]>
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. |
2e528b49 | 23 | * Copyright (c) 2013 by Delphix. All rights reserved. |
3a17a7a9 | 24 | * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. |
0ed212dc | 25 | * Copyright 2013 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 | |
107 | * protected from simultaneous callbacks from arc_buf_evict() | |
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 | * | |
121 | * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: | |
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> |
34dc7c2f BB |
138 | #ifdef _KERNEL |
139 | #include <sys/vmsystm.h> | |
140 | #include <vm/anon.h> | |
141 | #include <sys/fs/swapnode.h> | |
ab26409d | 142 | #include <sys/zpl.h> |
34dc7c2f BB |
143 | #endif |
144 | #include <sys/callb.h> | |
145 | #include <sys/kstat.h> | |
570827e1 | 146 | #include <sys/dmu_tx.h> |
428870ff | 147 | #include <zfs_fletcher.h> |
34dc7c2f | 148 | |
498877ba MA |
149 | #ifndef _KERNEL |
150 | /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ | |
151 | boolean_t arc_watch = B_FALSE; | |
152 | #endif | |
153 | ||
34dc7c2f BB |
154 | static kmutex_t arc_reclaim_thr_lock; |
155 | static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ | |
156 | static uint8_t arc_thread_exit; | |
157 | ||
ab26409d | 158 | /* number of bytes to prune from caches when at arc_meta_limit is reached */ |
bce45ec9 | 159 | int zfs_arc_meta_prune = 1048576; |
34dc7c2f BB |
160 | |
161 | typedef enum arc_reclaim_strategy { | |
162 | ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ | |
163 | ARC_RECLAIM_CONS /* Conservative reclaim strategy */ | |
164 | } arc_reclaim_strategy_t; | |
165 | ||
e8b96c60 MA |
166 | /* |
167 | * The number of iterations through arc_evict_*() before we | |
168 | * drop & reacquire the lock. | |
169 | */ | |
170 | int arc_evict_iterations = 100; | |
171 | ||
34dc7c2f | 172 | /* number of seconds before growing cache again */ |
bce45ec9 | 173 | int zfs_arc_grow_retry = 5; |
34dc7c2f | 174 | |
89c8cac4 PS |
175 | /* disable anon data aggressively growing arc_p */ |
176 | int zfs_arc_p_aggressive_disable = 1; | |
177 | ||
62422785 PS |
178 | /* disable arc_p adapt dampener in arc_adapt */ |
179 | int zfs_arc_p_dampener_disable = 1; | |
180 | ||
d164b209 | 181 | /* log2(fraction of arc to reclaim) */ |
bce45ec9 | 182 | int zfs_arc_shrink_shift = 5; |
d164b209 | 183 | |
34dc7c2f BB |
184 | /* |
185 | * minimum lifespan of a prefetch block in clock ticks | |
186 | * (initialized in arc_init()) | |
187 | */ | |
bce45ec9 BB |
188 | int zfs_arc_min_prefetch_lifespan = HZ; |
189 | ||
190 | /* disable arc proactive arc throttle due to low memory */ | |
191 | int zfs_arc_memory_throttle_disable = 1; | |
192 | ||
193 | /* disable duplicate buffer eviction */ | |
194 | int zfs_disable_dup_eviction = 0; | |
34dc7c2f | 195 | |
e8b96c60 MA |
196 | /* |
197 | * If this percent of memory is free, don't throttle. | |
198 | */ | |
199 | int arc_lotsfree_percent = 10; | |
200 | ||
34dc7c2f BB |
201 | static int arc_dead; |
202 | ||
bce45ec9 BB |
203 | /* expiration time for arc_no_grow */ |
204 | static clock_t arc_grow_time = 0; | |
205 | ||
b128c09f BB |
206 | /* |
207 | * The arc has filled available memory and has now warmed up. | |
208 | */ | |
209 | static boolean_t arc_warm; | |
210 | ||
34dc7c2f BB |
211 | /* |
212 | * These tunables are for performance analysis. | |
213 | */ | |
c28b2279 BB |
214 | unsigned long zfs_arc_max = 0; |
215 | unsigned long zfs_arc_min = 0; | |
216 | unsigned long zfs_arc_meta_limit = 0; | |
34dc7c2f BB |
217 | |
218 | /* | |
219 | * Note that buffers can be in one of 6 states: | |
220 | * ARC_anon - anonymous (discussed below) | |
221 | * ARC_mru - recently used, currently cached | |
222 | * ARC_mru_ghost - recentely used, no longer in cache | |
223 | * ARC_mfu - frequently used, currently cached | |
224 | * ARC_mfu_ghost - frequently used, no longer in cache | |
225 | * ARC_l2c_only - exists in L2ARC but not other states | |
226 | * When there are no active references to the buffer, they are | |
227 | * are linked onto a list in one of these arc states. These are | |
228 | * the only buffers that can be evicted or deleted. Within each | |
229 | * state there are multiple lists, one for meta-data and one for | |
230 | * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, | |
231 | * etc.) is tracked separately so that it can be managed more | |
232 | * explicitly: favored over data, limited explicitly. | |
233 | * | |
234 | * Anonymous buffers are buffers that are not associated with | |
235 | * a DVA. These are buffers that hold dirty block copies | |
236 | * before they are written to stable storage. By definition, | |
237 | * they are "ref'd" and are considered part of arc_mru | |
238 | * that cannot be freed. Generally, they will aquire a DVA | |
239 | * as they are written and migrate onto the arc_mru list. | |
240 | * | |
241 | * The ARC_l2c_only state is for buffers that are in the second | |
242 | * level ARC but no longer in any of the ARC_m* lists. The second | |
243 | * level ARC itself may also contain buffers that are in any of | |
244 | * the ARC_m* states - meaning that a buffer can exist in two | |
245 | * places. The reason for the ARC_l2c_only state is to keep the | |
246 | * buffer header in the hash table, so that reads that hit the | |
247 | * second level ARC benefit from these fast lookups. | |
248 | */ | |
249 | ||
250 | typedef struct arc_state { | |
251 | list_t arcs_list[ARC_BUFC_NUMTYPES]; /* list of evictable buffers */ | |
252 | uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ | |
253 | uint64_t arcs_size; /* total amount of data in this state */ | |
254 | kmutex_t arcs_mtx; | |
e0b0ca98 | 255 | arc_state_type_t arcs_state; |
34dc7c2f BB |
256 | } arc_state_t; |
257 | ||
258 | /* The 6 states: */ | |
259 | static arc_state_t ARC_anon; | |
260 | static arc_state_t ARC_mru; | |
261 | static arc_state_t ARC_mru_ghost; | |
262 | static arc_state_t ARC_mfu; | |
263 | static arc_state_t ARC_mfu_ghost; | |
264 | static arc_state_t ARC_l2c_only; | |
265 | ||
266 | typedef struct arc_stats { | |
267 | kstat_named_t arcstat_hits; | |
268 | kstat_named_t arcstat_misses; | |
269 | kstat_named_t arcstat_demand_data_hits; | |
270 | kstat_named_t arcstat_demand_data_misses; | |
271 | kstat_named_t arcstat_demand_metadata_hits; | |
272 | kstat_named_t arcstat_demand_metadata_misses; | |
273 | kstat_named_t arcstat_prefetch_data_hits; | |
274 | kstat_named_t arcstat_prefetch_data_misses; | |
275 | kstat_named_t arcstat_prefetch_metadata_hits; | |
276 | kstat_named_t arcstat_prefetch_metadata_misses; | |
277 | kstat_named_t arcstat_mru_hits; | |
278 | kstat_named_t arcstat_mru_ghost_hits; | |
279 | kstat_named_t arcstat_mfu_hits; | |
280 | kstat_named_t arcstat_mfu_ghost_hits; | |
281 | kstat_named_t arcstat_deleted; | |
282 | kstat_named_t arcstat_recycle_miss; | |
e49f1e20 WA |
283 | /* |
284 | * Number of buffers that could not be evicted because the hash lock | |
285 | * was held by another thread. The lock may not necessarily be held | |
286 | * by something using the same buffer, since hash locks are shared | |
287 | * by multiple buffers. | |
288 | */ | |
34dc7c2f | 289 | kstat_named_t arcstat_mutex_miss; |
e49f1e20 WA |
290 | /* |
291 | * Number of buffers skipped because they have I/O in progress, are | |
292 | * indrect prefetch buffers that have not lived long enough, or are | |
293 | * not from the spa we're trying to evict from. | |
294 | */ | |
34dc7c2f | 295 | kstat_named_t arcstat_evict_skip; |
428870ff BB |
296 | kstat_named_t arcstat_evict_l2_cached; |
297 | kstat_named_t arcstat_evict_l2_eligible; | |
298 | kstat_named_t arcstat_evict_l2_ineligible; | |
34dc7c2f BB |
299 | kstat_named_t arcstat_hash_elements; |
300 | kstat_named_t arcstat_hash_elements_max; | |
301 | kstat_named_t arcstat_hash_collisions; | |
302 | kstat_named_t arcstat_hash_chains; | |
303 | kstat_named_t arcstat_hash_chain_max; | |
304 | kstat_named_t arcstat_p; | |
305 | kstat_named_t arcstat_c; | |
306 | kstat_named_t arcstat_c_min; | |
307 | kstat_named_t arcstat_c_max; | |
308 | kstat_named_t arcstat_size; | |
309 | kstat_named_t arcstat_hdr_size; | |
d164b209 | 310 | kstat_named_t arcstat_data_size; |
cc7f677c | 311 | kstat_named_t arcstat_meta_size; |
d164b209 | 312 | kstat_named_t arcstat_other_size; |
13be560d BB |
313 | kstat_named_t arcstat_anon_size; |
314 | kstat_named_t arcstat_anon_evict_data; | |
315 | kstat_named_t arcstat_anon_evict_metadata; | |
316 | kstat_named_t arcstat_mru_size; | |
317 | kstat_named_t arcstat_mru_evict_data; | |
318 | kstat_named_t arcstat_mru_evict_metadata; | |
319 | kstat_named_t arcstat_mru_ghost_size; | |
320 | kstat_named_t arcstat_mru_ghost_evict_data; | |
321 | kstat_named_t arcstat_mru_ghost_evict_metadata; | |
322 | kstat_named_t arcstat_mfu_size; | |
323 | kstat_named_t arcstat_mfu_evict_data; | |
324 | kstat_named_t arcstat_mfu_evict_metadata; | |
325 | kstat_named_t arcstat_mfu_ghost_size; | |
326 | kstat_named_t arcstat_mfu_ghost_evict_data; | |
327 | kstat_named_t arcstat_mfu_ghost_evict_metadata; | |
34dc7c2f BB |
328 | kstat_named_t arcstat_l2_hits; |
329 | kstat_named_t arcstat_l2_misses; | |
330 | kstat_named_t arcstat_l2_feeds; | |
331 | kstat_named_t arcstat_l2_rw_clash; | |
d164b209 BB |
332 | kstat_named_t arcstat_l2_read_bytes; |
333 | kstat_named_t arcstat_l2_write_bytes; | |
34dc7c2f BB |
334 | kstat_named_t arcstat_l2_writes_sent; |
335 | kstat_named_t arcstat_l2_writes_done; | |
336 | kstat_named_t arcstat_l2_writes_error; | |
337 | kstat_named_t arcstat_l2_writes_hdr_miss; | |
338 | kstat_named_t arcstat_l2_evict_lock_retry; | |
339 | kstat_named_t arcstat_l2_evict_reading; | |
340 | kstat_named_t arcstat_l2_free_on_write; | |
341 | kstat_named_t arcstat_l2_abort_lowmem; | |
342 | kstat_named_t arcstat_l2_cksum_bad; | |
343 | kstat_named_t arcstat_l2_io_error; | |
344 | kstat_named_t arcstat_l2_size; | |
3a17a7a9 | 345 | kstat_named_t arcstat_l2_asize; |
34dc7c2f | 346 | kstat_named_t arcstat_l2_hdr_size; |
3a17a7a9 SK |
347 | kstat_named_t arcstat_l2_compress_successes; |
348 | kstat_named_t arcstat_l2_compress_zeros; | |
349 | kstat_named_t arcstat_l2_compress_failures; | |
34dc7c2f | 350 | kstat_named_t arcstat_memory_throttle_count; |
1eb5bfa3 GW |
351 | kstat_named_t arcstat_duplicate_buffers; |
352 | kstat_named_t arcstat_duplicate_buffers_size; | |
353 | kstat_named_t arcstat_duplicate_reads; | |
7cb67b45 BB |
354 | kstat_named_t arcstat_memory_direct_count; |
355 | kstat_named_t arcstat_memory_indirect_count; | |
1834f2d8 BB |
356 | kstat_named_t arcstat_no_grow; |
357 | kstat_named_t arcstat_tempreserve; | |
358 | kstat_named_t arcstat_loaned_bytes; | |
ab26409d | 359 | kstat_named_t arcstat_prune; |
1834f2d8 BB |
360 | kstat_named_t arcstat_meta_used; |
361 | kstat_named_t arcstat_meta_limit; | |
362 | kstat_named_t arcstat_meta_max; | |
34dc7c2f BB |
363 | } arc_stats_t; |
364 | ||
365 | static arc_stats_t arc_stats = { | |
366 | { "hits", KSTAT_DATA_UINT64 }, | |
367 | { "misses", KSTAT_DATA_UINT64 }, | |
368 | { "demand_data_hits", KSTAT_DATA_UINT64 }, | |
369 | { "demand_data_misses", KSTAT_DATA_UINT64 }, | |
370 | { "demand_metadata_hits", KSTAT_DATA_UINT64 }, | |
371 | { "demand_metadata_misses", KSTAT_DATA_UINT64 }, | |
372 | { "prefetch_data_hits", KSTAT_DATA_UINT64 }, | |
373 | { "prefetch_data_misses", KSTAT_DATA_UINT64 }, | |
374 | { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, | |
375 | { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, | |
376 | { "mru_hits", KSTAT_DATA_UINT64 }, | |
377 | { "mru_ghost_hits", KSTAT_DATA_UINT64 }, | |
378 | { "mfu_hits", KSTAT_DATA_UINT64 }, | |
379 | { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, | |
380 | { "deleted", KSTAT_DATA_UINT64 }, | |
381 | { "recycle_miss", KSTAT_DATA_UINT64 }, | |
382 | { "mutex_miss", KSTAT_DATA_UINT64 }, | |
383 | { "evict_skip", KSTAT_DATA_UINT64 }, | |
428870ff BB |
384 | { "evict_l2_cached", KSTAT_DATA_UINT64 }, |
385 | { "evict_l2_eligible", KSTAT_DATA_UINT64 }, | |
386 | { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, | |
34dc7c2f BB |
387 | { "hash_elements", KSTAT_DATA_UINT64 }, |
388 | { "hash_elements_max", KSTAT_DATA_UINT64 }, | |
389 | { "hash_collisions", KSTAT_DATA_UINT64 }, | |
390 | { "hash_chains", KSTAT_DATA_UINT64 }, | |
391 | { "hash_chain_max", KSTAT_DATA_UINT64 }, | |
392 | { "p", KSTAT_DATA_UINT64 }, | |
393 | { "c", KSTAT_DATA_UINT64 }, | |
394 | { "c_min", KSTAT_DATA_UINT64 }, | |
395 | { "c_max", KSTAT_DATA_UINT64 }, | |
396 | { "size", KSTAT_DATA_UINT64 }, | |
397 | { "hdr_size", KSTAT_DATA_UINT64 }, | |
d164b209 | 398 | { "data_size", KSTAT_DATA_UINT64 }, |
cc7f677c | 399 | { "meta_size", KSTAT_DATA_UINT64 }, |
d164b209 | 400 | { "other_size", KSTAT_DATA_UINT64 }, |
13be560d BB |
401 | { "anon_size", KSTAT_DATA_UINT64 }, |
402 | { "anon_evict_data", KSTAT_DATA_UINT64 }, | |
403 | { "anon_evict_metadata", KSTAT_DATA_UINT64 }, | |
404 | { "mru_size", KSTAT_DATA_UINT64 }, | |
405 | { "mru_evict_data", KSTAT_DATA_UINT64 }, | |
406 | { "mru_evict_metadata", KSTAT_DATA_UINT64 }, | |
407 | { "mru_ghost_size", KSTAT_DATA_UINT64 }, | |
408 | { "mru_ghost_evict_data", KSTAT_DATA_UINT64 }, | |
409 | { "mru_ghost_evict_metadata", KSTAT_DATA_UINT64 }, | |
410 | { "mfu_size", KSTAT_DATA_UINT64 }, | |
411 | { "mfu_evict_data", KSTAT_DATA_UINT64 }, | |
412 | { "mfu_evict_metadata", KSTAT_DATA_UINT64 }, | |
413 | { "mfu_ghost_size", KSTAT_DATA_UINT64 }, | |
414 | { "mfu_ghost_evict_data", KSTAT_DATA_UINT64 }, | |
415 | { "mfu_ghost_evict_metadata", KSTAT_DATA_UINT64 }, | |
34dc7c2f BB |
416 | { "l2_hits", KSTAT_DATA_UINT64 }, |
417 | { "l2_misses", KSTAT_DATA_UINT64 }, | |
418 | { "l2_feeds", KSTAT_DATA_UINT64 }, | |
419 | { "l2_rw_clash", KSTAT_DATA_UINT64 }, | |
d164b209 BB |
420 | { "l2_read_bytes", KSTAT_DATA_UINT64 }, |
421 | { "l2_write_bytes", KSTAT_DATA_UINT64 }, | |
34dc7c2f BB |
422 | { "l2_writes_sent", KSTAT_DATA_UINT64 }, |
423 | { "l2_writes_done", KSTAT_DATA_UINT64 }, | |
424 | { "l2_writes_error", KSTAT_DATA_UINT64 }, | |
425 | { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, | |
426 | { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, | |
427 | { "l2_evict_reading", KSTAT_DATA_UINT64 }, | |
428 | { "l2_free_on_write", KSTAT_DATA_UINT64 }, | |
429 | { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, | |
430 | { "l2_cksum_bad", KSTAT_DATA_UINT64 }, | |
431 | { "l2_io_error", KSTAT_DATA_UINT64 }, | |
432 | { "l2_size", KSTAT_DATA_UINT64 }, | |
3a17a7a9 | 433 | { "l2_asize", KSTAT_DATA_UINT64 }, |
34dc7c2f | 434 | { "l2_hdr_size", KSTAT_DATA_UINT64 }, |
3a17a7a9 SK |
435 | { "l2_compress_successes", KSTAT_DATA_UINT64 }, |
436 | { "l2_compress_zeros", KSTAT_DATA_UINT64 }, | |
437 | { "l2_compress_failures", KSTAT_DATA_UINT64 }, | |
1834f2d8 | 438 | { "memory_throttle_count", KSTAT_DATA_UINT64 }, |
1eb5bfa3 GW |
439 | { "duplicate_buffers", KSTAT_DATA_UINT64 }, |
440 | { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, | |
441 | { "duplicate_reads", KSTAT_DATA_UINT64 }, | |
7cb67b45 BB |
442 | { "memory_direct_count", KSTAT_DATA_UINT64 }, |
443 | { "memory_indirect_count", KSTAT_DATA_UINT64 }, | |
1834f2d8 BB |
444 | { "arc_no_grow", KSTAT_DATA_UINT64 }, |
445 | { "arc_tempreserve", KSTAT_DATA_UINT64 }, | |
446 | { "arc_loaned_bytes", KSTAT_DATA_UINT64 }, | |
ab26409d | 447 | { "arc_prune", KSTAT_DATA_UINT64 }, |
1834f2d8 BB |
448 | { "arc_meta_used", KSTAT_DATA_UINT64 }, |
449 | { "arc_meta_limit", KSTAT_DATA_UINT64 }, | |
450 | { "arc_meta_max", KSTAT_DATA_UINT64 }, | |
34dc7c2f BB |
451 | }; |
452 | ||
453 | #define ARCSTAT(stat) (arc_stats.stat.value.ui64) | |
454 | ||
455 | #define ARCSTAT_INCR(stat, val) \ | |
d3cc8b15 | 456 | atomic_add_64(&arc_stats.stat.value.ui64, (val)) |
34dc7c2f | 457 | |
428870ff | 458 | #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) |
34dc7c2f BB |
459 | #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) |
460 | ||
461 | #define ARCSTAT_MAX(stat, val) { \ | |
462 | uint64_t m; \ | |
463 | while ((val) > (m = arc_stats.stat.value.ui64) && \ | |
464 | (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ | |
465 | continue; \ | |
466 | } | |
467 | ||
468 | #define ARCSTAT_MAXSTAT(stat) \ | |
469 | ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) | |
470 | ||
471 | /* | |
472 | * We define a macro to allow ARC hits/misses to be easily broken down by | |
473 | * two separate conditions, giving a total of four different subtypes for | |
474 | * each of hits and misses (so eight statistics total). | |
475 | */ | |
476 | #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ | |
477 | if (cond1) { \ | |
478 | if (cond2) { \ | |
479 | ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ | |
480 | } else { \ | |
481 | ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ | |
482 | } \ | |
483 | } else { \ | |
484 | if (cond2) { \ | |
485 | ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ | |
486 | } else { \ | |
487 | ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ | |
488 | } \ | |
489 | } | |
490 | ||
491 | kstat_t *arc_ksp; | |
428870ff | 492 | static arc_state_t *arc_anon; |
34dc7c2f BB |
493 | static arc_state_t *arc_mru; |
494 | static arc_state_t *arc_mru_ghost; | |
495 | static arc_state_t *arc_mfu; | |
496 | static arc_state_t *arc_mfu_ghost; | |
497 | static arc_state_t *arc_l2c_only; | |
498 | ||
499 | /* | |
500 | * There are several ARC variables that are critical to export as kstats -- | |
501 | * but we don't want to have to grovel around in the kstat whenever we wish to | |
502 | * manipulate them. For these variables, we therefore define them to be in | |
503 | * terms of the statistic variable. This assures that we are not introducing | |
504 | * the possibility of inconsistency by having shadow copies of the variables, | |
505 | * while still allowing the code to be readable. | |
506 | */ | |
507 | #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ | |
508 | #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ | |
509 | #define arc_c ARCSTAT(arcstat_c) /* target size of cache */ | |
510 | #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ | |
511 | #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ | |
1834f2d8 BB |
512 | #define arc_no_grow ARCSTAT(arcstat_no_grow) |
513 | #define arc_tempreserve ARCSTAT(arcstat_tempreserve) | |
514 | #define arc_loaned_bytes ARCSTAT(arcstat_loaned_bytes) | |
23c0a133 GW |
515 | #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ |
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 | ||
34dc7c2f BB |
522 | typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; |
523 | ||
524 | typedef struct arc_callback arc_callback_t; | |
525 | ||
526 | struct arc_callback { | |
527 | void *acb_private; | |
528 | arc_done_func_t *acb_done; | |
34dc7c2f BB |
529 | arc_buf_t *acb_buf; |
530 | zio_t *acb_zio_dummy; | |
531 | arc_callback_t *acb_next; | |
532 | }; | |
533 | ||
534 | typedef struct arc_write_callback arc_write_callback_t; | |
535 | ||
536 | struct arc_write_callback { | |
537 | void *awcb_private; | |
538 | arc_done_func_t *awcb_ready; | |
e8b96c60 | 539 | arc_done_func_t *awcb_physdone; |
34dc7c2f BB |
540 | arc_done_func_t *awcb_done; |
541 | arc_buf_t *awcb_buf; | |
542 | }; | |
543 | ||
544 | struct arc_buf_hdr { | |
545 | /* protected by hash lock */ | |
546 | dva_t b_dva; | |
547 | uint64_t b_birth; | |
548 | uint64_t b_cksum0; | |
549 | ||
550 | kmutex_t b_freeze_lock; | |
551 | zio_cksum_t *b_freeze_cksum; | |
552 | ||
553 | arc_buf_hdr_t *b_hash_next; | |
