
6715 | void | |
b128c09f | 6716 | l2arc_fini(void) |
34dc7c2f | 6717 | { |
b128c09f BB |
6718 | /* |
6719 | * This is called from dmu_fini(), which is called from spa_fini(); | |
6720 | * Because of this, we can assume that all l2arc devices have | |
6721 | * already been removed when the pools themselves were removed. | |
6722 | */ | |
6723 | ||
6724 | l2arc_do_free_on_write(); | |
34dc7c2f BB |
6725 | |
6726 | mutex_destroy(&l2arc_feed_thr_lock); | |
6727 | cv_destroy(&l2arc_feed_thr_cv); | |
6728 | mutex_destroy(&l2arc_dev_mtx); | |
34dc7c2f BB |
6729 | mutex_destroy(&l2arc_free_on_write_mtx); |
6730 | ||
6731 | list_destroy(l2arc_dev_list); | |
6732 | list_destroy(l2arc_free_on_write); | |
6733 | } | |
b128c09f BB |
6734 | |
6735 | void | |
6736 | l2arc_start(void) | |
6737 | { | |
fb5f0bc8 | 6738 | if (!(spa_mode_global & FWRITE)) |
b128c09f BB |
6739 | return; |
6740 | ||
6741 | (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, | |
6742 | TS_RUN, minclsyspri); | |
6743 | } | |
6744 | ||
6745 | void | |
6746 | l2arc_stop(void) | |
6747 | { | |
fb5f0bc8 | 6748 | if (!(spa_mode_global & FWRITE)) |
b128c09f BB |
6749 | return; |
6750 | ||
6751 | mutex_enter(&l2arc_feed_thr_lock); | |
6752 | cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ | |
6753 | l2arc_thread_exit = 1; | |
6754 | while (l2arc_thread_exit != 0) | |
6755 | cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); | |
6756 | mutex_exit(&l2arc_feed_thr_lock); | |
6757 | } | |
c28b2279 BB |
6758 | |
6759 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
0f699108 AZ |
6760 | EXPORT_SYMBOL(arc_buf_size); |
6761 | EXPORT_SYMBOL(arc_write); | |
c28b2279 BB |
6762 | EXPORT_SYMBOL(arc_read); |
6763 | EXPORT_SYMBOL(arc_buf_remove_ref); | |
e0b0ca98 | 6764 | EXPORT_SYMBOL(arc_buf_info); |
c28b2279 | 6765 | EXPORT_SYMBOL(arc_getbuf_func); |
ab26409d BB |
6766 | EXPORT_SYMBOL(arc_add_prune_callback); |
6767 | EXPORT_SYMBOL(arc_remove_prune_callback); | |
c28b2279 | 6768 | |
bce45ec9 | 6769 | module_param(zfs_arc_min, ulong, 0644); |
c409e464 | 6770 | MODULE_PARM_DESC(zfs_arc_min, "Min arc size"); |
c28b2279 | 6771 | |
bce45ec9 | 6772 | module_param(zfs_arc_max, ulong, 0644); |
c409e464 | 6773 | MODULE_PARM_DESC(zfs_arc_max, "Max arc size"); |
c28b2279 | 6774 | |
bce45ec9 | 6775 | module_param(zfs_arc_meta_limit, ulong, 0644); |
c28b2279 | 6776 | MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size"); |
6a8f9b6b | 6777 | |
ca0bf58d PS |
6778 | module_param(zfs_arc_meta_min, ulong, 0644); |
6779 | MODULE_PARM_DESC(zfs_arc_meta_min, "Min arc metadata"); | |
6780 | ||
bce45ec9 | 6781 | module_param(zfs_arc_meta_prune, int, 0644); |
2cbb06b5 | 6782 | MODULE_PARM_DESC(zfs_arc_meta_prune, "Meta objects to scan for prune"); |
c409e464 | 6783 | |
ca67b33a | 6784 | module_param(zfs_arc_meta_adjust_restarts, int, 0644); |
bc888666 BB |
6785 | MODULE_PARM_DESC(zfs_arc_meta_adjust_restarts, |
6786 | "Limit number of restarts in arc_adjust_meta"); | |
6787 | ||
f6046738 BB |
6788 | module_param(zfs_arc_meta_strategy, int, 0644); |
6789 | MODULE_PARM_DESC(zfs_arc_meta_strategy, "Meta reclaim strategy"); | |
6790 | ||
bce45ec9 | 6791 | module_param(zfs_arc_grow_retry, int, 0644); |