554 | arc_buf_t *b_buf; | |
555 | uint32_t b_flags; | |
556 | uint32_t b_datacnt; | |
557 | ||
558 | arc_callback_t *b_acb; | |
559 | kcondvar_t b_cv; | |
560 | ||
561 | /* immutable */ | |
562 | arc_buf_contents_t b_type; | |
563 | uint64_t b_size; | |
d164b209 | 564 | uint64_t b_spa; |
34dc7c2f BB |
565 | |
566 | /* protected by arc state mutex */ | |
567 | arc_state_t *b_state; | |
568 | list_node_t b_arc_node; | |
569 | ||
570 | /* updated atomically */ | |
571 | clock_t b_arc_access; | |
e0b0ca98 BB |
572 | uint32_t b_mru_hits; |
573 | uint32_t b_mru_ghost_hits; | |
574 | uint32_t b_mfu_hits; | |
575 | uint32_t b_mfu_ghost_hits; | |
576 | uint32_t b_l2_hits; | |
34dc7c2f BB |
577 | |
578 | /* self protecting */ | |
579 | refcount_t b_refcnt; | |
580 | ||
581 | l2arc_buf_hdr_t *b_l2hdr; | |
582 | list_node_t b_l2node; | |
583 | }; | |
584 | ||
ab26409d BB |
585 | static list_t arc_prune_list; |
586 | static kmutex_t arc_prune_mtx; | |
34dc7c2f BB |
587 | static arc_buf_t *arc_eviction_list; |
588 | static kmutex_t arc_eviction_mtx; | |
589 | static arc_buf_hdr_t arc_eviction_hdr; | |
590 | static void arc_get_data_buf(arc_buf_t *buf); | |
591 | static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); | |
592 | static int arc_evict_needed(arc_buf_contents_t type); | |
68121a03 BB |
593 | static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes, |
594 | arc_buf_contents_t type); | |
498877ba | 595 | static void arc_buf_watch(arc_buf_t *buf); |
34dc7c2f | 596 | |
428870ff BB |
597 | static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); |
598 | ||
34dc7c2f BB |
599 | #define GHOST_STATE(state) \ |
600 | ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ | |
601 | (state) == arc_l2c_only) | |
602 | ||
603 | /* | |
604 | * Private ARC flags. These flags are private ARC only flags that will show up | |
605 | * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can | |
606 | * be passed in as arc_flags in things like arc_read. However, these flags | |
607 | * should never be passed and should only be set by ARC code. When adding new | |
608 | * public flags, make sure not to smash the private ones. | |
609 | */ | |
610 | ||
611 | #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ | |
612 | #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ | |
613 | #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ | |
614 | #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ | |
615 | #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ | |
616 | #define ARC_INDIRECT (1 << 14) /* this is an indirect block */ | |
617 | #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ | |
b128c09f BB |
618 | #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ |
619 | #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ | |
620 | #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ | |
34dc7c2f BB |
621 | |
622 | #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) | |
623 | #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) | |
624 | #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) | |
d164b209 | 625 | #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) |
34dc7c2f BB |
626 | #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) |
627 | #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) | |
628 | #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) | |
b128c09f BB |
629 | #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) |
630 | #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ | |
631 | (hdr)->b_l2hdr != NULL) | |
34dc7c2f BB |
632 | #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) |
633 | #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) | |
634 | #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) | |
635 | ||
636 | /* | |
637 | * Other sizes | |
638 | */ | |
639 | ||
640 | #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) | |
641 | #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) | |
642 | ||
643 | /* | |
644 | * Hash table routines | |
645 | */ | |
646 | ||
00b46022 BB |
647 | #define HT_LOCK_ALIGN 64 |
648 | #define HT_LOCK_PAD (P2NPHASE(sizeof (kmutex_t), (HT_LOCK_ALIGN))) | |
34dc7c2f BB |
649 | |
650 | struct ht_lock { | |
651 | kmutex_t ht_lock; | |
652 | #ifdef _KERNEL | |
00b46022 | 653 | unsigned char pad[HT_LOCK_PAD]; |
34dc7c2f BB |
654 | #endif |
655 | }; | |
656 | ||
657 | #define BUF_LOCKS 256 | |
658 | typedef struct buf_hash_table { | |
659 | uint64_t ht_mask; | |
660 | arc_buf_hdr_t **ht_table; | |
661 | struct ht_lock ht_locks[BUF_LOCKS]; | |
662 | } buf_hash_table_t; | |
663 | ||
664 | static buf_hash_table_t buf_hash_table; | |
665 | ||
666 | #define BUF_HASH_INDEX(spa, dva, birth) \ | |
667 | (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) | |
668 | #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) | |
669 | #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) | |
428870ff BB |
670 | #define HDR_LOCK(hdr) \ |
671 | (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) | |
34dc7c2f BB |
672 | |
673 | uint64_t zfs_crc64_table[256]; | |
674 | ||
675 | /* | |
676 | * Level 2 ARC | |
677 | */ | |
678 | ||
679 | #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ | |
3a17a7a9 SK |
680 | #define L2ARC_HEADROOM 2 /* num of writes */ |
681 | /* | |
682 | * If we discover during ARC scan any buffers to be compressed, we boost | |
683 | * our headroom for the next scanning cycle by this percentage multiple. | |
684 | */ | |
685 | #define L2ARC_HEADROOM_BOOST 200 | |
d164b209 BB |
686 | #define L2ARC_FEED_SECS 1 /* caching interval secs */ |
687 | #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ | |
34dc7c2f BB |
688 | |
689 | #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) | |
690 | #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) | |
691 | ||
d3cc8b15 | 692 | /* L2ARC Performance Tunables */ |
abd8610c BB |
693 | unsigned long l2arc_write_max = L2ARC_WRITE_SIZE; /* def max write size */ |
694 | unsigned long l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra warmup write */ | |
695 | unsigned long l2arc_headroom = L2ARC_HEADROOM; /* # of dev writes */ | |
3a17a7a9 | 696 | unsigned long l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; |
abd8610c BB |
697 | unsigned long l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ |
698 | unsigned long l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval msecs */ | |
699 | int l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ | |
3a17a7a9 | 700 | int l2arc_nocompress = B_FALSE; /* don't compress bufs */ |
abd8610c | 701 | int l2arc_feed_again = B_TRUE; /* turbo warmup */ |
c93504f0 | 702 | int l2arc_norw = B_FALSE; /* no reads during writes */ |
34dc7c2f BB |
703 | |
704 | /* | |
705 | * L2ARC Internals | |
706 | */ | |
707 | typedef struct l2arc_dev { | |
708 | vdev_t *l2ad_vdev; /* vdev */ | |
709 | spa_t *l2ad_spa; /* spa */ | |
710 | uint64_t l2ad_hand; /* next write location */ | |
34dc7c2f BB |
711 | uint64_t l2ad_start; /* first addr on device */ |
712 | uint64_t l2ad_end; /* last addr on device */ | |
713 | uint64_t l2ad_evict; /* last addr eviction reached */ | |
714 | boolean_t l2ad_first; /* first sweep through */ | |
d164b209 | 715 | boolean_t l2ad_writing; /* currently writing */ |
34dc7c2f BB |
716 | list_t *l2ad_buflist; /* buffer list */ |
717 | list_node_t l2ad_node; /* device list node */ | |
718 | } l2arc_dev_t; | |
719 | ||
720 | static list_t L2ARC_dev_list; /* device list */ | |
721 | static list_t *l2arc_dev_list; /* device list pointer */ | |
722 | static kmutex_t l2arc_dev_mtx; /* device list mutex */ | |
723 | static l2arc_dev_t *l2arc_dev_last; /* last device used */ | |
724 | static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ | |
725 | static list_t L2ARC_free_on_write; /* free after write buf list */ | |
726 | static list_t *l2arc_free_on_write; /* free after write list ptr */ | |
727 | static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ | |
728 | static uint64_t l2arc_ndev; /* number of devices */ | |
729 | ||
730 | typedef struct l2arc_read_callback { | |
3a17a7a9 SK |
731 | arc_buf_t *l2rcb_buf; /* read buffer */ |
732 | spa_t *l2rcb_spa; /* spa */ | |
733 | blkptr_t l2rcb_bp; /* original blkptr */ | |
734 | zbookmark_t l2rcb_zb; /* original bookmark */ | |
735 | int l2rcb_flags; /* original flags */ | |
736 | enum zio_compress l2rcb_compress; /* applied compress */ | |
34dc7c2f BB |
737 | } l2arc_read_callback_t; |
738 | ||
739 | typedef struct l2arc_write_callback { | |
740 | l2arc_dev_t *l2wcb_dev; /* device info */ | |
741 | arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ | |
742 | } l2arc_write_callback_t; | |
743 | ||
744 | struct l2arc_buf_hdr { | |
745 | /* protected by arc_buf_hdr mutex */ | |
3a17a7a9 SK |
746 | l2arc_dev_t *b_dev; /* L2ARC device */ |
747 | uint64_t b_daddr; /* disk address, offset byte */ | |
748 | /* compression applied to buffer data */ | |
749 | enum zio_compress b_compress; | |
750 | /* real alloc'd buffer size depending on b_compress applied */ | |
e0b0ca98 | 751 | uint32_t b_hits; |
c5cb66ad | 752 | uint64_t b_asize; |
3a17a7a9 SK |
753 | /* temporary buffer holder for in-flight compressed data */ |
754 | void *b_tmp_cdata; | |
34dc7c2f BB |
755 | }; |
756 | ||
757 | typedef struct l2arc_data_free { | |
758 | /* protected by l2arc_free_on_write_mtx */ | |
759 | void *l2df_data; | |
760 | size_t l2df_size; | |
761 | void (*l2df_func)(void *, size_t); | |
762 | list_node_t l2df_list_node; | |
763 | } l2arc_data_free_t; | |
764 | ||
765 | static kmutex_t l2arc_feed_thr_lock; | |
766 | static kcondvar_t l2arc_feed_thr_cv; | |
767 | static uint8_t l2arc_thread_exit; | |
768 | ||
769 | static void l2arc_read_done(zio_t *zio); | |
770 | static void l2arc_hdr_stat_add(void); | |
771 | static void l2arc_hdr_stat_remove(void); | |
772 | ||
3a17a7a9 SK |
773 | static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); |
774 | static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, | |
775 | enum zio_compress c); | |
776 | static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); | |
777 | ||
34dc7c2f | 778 | static uint64_t |
d164b209 | 779 | buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) |
34dc7c2f | 780 | { |
34dc7c2f BB |
781 | uint8_t *vdva = (uint8_t *)dva; |
782 | uint64_t crc = -1ULL; | |
783 | int i; | |
784 | ||
785 | ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); | |
786 | ||
787 | for (i = 0; i < sizeof (dva_t); i++) | |
788 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; | |
789 | ||
d164b209 | 790 | crc ^= (spa>>8) ^ birth; |
34dc7c2f BB |
791 | |
792 | return (crc); | |
793 | } | |
794 | ||
795 | #define BUF_EMPTY(buf) \ | |
796 | ((buf)->b_dva.dva_word[0] == 0 && \ | |
797 | (buf)->b_dva.dva_word[1] == 0 && \ | |
b0bc7a84 | 798 | (buf)->b_cksum0 == 0) |
34dc7c2f BB |
799 | |
800 | #define BUF_EQUAL(spa, dva, birth, buf) \ | |
801 | ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ | |
802 | ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ | |
803 | ((buf)->b_birth == birth) && ((buf)->b_spa == spa) | |
804 | ||
428870ff BB |
805 | static void |
806 | buf_discard_identity(arc_buf_hdr_t *hdr) | |
807 | { | |
808 | hdr->b_dva.dva_word[0] = 0; | |
809 | hdr->b_dva.dva_word[1] = 0; | |
810 | hdr->b_birth = 0; | |
811 | hdr->b_cksum0 = 0; | |
812 | } | |
813 | ||
34dc7c2f | 814 | static arc_buf_hdr_t * |
9b67f605 | 815 | buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) |
34dc7c2f | 816 | { |
9b67f605 MA |
817 | const dva_t *dva = BP_IDENTITY(bp); |
818 | uint64_t birth = BP_PHYSICAL_BIRTH(bp); | |
34dc7c2f BB |
819 | uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); |
820 | kmutex_t *hash_lock = BUF_HASH_LOCK(idx); | |
821 | arc_buf_hdr_t *buf; | |
822 | ||
823 | mutex_enter(hash_lock); | |
824 | for (buf = buf_hash_table.ht_table[idx]; buf != NULL; | |
825 | buf = buf->b_hash_next) { | |
826 | if (BUF_EQUAL(spa, dva, birth, buf)) { | |
827 | *lockp = hash_lock; | |
828 | return (buf); | |
829 | } | |
830 | } | |
831 | mutex_exit(hash_lock); | |
832 | *lockp = NULL; | |
833 | return (NULL); | |
834 | } | |
835 | ||
836 | /* | |
837 | * Insert an entry into the hash table. If there is already an element | |
838 | * equal to elem in the hash table, then the already existing element | |
839 | * will be returned and the new element will not be inserted. | |
840 | * Otherwise returns NULL. | |
841 | */ | |
842 | static arc_buf_hdr_t * | |
843 | buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) | |
844 | { | |
845 | uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); | |
846 | kmutex_t *hash_lock = BUF_HASH_LOCK(idx); | |
847 | arc_buf_hdr_t *fbuf; | |
848 | uint32_t i; | |
849 | ||
9b67f605 MA |
850 | ASSERT(!DVA_IS_EMPTY(&buf->b_dva)); |
851 | ASSERT(buf->b_birth != 0); | |
34dc7c2f BB |
852 | ASSERT(!HDR_IN_HASH_TABLE(buf)); |
853 | *lockp = hash_lock; | |
854 | mutex_enter(hash_lock); | |
855 | for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; | |
856 | fbuf = fbuf->b_hash_next, i++) { | |
857 | if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) | |
858 | return (fbuf); | |
859 | } | |
860 | ||
861 | buf->b_hash_next = buf_hash_table.ht_table[idx]; | |
862 | buf_hash_table.ht_table[idx] = buf; | |
863 | buf->b_flags |= ARC_IN_HASH_TABLE; | |
864 | ||
865 | /* collect some hash table performance data */ | |
866 | if (i > 0) { | |
867 | ARCSTAT_BUMP(arcstat_hash_collisions); | |
868 | if (i == 1) | |
869 | ARCSTAT_BUMP(arcstat_hash_chains); | |
870 | ||
871 | ARCSTAT_MAX(arcstat_hash_chain_max, i); | |
872 | } | |
873 | ||
874 | ARCSTAT_BUMP(arcstat_hash_elements); | |
875 | ARCSTAT_MAXSTAT(arcstat_hash_elements); | |
876 | ||
877 | return (NULL); | |
878 | } | |
879 | ||
880 | static void | |
881 | buf_hash_remove(arc_buf_hdr_t *buf) | |
882 | { | |
883 | arc_buf_hdr_t *fbuf, **bufp; | |
884 | uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); | |
885 | ||
886 | ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); | |
887 | ASSERT(HDR_IN_HASH_TABLE(buf)); | |
888 | ||
889 | bufp = &buf_hash_table.ht_table[idx]; | |
890 | while ((fbuf = *bufp) != buf) { | |
891 | ASSERT(fbuf != NULL); | |
892 | bufp = &fbuf->b_hash_next; | |
893 | } | |
894 | *bufp = buf->b_hash_next; | |
895 | buf->b_hash_next = NULL; | |
896 | buf->b_flags &= ~ARC_IN_HASH_TABLE; | |
897 | ||
898 | /* collect some hash table performance data */ | |
899 | ARCSTAT_BUMPDOWN(arcstat_hash_elements); | |
900 | ||
901 | if (buf_hash_table.ht_table[idx] && | |
902 | buf_hash_table.ht_table[idx]->b_hash_next == NULL) | |
903 | ARCSTAT_BUMPDOWN(arcstat_hash_chains); | |
904 | } | |
905 | ||
906 | /* | |
907 | * Global data structures and functions for the buf kmem cache. | |
908 | */ | |
909 | static kmem_cache_t *hdr_cache; | |
910 | static kmem_cache_t *buf_cache; | |
ecf3d9b8 | 911 | static kmem_cache_t *l2arc_hdr_cache; |
34dc7c2f BB |
912 | |
913 | static void | |
914 | buf_fini(void) | |
915 | { | |
916 | int i; | |
917 | ||
00b46022 | 918 | #if defined(_KERNEL) && defined(HAVE_SPL) |
d1d7e268 MK |
919 | /* |
920 | * Large allocations which do not require contiguous pages | |
921 | * should be using vmem_free() in the linux kernel\ | |
922 | */ | |
00b46022 BB |
923 | vmem_free(buf_hash_table.ht_table, |
924 | (buf_hash_table.ht_mask + 1) * sizeof (void *)); | |
925 | #else | |
34dc7c2f BB |
926 | kmem_free(buf_hash_table.ht_table, |
927 | (buf_hash_table.ht_mask + 1) * sizeof (void *)); | |
00b46022 | 928 | #endif |
34dc7c2f BB |
929 | for (i = 0; i < BUF_LOCKS; i++) |
930 | mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); | |
931 | kmem_cache_destroy(hdr_cache); | |
932 | kmem_cache_destroy(buf_cache); | |
ecf3d9b8 | 933 | kmem_cache_destroy(l2arc_hdr_cache); |
34dc7c2f BB |
934 | } |
935 | ||
936 | /* | |
937 | * Constructor callback - called when the cache is empty | |
938 | * and a new buf is requested. | |
939 | */ | |
940 | /* ARGSUSED */ | |
941 | static int | |
942 | hdr_cons(void *vbuf, void *unused, int kmflag) | |
943 | { | |
944 | arc_buf_hdr_t *buf = vbuf; | |
945 | ||
946 | bzero(buf, sizeof (arc_buf_hdr_t)); | |
947 | refcount_create(&buf->b_refcnt); | |
948 | cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); | |
949 | mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); | |
98f72a53 BB |
950 | list_link_init(&buf->b_arc_node); |
951 | list_link_init(&buf->b_l2node); | |
d164b209 | 952 | arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); |
34dc7c2f | 953 | |
34dc7c2f BB |
954 | return (0); |
955 | } | |
956 | ||
b128c09f BB |
957 | /* ARGSUSED */ |
958 | static int | |
959 | buf_cons(void *vbuf, void *unused, int kmflag) | |
960 | { | |
961 | arc_buf_t *buf = vbuf; | |
962 | ||
963 | bzero(buf, sizeof (arc_buf_t)); | |
428870ff | 964 | mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); |
d164b209 BB |
965 | arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); |
966 | ||
b128c09f BB |
967 | return (0); |
968 | } | |
969 | ||
34dc7c2f BB |
970 | /* |
971 | * Destructor callback - called when a cached buf is | |
972 | * no longer required. | |
973 | */ | |
974 | /* ARGSUSED */ | |
975 | static void | |
976 | hdr_dest(void *vbuf, void *unused) | |
977 | { | |
978 | arc_buf_hdr_t *buf = vbuf; | |
979 | ||
428870ff | 980 | ASSERT(BUF_EMPTY(buf)); |
34dc7c2f BB |
981 | refcount_destroy(&buf->b_refcnt); |
982 | cv_destroy(&buf->b_cv); | |
983 | mutex_destroy(&buf->b_freeze_lock); | |
d164b209 | 984 | arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); |
34dc7c2f BB |
985 | } |
986 | ||
b128c09f BB |
987 | /* ARGSUSED */ |
988 | static void | |
989 | buf_dest(void *vbuf, void *unused) | |
990 | { | |
991 | arc_buf_t *buf = vbuf; | |
992 | ||
428870ff | 993 | mutex_destroy(&buf->b_evict_lock); |
d164b209 | 994 | arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); |
b128c09f BB |
995 | } |
996 | ||
34dc7c2f BB |
997 | static void |
998 | buf_init(void) | |
999 | { | |
1000 | uint64_t *ct; | |
1001 | uint64_t hsize = 1ULL << 12; | |
1002 | int i, j; | |
1003 | ||
1004 | /* | |
1005 | * The hash table is big enough to fill all of physical memory | |
1006 | * with an average 64K block size. The table will take up | |
1007 | * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers). | |
1008 | */ | |
1009 | while (hsize * 65536 < physmem * PAGESIZE) | |
1010 | hsize <<= 1; | |
1011 | retry: | |
1012 | buf_hash_table.ht_mask = hsize - 1; | |
00b46022 | 1013 | #if defined(_KERNEL) && defined(HAVE_SPL) |
d1d7e268 MK |
1014 | /* |
1015 | * Large allocations which do not require contiguous pages | |
1016 | * should be using vmem_alloc() in the linux kernel | |
1017 | */ | |
00b46022 BB |
1018 | buf_hash_table.ht_table = |
1019 | vmem_zalloc(hsize * sizeof (void*), KM_SLEEP); | |
1020 | #else | |
34dc7c2f BB |
1021 | buf_hash_table.ht_table = |
1022 | kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); | |
00b46022 | 1023 | #endif |
34dc7c2f BB |
1024 | if (buf_hash_table.ht_table == NULL) { |
1025 | ASSERT(hsize > (1ULL << 8)); | |
1026 | hsize >>= 1; | |
1027 | goto retry; | |
1028 | } | |
1029 | ||
1030 | hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), | |
302f753f | 1031 | 0, hdr_cons, hdr_dest, NULL, NULL, NULL, 0); |
34dc7c2f | 1032 | buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), |
b128c09f | 1033 | 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); |
ecf3d9b8 JL |
1034 | l2arc_hdr_cache = kmem_cache_create("l2arc_buf_hdr_t", L2HDR_SIZE, |
1035 | 0, NULL, NULL, NULL, NULL, NULL, 0); | |
34dc7c2f BB |
1036 | |
1037 | for (i = 0; i < 256; i++) | |
1038 | for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) | |
1039 | *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); | |
1040 | ||
1041 | for (i = 0; i < BUF_LOCKS; i++) { | |
1042 | mutex_init(&buf_hash_table.ht_locks[i].ht_lock, | |
1043 | NULL, MUTEX_DEFAULT, NULL); | |
1044 | } | |
1045 | } | |
1046 | ||
1047 | #define ARC_MINTIME (hz>>4) /* 62 ms */ | |
1048 | ||
1049 | static void | |
1050 | arc_cksum_verify(arc_buf_t *buf) | |
1051 | { | |
1052 | zio_cksum_t zc; | |
1053 | ||
1054 | if (!(zfs_flags & ZFS_DEBUG_MODIFY)) | |
1055 | return; | |
1056 | ||
1057 | mutex_enter(&buf->b_hdr->b_freeze_lock); | |
1058 | if (buf->b_hdr->b_freeze_cksum == NULL || | |
1059 | (buf->b_hdr->b_flags & ARC_IO_ERROR)) { | |
1060 | mutex_exit(&buf->b_hdr->b_freeze_lock); | |
1061 | return; | |
1062 | } | |
1063 | fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); | |
1064 | if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) | |
1065 | panic("buffer modified while frozen!"); | |
1066 | mutex_exit(&buf->b_hdr->b_freeze_lock); | |
1067 | } | |
1068 | ||
1069 | static int | |
1070 | arc_cksum_equal(arc_buf_t *buf) | |
1071 | { | |
1072 | zio_cksum_t zc; | |
1073 | int equal; | |
1074 | ||
1075 | mutex_enter(&buf->b_hdr->b_freeze_lock); | |
1076 | fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); | |
1077 | equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); | |
1078 | mutex_exit(&buf->b_hdr->b_freeze_lock); | |
1079 | ||
1080 | return (equal); | |
1081 | } | |
1082 | ||
1083 | static void | |
1084 | arc_cksum_compute(arc_buf_t *buf, boolean_t force) | |
1085 | { | |
1086 | if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) | |
1087 | return; | |
1088 | ||
1089 | mutex_enter(&buf->b_hdr->b_freeze_lock); | |
1090 | if (buf->b_hdr->b_freeze_cksum != NULL) { | |
1091 | mutex_exit(&buf->b_hdr->b_freeze_lock); | |
1092 | return; | |
1093 | } | |
409dc1a5 | 1094 | buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), |
d1d7e268 | 1095 | KM_PUSHPAGE); |
34dc7c2f BB |
1096 | fletcher_2_native(buf->b_data, buf->b_hdr->b_size, |
1097 | buf->b_hdr->b_freeze_cksum); | |
1098 | mutex_exit(&buf->b_hdr->b_freeze_lock); | |
498877ba MA |
1099 | arc_buf_watch(buf); |
1100 | } | |
1101 | ||
1102 | #ifndef _KERNEL | |
1103 | void | |
1104 | arc_buf_sigsegv(int sig, siginfo_t *si, void *unused) | |
1105 | { | |
1106 | panic("Got SIGSEGV at address: 0x%lx\n", (long) si->si_addr); | |
1107 | } | |
1108 | #endif | |
1109 | ||
1110 | /* ARGSUSED */ | |
1111 | static void | |
1112 | arc_buf_unwatch(arc_buf_t *buf) | |
1113 | { | |
1114 | #ifndef _KERNEL | |
1115 | if (arc_watch) { | |
1116 | ASSERT0(mprotect(buf->b_data, buf->b_hdr->b_size, | |
1117 | PROT_READ | PROT_WRITE)); | |
1118 | } | |
1119 | #endif | |
1120 | } | |
1121 | ||
1122 | /* ARGSUSED */ | |
1123 | static void | |
1124 | arc_buf_watch(arc_buf_t *buf) | |
1125 | { | |
1126 | #ifndef _KERNEL | |
1127 | if (arc_watch) | |
1128 | ASSERT0(mprotect(buf->b_data, buf->b_hdr->b_size, PROT_READ)); | |
1129 | #endif | |
34dc7c2f BB |
1130 | } |
1131 | ||
1132 | void | |
1133 | arc_buf_thaw(arc_buf_t *buf) | |
1134 | { | |
1135 | if (zfs_flags & ZFS_DEBUG_MODIFY) { | |
1136 | if (buf->b_hdr->b_state != arc_anon) | |
1137 | panic("modifying non-anon buffer!"); | |
1138 | if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) | |
1139 | panic("modifying buffer while i/o in progress!"); | |
1140 | arc_cksum_verify(buf); | |
1141 | } | |
1142 | ||
1143 | mutex_enter(&buf->b_hdr->b_freeze_lock); | |
1144 | if (buf->b_hdr->b_freeze_cksum != NULL) { | |
1145 | kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); | |
1146 | buf->b_hdr->b_freeze_cksum = NULL; | |
1147 | } | |
428870ff | 1148 | |
34dc7c2f | 1149 | mutex_exit(&buf->b_hdr->b_freeze_lock); |
498877ba MA |
1150 | |
1151 | arc_buf_unwatch(buf); | |
34dc7c2f BB |
1152 | } |
1153 | ||
1154 | void | |
1155 | arc_buf_freeze(arc_buf_t *buf) | |
1156 | { | |
428870ff BB |
1157 | kmutex_t *hash_lock; |
1158 | ||
34dc7c2f BB |
1159 | if (!(zfs_flags & ZFS_DEBUG_MODIFY)) |
1160 | return; | |
1161 | ||
428870ff BB |
1162 | hash_lock = HDR_LOCK(buf->b_hdr); |
1163 | mutex_enter(hash_lock); | |
1164 | ||
34dc7c2f BB |
1165 | ASSERT(buf->b_hdr->b_freeze_cksum != NULL || |
1166 | buf->b_hdr->b_state == arc_anon); | |
1167 | arc_cksum_compute(buf, B_FALSE); | |
428870ff | 1168 | mutex_exit(hash_lock); |
498877ba | 1169 | |
34dc7c2f BB |
1170 | } |
1171 | ||
1172 | static void | |
1173 | add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) | |
1174 | { | |
1175 | ASSERT(MUTEX_HELD(hash_lock)); | |
1176 | ||
1177 | if ((refcount_add(&ab->b_refcnt, tag) == 1) && | |
1178 | (ab->b_state != arc_anon)) { | |
1179 | uint64_t delta = ab->b_size * ab->b_datacnt; | |
1180 | list_t *list = &ab->b_state->arcs_list[ab->b_type]; | |
1181 | uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; | |
1182 | ||
1183 | ASSERT(!MUTEX_HELD(&ab->b_state->arcs_mtx)); | |
1184 | mutex_enter(&ab->b_state->arcs_mtx); | |
1185 | ASSERT(list_link_active(&ab->b_arc_node)); | |
1186 | list_remove(list, ab); | |
1187 | if (GHOST_STATE(ab->b_state)) { | |
c99c9001 | 1188 | ASSERT0(ab->b_datacnt); |
34dc7c2f BB |
1189 | ASSERT3P(ab->b_buf, ==, NULL); |
1190 | delta = ab->b_size; | |
1191 | } | |
1192 | ASSERT(delta > 0); | |
1193 | ASSERT3U(*size, >=, delta); | |
1194 | atomic_add_64(size, -delta); | |
1195 | mutex_exit(&ab->b_state->arcs_mtx); | |
b128c09f | 1196 | /* remove the prefetch flag if we get a reference */ |
34dc7c2f BB |
1197 | if (ab->b_flags & ARC_PREFETCH) |
1198 | ab->b_flags &= ~ARC_PREFETCH; | |
1199 | } | |
1200 | } | |
1201 | ||
1202 | static int | |
1203 | remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) | |
1204 | { | |
1205 | int cnt; | |
1206 | arc_state_t *state = ab->b_state; | |
1207 | ||
1208 | ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); | |
1209 | ASSERT(!GHOST_STATE(state)); | |
1210 | ||
1211 | if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && | |
1212 | (state != arc_anon)) { | |
1213 | uint64_t *size = &state->arcs_lsize[ab->b_type]; | |
1214 | ||
1215 | ASSERT(!MUTEX_HELD(&state->arcs_mtx)); | |
1216 | mutex_enter(&state->arcs_mtx); | |
1217 | ASSERT(!list_link_active(&ab->b_arc_node)); | |
1218 | list_insert_head(&state->arcs_list[ab->b_type], ab); | |
1219 | ASSERT(ab->b_datacnt > 0); | |
1220 | atomic_add_64(size, ab->b_size * ab->b_datacnt); | |
1221 | mutex_exit(&state->arcs_mtx); | |
1222 | } | |
1223 | return (cnt); | |
1224 | } | |
1225 | ||
e0b0ca98 BB |
1226 | /* |
1227 | * Returns detailed information about a specific arc buffer. When the | |
1228 | * state_index argument is set the function will calculate the arc header | |
1229 | * list position for its arc state. Since this requires a linear traversal | |
1230 | * callers are strongly encourage not to do this. However, it can be helpful | |
1231 | * for targeted analysis so the functionality is provided. | |
1232 | */ | |
1233 | void | |
1234 | arc_buf_info(arc_buf_t *ab, arc_buf_info_t *abi, int state_index) | |
1235 | { | |
1236 | arc_buf_hdr_t *hdr = ab->b_hdr; | |
1237 | arc_state_t *state = hdr->b_state; | |
1238 | ||
d1d7e268 | 1239 | memset(abi, 0, sizeof (arc_buf_info_t)); |
e0b0ca98 BB |
1240 | abi->abi_flags = hdr->b_flags; |
1241 | abi->abi_datacnt = hdr->b_datacnt; | |
1242 | abi->abi_state_type = state ? state->arcs_state : ARC_STATE_ANON; | |
1243 | abi->abi_state_contents = hdr->b_type; | |
1244 | abi->abi_state_index = -1; | |
1245 | abi->abi_size = hdr->b_size; | |
1246 | abi->abi_access = hdr->b_arc_access; | |
1247 | abi->abi_mru_hits = hdr->b_mru_hits; | |
1248 | abi->abi_mru_ghost_hits = hdr->b_mru_ghost_hits; | |
1249 | abi->abi_mfu_hits = hdr->b_mfu_hits; | |
1250 | abi->abi_mfu_ghost_hits = hdr->b_mfu_ghost_hits; | |
1251 | abi->abi_holds = refcount_count(&hdr->b_refcnt); | |
1252 | ||
1253 | if (hdr->b_l2hdr) { | |
1254 | abi->abi_l2arc_dattr = hdr->b_l2hdr->b_daddr; | |
1255 | abi->abi_l2arc_asize = hdr->b_l2hdr->b_asize; | |
1256 | abi->abi_l2arc_compress = hdr->b_l2hdr->b_compress; | |
1257 | abi->abi_l2arc_hits = hdr->b_l2hdr->b_hits; | |
1258 | } | |
1259 | ||
1260 | if (state && state_index && list_link_active(&hdr->b_arc_node)) { | |
1261 | list_t *list = &state->arcs_list[hdr->b_type]; | |
1262 | arc_buf_hdr_t *h; | |
1263 | ||
1264 | mutex_enter(&state->arcs_mtx); | |
1265 | for (h = list_head(list); h != NULL; h = list_next(list, h)) { | |
1266 | abi->abi_state_index++; | |
1267 | if (h == hdr) | |
1268 | break; | |
1269 | } | |
1270 | mutex_exit(&state->arcs_mtx); | |
1271 | } | |
1272 | } | |
1273 | ||
34dc7c2f BB |
1274 | /* |
1275 | * Move the supplied buffer to the indicated state. The mutex | |
1276 | * for the buffer must be held by the caller. | |
1277 | */ | |
1278 | static void | |
1279 | arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) | |
1280 | { | |
1281 | arc_state_t *old_state = ab->b_state; | |
1282 | int64_t refcnt = refcount_count(&ab->b_refcnt); | |
1283 | uint64_t from_delta, to_delta; | |
1284 | ||
1285 | ASSERT(MUTEX_HELD(hash_lock)); | |
e8b96c60 | 1286 | ASSERT3P(new_state, !=, old_state); |
34dc7c2f BB |
1287 | ASSERT(refcnt == 0 || ab->b_datacnt > 0); |
1288 | ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); | |
428870ff | 1289 | ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); |
34dc7c2f BB |
1290 | |
1291 | from_delta = to_delta = ab->b_datacnt * ab->b_size; | |
1292 | ||
1293 | /* | |
1294 | * If this buffer is evictable, transfer it from the | |
1295 | * old state list to the new state list. | |
1296 | */ | |
1297 | if (refcnt == 0) { | |
1298 | if (old_state != arc_anon) { | |
1299 | int use_mutex = !MUTEX_HELD(&old_state->arcs_mtx); | |
1300 | uint64_t *size = &old_state->arcs_lsize[ab->b_type]; | |
1301 | ||
1302 | if (use_mutex) | |
1303 | mutex_enter(&old_state->arcs_mtx); | |
1304 | ||
1305 | ASSERT(list_link_active(&ab->b_arc_node)); | |
1306 | list_remove(&old_state->arcs_list[ab->b_type], ab); | |
1307 | ||
1308 | /* | |
1309 | * If prefetching out of the ghost cache, | |
428870ff | 1310 | * we will have a non-zero datacnt. |
34dc7c2f BB |
1311 | */ |
1312 | if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { | |
1313 | /* ghost elements have a ghost size */ | |
1314 | ASSERT(ab->b_buf == NULL); | |
1315 | from_delta = ab->b_size; | |
1316 | } | |
1317 | ASSERT3U(*size, >=, from_delta); | |
1318 | atomic_add_64(size, -from_delta); | |
1319 | ||
1320 | if (use_mutex) | |
1321 | mutex_exit(&old_state->arcs_mtx); | |
1322 | } | |
1323 | if (new_state != arc_anon) { | |
1324 | int use_mutex = !MUTEX_HELD(&new_state->arcs_mtx); | |
1325 | uint64_t *size = &new_state->arcs_lsize[ab->b_type]; | |
1326 | ||
1327 | if (use_mutex) | |
1328 | mutex_enter(&new_state->arcs_mtx); | |
1329 | ||
1330 | list_insert_head(&new_state->arcs_list[ab->b_type], ab); | |
1331 | ||
1332 | /* ghost elements have a ghost size */ | |
1333 | if (GHOST_STATE(new_state)) { | |
1334 | ASSERT(ab->b_datacnt == 0); | |
1335 | ASSERT(ab->b_buf == NULL); | |
1336 | to_delta = ab->b_size; | |
1337 | } | |
1338 | atomic_add_64(size, to_delta); | |
1339 | ||
1340 | if (use_mutex) | |
1341 | mutex_exit(&new_state->arcs_mtx); | |
1342 | } | |
1343 | } | |
1344 | ||
1345 | ASSERT(!BUF_EMPTY(ab)); | |
428870ff | 1346 | if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) |
34dc7c2f | 1347 | buf_hash_remove(ab); |
34dc7c2f BB |
1348 | |
1349 | /* adjust state sizes */ | |
1350 | if (to_delta) | |
1351 | atomic_add_64(&new_state->arcs_size, to_delta); | |
1352 | if (from_delta) { | |
1353 | ASSERT3U(old_state->arcs_size, >=, from_delta); | |
1354 | atomic_add_64(&old_state->arcs_size, -from_delta); | |
1355 | } | |
1356 | ab->b_state = new_state; | |
1357 | ||
1358 | /* adjust l2arc hdr stats */ | |
1359 | if (new_state == arc_l2c_only) | |
1360 | l2arc_hdr_stat_add(); | |
1361 | else if (old_state == arc_l2c_only) | |
1362 | l2arc_hdr_stat_remove(); | |
1363 | } | |
1364 | ||
1365 | void | |
d164b209 | 1366 | arc_space_consume(uint64_t space, arc_space_type_t type) |
34dc7c2f | 1367 | { |
d164b209 BB |
1368 | ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); |
1369 | ||
1370 | switch (type) { | |
e75c13c3 BB |
1371 | default: |
1372 | break; | |
d164b209 BB |
1373 | case ARC_SPACE_DATA: |
1374 | ARCSTAT_INCR(arcstat_data_size, space); | |
1375 | break; | |
cc7f677c PS |
1376 | case ARC_SPACE_META: |
1377 | ARCSTAT_INCR(arcstat_meta_size, space); | |
1378 | break; | |
d164b209 BB |
1379 | case ARC_SPACE_OTHER: |
1380 | ARCSTAT_INCR(arcstat_other_size, space); | |
1381 | break; | |
1382 | case ARC_SPACE_HDRS: | |
1383 | ARCSTAT_INCR(arcstat_hdr_size, space); | |
1384 | break; | |
1385 | case ARC_SPACE_L2HDRS: | |
1386 | ARCSTAT_INCR(arcstat_l2_hdr_size, space); | |
1387 | break; | |
1388 | } | |
1389 | ||
cc7f677c PS |
1390 | if (type != ARC_SPACE_DATA) |
1391 | ARCSTAT_INCR(arcstat_meta_used, space); | |
1392 | ||
34dc7c2f BB |
1393 | atomic_add_64(&arc_size, space); |
1394 | } | |
1395 | ||
1396 | void | |
d164b209 | 1397 | arc_space_return(uint64_t space, arc_space_type_t type) |
34dc7c2f | 1398 | { |
d164b209 BB |
1399 | ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); |
1400 | ||
1401 | switch (type) { | |
e75c13c3 BB |
1402 | default: |
1403 | break; | |
d164b209 BB |
1404 | case ARC_SPACE_DATA: |
1405 | ARCSTAT_INCR(arcstat_data_size, -space); | |
1406 | break; | |
cc7f677c PS |
1407 | case ARC_SPACE_META: |
1408 | ARCSTAT_INCR(arcstat_meta_size, -space); | |
1409 | break; | |
d164b209 BB |
1410 | case ARC_SPACE_OTHER: |
1411 | ARCSTAT_INCR(arcstat_other_size, -space); | |
1412 | break; | |
1413 | case ARC_SPACE_HDRS: | |
1414 | ARCSTAT_INCR(arcstat_hdr_size, -space); | |
1415 | break; | |
1416 | case ARC_SPACE_L2HDRS: | |
1417 | ARCSTAT_INCR(arcstat_l2_hdr_size, -space); | |
1418 | break; | |
1419 | } | |
1420 | ||
cc7f677c PS |
1421 | if (type != ARC_SPACE_DATA) { |
1422 | ASSERT(arc_meta_used >= space); | |
1423 | if (arc_meta_max < arc_meta_used) | |
1424 | arc_meta_max = arc_meta_used; | |
1425 | ARCSTAT_INCR(arcstat_meta_used, -space); | |
1426 | } | |
1427 | ||
34dc7c2f BB |
1428 | ASSERT(arc_size >= space); |
1429 | atomic_add_64(&arc_size, -space); | |
1430 | } | |
1431 | ||
34dc7c2f BB |
1432 | arc_buf_t * |
1433 | arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) | |
1434 | { | |
1435 | arc_buf_hdr_t *hdr; | |
1436 | arc_buf_t *buf; | |
1437 | ||
1438 | ASSERT3U(size, >, 0); | |
1439 | hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); | |
1440 | ASSERT(BUF_EMPTY(hdr)); | |
1441 | hdr->b_size = size; | |
1442 | hdr->b_type = type; | |
3541dc6d | 1443 | hdr->b_spa = spa_load_guid(spa); |
34dc7c2f BB |
1444 | hdr->b_state = arc_anon; |
1445 | hdr->b_arc_access = 0; | |
e0b0ca98 BB |
1446 | hdr->b_mru_hits = 0; |
1447 | hdr->b_mru_ghost_hits = 0; | |
1448 | hdr->b_mfu_hits = 0; | |
1449 | hdr->b_mfu_ghost_hits = 0; | |
1450 | hdr->b_l2_hits = 0; | |
34dc7c2f BB |
1451 | buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); |
1452 | buf->b_hdr = hdr; | |
1453 | buf->b_data = NULL; | |
1454 | buf->b_efunc = NULL; | |
1455 | buf->b_private = NULL; | |
1456 | buf->b_next = NULL; | |
1457 | hdr->b_buf = buf; | |
1458 | arc_get_data_buf(buf); | |
1459 | hdr->b_datacnt = 1; | |
1460 | hdr->b_flags = 0; | |
1461 | ASSERT(refcount_is_zero(&hdr->b_refcnt)); | |
1462 | (void) refcount_add(&hdr->b_refcnt, tag); | |
1463 | ||
1464 | return (buf); | |
1465 | } | |
1466 | ||
9babb374 BB |
1467 | static char *arc_onloan_tag = "onloan"; |
1468 | ||
1469 | /* | |
1470 | * Loan out an anonymous arc buffer. Loaned buffers are not counted as in | |
1471 | * flight data by arc_tempreserve_space() until they are "returned". Loaned | |
1472 | * buffers must be returned to the arc before they can be used by the DMU or | |
1473 | * freed. | |
1474 | */ | |
1475 | arc_buf_t * | |
1476 | arc_loan_buf(spa_t *spa, int size) | |
1477 | { | |
1478 | arc_buf_t *buf; | |
1479 | ||
1480 | buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); | |
1481 | ||
1482 | atomic_add_64(&arc_loaned_bytes, size); | |
1483 | return (buf); | |
1484 | } | |
1485 | ||
1486 | /* | |
1487 | * Return a loaned arc buffer to the arc. | |
1488 | */ | |
1489 | void | |
1490 | arc_return_buf(arc_buf_t *buf, void *tag) | |
1491 | { | |
1492 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
1493 | ||
9babb374 | 1494 | ASSERT(buf->b_data != NULL); |
428870ff BB |
1495 | (void) refcount_add(&hdr->b_refcnt, tag); |
1496 | (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); | |
9babb374 BB |
1497 | |
1498 | atomic_add_64(&arc_loaned_bytes, -hdr->b_size); | |
1499 | } | |
1500 | ||
428870ff BB |
1501 | /* Detach an arc_buf from a dbuf (tag) */ |
1502 | void | |
1503 | arc_loan_inuse_buf(arc_buf_t *buf, void *tag) | |
1504 | { | |
1505 | arc_buf_hdr_t *hdr; | |
1506 | ||
1507 | ASSERT(buf->b_data != NULL); | |
1508 | hdr = buf->b_hdr; | |
1509 | (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); | |
1510 | (void) refcount_remove(&hdr->b_refcnt, tag); | |
1511 | buf->b_efunc = NULL; | |
1512 | buf->b_private = NULL; | |
1513 | ||
1514 | atomic_add_64(&arc_loaned_bytes, hdr->b_size); | |
1515 | } | |
1516 | ||
34dc7c2f BB |
1517 | static arc_buf_t * |
1518 | arc_buf_clone(arc_buf_t *from) | |
1519 | { | |
1520 | arc_buf_t *buf; | |
1521 | arc_buf_hdr_t *hdr = from->b_hdr; | |
1522 | uint64_t size = hdr->b_size; | |
1523 | ||
428870ff BB |
1524 | ASSERT(hdr->b_state != arc_anon); |
1525 | ||
34dc7c2f BB |
1526 | buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); |
1527 | buf->b_hdr = hdr; | |
1528 | buf->b_data = NULL; | |
1529 | buf->b_efunc = NULL; | |
1530 | buf->b_private = NULL; | |
1531 | buf->b_next = hdr->b_buf; | |
1532 | hdr->b_buf = buf; | |
1533 | arc_get_data_buf(buf); | |
1534 | bcopy(from->b_data, buf->b_data, size); | |
1eb5bfa3 GW |
1535 | |
1536 | /* | |
1537 | * This buffer already exists in the arc so create a duplicate | |
1538 | * copy for the caller. If the buffer is associated with user data | |
1539 | * then track the size and number of duplicates. These stats will be | |
1540 | * updated as duplicate buffers are created and destroyed. | |
1541 | */ | |
1542 | if (hdr->b_type == ARC_BUFC_DATA) { | |
1543 | ARCSTAT_BUMP(arcstat_duplicate_buffers); | |
1544 | ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); | |
1545 | } | |
34dc7c2f BB |
1546 | hdr->b_datacnt += 1; |
1547 | return (buf); | |
1548 | } | |
1549 | ||
1550 | void | |
1551 | arc_buf_add_ref(arc_buf_t *buf, void* tag) | |
1552 | { | |
1553 | arc_buf_hdr_t *hdr; | |
1554 | kmutex_t *hash_lock; | |
1555 | ||
1556 | /* | |
b128c09f BB |
1557 | * Check to see if this buffer is evicted. Callers |
1558 | * must verify b_data != NULL to know if the add_ref | |
1559 | * was successful. | |
34dc7c2f | 1560 | */ |
428870ff | 1561 | mutex_enter(&buf->b_evict_lock); |
b128c09f | 1562 | if (buf->b_data == NULL) { |
428870ff | 1563 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
1564 | return; |
1565 | } | |
428870ff | 1566 | hash_lock = HDR_LOCK(buf->b_hdr); |
34dc7c2f | 1567 | mutex_enter(hash_lock); |
428870ff BB |
1568 | hdr = buf->b_hdr; |
1569 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
1570 | mutex_exit(&buf->b_evict_lock); | |
34dc7c2f | 1571 | |
34dc7c2f BB |
1572 | ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); |
1573 | add_reference(hdr, hash_lock, tag); | |
d164b209 | 1574 | DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); |
34dc7c2f BB |
1575 | arc_access(hdr, hash_lock); |
1576 | mutex_exit(hash_lock); | |
1577 | ARCSTAT_BUMP(arcstat_hits); | |
1578 | ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), | |