c409e464 BB |
6792 | MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size"); |
6793 | ||
89c8cac4 PS |
6794 | module_param(zfs_arc_p_aggressive_disable, int, 0644); |
6795 | MODULE_PARM_DESC(zfs_arc_p_aggressive_disable, "disable aggressive arc_p grow"); | |
6796 | ||
62422785 PS |
6797 | module_param(zfs_arc_p_dampener_disable, int, 0644); |
6798 | MODULE_PARM_DESC(zfs_arc_p_dampener_disable, "disable arc_p adapt dampener"); | |
6799 | ||
bce45ec9 | 6800 | module_param(zfs_arc_shrink_shift, int, 0644); |
c409e464 BB |
6801 | MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)"); |
6802 | ||
1f7c30df BB |
6803 | module_param(zfs_disable_dup_eviction, int, 0644); |
6804 | MODULE_PARM_DESC(zfs_disable_dup_eviction, "disable duplicate buffer eviction"); | |
6805 | ||
49ddb315 MA |
6806 | module_param(zfs_arc_average_blocksize, int, 0444); |
6807 | MODULE_PARM_DESC(zfs_arc_average_blocksize, "Target average block size"); | |
6808 | ||
0c5493d4 BB |
6809 | module_param(zfs_arc_memory_throttle_disable, int, 0644); |
6810 | MODULE_PARM_DESC(zfs_arc_memory_throttle_disable, "disable memory throttle"); | |
6811 | ||
bce45ec9 BB |
6812 | module_param(zfs_arc_min_prefetch_lifespan, int, 0644); |
6813 | MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan, "Min life of prefetch block"); | |
6814 | ||
ca0bf58d PS |
6815 | module_param(zfs_arc_num_sublists_per_state, int, 0644); |
6816 | MODULE_PARM_DESC(zfs_arc_num_sublists_per_state, | |
6817 | "Number of sublists used in each of the ARC state lists"); | |
6818 | ||
bce45ec9 | 6819 | module_param(l2arc_write_max, ulong, 0644); |
abd8610c BB |
6820 | MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval"); |
6821 | ||
bce45ec9 | 6822 | module_param(l2arc_write_boost, ulong, 0644); |
abd8610c BB |
6823 | MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup"); |
6824 | ||
bce45ec9 | 6825 | module_param(l2arc_headroom, ulong, 0644); |
abd8610c BB |
6826 | MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache"); |
6827 | ||
3a17a7a9 SK |
6828 | module_param(l2arc_headroom_boost, ulong, 0644); |
6829 | MODULE_PARM_DESC(l2arc_headroom_boost, "Compressed l2arc_headroom multiplier"); | |
6830 | ||
bce45ec9 | 6831 | module_param(l2arc_feed_secs, ulong, 0644); |
abd8610c BB |
6832 | MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing"); |
6833 | ||
bce45ec9 | 6834 | module_param(l2arc_feed_min_ms, ulong, 0644); |
abd8610c BB |
6835 | MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds"); |
6836 | ||
bce45ec9 | 6837 | module_param(l2arc_noprefetch, int, 0644); |
abd8610c BB |
6838 | MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers"); |
6839 | ||
3a17a7a9 SK |
6840 | module_param(l2arc_nocompress, int, 0644); |
6841 | MODULE_PARM_DESC(l2arc_nocompress, "Skip compressing L2ARC buffers"); | |
6842 | ||
bce45ec9 | 6843 | module_param(l2arc_feed_again, int, 0644); |
abd8610c BB |
6844 | MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup"); |
6845 | ||
bce45ec9 | 6846 | module_param(l2arc_norw, int, 0644); |
abd8610c BB |
6847 | MODULE_PARM_DESC(l2arc_norw, "No reads during writes"); |
6848 | ||
c28b2279 | 6849 | #endif |