1579 | demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, | |
1580 | data, metadata, hits); | |
1581 | } | |
1582 | ||
1583 | /* | |
1584 | * Free the arc data buffer. If it is an l2arc write in progress, | |
1585 | * the buffer is placed on l2arc_free_on_write to be freed later. | |
1586 | */ | |
1587 | static void | |
498877ba | 1588 | arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) |
34dc7c2f | 1589 | { |
498877ba MA |
1590 | arc_buf_hdr_t *hdr = buf->b_hdr; |
1591 | ||
34dc7c2f BB |
1592 | if (HDR_L2_WRITING(hdr)) { |
1593 | l2arc_data_free_t *df; | |
594b4dd8 | 1594 | df = kmem_alloc(sizeof (l2arc_data_free_t), KM_PUSHPAGE); |
498877ba MA |
1595 | df->l2df_data = buf->b_data; |
1596 | df->l2df_size = hdr->b_size; | |
34dc7c2f BB |
1597 | df->l2df_func = free_func; |
1598 | mutex_enter(&l2arc_free_on_write_mtx); | |
1599 | list_insert_head(l2arc_free_on_write, df); | |
1600 | mutex_exit(&l2arc_free_on_write_mtx); | |
1601 | ARCSTAT_BUMP(arcstat_l2_free_on_write); | |
1602 | } else { | |
498877ba | 1603 | free_func(buf->b_data, hdr->b_size); |
34dc7c2f BB |
1604 | } |
1605 | } | |
1606 | ||
1607 | static void | |
1608 | arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all) | |
1609 | { | |
1610 | arc_buf_t **bufp; | |
1611 | ||
1612 | /* free up data associated with the buf */ | |
1613 | if (buf->b_data) { | |
1614 | arc_state_t *state = buf->b_hdr->b_state; | |
1615 | uint64_t size = buf->b_hdr->b_size; | |
1616 | arc_buf_contents_t type = buf->b_hdr->b_type; | |
1617 | ||
1618 | arc_cksum_verify(buf); | |
498877ba | 1619 | arc_buf_unwatch(buf); |
428870ff | 1620 | |
34dc7c2f BB |
1621 | if (!recycle) { |
1622 | if (type == ARC_BUFC_METADATA) { | |
498877ba | 1623 | arc_buf_data_free(buf, zio_buf_free); |
cc7f677c | 1624 | arc_space_return(size, ARC_SPACE_META); |
34dc7c2f BB |
1625 | } else { |
1626 | ASSERT(type == ARC_BUFC_DATA); | |
498877ba | 1627 | arc_buf_data_free(buf, zio_data_buf_free); |
cc7f677c | 1628 | arc_space_return(size, ARC_SPACE_DATA); |
34dc7c2f BB |
1629 | } |
1630 | } | |
1631 | if (list_link_active(&buf->b_hdr->b_arc_node)) { | |
1632 | uint64_t *cnt = &state->arcs_lsize[type]; | |
1633 | ||
1634 | ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); | |
1635 | ASSERT(state != arc_anon); | |
1636 | ||
1637 | ASSERT3U(*cnt, >=, size); | |
1638 | atomic_add_64(cnt, -size); | |
1639 | } | |
1640 | ASSERT3U(state->arcs_size, >=, size); | |
1641 | atomic_add_64(&state->arcs_size, -size); | |
1642 | buf->b_data = NULL; | |
1eb5bfa3 GW |
1643 | |
1644 | /* | |
1645 | * If we're destroying a duplicate buffer make sure | |
1646 | * that the appropriate statistics are updated. | |
1647 | */ | |
1648 | if (buf->b_hdr->b_datacnt > 1 && | |
1649 | buf->b_hdr->b_type == ARC_BUFC_DATA) { | |
1650 | ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); | |
1651 | ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); | |
1652 | } | |
34dc7c2f BB |
1653 | ASSERT(buf->b_hdr->b_datacnt > 0); |
1654 | buf->b_hdr->b_datacnt -= 1; | |
1655 | } | |
1656 | ||
1657 | /* only remove the buf if requested */ | |
1658 | if (!all) | |
1659 | return; | |
1660 | ||
1661 | /* remove the buf from the hdr list */ | |
1662 | for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) | |
1663 | continue; | |
1664 | *bufp = buf->b_next; | |
428870ff | 1665 | buf->b_next = NULL; |
34dc7c2f BB |
1666 | |
1667 | ASSERT(buf->b_efunc == NULL); | |
1668 | ||
1669 | /* clean up the buf */ | |
1670 | buf->b_hdr = NULL; | |
1671 | kmem_cache_free(buf_cache, buf); | |
1672 | } | |
1673 | ||
1674 | static void | |
1675 | arc_hdr_destroy(arc_buf_hdr_t *hdr) | |
1676 | { | |
d6320ddb BB |
1677 | l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; |
1678 | ||
34dc7c2f BB |
1679 | ASSERT(refcount_is_zero(&hdr->b_refcnt)); |
1680 | ASSERT3P(hdr->b_state, ==, arc_anon); | |
1681 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
1682 | ||
428870ff BB |
1683 | if (l2hdr != NULL) { |
1684 | boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); | |
1685 | /* | |
1686 | * To prevent arc_free() and l2arc_evict() from | |
1687 | * attempting to free the same buffer at the same time, | |
1688 | * a FREE_IN_PROGRESS flag is given to arc_free() to | |
1689 | * give it priority. l2arc_evict() can't destroy this | |
1690 | * header while we are waiting on l2arc_buflist_mtx. | |
1691 | * | |
1692 | * The hdr may be removed from l2ad_buflist before we | |
1693 | * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. | |
1694 | */ | |
1695 | if (!buflist_held) { | |
34dc7c2f | 1696 | mutex_enter(&l2arc_buflist_mtx); |
428870ff | 1697 | l2hdr = hdr->b_l2hdr; |
34dc7c2f | 1698 | } |
428870ff BB |
1699 | |
1700 | if (l2hdr != NULL) { | |
1701 | list_remove(l2hdr->b_dev->l2ad_buflist, hdr); | |
1702 | ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); | |
3a17a7a9 | 1703 | ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); |
ecf3d9b8 | 1704 | kmem_cache_free(l2arc_hdr_cache, l2hdr); |
6e1d7276 | 1705 | arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS); |
428870ff BB |
1706 | if (hdr->b_state == arc_l2c_only) |
1707 | l2arc_hdr_stat_remove(); | |
1708 | hdr->b_l2hdr = NULL; | |
1709 | } | |
1710 | ||
1711 | if (!buflist_held) | |
1712 | mutex_exit(&l2arc_buflist_mtx); | |
34dc7c2f BB |
1713 | } |
1714 | ||
1715 | if (!BUF_EMPTY(hdr)) { | |
1716 | ASSERT(!HDR_IN_HASH_TABLE(hdr)); | |
428870ff | 1717 | buf_discard_identity(hdr); |
34dc7c2f BB |
1718 | } |
1719 | while (hdr->b_buf) { | |
1720 | arc_buf_t *buf = hdr->b_buf; | |
1721 | ||
1722 | if (buf->b_efunc) { | |
1723 | mutex_enter(&arc_eviction_mtx); | |
428870ff | 1724 | mutex_enter(&buf->b_evict_lock); |
34dc7c2f BB |
1725 | ASSERT(buf->b_hdr != NULL); |
1726 | arc_buf_destroy(hdr->b_buf, FALSE, FALSE); | |
1727 | hdr->b_buf = buf->b_next; | |
1728 | buf->b_hdr = &arc_eviction_hdr; | |
1729 | buf->b_next = arc_eviction_list; | |
1730 | arc_eviction_list = buf; | |
428870ff | 1731 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
1732 | mutex_exit(&arc_eviction_mtx); |
1733 | } else { | |
1734 | arc_buf_destroy(hdr->b_buf, FALSE, TRUE); | |
1735 | } | |
1736 | } | |
1737 | if (hdr->b_freeze_cksum != NULL) { | |
1738 | kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); | |
1739 | hdr->b_freeze_cksum = NULL; | |
1740 | } | |
1741 | ||
1742 | ASSERT(!list_link_active(&hdr->b_arc_node)); | |
1743 | ASSERT3P(hdr->b_hash_next, ==, NULL); | |
1744 | ASSERT3P(hdr->b_acb, ==, NULL); | |
1745 | kmem_cache_free(hdr_cache, hdr); | |
1746 | } | |
1747 | ||
1748 | void | |
1749 | arc_buf_free(arc_buf_t *buf, void *tag) | |
1750 | { | |
1751 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
1752 | int hashed = hdr->b_state != arc_anon; | |
1753 | ||
1754 | ASSERT(buf->b_efunc == NULL); | |
1755 | ASSERT(buf->b_data != NULL); | |
1756 | ||
1757 | if (hashed) { | |
1758 | kmutex_t *hash_lock = HDR_LOCK(hdr); | |
1759 | ||
1760 | mutex_enter(hash_lock); | |
428870ff BB |
1761 | hdr = buf->b_hdr; |
1762 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
1763 | ||
34dc7c2f | 1764 | (void) remove_reference(hdr, hash_lock, tag); |
428870ff | 1765 | if (hdr->b_datacnt > 1) { |
34dc7c2f | 1766 | arc_buf_destroy(buf, FALSE, TRUE); |
428870ff BB |
1767 | } else { |
1768 | ASSERT(buf == hdr->b_buf); | |
1769 | ASSERT(buf->b_efunc == NULL); | |
34dc7c2f | 1770 | hdr->b_flags |= ARC_BUF_AVAILABLE; |
428870ff | 1771 | } |
34dc7c2f BB |
1772 | mutex_exit(hash_lock); |
1773 | } else if (HDR_IO_IN_PROGRESS(hdr)) { | |
1774 | int destroy_hdr; | |
1775 | /* | |
1776 | * We are in the middle of an async write. Don't destroy | |
1777 | * this buffer unless the write completes before we finish | |
1778 | * decrementing the reference count. | |
1779 | */ | |
1780 | mutex_enter(&arc_eviction_mtx); | |
1781 | (void) remove_reference(hdr, NULL, tag); | |
1782 | ASSERT(refcount_is_zero(&hdr->b_refcnt)); | |
1783 | destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); | |
1784 | mutex_exit(&arc_eviction_mtx); | |
1785 | if (destroy_hdr) | |
1786 | arc_hdr_destroy(hdr); | |
1787 | } else { | |
428870ff | 1788 | if (remove_reference(hdr, NULL, tag) > 0) |
34dc7c2f | 1789 | arc_buf_destroy(buf, FALSE, TRUE); |
428870ff | 1790 | else |
34dc7c2f | 1791 | arc_hdr_destroy(hdr); |
34dc7c2f BB |
1792 | } |
1793 | } | |
1794 | ||
13fe0198 | 1795 | boolean_t |
34dc7c2f BB |
1796 | arc_buf_remove_ref(arc_buf_t *buf, void* tag) |
1797 | { | |
1798 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
b4f7f105 | 1799 | kmutex_t *hash_lock = NULL; |
13fe0198 | 1800 | boolean_t no_callback = (buf->b_efunc == NULL); |
34dc7c2f BB |
1801 | |
1802 | if (hdr->b_state == arc_anon) { | |
428870ff | 1803 | ASSERT(hdr->b_datacnt == 1); |
34dc7c2f BB |
1804 | arc_buf_free(buf, tag); |
1805 | return (no_callback); | |
1806 | } | |
1807 | ||
b4f7f105 | 1808 | hash_lock = HDR_LOCK(hdr); |
34dc7c2f | 1809 | mutex_enter(hash_lock); |
428870ff BB |
1810 | hdr = buf->b_hdr; |
1811 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
34dc7c2f BB |
1812 | ASSERT(hdr->b_state != arc_anon); |
1813 | ASSERT(buf->b_data != NULL); | |
1814 | ||
1815 | (void) remove_reference(hdr, hash_lock, tag); | |
1816 | if (hdr->b_datacnt > 1) { | |
1817 | if (no_callback) | |
1818 | arc_buf_destroy(buf, FALSE, TRUE); | |
1819 | } else if (no_callback) { | |
1820 | ASSERT(hdr->b_buf == buf && buf->b_next == NULL); | |
428870ff | 1821 | ASSERT(buf->b_efunc == NULL); |
34dc7c2f BB |
1822 | hdr->b_flags |= ARC_BUF_AVAILABLE; |
1823 | } | |
1824 | ASSERT(no_callback || hdr->b_datacnt > 1 || | |
1825 | refcount_is_zero(&hdr->b_refcnt)); | |
1826 | mutex_exit(hash_lock); | |
1827 | return (no_callback); | |
1828 | } | |
1829 | ||
1830 | int | |
1831 | arc_buf_size(arc_buf_t *buf) | |
1832 | { | |
1833 | return (buf->b_hdr->b_size); | |
1834 | } | |
1835 | ||
1eb5bfa3 GW |
1836 | /* |
1837 | * Called from the DMU to determine if the current buffer should be | |
1838 | * evicted. In order to ensure proper locking, the eviction must be initiated | |
1839 | * from the DMU. Return true if the buffer is associated with user data and | |
1840 | * duplicate buffers still exist. | |
1841 | */ | |
1842 | boolean_t | |
1843 | arc_buf_eviction_needed(arc_buf_t *buf) | |
1844 | { | |
1845 | arc_buf_hdr_t *hdr; | |
1846 | boolean_t evict_needed = B_FALSE; | |
1847 | ||
1848 | if (zfs_disable_dup_eviction) | |
1849 | return (B_FALSE); | |
1850 | ||
1851 | mutex_enter(&buf->b_evict_lock); | |
1852 | hdr = buf->b_hdr; | |
1853 | if (hdr == NULL) { | |
1854 | /* | |
1855 | * We are in arc_do_user_evicts(); let that function | |
1856 | * perform the eviction. | |
1857 | */ | |
1858 | ASSERT(buf->b_data == NULL); | |
1859 | mutex_exit(&buf->b_evict_lock); | |
1860 | return (B_FALSE); | |
1861 | } else if (buf->b_data == NULL) { | |
1862 | /* | |
1863 | * We have already been added to the arc eviction list; | |
1864 | * recommend eviction. | |
1865 | */ | |
1866 | ASSERT3P(hdr, ==, &arc_eviction_hdr); | |
1867 | mutex_exit(&buf->b_evict_lock); | |
1868 | return (B_TRUE); | |
1869 | } | |
1870 | ||
1871 | if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) | |
1872 | evict_needed = B_TRUE; | |
1873 | ||
1874 | mutex_exit(&buf->b_evict_lock); | |
1875 | return (evict_needed); | |
1876 | } | |
1877 | ||
34dc7c2f BB |
1878 | /* |
1879 | * Evict buffers from list until we've removed the specified number of | |
1880 | * bytes. Move the removed buffers to the appropriate evict state. | |
1881 | * If the recycle flag is set, then attempt to "recycle" a buffer: | |
1882 | * - look for a buffer to evict that is `bytes' long. | |
1883 | * - return the data block from this buffer rather than freeing it. | |
1884 | * This flag is used by callers that are trying to make space for a | |
1885 | * new buffer in a full arc cache. | |
1886 | * | |
1887 | * This function makes a "best effort". It skips over any buffers | |
1888 | * it can't get a hash_lock on, and so may not catch all candidates. | |
1889 | * It may also return without evicting as much space as requested. | |
1890 | */ | |
1891 | static void * | |
d164b209 | 1892 | arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, |
34dc7c2f BB |
1893 | arc_buf_contents_t type) |
1894 | { | |
1895 | arc_state_t *evicted_state; | |
1896 | uint64_t bytes_evicted = 0, skipped = 0, missed = 0; | |
1897 | arc_buf_hdr_t *ab, *ab_prev = NULL; | |
1898 | list_t *list = &state->arcs_list[type]; | |
1899 | kmutex_t *hash_lock; | |
1900 | boolean_t have_lock; | |
1901 | void *stolen = NULL; | |
e8b96c60 MA |
1902 | arc_buf_hdr_t marker = {{{ 0 }}}; |
1903 | int count = 0; | |
34dc7c2f BB |
1904 | |
1905 | ASSERT(state == arc_mru || state == arc_mfu); | |
1906 | ||
1907 | evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; | |
1908 | ||
da8ccd0e | 1909 | top: |
34dc7c2f BB |
1910 | mutex_enter(&state->arcs_mtx); |
1911 | mutex_enter(&evicted_state->arcs_mtx); | |
1912 | ||
1913 | for (ab = list_tail(list); ab; ab = ab_prev) { | |
1914 | ab_prev = list_prev(list, ab); | |
1915 | /* prefetch buffers have a minimum lifespan */ | |
1916 | if (HDR_IO_IN_PROGRESS(ab) || | |
1917 | (spa && ab->b_spa != spa) || | |
1918 | (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && | |
428870ff | 1919 | ddi_get_lbolt() - ab->b_arc_access < |
bce45ec9 | 1920 | zfs_arc_min_prefetch_lifespan)) { |
34dc7c2f BB |
1921 | skipped++; |
1922 | continue; | |
1923 | } | |
1924 | /* "lookahead" for better eviction candidate */ | |
1925 | if (recycle && ab->b_size != bytes && | |
1926 | ab_prev && ab_prev->b_size == bytes) | |
1927 | continue; | |
e8b96c60 MA |
1928 | |
1929 | /* ignore markers */ | |
1930 | if (ab->b_spa == 0) | |
1931 | continue; | |
1932 | ||
1933 | /* | |
1934 | * It may take a long time to evict all the bufs requested. | |
1935 | * To avoid blocking all arc activity, periodically drop | |
1936 | * the arcs_mtx and give other threads a chance to run | |
1937 | * before reacquiring the lock. | |
1938 | * | |
1939 | * If we are looking for a buffer to recycle, we are in | |
1940 | * the hot code path, so don't sleep. | |
1941 | */ | |
1942 | if (!recycle && count++ > arc_evict_iterations) { | |
1943 | list_insert_after(list, ab, &marker); | |
1944 | mutex_exit(&evicted_state->arcs_mtx); | |
1945 | mutex_exit(&state->arcs_mtx); | |
1946 | kpreempt(KPREEMPT_SYNC); | |
1947 | mutex_enter(&state->arcs_mtx); | |
1948 | mutex_enter(&evicted_state->arcs_mtx); | |
1949 | ab_prev = list_prev(list, &marker); | |
1950 | list_remove(list, &marker); | |
1951 | count = 0; | |
1952 | continue; | |
1953 | } | |
1954 | ||
34dc7c2f BB |
1955 | hash_lock = HDR_LOCK(ab); |
1956 | have_lock = MUTEX_HELD(hash_lock); | |
1957 | if (have_lock || mutex_tryenter(hash_lock)) { | |
c99c9001 | 1958 | ASSERT0(refcount_count(&ab->b_refcnt)); |
34dc7c2f BB |
1959 | ASSERT(ab->b_datacnt > 0); |
1960 | while (ab->b_buf) { | |
1961 | arc_buf_t *buf = ab->b_buf; | |
428870ff | 1962 | if (!mutex_tryenter(&buf->b_evict_lock)) { |
b128c09f BB |
1963 | missed += 1; |
1964 | break; | |
1965 | } | |
34dc7c2f BB |
1966 | if (buf->b_data) { |
1967 | bytes_evicted += ab->b_size; | |
1968 | if (recycle && ab->b_type == type && | |
1969 | ab->b_size == bytes && | |
1970 | !HDR_L2_WRITING(ab)) { | |
1971 | stolen = buf->b_data; | |
1972 | recycle = FALSE; | |
1973 | } | |
1974 | } | |
1975 | if (buf->b_efunc) { | |
1976 | mutex_enter(&arc_eviction_mtx); | |
1977 | arc_buf_destroy(buf, | |
1978 | buf->b_data == stolen, FALSE); | |
1979 | ab->b_buf = buf->b_next; | |
1980 | buf->b_hdr = &arc_eviction_hdr; | |
1981 | buf->b_next = arc_eviction_list; | |
1982 | arc_eviction_list = buf; | |
1983 | mutex_exit(&arc_eviction_mtx); | |
428870ff | 1984 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f | 1985 | } else { |
428870ff | 1986 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
1987 | arc_buf_destroy(buf, |
1988 | buf->b_data == stolen, TRUE); | |
1989 | } | |
1990 | } | |
428870ff BB |
1991 | |
1992 | if (ab->b_l2hdr) { | |
1993 | ARCSTAT_INCR(arcstat_evict_l2_cached, | |
1994 | ab->b_size); | |
1995 | } else { | |
1996 | if (l2arc_write_eligible(ab->b_spa, ab)) { | |
1997 | ARCSTAT_INCR(arcstat_evict_l2_eligible, | |
1998 | ab->b_size); | |
1999 | } else { | |
2000 | ARCSTAT_INCR( | |
2001 | arcstat_evict_l2_ineligible, | |
2002 | ab->b_size); | |
2003 | } | |
2004 | } | |
2005 | ||
b128c09f BB |
2006 | if (ab->b_datacnt == 0) { |
2007 | arc_change_state(evicted_state, ab, hash_lock); | |
2008 | ASSERT(HDR_IN_HASH_TABLE(ab)); | |
2009 | ab->b_flags |= ARC_IN_HASH_TABLE; | |
2010 | ab->b_flags &= ~ARC_BUF_AVAILABLE; | |
2011 | DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); | |
2012 | } | |
34dc7c2f BB |
2013 | if (!have_lock) |
2014 | mutex_exit(hash_lock); | |
2015 | if (bytes >= 0 && bytes_evicted >= bytes) | |
2016 | break; | |
2017 | } else { | |
2018 | missed += 1; | |
2019 | } | |
2020 | } | |
2021 | ||
2022 | mutex_exit(&evicted_state->arcs_mtx); | |
2023 | mutex_exit(&state->arcs_mtx); | |
2024 | ||
da8ccd0e PS |
2025 | if (list == &state->arcs_list[ARC_BUFC_DATA] && |
2026 | (bytes < 0 || bytes_evicted < bytes)) { | |
2027 | /* Prevent second pass from recycling metadata into data */ | |
2028 | recycle = FALSE; | |
2029 | type = ARC_BUFC_METADATA; | |
2030 | list = &state->arcs_list[type]; | |
2031 | goto top; | |
2032 | } | |
2033 | ||
34dc7c2f | 2034 | if (bytes_evicted < bytes) |
3f504482 | 2035 | dprintf("only evicted %lld bytes from %x\n", |
34dc7c2f BB |
2036 | (longlong_t)bytes_evicted, state); |
2037 | ||
2038 | if (skipped) | |
2039 | ARCSTAT_INCR(arcstat_evict_skip, skipped); | |
2040 | ||
2041 | if (missed) | |
2042 | ARCSTAT_INCR(arcstat_mutex_miss, missed); | |
2043 | ||
2044 | /* | |
e8b96c60 MA |
2045 | * Note: we have just evicted some data into the ghost state, |
2046 | * potentially putting the ghost size over the desired size. Rather | |
2047 | * that evicting from the ghost list in this hot code path, leave | |
2048 | * this chore to the arc_reclaim_thread(). | |
34dc7c2f | 2049 | */ |
34dc7c2f BB |
2050 | |
2051 | return (stolen); | |
2052 | } | |
2053 | ||
2054 | /* | |
2055 | * Remove buffers from list until we've removed the specified number of | |
2056 | * bytes. Destroy the buffers that are removed. | |
2057 | */ | |
2058 | static void | |
68121a03 BB |
2059 | arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes, |
2060 | arc_buf_contents_t type) | |
34dc7c2f BB |
2061 | { |
2062 | arc_buf_hdr_t *ab, *ab_prev; | |
2598c001 | 2063 | arc_buf_hdr_t marker; |
68121a03 | 2064 | list_t *list = &state->arcs_list[type]; |
34dc7c2f BB |
2065 | kmutex_t *hash_lock; |
2066 | uint64_t bytes_deleted = 0; | |
2067 | uint64_t bufs_skipped = 0; | |
e8b96c60 | 2068 | int count = 0; |
34dc7c2f BB |
2069 | |
2070 | ASSERT(GHOST_STATE(state)); | |
d1d7e268 | 2071 | bzero(&marker, sizeof (marker)); |
34dc7c2f BB |
2072 | top: |
2073 | mutex_enter(&state->arcs_mtx); | |
2074 | for (ab = list_tail(list); ab; ab = ab_prev) { | |
2075 | ab_prev = list_prev(list, ab); | |
e8b96c60 MA |
2076 | if (ab->b_type > ARC_BUFC_NUMTYPES) |
2077 | panic("invalid ab=%p", (void *)ab); | |
34dc7c2f BB |
2078 | if (spa && ab->b_spa != spa) |
2079 | continue; | |
572e2857 BB |
2080 | |
2081 | /* ignore markers */ | |
2082 | if (ab->b_spa == 0) | |
2083 | continue; | |
2084 | ||
34dc7c2f | 2085 | hash_lock = HDR_LOCK(ab); |
428870ff BB |
2086 | /* caller may be trying to modify this buffer, skip it */ |
2087 | if (MUTEX_HELD(hash_lock)) | |
2088 | continue; | |
e8b96c60 MA |
2089 | |
2090 | /* | |
2091 | * It may take a long time to evict all the bufs requested. | |
2092 | * To avoid blocking all arc activity, periodically drop | |
2093 | * the arcs_mtx and give other threads a chance to run | |
2094 | * before reacquiring the lock. | |
2095 | */ | |
2096 | if (count++ > arc_evict_iterations) { | |
2097 | list_insert_after(list, ab, &marker); | |
2098 | mutex_exit(&state->arcs_mtx); | |
2099 | kpreempt(KPREEMPT_SYNC); | |
2100 | mutex_enter(&state->arcs_mtx); | |
2101 | ab_prev = list_prev(list, &marker); | |
2102 | list_remove(list, &marker); | |
2103 | count = 0; | |
2104 | continue; | |
2105 | } | |
34dc7c2f BB |
2106 | if (mutex_tryenter(hash_lock)) { |
2107 | ASSERT(!HDR_IO_IN_PROGRESS(ab)); | |
2108 | ASSERT(ab->b_buf == NULL); | |
2109 | ARCSTAT_BUMP(arcstat_deleted); | |
2110 | bytes_deleted += ab->b_size; | |
2111 | ||
2112 | if (ab->b_l2hdr != NULL) { | |
2113 | /* | |
2114 | * This buffer is cached on the 2nd Level ARC; | |
2115 | * don't destroy the header. | |
2116 | */ | |
2117 | arc_change_state(arc_l2c_only, ab, hash_lock); | |
2118 | mutex_exit(hash_lock); | |
2119 | } else { | |
2120 | arc_change_state(arc_anon, ab, hash_lock); | |
2121 | mutex_exit(hash_lock); | |
2122 | arc_hdr_destroy(ab); | |
2123 | } | |
2124 | ||
2125 | DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); | |
2126 | if (bytes >= 0 && bytes_deleted >= bytes) | |
2127 | break; | |
572e2857 BB |
2128 | } else if (bytes < 0) { |
2129 | /* | |
2130 | * Insert a list marker and then wait for the | |
2131 | * hash lock to become available. Once its | |
2132 | * available, restart from where we left off. | |
2133 | */ | |
2134 | list_insert_after(list, ab, &marker); | |
2135 | mutex_exit(&state->arcs_mtx); | |
2136 | mutex_enter(hash_lock); | |
2137 | mutex_exit(hash_lock); | |
2138 | mutex_enter(&state->arcs_mtx); | |
2139 | ab_prev = list_prev(list, &marker); | |
2140 | list_remove(list, &marker); | |
e8b96c60 | 2141 | } else { |
34dc7c2f | 2142 | bufs_skipped += 1; |
e8b96c60 | 2143 | } |
34dc7c2f BB |
2144 | } |
2145 | mutex_exit(&state->arcs_mtx); | |
2146 | ||
2147 | if (list == &state->arcs_list[ARC_BUFC_DATA] && | |
2148 | (bytes < 0 || bytes_deleted < bytes)) { | |
2149 | list = &state->arcs_list[ARC_BUFC_METADATA]; | |
2150 | goto top; | |
2151 | } | |
2152 | ||
2153 | if (bufs_skipped) { | |
2154 | ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); | |
2155 | ASSERT(bytes >= 0); | |
2156 | } | |
2157 | ||
2158 | if (bytes_deleted < bytes) | |
3f504482 | 2159 | dprintf("only deleted %lld bytes from %p\n", |
34dc7c2f BB |
2160 | (longlong_t)bytes_deleted, state); |
2161 | } | |
2162 | ||
2163 | static void | |
2164 | arc_adjust(void) | |
2165 | { | |
d164b209 BB |
2166 | int64_t adjustment, delta; |
2167 | ||
2168 | /* | |
2169 | * Adjust MRU size | |
2170 | */ | |
34dc7c2f | 2171 | |
572e2857 | 2172 | adjustment = MIN((int64_t)(arc_size - arc_c), |
77765b54 | 2173 | (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size - arc_p)); |
34dc7c2f | 2174 | |
da8ccd0e PS |
2175 | if (adjustment > 0 && arc_mru->arcs_size > 0) { |
2176 | delta = MIN(arc_mru->arcs_size, adjustment); | |
b8864a23 | 2177 | (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); |
34dc7c2f BB |
2178 | } |
2179 | ||
d164b209 BB |
2180 | /* |
2181 | * Adjust MFU size | |
2182 | */ | |
34dc7c2f | 2183 | |
d164b209 BB |
2184 | adjustment = arc_size - arc_c; |
2185 | ||
da8ccd0e PS |
2186 | if (adjustment > 0 && arc_mfu->arcs_size > 0) { |
2187 | delta = MIN(arc_mfu->arcs_size, adjustment); | |
b8864a23 | 2188 | (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); |
d164b209 | 2189 | } |
34dc7c2f | 2190 | |
d164b209 BB |
2191 | /* |
2192 | * Adjust ghost lists | |
2193 | */ | |
34dc7c2f | 2194 | |
d164b209 BB |
2195 | adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; |
2196 | ||
2197 | if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { | |
2198 | delta = MIN(arc_mru_ghost->arcs_size, adjustment); | |
68121a03 | 2199 | arc_evict_ghost(arc_mru_ghost, 0, delta, ARC_BUFC_DATA); |
d164b209 | 2200 | } |
34dc7c2f | 2201 | |
d164b209 BB |
2202 | adjustment = |
2203 | arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; | |
34dc7c2f | 2204 | |
d164b209 BB |
2205 | if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { |
2206 | delta = MIN(arc_mfu_ghost->arcs_size, adjustment); | |
68121a03 | 2207 | arc_evict_ghost(arc_mfu_ghost, 0, delta, ARC_BUFC_DATA); |
34dc7c2f BB |
2208 | } |
2209 | } | |
2210 | ||
ab26409d BB |
2211 | /* |
2212 | * Request that arc user drop references so that N bytes can be released | |
2213 | * from the cache. This provides a mechanism to ensure the arc can honor | |
2214 | * the arc_meta_limit and reclaim buffers which are pinned in the cache | |
2215 | * by higher layers. (i.e. the zpl) | |
2216 | */ | |
2217 | static void | |
2218 | arc_do_user_prune(int64_t adjustment) | |
2219 | { | |
2220 | arc_prune_func_t *func; | |
2221 | void *private; | |
2222 | arc_prune_t *cp, *np; | |
2223 | ||
2224 | mutex_enter(&arc_prune_mtx); | |
2225 | ||
2226 | cp = list_head(&arc_prune_list); | |
2227 | while (cp != NULL) { | |
2228 | func = cp->p_pfunc; | |
2229 | private = cp->p_private; | |
2230 | np = list_next(&arc_prune_list, cp); | |
2231 | refcount_add(&cp->p_refcnt, func); | |
2232 | mutex_exit(&arc_prune_mtx); | |
2233 | ||
2234 | if (func != NULL) | |
2235 | func(adjustment, private); | |
2236 | ||
2237 | mutex_enter(&arc_prune_mtx); | |
2238 | ||
2239 | /* User removed prune callback concurrently with execution */ | |
2240 | if (refcount_remove(&cp->p_refcnt, func) == 0) { | |
2241 | ASSERT(!list_link_active(&cp->p_node)); | |
2242 | refcount_destroy(&cp->p_refcnt); | |
2243 | kmem_free(cp, sizeof (*cp)); | |
2244 | } | |
2245 | ||
2246 | cp = np; | |
2247 | } | |
2248 | ||
2249 | ARCSTAT_BUMP(arcstat_prune); | |
2250 | mutex_exit(&arc_prune_mtx); | |
2251 | } | |
2252 | ||
34dc7c2f BB |
2253 | static void |
2254 | arc_do_user_evicts(void) | |
2255 | { | |
2256 | mutex_enter(&arc_eviction_mtx); | |
2257 | while (arc_eviction_list != NULL) { | |
2258 | arc_buf_t *buf = arc_eviction_list; | |
2259 | arc_eviction_list = buf->b_next; | |
428870ff | 2260 | mutex_enter(&buf->b_evict_lock); |
34dc7c2f | 2261 | buf->b_hdr = NULL; |
428870ff | 2262 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
2263 | mutex_exit(&arc_eviction_mtx); |
2264 | ||
2265 | if (buf->b_efunc != NULL) | |
2266 | VERIFY(buf->b_efunc(buf) == 0); | |
2267 | ||
2268 | buf->b_efunc = NULL; | |
2269 | buf->b_private = NULL; | |
2270 | kmem_cache_free(buf_cache, buf); | |
2271 | mutex_enter(&arc_eviction_mtx); | |
2272 | } | |
2273 | mutex_exit(&arc_eviction_mtx); | |
2274 | } | |
2275 | ||
ab26409d BB |
2276 | /* |
2277 | * Evict only meta data objects from the cache leaving the data objects. | |
2278 | * This is only used to enforce the tunable arc_meta_limit, if we are | |
2279 | * unable to evict enough buffers notify the user via the prune callback. | |
2280 | */ | |
94520ca4 PS |
2281 | static void |
2282 | arc_adjust_meta(void) | |
ab26409d | 2283 | { |
94520ca4 | 2284 | int64_t adjustmnt, delta; |
ab26409d | 2285 | |
94520ca4 PS |
2286 | /* |
2287 | * This slightly differs than the way we evict from the mru in | |
2288 | * arc_adjust because we don't have a "target" value (i.e. no | |
2289 | * "meta" arc_p). As a result, I think we can completely | |
2290 | * cannibalize the metadata in the MRU before we evict the | |
2291 | * metadata from the MFU. I think we probably need to implement a | |
2292 | * "metadata arc_p" value to do this properly. | |
2293 | */ | |
2294 | adjustmnt = arc_meta_used - arc_meta_limit; | |
2295 | ||
2296 | if (adjustmnt > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { | |
2297 | delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustmnt); | |
ab26409d | 2298 | arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_METADATA); |
94520ca4 | 2299 | adjustmnt -= delta; |
ab26409d BB |
2300 | } |
2301 | ||
94520ca4 PS |
2302 | /* |
2303 | * We can't afford to recalculate adjustmnt here. If we do, | |
2304 | * new metadata buffers can sneak into the MRU or ANON lists, | |
2305 | * thus penalize the MFU metadata. Although the fudge factor is | |
2306 | * small, it has been empirically shown to be significant for | |
2307 | * certain workloads (e.g. creating many empty directories). As | |
2308 | * such, we use the original calculation for adjustmnt, and | |
2309 | * simply decrement the amount of data evicted from the MRU. | |
2310 | */ | |
2311 | ||
2312 | if (adjustmnt > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { | |
2313 | delta = MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA], adjustmnt); | |
ab26409d | 2314 | arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_METADATA); |
ab26409d BB |
2315 | } |
2316 | ||
94520ca4 PS |
2317 | adjustmnt = arc_mru->arcs_lsize[ARC_BUFC_METADATA] + |
2318 | arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] - arc_meta_limit; | |
2319 | ||
2320 | if (adjustmnt > 0 && arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] > 0) { | |
2321 | delta = MIN(adjustmnt, | |
2322 | arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA]); | |
2323 | arc_evict_ghost(arc_mru_ghost, 0, delta, ARC_BUFC_METADATA); | |
2324 | } | |
2325 | ||
2326 | adjustmnt = arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] + | |
2327 | arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA] - arc_meta_limit; | |
2328 | ||
2329 | if (adjustmnt > 0 && arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA] > 0) { | |
2330 | delta = MIN(adjustmnt, | |
2331 | arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA]); | |
2332 | arc_evict_ghost(arc_mfu_ghost, 0, delta, ARC_BUFC_METADATA); | |
2333 | } | |
2334 | ||
2335 | if (arc_meta_used > arc_meta_limit) | |
bce45ec9 | 2336 | arc_do_user_prune(zfs_arc_meta_prune); |
ab26409d BB |
2337 | } |
2338 | ||
34dc7c2f BB |
2339 | /* |
2340 | * Flush all *evictable* data from the cache for the given spa. | |
2341 | * NOTE: this will not touch "active" (i.e. referenced) data. | |
2342 | */ | |
2343 | void | |
2344 | arc_flush(spa_t *spa) | |
2345 | { | |
d164b209 BB |
2346 | uint64_t guid = 0; |
2347 | ||
2348 | if (spa) | |
3541dc6d | 2349 | guid = spa_load_guid(spa); |
d164b209 | 2350 | |
34dc7c2f | 2351 | while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) { |
d164b209 | 2352 | (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); |
34dc7c2f BB |
2353 | if (spa) |
2354 | break; | |
2355 | } | |
2356 | while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) { | |
d164b209 | 2357 | (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); |
34dc7c2f BB |
2358 | if (spa) |
2359 | break; | |
2360 | } | |
2361 | while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) { | |
d164b209 | 2362 | (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); |
34dc7c2f BB |
2363 | if (spa) |
2364 | break; | |
2365 | } | |
2366 | while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) { | |
d164b209 | 2367 | (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); |
34dc7c2f BB |
2368 | if (spa) |
2369 | break; | |
2370 | } | |
2371 | ||
68121a03 BB |
2372 | arc_evict_ghost(arc_mru_ghost, guid, -1, ARC_BUFC_DATA); |
2373 | arc_evict_ghost(arc_mfu_ghost, guid, -1, ARC_BUFC_DATA); | |
34dc7c2f BB |
2374 | |
2375 | mutex_enter(&arc_reclaim_thr_lock); | |
2376 | arc_do_user_evicts(); | |
2377 | mutex_exit(&arc_reclaim_thr_lock); | |
2378 | ASSERT(spa || arc_eviction_list == NULL); | |
2379 | } | |
2380 | ||
34dc7c2f | 2381 | void |
302f753f | 2382 | arc_shrink(uint64_t bytes) |
34dc7c2f BB |
2383 | { |
2384 | if (arc_c > arc_c_min) { | |
2385 | uint64_t to_free; | |
2386 | ||
bce45ec9 | 2387 | to_free = bytes ? bytes : arc_c >> zfs_arc_shrink_shift; |
302f753f | 2388 | |
34dc7c2f BB |
2389 | if (arc_c > arc_c_min + to_free) |
2390 | atomic_add_64(&arc_c, -to_free); | |
2391 | else | |
2392 | arc_c = arc_c_min; | |
2393 | ||
39e055c4 PS |
2394 | to_free = bytes ? bytes : arc_p >> zfs_arc_shrink_shift; |
2395 | ||
f521ce1b | 2396 | if (arc_p > to_free) |
39e055c4 PS |
2397 | atomic_add_64(&arc_p, -to_free); |
2398 | else | |
f521ce1b | 2399 | arc_p = 0; |
39e055c4 | 2400 | |
34dc7c2f BB |
2401 | if (arc_c > arc_size) |
2402 | arc_c = MAX(arc_size, arc_c_min); | |
2403 | if (arc_p > arc_c) | |
2404 | arc_p = (arc_c >> 1); | |
2405 | ASSERT(arc_c >= arc_c_min); | |
2406 | ASSERT((int64_t)arc_p >= 0); | |
2407 | } | |
2408 | ||
2409 | if (arc_size > arc_c) | |
2410 | arc_adjust(); | |
2411 | } | |
2412 | ||
34dc7c2f | 2413 | static void |
302f753f | 2414 | arc_kmem_reap_now(arc_reclaim_strategy_t strat, uint64_t bytes) |
34dc7c2f BB |
2415 | { |
2416 | size_t i; | |
2417 | kmem_cache_t *prev_cache = NULL; | |
2418 | kmem_cache_t *prev_data_cache = NULL; | |
2419 | extern kmem_cache_t *zio_buf_cache[]; | |
2420 | extern kmem_cache_t *zio_data_buf_cache[]; | |
34dc7c2f BB |
2421 | |
2422 | /* | |
2423 | * An aggressive reclamation will shrink the cache size as well as | |
2424 | * reap free buffers from the arc kmem caches. | |
2425 | */ | |
2426 | if (strat == ARC_RECLAIM_AGGR) | |
302f753f | 2427 | arc_shrink(bytes); |
34dc7c2f BB |
2428 | |
2429 | for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { | |
2430 | if (zio_buf_cache[i] != prev_cache) { | |
2431 | prev_cache = zio_buf_cache[i]; | |
2432 | kmem_cache_reap_now(zio_buf_cache[i]); | |
2433 | } | |
2434 | if (zio_data_buf_cache[i] != prev_data_cache) { | |
2435 | prev_data_cache = zio_data_buf_cache[i]; | |
2436 | kmem_cache_reap_now(zio_data_buf_cache[i]); | |
2437 | } | |
2438 | } | |
ab26409d | 2439 | |
34dc7c2f BB |
2440 | kmem_cache_reap_now(buf_cache); |
2441 | kmem_cache_reap_now(hdr_cache); | |
2442 | } | |
2443 | ||
302f753f BB |
2444 | /* |
2445 | * Unlike other ZFS implementations this thread is only responsible for | |
2446 | * adapting the target ARC size on Linux. The responsibility for memory | |
2447 | * reclamation has been entirely delegated to the arc_shrinker_func() | |
2448 | * which is registered with the VM. To reflect this change in behavior | |
2449 | * the arc_reclaim thread has been renamed to arc_adapt. | |
2450 | */ | |
34dc7c2f | 2451 | static void |
302f753f | 2452 | arc_adapt_thread(void) |
34dc7c2f | 2453 | { |
34dc7c2f BB |
2454 | callb_cpr_t cpr; |
2455 | ||
2456 | CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); | |
2457 | ||
2458 | mutex_enter(&arc_reclaim_thr_lock); | |
2459 | while (arc_thread_exit == 0) { | |
302f753f BB |
2460 | #ifndef _KERNEL |
2461 | arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; | |
2462 | ||
2463 | if (spa_get_random(100) == 0) { | |
34dc7c2f BB |
2464 | |
2465 | if (arc_no_grow) { | |
2466 | if (last_reclaim == ARC_RECLAIM_CONS) { | |
2467 | last_reclaim = ARC_RECLAIM_AGGR; | |
2468 | } else { | |
2469 | last_reclaim = ARC_RECLAIM_CONS; | |
2470 | } | |
2471 | } else { | |
2472 | arc_no_grow = TRUE; | |
2473 | last_reclaim = ARC_RECLAIM_AGGR; | |
2474 | membar_producer(); | |
2475 | } | |
2476 | ||
2477 | /* reset the growth delay for every reclaim */ | |
d1d7e268 MK |
2478 | arc_grow_time = ddi_get_lbolt() + |
2479 | (zfs_arc_grow_retry * hz); | |
34dc7c2f | 2480 | |
302f753f | 2481 | arc_kmem_reap_now(last_reclaim, 0); |
b128c09f | 2482 | arc_warm = B_TRUE; |
302f753f BB |
2483 | } |
2484 | #endif /* !_KERNEL */ | |
34dc7c2f | 2485 | |
302f753f | 2486 | /* No recent memory pressure allow the ARC to grow. */ |
0b75bdb3 CC |
2487 | if (arc_no_grow && |
2488 | ddi_time_after_eq(ddi_get_lbolt(), arc_grow_time)) | |
34dc7c2f | 2489 | arc_no_grow = FALSE; |
34dc7c2f | 2490 | |
94520ca4 | 2491 | arc_adjust_meta(); |
6a8f9b6b | 2492 | |
572e2857 | 2493 | arc_adjust(); |
34dc7c2f BB |
2494 | |
2495 | if (arc_eviction_list != NULL) | |
2496 | arc_do_user_evicts(); | |
2497 | ||
2498 | /* block until needed, or one second, whichever is shorter */ | |
2499 | CALLB_CPR_SAFE_BEGIN(&cpr); | |
5b63b3eb | 2500 | (void) cv_timedwait_interruptible(&arc_reclaim_thr_cv, |
428870ff | 2501 | &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz)); |
34dc7c2f | 2502 | CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); |
bce45ec9 BB |
2503 | |
2504 | ||
2505 | /* Allow the module options to be changed */ | |
2506 | if (zfs_arc_max > 64 << 20 && | |
2507 | zfs_arc_max < physmem * PAGESIZE && | |
2508 | zfs_arc_max != arc_c_max) | |
2509 | arc_c_max = zfs_arc_max; | |
2510 | ||
2511 | if (zfs_arc_min > 0 && | |
2512 | zfs_arc_min < arc_c_max && | |
2513 | zfs_arc_min != arc_c_min) | |
2514 | arc_c_min = zfs_arc_min; | |
2515 | ||
2516 | if (zfs_arc_meta_limit > 0 && | |
2517 | zfs_arc_meta_limit <= arc_c_max && | |
2518 | zfs_arc_meta_limit != arc_meta_limit) | |
2519 | arc_meta_limit = zfs_arc_meta_limit; | |
2520 | ||
2521 | ||
2522 | ||
34dc7c2f BB |
2523 | } |
2524 | ||
2525 | arc_thread_exit = 0; | |
2526 | cv_broadcast(&arc_reclaim_thr_cv); | |
2527 | CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ | |
2528 | thread_exit(); | |
2529 | } | |
2530 | ||
7cb67b45 BB |
2531 | #ifdef _KERNEL |
2532 | /* | |
302f753f BB |
2533 | * Determine the amount of memory eligible for eviction contained in the |
2534 | * ARC. All clean data reported by the ghost lists can always be safely | |
2535 | * evicted. Due to arc_c_min, the same does not hold for all clean data | |
2536 | * contained by the regular mru and mfu lists. | |
2537 | * | |
2538 | * In the case of the regular mru and mfu lists, we need to report as | |
2539 | * much clean data as possible, such that evicting that same reported | |
2540 | * data will not bring arc_size below arc_c_min. Thus, in certain | |
2541 | * circumstances, the total amount of clean data in the mru and mfu | |
2542 | * lists might not actually be evictable. | |
2543 | * | |
2544 | * The following two distinct cases are accounted for: | |
2545 | * | |
2546 | * 1. The sum of the amount of dirty data contained by both the mru and | |
2547 | * mfu lists, plus the ARC's other accounting (e.g. the anon list), | |
2548 | * is greater than or equal to arc_c_min. | |
2549 | * (i.e. amount of dirty data >= arc_c_min) | |
2550 | * | |
2551 | * This is the easy case; all clean data contained by the mru and mfu | |
2552 | * lists is evictable. Evicting all clean data can only drop arc_size | |
2553 | * to the amount of dirty data, which is greater than arc_c_min. | |
2554 | * | |
2555 | * 2. The sum of the amount of dirty data contained by both the mru and | |
2556 | * mfu lists, plus the ARC's other accounting (e.g. the anon list), | |
2557 | * is less than arc_c_min. | |
2558 | * (i.e. arc_c_min > amount of dirty data) | |
2559 | * | |
2560 | * 2.1. arc_size is greater than or equal arc_c_min. | |
2561 | * (i.e. arc_size >= arc_c_min > amount of dirty data) | |
2562 | * | |
2563 | * In this case, not all clean data from the regular mru and mfu | |
2564 | * lists is actually evictable; we must leave enough clean data | |
2565 | * to keep arc_size above arc_c_min. Thus, the maximum amount of | |
2566 | * evictable data from the two lists combined, is exactly the | |
2567 | * difference between arc_size and arc_c_min. | |
2568 | * | |
2569 | * 2.2. arc_size is less than arc_c_min | |
2570 | * (i.e. arc_c_min > arc_size > amount of dirty data) | |
2571 | * | |
2572 | * In this case, none of the data contained in the mru and mfu | |
2573 | * lists is evictable, even if it's clean. Since arc_size is | |
2574 | * already below arc_c_min, evicting any more would only | |
2575 | * increase this negative difference. | |
7cb67b45 | 2576 | */ |
302f753f BB |
2577 | static uint64_t |
2578 | arc_evictable_memory(void) { | |
2579 | uint64_t arc_clean = | |
2580 | arc_mru->arcs_lsize[ARC_BUFC_DATA] + | |
2581 | arc_mru->arcs_lsize[ARC_BUFC_METADATA] + | |
2582 | arc_mfu->arcs_lsize[ARC_BUFC_DATA] + | |
2583 | arc_mfu->arcs_lsize[ARC_BUFC_METADATA]; | |
2584 | uint64_t ghost_clean = | |
2585 | arc_mru_ghost->arcs_lsize[ARC_BUFC_DATA] + | |
2586 | arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] + | |
2587 | arc_mfu_ghost->arcs_lsize[ARC_BUFC_DATA] + | |
2588 | arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA]; | |
2589 | uint64_t arc_dirty = MAX((int64_t)arc_size - (int64_t)arc_clean, 0); | |
2590 | ||
2591 | if (arc_dirty >= arc_c_min) | |
2592 | return (ghost_clean + arc_clean); | |
2593 | ||
2594 | return (ghost_clean + MAX((int64_t)arc_size - (int64_t)arc_c_min, 0)); | |
2595 | } | |
2596 | ||
7e7baeca BB |
2597 | static int |
2598 | __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc) | |
7cb67b45 | 2599 | { |
302f753f | 2600 | uint64_t pages; |
7cb67b45 | 2601 | |
302f753f BB |
2602 | /* The arc is considered warm once reclaim has occurred */ |
2603 | if (unlikely(arc_warm == B_FALSE)) | |
2604 | arc_warm = B_TRUE; | |
7cb67b45 | 2605 | |
302f753f BB |
2606 | /* Return the potential number of reclaimable pages */ |
2607 | pages = btop(arc_evictable_memory()); | |
2608 | if (sc->nr_to_scan == 0) | |
2609 | return (pages); | |
3fd70ee6 BB |
2610 | |
2611 | /* Not allowed to perform filesystem reclaim */ | |
7e7baeca | 2612 | if (!(sc->gfp_mask & __GFP_FS)) |
3fd70ee6 BB |
2613 | return (-1); |
2614 | ||
7cb67b45 BB |
2615 | /* Reclaim in progress */ |
2616 | if (mutex_tryenter(&arc_reclaim_thr_lock) == 0) | |
2617 | return (-1); | |
2618 | ||
302f753f BB |
2619 | /* |
2620 | * Evict the requested number of pages by shrinking arc_c the | |
2621 | * requested amount. If there is nothing left to evict just | |
2622 | * reap whatever we can from the various arc slabs. | |
2623 | */ | |
2624 | if (pages > 0) { | |
2625 | arc_kmem_reap_now(ARC_RECLAIM_AGGR, ptob(sc->nr_to_scan)); | |
1e3cb67b | 2626 | pages = btop(arc_evictable_memory()); |
302f753f BB |
2627 | } else { |
2628 | arc_kmem_reap_now(ARC_RECLAIM_CONS, ptob(sc->nr_to_scan)); | |
1e3cb67b | 2629 | pages = -1; |
302f753f BB |
2630 | } |
2631 | ||
2632 | /* | |
2633 | * When direct reclaim is observed it usually indicates a rapid | |
2634 | * increase in memory pressure. This occurs because the kswapd | |
2635 | * threads were unable to asynchronously keep enough free memory | |
2636 | * available. In this case set arc_no_grow to briefly pause arc | |
2637 | * growth to avoid compounding the memory pressure. | |
2638 | */ | |
7cb67b45 | 2639 | if (current_is_kswapd()) { |
302f753f | 2640 | ARCSTAT_BUMP(arcstat_memory_indirect_count); |
7cb67b45 | 2641 | } else { |
302f753f | 2642 | arc_no_grow = B_TRUE; |
bce45ec9 | 2643 | arc_grow_time = ddi_get_lbolt() + (zfs_arc_grow_retry * hz); |
302f753f | 2644 | ARCSTAT_BUMP(arcstat_memory_direct_count); |
7cb67b45 BB |
2645 | } |
2646 | ||
7cb67b45 BB |
2647 | mutex_exit(&arc_reclaim_thr_lock); |
2648 | ||
1e3cb67b | 2649 | return (pages); |
7cb67b45 | 2650 | } |
7e7baeca | 2651 | SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func); |
7cb67b45 BB |
2652 | |
2653 | SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS); | |
2654 | #endif /* _KERNEL */ | |
2655 | ||
34dc7c2f BB |
2656 | /* |
2657 | * Adapt arc info given the number of bytes we are trying to add and | |
2658 | * the state that we are comming from. This function is only called | |
2659 | * when we are adding new content to the cache. | |
2660 | */ | |
2661 | static void | |
2662 | arc_adapt(int bytes, arc_state_t *state) | |
2663 | { | |
2664 | int mult; | |
2665 | ||
2666 | if (state == arc_l2c_only) | |
2667 | return; | |
2668 | ||
2669 | ASSERT(bytes > 0); | |
2670 | /* | |
2671 | * Adapt the target size of the MRU list: | |
2672 | * - if we just hit in the MRU ghost list, then increase | |
2673 | * the target size of the MRU list. | |
2674 | * - if we just hit in the MFU ghost list, then increase | |
2675 | * the target size of the MFU list by decreasing the | |
2676 | * target size of the MRU list. | |
2677 | */ | |
2678 | if (state == arc_mru_ghost) { | |
2679 | mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? | |
2680 | 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); | |
62422785 PS |
2681 | |
2682 | if (!zfs_arc_p_dampener_disable) | |
2683 | mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ | |
34dc7c2f | 2684 | |
f521ce1b | 2685 | arc_p = MIN(arc_c, arc_p + bytes * mult); |
34dc7c2f | 2686 | } else if (state == arc_mfu_ghost) { |
d164b209 BB |
2687 | uint64_t delta; |
2688 | ||
34dc7c2f BB |
2689 | mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? |
2690 | 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); | |
62422785 PS |
2691 | |
2692 | if (!zfs_arc_p_dampener_disable) | |
2693 | mult = MIN(mult, 10); | |
34dc7c2f | 2694 | |
d164b209 | 2695 | delta = MIN(bytes * mult, arc_p); |
f521ce1b | 2696 | arc_p = MAX(0, arc_p - delta); |
34dc7c2f BB |
2697 | } |
2698 | ASSERT((int64_t)arc_p >= 0); | |
2699 | ||
34dc7c2f BB |
2700 | if (arc_no_grow) |
2701 | return; | |
2702 | ||
2703 | if (arc_c >= arc_c_max) | |
2704 | return; | |
2705 | ||
2706 | /* | |
2707 | * If we're within (2 * maxblocksize) bytes of the target | |
2708 | * cache size, increment the target cache size | |
2709 | */ | |
2710 | if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { | |
2711 | atomic_add_64(&arc_c, (int64_t)bytes); | |
2712 | if (arc_c > arc_c_max) | |
2713 | arc_c = arc_c_max; | |
2714 | else if (state == arc_anon) | |
2715 | atomic_add_64(&arc_p, (int64_t)bytes); | |
2716 | if (arc_p > arc_c) | |
2717 | arc_p = arc_c; | |
2718 | } | |
2719 | ASSERT((int64_t)arc_p >= 0); | |
2720 | } | |
2721 | ||
2722 | /* | |
2723 | * Check if the cache has reached its limits and eviction is required | |
2724 | * prior to insert. | |
2725 | */ | |
2726 | static int | |
2727 | arc_evict_needed(arc_buf_contents_t type) | |
2728 | { | |
2729 | if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) | |
2730 | return (1); | |
2731 | ||
302f753f | 2732 | if (arc_no_grow) |
34dc7c2f BB |
2733 | return (1); |
2734 | ||
2735 | return (arc_size > arc_c); | |
2736 | } | |
2737 | ||
2738 | /* | |
2739 | * The buffer, supplied as the first argument, needs a data block. | |
2740 | * So, if we are at cache max, determine which cache should be victimized. | |
2741 | * We have the following cases: | |
2742 | * | |
2743 | * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> | |
2744 | * In this situation if we're out of space, but the resident size of the MFU is | |
2745 | * under the limit, victimize the MFU cache to satisfy this insertion request. | |
2746 | * | |
2747 | * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> | |
2748 | * Here, we've used up all of the available space for the MRU, so we need to | |
2749 | * evict from our own cache instead. Evict from the set of resident MRU | |
2750 | * entries. | |
2751 | * | |
2752 | * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> | |
2753 | * c minus p represents the MFU space in the cache, since p is the size of the | |
2754 | * cache that is dedicated to the MRU. In this situation there's still space on | |
2755 | * the MFU side, so the MRU side needs to be victimized. | |
2756 | * | |
2757 | * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> | |
2758 | * MFU's resident set is consuming more space than it has been allotted. In | |
2759 | * this situation, we must victimize our own cache, the MFU, for this insertion. | |
2760 | */ | |
2761 | static void | |
2762 | arc_get_data_buf(arc_buf_t *buf) | |
2763 | { | |
2764 | arc_state_t *state = buf->b_hdr->b_state; | |
2765 | uint64_t size = buf->b_hdr->b_size; | |
2766 | arc_buf_contents_t type = buf->b_hdr->b_type; | |
da8ccd0e PS |
2767 | arc_buf_contents_t evict = ARC_BUFC_DATA; |
2768 | boolean_t recycle = TRUE; | |
34dc7c2f BB |
2769 | |
2770 | arc_adapt(size, state); | |
2771 | ||
2772 | /* | |
2773 | * We have not yet reached cache maximum size, | |
2774 | * just allocate a new buffer. | |
2775 | */ | |
2776 | if (!arc_evict_needed(type)) { | |
2777 | if (type == ARC_BUFC_METADATA) { | |
2778 | buf->b_data = zio_buf_alloc(size); | |
cc7f677c | 2779 | arc_space_consume(size, ARC_SPACE_META); |
34dc7c2f BB |
2780 | } else { |
2781 | ASSERT(type == ARC_BUFC_DATA); | |
2782 | buf->b_data = zio_data_buf_alloc(size); | |
cc7f677c | 2783 | arc_space_consume(size, ARC_SPACE_DATA); |
34dc7c2f BB |
2784 | } |
2785 | goto out; | |
2786 | } | |
2787 | ||
2788 | /* | |
2789 | * If we are prefetching from the mfu ghost list, this buffer | |
2790 | * will end up on the mru list; so steal space from there. | |
2791 | */ | |
2792 | if (state == arc_mfu_ghost) | |
2793 | state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; | |
2794 | else if (state == arc_mru_ghost) | |
2795 | state = arc_mru; | |
2796 | ||
2797 | if (state == arc_mru || state == arc_anon) { | |
2798 | uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; | |
d164b209 | 2799 | state = (arc_mfu->arcs_lsize[type] >= size && |
34dc7c2f BB |
2800 | arc_p > mru_used) ? arc_mfu : arc_mru; |
2801 | } else { | |
2802 | /* MFU cases */ | |
2803 | uint64_t mfu_space = arc_c - arc_p; | |
d164b209 | 2804 | state = (arc_mru->arcs_lsize[type] >= size && |
34dc7c2f BB |
2805 | mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; |
2806 | } | |
ab26409d | 2807 | |
da8ccd0e PS |
2808 | /* |
2809 | * Evict data buffers prior to metadata buffers, unless we're | |
2810 | * over the metadata limit and adding a metadata buffer. | |
2811 | */ | |
2812 | if (type == ARC_BUFC_METADATA) { | |
2813 | if (arc_meta_used >= arc_meta_limit) | |
2814 | evict = ARC_BUFC_METADATA; | |
2815 | else | |
2816 | /* | |
2817 | * In this case, we're evicting data while | |
2818 | * adding metadata. Thus, to prevent recycling a | |
2819 | * data buffer into a metadata buffer, recycling | |
2820 | * is disabled in the following arc_evict call. | |
2821 | */ | |
2822 | recycle = FALSE; | |
2823 | } | |
2824 | ||
2825 | if ((buf->b_data = arc_evict(state, 0, size, recycle, evict)) == NULL) { | |
34dc7c2f BB |
2826 | if (type == ARC_BUFC_METADATA) { |
2827 | buf->b_data = zio_buf_alloc(size); | |
cc7f677c | 2828 | arc_space_consume(size, ARC_SPACE_META); |
ab26409d BB |
2829 | |
2830 | /* | |
2831 | * If we are unable to recycle an existing meta buffer | |
2832 | * signal the reclaim thread. It will notify users | |
2833 | * via the prune callback to drop references. The | |
2834 | * prune callback in run in the context of the reclaim | |
2835 | * thread to avoid deadlocking on the hash_lock. | |
da8ccd0e | 2836 | * Of course, only do this when recycle is true. |
ab26409d | 2837 | */ |
da8ccd0e PS |
2838 | if (recycle) |
2839 | cv_signal(&arc_reclaim_thr_cv); | |
34dc7c2f BB |
2840 | } else { |
2841 | ASSERT(type == ARC_BUFC_DATA); | |
2842 | buf->b_data = zio_data_buf_alloc(size); | |
cc7f677c | 2843 | arc_space_consume(size, ARC_SPACE_DATA); |
34dc7c2f | 2844 | } |
ab26409d | 2845 | |
da8ccd0e PS |
2846 | /* Only bump this if we tried to recycle and failed */ |
2847 | if (recycle) | |
2848 | ARCSTAT_BUMP(arcstat_recycle_miss); | |
34dc7c2f BB |
2849 | } |
2850 | ASSERT(buf->b_data != NULL); | |
2851 | out: | |
2852 | /* | |
2853 | * Update the state size. Note that ghost states have a | |
2854 | * "ghost size" and so don't need to be updated. | |
2855 | */ | |
2856 | if (!GHOST_STATE(buf->b_hdr->b_state)) { | |
2857 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
2858 | ||
2859 | atomic_add_64(&hdr->b_state->arcs_size, size); | |
2860 | if (list_link_active(&hdr->b_arc_node)) { | |
2861 | ASSERT(refcount_is_zero(&hdr->b_refcnt)); | |
2862 | atomic_add_64(&hdr->b_state->arcs_lsize[type], size); | |
2863 | } | |
2864 | /* | |
2865 | * If we are growing the cache, and we are adding anonymous | |
2866 | * data, and we have outgrown arc_p, update arc_p | |
2867 | */ | |
89c8cac4 PS |
2868 | if (!zfs_arc_p_aggressive_disable && |
2869 | arc_size < arc_c && hdr->b_state == arc_anon && | |
34dc7c2f BB |
2870 | arc_anon->arcs_size + arc_mru->arcs_size > arc_p) |
2871 | arc_p = MIN(arc_c, arc_p + size); | |
2872 | } | |
2873 | } | |
2874 | ||
2875 | /* | |
2876 | * This routine is called whenever a buffer is accessed. | |
2877 | * NOTE: the hash lock is dropped in this function. | |
2878 | */ | |
2879 | static void | |
2880 | arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) | |
2881 | { | |
428870ff BB |
2882 | clock_t now; |
2883 | ||
34dc7c2f BB |
2884 | ASSERT(MUTEX_HELD(hash_lock)); |
2885 | ||
2886 | if (buf->b_state == arc_anon) { | |
2887 | /* | |
2888 | * This buffer is not in the cache, and does not | |
2889 | * appear in our "ghost" list. Add the new buffer | |
2890 | * to the MRU state. | |
2891 | */ | |
2892 | ||
2893 | ASSERT(buf->b_arc_access == 0); | |
428870ff | 2894 | buf->b_arc_access = ddi_get_lbolt(); |
34dc7c2f BB |
2895 | DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); |
2896 | arc_change_state(arc_mru, buf, hash_lock); | |
2897 | ||
2898 | } else if (buf->b_state == arc_mru) { | |
428870ff BB |
2899 | now = ddi_get_lbolt(); |
2900 | ||
34dc7c2f BB |
2901 | /* |
2902 | * If this buffer is here because of a prefetch, then either: | |
2903 | * - clear the flag if this is a "referencing" read | |
2904 | * (any subsequent access will bump this into the MFU state). | |
2905 | * or | |
2906 | * - move the buffer to the head of the list if this is | |
2907 | * another prefetch (to make it less likely to be evicted). | |
2908 | */ | |
2909 | if ((buf->b_flags & ARC_PREFETCH) != 0) { | |
2910 | if (refcount_count(&buf->b_refcnt) == 0) { | |
2911 | ASSERT(list_link_active(&buf->b_arc_node)); | |
2912 | } else { | |
2913 | buf->b_flags &= ~ARC_PREFETCH; | |
e0b0ca98 | 2914 | atomic_inc_32(&buf->b_mru_hits); |
34dc7c2f BB |
2915 | ARCSTAT_BUMP(arcstat_mru_hits); |
2916 | } | |
428870ff | 2917 | buf->b_arc_access = now; |
34dc7c2f BB |
2918 | return; |
2919 | } | |
2920 | ||
2921 | /* | |
2922 | * This buffer has been "accessed" only once so far, | |
2923 | * but it is still in the cache. Move it to the MFU | |
2924 | * state. | |
2925 | */ | |
0b75bdb3 | 2926 | if (ddi_time_after(now, buf->b_arc_access + ARC_MINTIME)) { |
34dc7c2f BB |
2927 | /* |
2928 | * More than 125ms have passed since we | |
2929 | * instantiated this buffer. Move it to the | |
2930 | * most frequently used state. | |
2931 | */ | |
428870ff | 2932 | buf->b_arc_access = now; |
34dc7c2f BB |
2933 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); |
2934 | arc_change_state(arc_mfu, buf, hash_lock); | |
2935 | } | |
e0b0ca98 | 2936 | atomic_inc_32(&buf->b_mru_hits); |
34dc7c2f BB |
2937 | ARCSTAT_BUMP(arcstat_mru_hits); |
2938 | } else if (buf->b_state == arc_mru_ghost) { | |
2939 | arc_state_t *new_state; | |
2940 | /* | |
2941 | * This buffer has been "accessed" recently, but | |
2942 | * was evicted from the cache. Move it to the | |
2943 | * MFU state. | |
2944 | */ | |
2945 | ||
2946 | if (buf->b_flags & ARC_PREFETCH) { | |
2947 | new_state = arc_mru; | |
2948 | if (refcount_count(&buf->b_refcnt) > 0) | |
2949 | buf->b_flags &= ~ARC_PREFETCH; | |
2950 | DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); | |
2951 | } else { | |
2952 | new_state = arc_mfu; | |
2953 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); | |
2954 | } | |
2955 | ||
428870ff | 2956 | buf->b_arc_access = ddi_get_lbolt(); |
34dc7c2f BB |
2957 | arc_change_state(new_state, buf, hash_lock); |
2958 | ||
e0b0ca98 | 2959 | atomic_inc_32(&buf->b_mru_ghost_hits); |
34dc7c2f BB |
2960 | ARCSTAT_BUMP(arcstat_mru_ghost_hits); |
2961 | } else if (buf->b_state == arc_mfu) { | |
2962 | /* | |
2963 | * This buffer has been accessed more than once and is | |
2964 | * still in the cache. Keep it in the MFU state. | |
2965 | * | |
2966 | * NOTE: an add_reference() that occurred when we did | |
2967 | * the arc_read() will have kicked this off the list. | |
2968 | * If it was a prefetch, we will explicitly move it to | |
2969 | * the head of the list now. | |
2970 | */ | |
2971 | if ((buf->b_flags & ARC_PREFETCH) != 0) { | |
2972 | ASSERT(refcount_count(&buf->b_refcnt) == 0); | |
2973 | ASSERT(list_link_active(&buf->b_arc_node)); | |
2974 | } | |
e0b0ca98 | 2975 | atomic_inc_32(&buf->b_mfu_hits); |
34dc7c2f | 2976 | ARCSTAT_BUMP(arcstat_mfu_hits); |
428870ff | 2977 | buf->b_arc_access = ddi_get_lbolt(); |
34dc7c2f BB |
2978 | } else if (buf->b_state == arc_mfu_ghost) { |
2979 | arc_state_t *new_state = arc_mfu; | |
2980 | /* | |
2981 | * This buffer has been accessed more than once but has | |
2982 | * been evicted from the cache. Move it back to the | |
2983 | * MFU state. | |
2984 | */ | |
2985 | ||
2986 | if (buf->b_flags & ARC_PREFETCH) { | |
2987 | /* | |
2988 | * This is a prefetch access... | |
2989 | * move this block back to the MRU state. | |
2990 | */ | |
c99c9001 | 2991 | ASSERT0(refcount_count(&buf->b_refcnt)); |
34dc7c2f BB |
2992 | new_state = arc_mru; |
2993 | } | |
2994 | ||
428870ff | 2995 | buf->b_arc_access = ddi_get_lbolt(); |
34dc7c2f BB |
2996 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); |
2997 | arc_change_state(new_state, buf, hash_lock); | |
2998 | ||
e0b0ca98 | 2999 | atomic_inc_32(&buf->b_mfu_ghost_hits); |
34dc7c2f BB |
3000 | ARCSTAT_BUMP(arcstat_mfu_ghost_hits); |
3001 | } else if (buf->b_state == arc_l2c_only) { | |
3002 | /* | |
3003 | * This buffer is on the 2nd Level ARC. | |
3004 | */ | |
3005 | ||
428870ff | 3006 | buf->b_arc_access = ddi_get_lbolt(); |
34dc7c2f BB |
3007 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); |
3008 | arc_change_state(arc_mfu, buf, hash_lock); | |
3009 | } else { | |
3010 | ASSERT(!"invalid arc state"); | |
3011 | } | |
3012 | } | |
3013 | ||
3014 | /* a generic arc_done_func_t which you can use */ | |
3015 | /* ARGSUSED */ | |
3016 | void | |
3017 | arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) | |
3018 | { | |
428870ff BB |
3019 | if (zio == NULL || zio->io_error == 0) |
3020 | bcopy(buf->b_data, arg, buf->b_hdr->b_size); | |
13fe0198 | 3021 | VERIFY(arc_buf_remove_ref(buf, arg)); |
34dc7c2f BB |
3022 | } |
3023 | ||
3024 | /* a generic arc_done_func_t */ | |
3025 | void | |
3026 | arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) | |
3027 | { | |
3028 | arc_buf_t **bufp = arg; | |
3029 | if (zio && zio->io_error) { | |
13fe0198 | 3030 | VERIFY(arc_buf_remove_ref(buf, arg)); |
34dc7c2f BB |
3031 | *bufp = NULL; |
3032 | } else { | |
3033 | *bufp = buf; | |
428870ff | 3034 | ASSERT(buf->b_data); |
34dc7c2f BB |
3035 | } |
3036 | } | |
3037 | ||
3038 | static void | |
3039 | arc_read_done(zio_t *zio) | |
3040 | { | |
9b67f605 | 3041 | arc_buf_hdr_t *hdr; |
34dc7c2f BB |
3042 | arc_buf_t *buf; |
3043 | arc_buf_t *abuf; /* buffer we're assigning to callback */ | |
9b67f605 | 3044 | kmutex_t *hash_lock = NULL; |
34dc7c2f BB |
3045 | arc_callback_t *callback_list, *acb; |
3046 | int freeable = FALSE; | |
3047 | ||
3048 | buf = zio->io_private; | |
3049 | hdr = buf->b_hdr; | |
3050 | ||
3051 | /* | |
3052 | * The hdr was inserted into hash-table and removed from lists | |
3053 | * prior to starting I/O. We should find this header, since | |
3054 | * it's in the hash table, and it should be legit since it's | |
3055 | * not possible to evict it during the I/O. The only possible | |
3056 | * reason for it not to be found is if we were freed during the | |
3057 | * read. | |
3058 | */ | |
9b67f605 MA |
3059 | if (HDR_IN_HASH_TABLE(hdr)) { |
3060 | arc_buf_hdr_t *found; | |
3061 | ||
3062 | ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); | |
3063 | ASSERT3U(hdr->b_dva.dva_word[0], ==, | |
3064 | BP_IDENTITY(zio->io_bp)->dva_word[0]); | |
3065 | ASSERT3U(hdr->b_dva.dva_word[1], ==, | |
3066 | BP_IDENTITY(zio->io_bp)->dva_word[1]); | |
3067 | ||
3068 | found = buf_hash_find(hdr->b_spa, zio->io_bp, | |
3069 | &hash_lock); | |
3070 | ||
3071 | ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && | |
3072 | hash_lock == NULL) || | |
3073 | (found == hdr && | |
3074 | DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || | |
3075 | (found == hdr && HDR_L2_READING(hdr))); | |
3076 | } | |
34dc7c2f | 3077 | |
b128c09f | 3078 | hdr->b_flags &= ~ARC_L2_EVICTED; |
34dc7c2f | 3079 | if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) |
b128c09f | 3080 | hdr->b_flags &= ~ARC_L2CACHE; |
34dc7c2f BB |
3081 | |
3082 | /* byteswap if necessary */ | |
3083 | callback_list = hdr->b_acb; | |
3084 | ASSERT(callback_list != NULL); | |
428870ff | 3085 | if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { |
9ae529ec CS |
3086 | dmu_object_byteswap_t bswap = |
3087 | DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); | |
b01615d5 RY |
3088 | if (BP_GET_LEVEL(zio->io_bp) > 0) |
3089 | byteswap_uint64_array(buf->b_data, hdr->b_size); | |
3090 | else | |
3091 | dmu_ot_byteswap[bswap].ob_func(buf->b_data, hdr->b_size); | |
b128c09f | 3092 | } |
34dc7c2f BB |
3093 | |
3094 | arc_cksum_compute(buf, B_FALSE); | |
498877ba | 3095 | arc_buf_watch(buf); |
34dc7c2f | 3096 | |
428870ff BB |
3097 | if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { |
3098 | /* | |
3099 | * Only call arc_access on anonymous buffers. This is because | |
3100 | * if we've issued an I/O for an evicted buffer, we've already | |
3101 | * called arc_access (to prevent any simultaneous readers from | |
3102 | * getting confused). | |
3103 | */ | |
3104 | arc_access(hdr, hash_lock); | |
3105 | } | |
3106 | ||
34dc7c2f BB |
3107 | /* create copies of the data buffer for the callers */ |
3108 | abuf = buf; | |
3109 | for (acb = callback_list; acb; acb = acb->acb_next) { | |
3110 | if (acb->acb_done) { | |
1eb5bfa3 GW |
3111 | if (abuf == NULL) { |
3112 | ARCSTAT_BUMP(arcstat_duplicate_reads); | |
34dc7c2f | 3113 | abuf = arc_buf_clone(buf); |
1eb5bfa3 | 3114 | } |
34dc7c2f BB |
3115 | acb->acb_buf = abuf; |
3116 | abuf = NULL; | |
3117 | } | |
3118 | } | |
3119 | hdr->b_acb = NULL; | |
3120 | hdr->b_flags &= ~ARC_IO_IN_PROGRESS; | |
3121 | ASSERT(!HDR_BUF_AVAILABLE(hdr)); | |
428870ff BB |
3122 | if (abuf == buf) { |
3123 | ASSERT(buf->b_efunc == NULL); | |
3124 | ASSERT(hdr->b_datacnt == 1); | |
34dc7c2f | 3125 | hdr->b_flags |= ARC_BUF_AVAILABLE; |
428870ff | 3126 | } |
34dc7c2f BB |
3127 | |
3128 | ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); | |
3129 | ||
3130 | if (zio->io_error != 0) { | |
3131 | hdr->b_flags |= ARC_IO_ERROR; | |
3132 | if (hdr->b_state != arc_anon) | |
3133 | arc_change_state(arc_anon, hdr, hash_lock); | |
3134 | if (HDR_IN_HASH_TABLE(hdr)) | |
3135 | buf_hash_remove(hdr); | |
3136 | freeable = refcount_is_zero(&hdr->b_refcnt); | |
34dc7c2f BB |
3137 | } |
3138 | ||
3139 | /* | |
3140 | * Broadcast before we drop the hash_lock to avoid the possibility | |
3141 | * that the hdr (and hence the cv) might be freed before we get to | |
3142 | * the cv_broadcast(). | |
3143 | */ | |
3144 | cv_broadcast(&hdr->b_cv); | |
3145 | ||
3146 | if (hash_lock) { | |
34dc7c2f BB |
3147 | mutex_exit(hash_lock); |
3148 | } else { | |
3149 | /* | |
3150 | * This block was freed while we waited for the read to | |
3151 | * complete. It has been removed from the hash table and | |
3152 | * moved to the anonymous state (so that it won't show up | |
3153 | * in the cache). | |
3154 | */ | |
3155 | ASSERT3P(hdr->b_state, ==, arc_anon); | |
3156 | freeable = refcount_is_zero(&hdr->b_refcnt); | |
3157 | } | |
3158 | ||
3159 | /* execute each callback and free its structure */ | |
3160 | while ((acb = callback_list) != NULL) { | |
3161 | if (acb->acb_done) | |
3162 | acb->acb_done(zio, acb->acb_buf, acb->acb_private); | |
3163 | ||
3164 | if (acb->acb_zio_dummy != NULL) { | |
3165 | acb->acb_zio_dummy->io_error = zio->io_error; | |
3166 | zio_nowait(acb->acb_zio_dummy); | |
3167 | } | |
3168 | ||
3169 | callback_list = acb->acb_next; | |
3170 | kmem_free(acb, sizeof (arc_callback_t)); | |
3171 | } | |
3172 | ||
3173 | if (freeable) | |
3174 | arc_hdr_destroy(hdr); | |
3175 | } | |
3176 | ||
3177 | /* | |
5c839890 | 3178 | * "Read" the block at the specified DVA (in bp) via the |
34dc7c2f BB |
3179 | * cache. If the block is found in the cache, invoke the provided |
3180 | * callback immediately and return. Note that the `zio' parameter | |
3181 | * in the callback will be NULL in this case, since no IO was | |
3182 | * required. If the block is not in the cache pass the read request | |
3183 | * on to the spa with a substitute callback function, so that the | |
3184 | * requested block will be added to the cache. | |
3185 | * | |
3186 | * If a read request arrives for a block that has a read in-progress, | |
3187 | * either wait for the in-progress read to complete (and return the | |
3188 | * results); or, if this is a read with a "done" func, add a record | |
3189 | * to the read to invoke the "done" func when the read completes, | |
3190 | * and return; or just return. | |
3191 | * | |
3192 | * arc_read_done() will invoke all the requested "done" functions | |
3193 | * for readers of this block. | |
3194 | */ | |
3195 | int | |
294f6806 | 3196 | arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, |
e8b96c60 | 3197 | void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, |
294f6806 | 3198 | const zbookmark_t *zb) |
34dc7c2f | 3199 | { |
9b67f605 | 3200 | arc_buf_hdr_t *hdr = NULL; |
d4ed6673 | 3201 | arc_buf_t *buf = NULL; |
9b67f605 | 3202 | kmutex_t *hash_lock = NULL; |
34dc7c2f | 3203 | zio_t *rzio; |
3541dc6d | 3204 | uint64_t guid = spa_load_guid(spa); |
1421c891 | 3205 | int rc = 0; |
34dc7c2f | 3206 | |
9b67f605 MA |
3207 | ASSERT(!BP_IS_EMBEDDED(bp) || |
3208 | BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); | |
3209 | ||
34dc7c2f | 3210 | top: |
9b67f605 MA |
3211 | if (!BP_IS_EMBEDDED(bp)) { |
3212 | /* | |
3213 | * Embedded BP's have no DVA and require no I/O to "read". | |
3214 | * Create an anonymous arc buf to back it. | |
3215 | */ | |
3216 | hdr = buf_hash_find(guid, bp, &hash_lock); | |
3217 | } | |
3218 | ||
3219 | if (hdr != NULL && hdr->b_datacnt > 0) { | |
34dc7c2f BB |
3220 | |
3221 | *arc_flags |= ARC_CACHED; | |
3222 | ||
3223 | if (HDR_IO_IN_PROGRESS(hdr)) { | |
3224 | ||
3225 | if (*arc_flags & ARC_WAIT) { | |
3226 | cv_wait(&hdr->b_cv, hash_lock); | |
3227 | mutex_exit(hash_lock); | |
3228 | goto top; | |
3229 | } | |
3230 | ASSERT(*arc_flags & ARC_NOWAIT); | |
3231 | ||
3232 | if (done) { | |
3233 | arc_callback_t *acb = NULL; | |
3234 | ||
3235 | acb = kmem_zalloc(sizeof (arc_callback_t), | |
691f6ac4 | 3236 | KM_PUSHPAGE); |
34dc7c2f BB |
3237 | acb->acb_done = done; |
3238 | acb->acb_private = private; | |
34dc7c2f BB |
3239 | if (pio != NULL) |
3240 | acb->acb_zio_dummy = zio_null(pio, | |
d164b209 | 3241 | spa, NULL, NULL, NULL, zio_flags); |
34dc7c2f BB |
3242 | |
3243 | ASSERT(acb->acb_done != NULL); | |
3244 | acb->acb_next = hdr->b_acb; | |
3245 | hdr->b_acb = acb; | |
3246 | add_reference(hdr, hash_lock, private); | |
3247 | mutex_exit(hash_lock); | |
1421c891 | 3248 | goto out; |
34dc7c2f BB |
3249 | } |
3250 | mutex_exit(hash_lock); | |
1421c891 | 3251 | goto out; |
34dc7c2f BB |
3252 | } |
3253 | ||
3254 | ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); | |
3255 | ||
3256 | if (done) { | |
3257 | add_reference(hdr, hash_lock, private); | |
3258 | /* | |
3259 | * If this block is already in use, create a new | |
3260 | * copy of the data so that we will be guaranteed | |
3261 | * that arc_release() will always succeed. | |
3262 | */ | |
3263 | buf = hdr->b_buf; | |
3264 | ASSERT(buf); | |
3265 | ASSERT(buf->b_data); | |
3266 | if (HDR_BUF_AVAILABLE(hdr)) { | |
3267 | ASSERT(buf->b_efunc == NULL); | |
3268 | hdr->b_flags &= ~ARC_BUF_AVAILABLE; | |
3269 | } else { | |
3270 | buf = arc_buf_clone(buf); | |
3271 | } | |
428870ff | 3272 | |
34dc7c2f BB |
3273 | } else if (*arc_flags & ARC_PREFETCH && |
3274 | refcount_count(&hdr->b_refcnt) == 0) { | |
3275 | hdr->b_flags |= ARC_PREFETCH; | |
3276 | } | |
3277 | DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); | |
3278 | arc_access(hdr, hash_lock); | |
b128c09f BB |
3279 | if (*arc_flags & ARC_L2CACHE) |
3280 | hdr->b_flags |= ARC_L2CACHE; | |
3a17a7a9 SK |
3281 | if (*arc_flags & ARC_L2COMPRESS) |
3282 | hdr->b_flags |= ARC_L2COMPRESS; | |
34dc7c2f BB |
3283 | mutex_exit(hash_lock); |
3284 | ARCSTAT_BUMP(arcstat_hits); | |
3285 | ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), | |
3286 | demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, | |
3287 | data, metadata, hits); | |
3288 | ||
3289 | if (done) | |
3290 | done(NULL, buf, private); | |
3291 | } else { | |
3292 | uint64_t size = BP_GET_LSIZE(bp); | |
9b67f605 | 3293 | arc_callback_t *acb; |
b128c09f | 3294 | vdev_t *vd = NULL; |
a117a6d6 | 3295 | uint64_t addr = 0; |
d164b209 | 3296 | boolean_t devw = B_FALSE; |
0ed212dc BP |
3297 | enum zio_compress b_compress = ZIO_COMPRESS_OFF; |
3298 | uint64_t b_asize = 0; | |
34dc7c2f BB |
3299 | |
3300 | if (hdr == NULL) { | |
3301 | /* this block is not in the cache */ | |
9b67f605 | 3302 | arc_buf_hdr_t *exists = NULL; |
34dc7c2f BB |
3303 | arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); |
3304 | buf = arc_buf_alloc(spa, size, private, type); | |
3305 | hdr = buf->b_hdr; | |
9b67f605 MA |
3306 | if (!BP_IS_EMBEDDED(bp)) { |
3307 | hdr->b_dva = *BP_IDENTITY(bp); | |
3308 | hdr->b_birth = BP_PHYSICAL_BIRTH(bp); | |
3309 | hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; | |
3310 | exists = buf_hash_insert(hdr, &hash_lock); | |
3311 | } | |
3312 | if (exists != NULL) { | |
34dc7c2f BB |
3313 | /* somebody beat us to the hash insert */ |
3314 | mutex_exit(hash_lock); | |
428870ff | 3315 | buf_discard_identity(hdr); |
34dc7c2f BB |
3316 | (void) arc_buf_remove_ref(buf, private); |
3317 | goto top; /* restart the IO request */ | |
3318 | } | |
3319 | /* if this is a prefetch, we don't have a reference */ | |
3320 | if (*arc_flags & ARC_PREFETCH) { | |
3321 | (void) remove_reference(hdr, hash_lock, | |
3322 | private); | |
3323 | hdr->b_flags |= ARC_PREFETCH; | |
3324 | } | |
b128c09f BB |
3325 | if (*arc_flags & ARC_L2CACHE) |
3326 | hdr->b_flags |= ARC_L2CACHE; | |
3a17a7a9 SK |
3327 | if (*arc_flags & ARC_L2COMPRESS) |
3328 | hdr->b_flags |= ARC_L2COMPRESS; | |
34dc7c2f BB |
3329 | if (BP_GET_LEVEL(bp) > 0) |
3330 | hdr->b_flags |= ARC_INDIRECT; | |
3331 | } else { | |
3332 | /* this block is in the ghost cache */ | |
3333 | ASSERT(GHOST_STATE(hdr->b_state)); | |
3334 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
c99c9001 | 3335 | ASSERT0(refcount_count(&hdr->b_refcnt)); |
34dc7c2f BB |
3336 | ASSERT(hdr->b_buf == NULL); |
3337 | ||
3338 | /* if this is a prefetch, we don't have a reference */ | |
3339 | if (*arc_flags & ARC_PREFETCH) | |
3340 | hdr->b_flags |= ARC_PREFETCH; | |
3341 | else | |
3342 | add_reference(hdr, hash_lock, private); | |
b128c09f BB |
3343 | if (*arc_flags & ARC_L2CACHE) |
3344 | hdr->b_flags |= ARC_L2CACHE; | |
3a17a7a9 SK |
3345 | if (*arc_flags & ARC_L2COMPRESS) |
3346 | hdr->b_flags |= ARC_L2COMPRESS; | |
34dc7c2f BB |
3347 | buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); |
3348 | buf->b_hdr = hdr; | |
3349 | buf->b_data = NULL; | |
3350 | buf->b_efunc = NULL; | |
3351 | buf->b_private = NULL; | |
3352 | buf->b_next = NULL; | |
3353 | hdr->b_buf = buf; | |
34dc7c2f BB |
3354 | ASSERT(hdr->b_datacnt == 0); |
3355 | hdr->b_datacnt = 1; | |
428870ff BB |
3356 | arc_get_data_buf(buf); |
3357 | arc_access(hdr, hash_lock); | |
34dc7c2f BB |
3358 | } |
3359 | ||
428870ff BB |
3360 | ASSERT(!GHOST_STATE(hdr->b_state)); |
3361 | ||
691f6ac4 | 3362 | acb = kmem_zalloc(sizeof (arc_callback_t), KM_PUSHPAGE); |
34dc7c2f BB |
3363 | acb->acb_done = done; |
3364 | acb->acb_private = private; | |
34dc7c2f BB |
3365 | |
3366 | ASSERT(hdr->b_acb == NULL); | |
3367 | hdr->b_acb = acb; | |
3368 | hdr->b_flags |= ARC_IO_IN_PROGRESS; | |
3369 | ||
0ed212dc | 3370 | if (hdr->b_l2hdr != NULL && |
b128c09f | 3371 | (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { |
d164b209 | 3372 | devw = hdr->b_l2hdr->b_dev->l2ad_writing; |
b128c09f | 3373 | addr = hdr->b_l2hdr->b_daddr; |
0ed212dc BP |
3374 | b_compress = hdr->b_l2hdr->b_compress; |
3375 | b_asize = hdr->b_l2hdr->b_asize; | |
b128c09f BB |
3376 | /* |
3377 | * Lock out device removal. | |
3378 | */ | |
3379 | if (vdev_is_dead(vd) || | |
3380 | !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) | |
3381 | vd = NULL; | |
3382 | } | |
3383 | ||
9b67f605 MA |
3384 | if (hash_lock != NULL) |
3385 | mutex_exit(hash_lock); | |
b128c09f | 3386 | |
e49f1e20 WA |
3387 | /* |
3388 | * At this point, we have a level 1 cache miss. Try again in | |
3389 | * L2ARC if possible. | |
3390 | */ | |
34dc7c2f | 3391 | ASSERT3U(hdr->b_size, ==, size); |
428870ff BB |
3392 | DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, |
3393 | uint64_t, size, zbookmark_t *, zb); | |
34dc7c2f BB |
3394 | ARCSTAT_BUMP(arcstat_misses); |
3395 | ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), | |
3396 | demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, | |
3397 | data, metadata, misses); | |
3398 | ||
d164b209 | 3399 | if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { |
34dc7c2f BB |
3400 | /* |
3401 | * Read from the L2ARC if the following are true: | |
b128c09f BB |
3402 | * 1. The L2ARC vdev was previously cached. |
3403 | * 2. This buffer still has L2ARC metadata. | |
3404 | * 3. This buffer isn't currently writing to the L2ARC. | |
3405 | * 4. The L2ARC entry wasn't evicted, which may | |
3406 | * also have invalidated the vdev. | |
d164b209 | 3407 | * 5. This isn't prefetch and l2arc_noprefetch is set. |
34dc7c2f | 3408 | */ |
b128c09f | 3409 | if (hdr->b_l2hdr != NULL && |
d164b209 BB |
3410 | !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && |
3411 | !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { | |
34dc7c2f BB |
3412 | l2arc_read_callback_t *cb; |
3413 | ||
3414 | DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); | |
3415 | ARCSTAT_BUMP(arcstat_l2_hits); | |
e0b0ca98 | 3416 | atomic_inc_32(&hdr->b_l2hdr->b_hits); |
34dc7c2f | 3417 | |
34dc7c2f | 3418 | cb = kmem_zalloc(sizeof (l2arc_read_callback_t), |
691f6ac4 | 3419 | KM_PUSHPAGE); |
34dc7c2f BB |
3420 | cb->l2rcb_buf = buf; |
3421 | cb->l2rcb_spa = spa; | |
3422 | cb->l2rcb_bp = *bp; | |
3423 | cb->l2rcb_zb = *zb; | |
b128c09f | 3424 | cb->l2rcb_flags = zio_flags; |
0ed212dc | 3425 | cb->l2rcb_compress = b_compress; |
34dc7c2f | 3426 | |
a117a6d6 GW |
3427 | ASSERT(addr >= VDEV_LABEL_START_SIZE && |
3428 | addr + size < vd->vdev_psize - | |
3429 | VDEV_LABEL_END_SIZE); | |
3430 | ||
34dc7c2f | 3431 | /* |
b128c09f BB |
3432 | * l2arc read. The SCL_L2ARC lock will be |
3433 | * released by l2arc_read_done(). | |
3a17a7a9 SK |
3434 | * Issue a null zio if the underlying buffer |
3435 | * was squashed to zero size by compression. | |
34dc7c2f | 3436 | */ |
0ed212dc | 3437 | if (b_compress == ZIO_COMPRESS_EMPTY) { |
3a17a7a9 SK |
3438 | rzio = zio_null(pio, spa, vd, |
3439 | l2arc_read_done, cb, | |
3440 | zio_flags | ZIO_FLAG_DONT_CACHE | | |
3441 | ZIO_FLAG_CANFAIL | | |
3442 | ZIO_FLAG_DONT_PROPAGATE | | |
3443 | ZIO_FLAG_DONT_RETRY); | |
3444 | } else { | |
3445 | rzio = zio_read_phys(pio, vd, addr, | |
0ed212dc BP |
3446 | b_asize, buf->b_data, |
3447 | ZIO_CHECKSUM_OFF, | |
3a17a7a9 SK |
3448 | l2arc_read_done, cb, priority, |
3449 | zio_flags | ZIO_FLAG_DONT_CACHE | | |
3450 | ZIO_FLAG_CANFAIL | | |
3451 | ZIO_FLAG_DONT_PROPAGATE | | |
3452 | ZIO_FLAG_DONT_RETRY, B_FALSE); | |
3453 | } | |
34dc7c2f BB |
3454 | DTRACE_PROBE2(l2arc__read, vdev_t *, vd, |
3455 | zio_t *, rzio); | |
0ed212dc | 3456 | ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); |
34dc7c2f | 3457 | |
b128c09f BB |
3458 | if (*arc_flags & ARC_NOWAIT) { |
3459 | zio_nowait(rzio); | |
1421c891 | 3460 | goto out; |
b128c09f | 3461 | } |
34dc7c2f | 3462 | |
b128c09f BB |
3463 | ASSERT(*arc_flags & ARC_WAIT); |
3464 | if (zio_wait(rzio) == 0) | |
1421c891 | 3465 | goto out; |
b128c09f BB |
3466 | |
3467 | /* l2arc read error; goto zio_read() */ | |
34dc7c2f BB |
3468 | } else { |
3469 | DTRACE_PROBE1(l2arc__miss, | |
3470 | arc_buf_hdr_t *, hdr); | |
3471 | ARCSTAT_BUMP(arcstat_l2_misses); | |
3472 | if (HDR_L2_WRITING(hdr)) | |
3473 | ARCSTAT_BUMP(arcstat_l2_rw_clash); | |
b128c09f | 3474 | spa_config_exit(spa, SCL_L2ARC, vd); |
34dc7c2f | 3475 | } |
d164b209 BB |
3476 | } else { |
3477 | if (vd != NULL) | |
3478 | spa_config_exit(spa, SCL_L2ARC, vd); | |
3479 | if (l2arc_ndev != 0) { | |
3480 | DTRACE_PROBE1(l2arc__miss, | |
3481 | arc_buf_hdr_t *, hdr); | |
3482 | ARCSTAT_BUMP(arcstat_l2_misses); | |
3483 | } | |
34dc7c2f | 3484 | } |
34dc7c2f BB |
3485 | |
3486 | rzio = zio_read(pio, spa, bp, buf->b_data, size, | |
b128c09f | 3487 | arc_read_done, buf, priority, zio_flags, zb); |
34dc7c2f | 3488 | |
1421c891 PS |
3489 | if (*arc_flags & ARC_WAIT) { |
3490 | rc = zio_wait(rzio); | |
3491 | goto out; | |
3492 | } | |
34dc7c2f BB |
3493 | |
3494 | ASSERT(*arc_flags & ARC_NOWAIT); | |
3495 | zio_nowait(rzio); | |
3496 | } | |
1421c891 PS |
3497 | |
3498 | out: | |
3499 | spa_read_history_add(spa, zb, *arc_flags); | |
3500 | return (rc); | |
34dc7c2f BB |
3501 | } |
3502 | ||
ab26409d BB |
3503 | arc_prune_t * |
3504 | arc_add_prune_callback(arc_prune_func_t *func, void *private) | |
3505 | { | |
3506 | arc_prune_t *p; | |
3507 | ||
d1d7e268 | 3508 | p = kmem_alloc(sizeof (*p), KM_SLEEP); |
ab26409d BB |
3509 | p->p_pfunc = func; |
3510 | p->p_private = private; | |
3511 | list_link_init(&p->p_node); | |
3512 | refcount_create(&p->p_refcnt); | |
3513 | ||
3514 | mutex_enter(&arc_prune_mtx); | |
3515 | refcount_add(&p->p_refcnt, &arc_prune_list); | |
3516 | list_insert_head(&arc_prune_list, p); | |
3517 | mutex_exit(&arc_prune_mtx); | |
3518 | ||
3519 | return (p); | |
3520 | } | |
3521 | ||
3522 | void | |
3523 | arc_remove_prune_callback(arc_prune_t *p) | |
3524 | { | |
3525 | mutex_enter(&arc_prune_mtx); | |
3526 | list_remove(&arc_prune_list, p); | |
3527 | if (refcount_remove(&p->p_refcnt, &arc_prune_list) == 0) { | |
3528 | refcount_destroy(&p->p_refcnt); | |
3529 | kmem_free(p, sizeof (*p)); | |
3530 | } | |
3531 | mutex_exit(&arc_prune_mtx); | |
3532 | } | |
3533 | ||
34dc7c2f BB |
3534 | void |
3535 | arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) | |
3536 | { | |
3537 | ASSERT(buf->b_hdr != NULL); | |
3538 | ASSERT(buf->b_hdr->b_state != arc_anon); | |
3539 | ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); | |
428870ff BB |
3540 | ASSERT(buf->b_efunc == NULL); |
3541 | ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); | |
3542 | ||
34dc7c2f BB |
3543 | buf->b_efunc = func; |
3544 | buf->b_private = private; | |
3545 | } | |
3546 | ||
df4474f9 MA |
3547 | /* |
3548 | * Notify the arc that a block was freed, and thus will never be used again. | |
3549 | */ | |
3550 | void | |
3551 | arc_freed(spa_t *spa, const blkptr_t *bp) | |
3552 | { | |
3553 | arc_buf_hdr_t *hdr; | |
3554 | kmutex_t *hash_lock; | |
3555 | uint64_t guid = spa_load_guid(spa); | |
3556 | ||
9b67f605 MA |
3557 | ASSERT(!BP_IS_EMBEDDED(bp)); |
3558 | ||
3559 | hdr = buf_hash_find(guid, bp, &hash_lock); | |
df4474f9 MA |
3560 | if (hdr == NULL) |
3561 | return; | |
3562 | if (HDR_BUF_AVAILABLE(hdr)) { | |
3563 | arc_buf_t *buf = hdr->b_buf; | |
3564 | add_reference(hdr, hash_lock, FTAG); | |
3565 | hdr->b_flags &= ~ARC_BUF_AVAILABLE; | |
3566 | mutex_exit(hash_lock); | |
3567 | ||
3568 | arc_release(buf, FTAG); | |
3569 | (void) arc_buf_remove_ref(buf, FTAG); | |
3570 | } else { | |
3571 | mutex_exit(hash_lock); | |
3572 | } | |
3573 | ||
3574 | } | |
3575 | ||
34dc7c2f BB |
3576 | /* |
3577 | * This is used by the DMU to let the ARC know that a buffer is | |
3578 | * being evicted, so the ARC should clean up. If this arc buf | |
3579 | * is not yet in the evicted state, it will be put there. | |
3580 | */ | |
3581 | int | |
3582 | arc_buf_evict(arc_buf_t *buf) | |
3583 | { | |
3584 | arc_buf_hdr_t *hdr; | |
3585 | kmutex_t *hash_lock; | |
3586 | arc_buf_t **bufp; | |
3587 | ||
428870ff | 3588 | mutex_enter(&buf->b_evict_lock); |
34dc7c2f BB |
3589 | hdr = buf->b_hdr; |
3590 | if (hdr == NULL) { | |
3591 | /* | |
3592 | * We are in arc_do_user_evicts(). | |
3593 | */ | |
3594 | ASSERT(buf->b_data == NULL); | |
428870ff | 3595 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f | 3596 | return (0); |
b128c09f BB |
3597 | } else if (buf->b_data == NULL) { |
3598 | arc_buf_t copy = *buf; /* structure assignment */ | |
34dc7c2f | 3599 | /* |
b128c09f BB |
3600 | * We are on the eviction list; process this buffer now |
3601 | * but let arc_do_user_evicts() do the reaping. | |
34dc7c2f | 3602 | */ |
b128c09f | 3603 | buf->b_efunc = NULL; |
428870ff | 3604 | mutex_exit(&buf->b_evict_lock); |
b128c09f BB |
3605 | VERIFY(copy.b_efunc(©) == 0); |
3606 | return (1); | |
34dc7c2f | 3607 | } |
b128c09f BB |
3608 | hash_lock = HDR_LOCK(hdr); |
3609 | mutex_enter(hash_lock); | |
428870ff BB |
3610 | hdr = buf->b_hdr; |
3611 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
34dc7c2f | 3612 | |
34dc7c2f BB |
3613 | ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); |
3614 | ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); | |
3615 | ||
3616 | /* | |
3617 | * Pull this buffer off of the hdr | |
3618 | */ | |
3619 | bufp = &hdr->b_buf; | |
3620 | while (*bufp != buf) | |
3621 | bufp = &(*bufp)->b_next; | |
3622 | *bufp = buf->b_next; | |
3623 | ||
3624 | ASSERT(buf->b_data != NULL); | |
3625 | arc_buf_destroy(buf, FALSE, FALSE); | |
3626 | ||
3627 | if (hdr->b_datacnt == 0) { | |
3628 | arc_state_t *old_state = hdr->b_state; | |
3629 | arc_state_t *evicted_state; | |
3630 | ||
428870ff | 3631 | ASSERT(hdr->b_buf == NULL); |
34dc7c2f BB |
3632 | ASSERT(refcount_is_zero(&hdr->b_refcnt)); |
3633 | ||
3634 | evicted_state = | |
3635 | (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; | |
3636 | ||
3637 | mutex_enter(&old_state->arcs_mtx); | |
3638 | mutex_enter(&evicted_state->arcs_mtx); | |
3639 | ||
3640 | arc_change_state(evicted_state, hdr, hash_lock); | |
3641 | ASSERT(HDR_IN_HASH_TABLE(hdr)); | |
3642 | hdr->b_flags |= ARC_IN_HASH_TABLE; | |
3643 | hdr->b_flags &= ~ARC_BUF_AVAILABLE; | |
3644 | ||
3645 | mutex_exit(&evicted_state->arcs_mtx); | |
3646 | mutex_exit(&old_state->arcs_mtx); | |
3647 | } | |
3648 | mutex_exit(hash_lock); | |
428870ff | 3649 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
3650 | |
3651 | VERIFY(buf->b_efunc(buf) == 0); | |
3652 | buf->b_efunc = NULL; | |
3653 | buf->b_private = NULL; | |
3654 | buf->b_hdr = NULL; | |
428870ff | 3655 | buf->b_next = NULL; |
34dc7c2f BB |
3656 | kmem_cache_free(buf_cache, buf); |
3657 | return (1); | |
3658 | } | |
3659 | ||
3660 | /* | |
e49f1e20 WA |
3661 | * Release this buffer from the cache, making it an anonymous buffer. This |
3662 | * must be done after a read and prior to modifying the buffer contents. | |
34dc7c2f | 3663 | * If the buffer has more than one reference, we must make |
b128c09f | 3664 | * a new hdr for the buffer. |
34dc7c2f BB |
3665 | */ |
3666 | void | |
3667 | arc_release(arc_buf_t *buf, void *tag) | |
3668 | { | |
b128c09f | 3669 | arc_buf_hdr_t *hdr; |
428870ff | 3670 | kmutex_t *hash_lock = NULL; |
b128c09f | 3671 | l2arc_buf_hdr_t *l2hdr; |
d4ed6673 | 3672 | uint64_t buf_size = 0; |
34dc7c2f | 3673 | |
428870ff BB |
3674 | /* |
3675 | * It would be nice to assert that if it's DMU metadata (level > | |
3676 | * 0 || it's the dnode file), then it must be syncing context. | |
3677 | * But we don't know that information at this level. | |
3678 | */ | |
3679 | ||
3680 | mutex_enter(&buf->b_evict_lock); | |
b128c09f BB |
3681 | hdr = buf->b_hdr; |
3682 | ||
34dc7c2f BB |
3683 | /* this buffer is not on any list */ |
3684 | ASSERT(refcount_count(&hdr->b_refcnt) > 0); | |
3685 | ||
3686 | if (hdr->b_state == arc_anon) { | |
3687 | /* this buffer is already released */ | |
34dc7c2f | 3688 | ASSERT(buf->b_efunc == NULL); |
9babb374 BB |
3689 | } else { |
3690 | hash_lock = HDR_LOCK(hdr); | |
3691 | mutex_enter(hash_lock); | |
428870ff BB |
3692 | hdr = buf->b_hdr; |
3693 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
34dc7c2f BB |
3694 | } |
3695 | ||
b128c09f BB |
3696 | l2hdr = hdr->b_l2hdr; |
3697 | if (l2hdr) { | |
3698 | mutex_enter(&l2arc_buflist_mtx); | |
3699 | hdr->b_l2hdr = NULL; | |
47fe91b5 | 3700 | list_remove(l2hdr->b_dev->l2ad_buflist, hdr); |
b128c09f | 3701 | } |
a117a6d6 | 3702 | buf_size = hdr->b_size; |
b128c09f | 3703 | |
34dc7c2f BB |
3704 | /* |
3705 | * Do we have more than one buf? | |
3706 | */ | |
b128c09f | 3707 | if (hdr->b_datacnt > 1) { |
34dc7c2f BB |
3708 | arc_buf_hdr_t *nhdr; |
3709 | arc_buf_t **bufp; | |
3710 | uint64_t blksz = hdr->b_size; | |
d164b209 | 3711 | uint64_t spa = hdr->b_spa; |
34dc7c2f BB |
3712 | arc_buf_contents_t type = hdr->b_type; |
3713 | uint32_t flags = hdr->b_flags; | |
3714 | ||
b128c09f | 3715 | ASSERT(hdr->b_buf != buf || buf->b_next != NULL); |
34dc7c2f | 3716 | /* |
428870ff BB |
3717 | * Pull the data off of this hdr and attach it to |
3718 | * a new anonymous hdr. | |
34dc7c2f BB |
3719 | */ |
3720 | (void) remove_reference(hdr, hash_lock, tag); | |
3721 | bufp = &hdr->b_buf; | |
3722 | while (*bufp != buf) | |
3723 | bufp = &(*bufp)->b_next; | |
428870ff | 3724 | *bufp = buf->b_next; |
34dc7c2f BB |
3725 | buf->b_next = NULL; |
3726 | ||
3727 | ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); | |
3728 | atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); | |
3729 | if (refcount_is_zero(&hdr->b_refcnt)) { | |
3730 | uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; | |
3731 | ASSERT3U(*size, >=, hdr->b_size); | |
3732 | atomic_add_64(size, -hdr->b_size); | |
3733 | } | |
1eb5bfa3 GW |
3734 | |
3735 | /* | |
3736 | * We're releasing a duplicate user data buffer, update | |
3737 | * our statistics accordingly. | |
3738 | */ | |
3739 | if (hdr->b_type == ARC_BUFC_DATA) { | |
3740 | ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); | |
3741 | ARCSTAT_INCR(arcstat_duplicate_buffers_size, | |
3742 | -hdr->b_size); | |
3743 | } | |
34dc7c2f | 3744 | hdr->b_datacnt -= 1; |
34dc7c2f | 3745 | arc_cksum_verify(buf); |
498877ba | 3746 | arc_buf_unwatch(buf); |
34dc7c2f BB |
3747 | |
3748 | mutex_exit(hash_lock); | |
3749 | ||
3750 | nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); | |
3751 | nhdr->b_size = blksz; | |
3752 | nhdr->b_spa = spa; | |
3753 | nhdr->b_type = type; | |
3754 | nhdr->b_buf = buf; | |
3755 | nhdr->b_state = arc_anon; | |
3756 | nhdr->b_arc_access = 0; | |
e0b0ca98 BB |
3757 | nhdr->b_mru_hits = 0; |
3758 | nhdr->b_mru_ghost_hits = 0; | |
3759 | nhdr->b_mfu_hits = 0; | |
3760 | nhdr->b_mfu_ghost_hits = 0; | |
3761 | nhdr->b_l2_hits = 0; | |
34dc7c2f BB |
3762 | nhdr->b_flags = flags & ARC_L2_WRITING; |
3763 | nhdr->b_l2hdr = NULL; | |
3764 | nhdr->b_datacnt = 1; | |
3765 | nhdr->b_freeze_cksum = NULL; | |
3766 | (void) refcount_add(&nhdr->b_refcnt, tag); | |
3767 | buf->b_hdr = nhdr; | |
428870ff | 3768 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
3769 | atomic_add_64(&arc_anon->arcs_size, blksz); |
3770 | } else { | |
428870ff | 3771 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
3772 | ASSERT(refcount_count(&hdr->b_refcnt) == 1); |
3773 | ASSERT(!list_link_active(&hdr->b_arc_node)); | |
3774 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
428870ff BB |
3775 | if (hdr->b_state != arc_anon) |
3776 | arc_change_state(arc_anon, hdr, hash_lock); | |
34dc7c2f | 3777 | hdr->b_arc_access = 0; |
e0b0ca98 BB |
3778 | hdr->b_mru_hits = 0; |
3779 | hdr->b_mru_ghost_hits = 0; | |
3780 | hdr->b_mfu_hits = 0; | |
3781 | hdr->b_mfu_ghost_hits = 0; | |
3782 | hdr->b_l2_hits = 0; | |
428870ff BB |
3783 | if (hash_lock) |
3784 | mutex_exit(hash_lock); | |
34dc7c2f | 3785 | |
428870ff | 3786 | buf_discard_identity(hdr); |
34dc7c2f BB |
3787 | arc_buf_thaw(buf); |
3788 | } | |
3789 | buf->b_efunc = NULL; | |
3790 | buf->b_private = NULL; | |
3791 | ||
3792 | if (l2hdr) { | |
3a17a7a9 | 3793 | ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); |
ecf3d9b8 | 3794 | kmem_cache_free(l2arc_hdr_cache, l2hdr); |
6e1d7276 | 3795 | arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS); |
34dc7c2f | 3796 | ARCSTAT_INCR(arcstat_l2_size, -buf_size); |
34dc7c2f | 3797 | mutex_exit(&l2arc_buflist_mtx); |
b128c09f | 3798 | } |
34dc7c2f BB |
3799 | } |
3800 | ||
3801 | int | |
3802 | arc_released(arc_buf_t *buf) | |
3803 | { | |
b128c09f BB |
3804 | int released; |
3805 | ||
428870ff | 3806 | mutex_enter(&buf->b_evict_lock); |
b128c09f | 3807 | released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); |
428870ff | 3808 | mutex_exit(&buf->b_evict_lock); |
b128c09f | 3809 | return (released); |
34dc7c2f BB |
3810 | } |
3811 | ||
3812 | int | |
3813 | arc_has_callback(arc_buf_t *buf) | |
3814 | { | |
b128c09f BB |
3815 | int callback; |
3816 | ||
428870ff | 3817 | mutex_enter(&buf->b_evict_lock); |
b128c09f | 3818 | callback = (buf->b_efunc != NULL); |
428870ff | 3819 | mutex_exit(&buf->b_evict_lock); |
b128c09f | 3820 | return (callback); |
34dc7c2f BB |
3821 | } |
3822 | ||
3823 | #ifdef ZFS_DEBUG | |
3824 | int | |
3825 | arc_referenced(arc_buf_t *buf) | |
3826 | { | |
b128c09f BB |
3827 | int referenced; |
3828 | ||
428870ff | 3829 | mutex_enter(&buf->b_evict_lock); |
b128c09f | 3830 | referenced = (refcount_count(&buf->b_hdr->b_refcnt)); |
428870ff | 3831 | mutex_exit(&buf->b_evict_lock); |
b128c09f | 3832 | return (referenced); |
34dc7c2f BB |
3833 | } |
3834 | #endif | |
3835 | ||
3836 | static void | |
3837 | arc_write_ready(zio_t *zio) | |
3838 | { | |
3839 | arc_write_callback_t *callback = zio->io_private; | |
3840 | arc_buf_t *buf = callback->awcb_buf; | |
3841 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
3842 | ||
b128c09f BB |
3843 | ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); |
3844 | callback->awcb_ready(zio, buf, callback->awcb_private); | |
3845 | ||
34dc7c2f BB |
3846 | /* |
3847 | * If the IO is already in progress, then this is a re-write | |
b128c09f BB |
3848 | * attempt, so we need to thaw and re-compute the cksum. |
3849 | * It is the responsibility of the callback to handle the | |
3850 | * accounting for any re-write attempt. | |
34dc7c2f BB |
3851 | */ |
3852 | if (HDR_IO_IN_PROGRESS(hdr)) { | |
34dc7c2f BB |
3853 | mutex_enter(&hdr->b_freeze_lock); |
3854 | if (hdr->b_freeze_cksum != NULL) { | |
3855 | kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); | |
3856 | hdr->b_freeze_cksum = NULL; | |
3857 | } | |
3858 | mutex_exit(&hdr->b_freeze_lock); | |
3859 | } | |
3860 | arc_cksum_compute(buf, B_FALSE); | |
3861 | hdr->b_flags |= ARC_IO_IN_PROGRESS; | |
3862 | } | |
3863 | ||
e8b96c60 MA |
3864 | /* |
3865 | * The SPA calls this callback for each physical write that happens on behalf | |
3866 | * of a logical write. See the comment in dbuf_write_physdone() for details. | |
3867 | */ | |
3868 | static void | |
3869 | arc_write_physdone(zio_t *zio) | |
3870 | { | |
3871 | arc_write_callback_t *cb = zio->io_private; | |
3872 | if (cb->awcb_physdone != NULL) | |
3873 | cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); | |
3874 | } | |
3875 | ||
34dc7c2f BB |
3876 | static void |
3877 | arc_write_done(zio_t *zio) | |
3878 | { | |
3879 | arc_write_callback_t *callback = zio->io_private; | |
3880 | arc_buf_t *buf = callback->awcb_buf; | |
3881 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
3882 | ||
428870ff BB |
3883 | ASSERT(hdr->b_acb == NULL); |
3884 | ||
3885 | if (zio->io_error == 0) { | |
9b67f605 | 3886 | if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { |
b0bc7a84 MG |
3887 | buf_discard_identity(hdr); |
3888 | } else { | |
3889 | hdr->b_dva = *BP_IDENTITY(zio->io_bp); | |
3890 | hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); | |
3891 | hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; | |
3892 | } | |
428870ff BB |
3893 | } else { |
3894 | ASSERT(BUF_EMPTY(hdr)); | |
3895 | } | |
34dc7c2f | 3896 | |
34dc7c2f | 3897 | /* |
9b67f605 MA |
3898 | * If the block to be written was all-zero or compressed enough to be |
3899 | * embedded in the BP, no write was performed so there will be no | |
3900 | * dva/birth/checksum. The buffer must therefore remain anonymous | |
3901 | * (and uncached). | |
34dc7c2f BB |
3902 | */ |
3903 | if (!BUF_EMPTY(hdr)) { | |
3904 | arc_buf_hdr_t *exists; | |
3905 | kmutex_t *hash_lock; | |
3906 | ||
428870ff BB |
3907 | ASSERT(zio->io_error == 0); |
3908 | ||
34dc7c2f BB |
3909 | arc_cksum_verify(buf); |
3910 | ||
3911 | exists = buf_hash_insert(hdr, &hash_lock); | |
3912 | if (exists) { | |
3913 | /* | |
3914 | * This can only happen if we overwrite for | |
3915 | * sync-to-convergence, because we remove | |
3916 | * buffers from the hash table when we arc_free(). | |
3917 | */ | |
428870ff BB |
3918 | if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { |
3919 | if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) | |
3920 | panic("bad overwrite, hdr=%p exists=%p", | |
3921 | (void *)hdr, (void *)exists); | |
3922 | ASSERT(refcount_is_zero(&exists->b_refcnt)); | |
3923 | arc_change_state(arc_anon, exists, hash_lock); | |
3924 | mutex_exit(hash_lock); | |
3925 | arc_hdr_destroy(exists); | |
3926 | exists = buf_hash_insert(hdr, &hash_lock); | |
3927 | ASSERT3P(exists, ==, NULL); | |
03c6040b GW |
3928 | } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { |
3929 | /* nopwrite */ | |
3930 | ASSERT(zio->io_prop.zp_nopwrite); | |
3931 | if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) | |
3932 | panic("bad nopwrite, hdr=%p exists=%p", | |
3933 | (void *)hdr, (void *)exists); | |
428870ff BB |
3934 | } else { |
3935 | /* Dedup */ | |
3936 | ASSERT(hdr->b_datacnt == 1); | |
3937 | ASSERT(hdr->b_state == arc_anon); | |
3938 | ASSERT(BP_GET_DEDUP(zio->io_bp)); | |
3939 | ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); | |
3940 | } | |
34dc7c2f BB |
3941 | } |
3942 | hdr->b_flags &= ~ARC_IO_IN_PROGRESS; | |
b128c09f | 3943 | /* if it's not anon, we are doing a scrub */ |
428870ff | 3944 | if (!exists && hdr->b_state == arc_anon) |
b128c09f | 3945 | arc_access(hdr, hash_lock); |
34dc7c2f | 3946 | mutex_exit(hash_lock); |
34dc7c2f BB |
3947 | } else { |
3948 | hdr->b_flags &= ~ARC_IO_IN_PROGRESS; | |
3949 | } | |
3950 | ||
428870ff BB |
3951 | ASSERT(!refcount_is_zero(&hdr->b_refcnt)); |
3952 | callback->awcb_done(zio, buf, callback->awcb_private); | |
34dc7c2f BB |
3953 | |
3954 | kmem_free(callback, sizeof (arc_write_callback_t)); | |
3955 | } | |
3956 | ||
3957 | zio_t * | |
428870ff | 3958 | arc_write(zio_t *pio, spa_t *spa, uint64_t txg, |
3a17a7a9 | 3959 | blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, |
e8b96c60 MA |
3960 | const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, |
3961 | arc_done_func_t *done, void *private, zio_priority_t priority, | |
3962 | int zio_flags, const zbookmark_t *zb) | |
34dc7c2f BB |
3963 | { |
3964 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
3965 | arc_write_callback_t *callback; | |
b128c09f | 3966 | zio_t *zio; |
34dc7c2f | 3967 | |
b128c09f | 3968 | ASSERT(ready != NULL); |
428870ff | 3969 | ASSERT(done != NULL); |
34dc7c2f BB |
3970 | ASSERT(!HDR_IO_ERROR(hdr)); |
3971 | ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); | |
428870ff | 3972 | ASSERT(hdr->b_acb == NULL); |
b128c09f BB |
3973 | if (l2arc) |
3974 | hdr->b_flags |= ARC_L2CACHE; | |
3a17a7a9 SK |
3975 | if (l2arc_compress) |
3976 | hdr->b_flags |= ARC_L2COMPRESS; | |
b8d06fca | 3977 | callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_PUSHPAGE); |
34dc7c2f | 3978 | callback->awcb_ready = ready; |
e8b96c60 | 3979 | callback->awcb_physdone = physdone; |
34dc7c2f BB |
3980 | callback->awcb_done = done; |
3981 | callback->awcb_private = private; | |
3982 | callback->awcb_buf = buf; | |
b128c09f | 3983 | |
428870ff | 3984 | zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, |
e8b96c60 MA |
3985 | arc_write_ready, arc_write_physdone, arc_write_done, callback, |
3986 | priority, zio_flags, zb); | |
34dc7c2f BB |
3987 | |
3988 | return (zio); | |
3989 | } | |
3990 | ||
34dc7c2f | 3991 | static int |
e8b96c60 | 3992 | arc_memory_throttle(uint64_t reserve, uint64_t txg) |
34dc7c2f BB |
3993 | { |
3994 | #ifdef _KERNEL | |
0c5493d4 BB |
3995 | if (zfs_arc_memory_throttle_disable) |
3996 | return (0); | |
3997 | ||
e8b96c60 | 3998 | if (freemem <= physmem * arc_lotsfree_percent / 100) { |
34dc7c2f | 3999 | ARCSTAT_INCR(arcstat_memory_throttle_count, 1); |
570827e1 | 4000 | DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim); |
2e528b49 | 4001 | return (SET_ERROR(EAGAIN)); |
34dc7c2f | 4002 | } |
34dc7c2f BB |
4003 | #endif |
4004 | return (0); | |
4005 | } | |
4006 | ||
4007 | void | |
4008 | arc_tempreserve_clear(uint64_t reserve) | |
4009 | { | |
4010 | atomic_add_64(&arc_tempreserve, -reserve); | |
4011 | ASSERT((int64_t)arc_tempreserve >= 0); | |
4012 | } | |
4013 | ||
4014 | int | |
4015 | arc_tempreserve_space(uint64_t reserve, uint64_t txg) | |
4016 | { | |
4017 | int error; | |
9babb374 | 4018 | uint64_t anon_size; |
34dc7c2f | 4019 | |
34dc7c2f BB |
4020 | if (reserve > arc_c/4 && !arc_no_grow) |
4021 | arc_c = MIN(arc_c_max, reserve * 4); | |
12f9a6a3 BB |
4022 | |
4023 | /* | |
4024 | * Throttle when the calculated memory footprint for the TXG | |
4025 | * exceeds the target ARC size. | |
4026 | */ | |
570827e1 BB |
4027 | if (reserve > arc_c) { |
4028 | DMU_TX_STAT_BUMP(dmu_tx_memory_reserve); | |
12f9a6a3 | 4029 | return (SET_ERROR(ERESTART)); |
570827e1 | 4030 | } |
34dc7c2f | 4031 | |
9babb374 BB |
4032 | /* |
4033 | * Don't count loaned bufs as in flight dirty data to prevent long | |
4034 | * network delays from blocking transactions that are ready to be | |
4035 | * assigned to a txg. | |
4036 | */ | |
4037 | anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); | |
4038 | ||
34dc7c2f BB |
4039 | /* |
4040 | * Writes will, almost always, require additional memory allocations | |
d3cc8b15 | 4041 | * in order to compress/encrypt/etc the data. We therefore need to |
34dc7c2f BB |
4042 | * make sure that there is sufficient available memory for this. |
4043 | */ | |
e8b96c60 MA |
4044 | error = arc_memory_throttle(reserve, txg); |
4045 | if (error != 0) | |
34dc7c2f BB |
4046 | return (error); |
4047 | ||
4048 | /* | |
4049 | * Throttle writes when the amount of dirty data in the cache | |
4050 | * gets too large. We try to keep the cache less than half full | |
4051 | * of dirty blocks so that our sync times don't grow too large. | |
4052 | * Note: if two requests come in concurrently, we might let them | |
4053 | * both succeed, when one of them should fail. Not a huge deal. | |
4054 | */ | |
9babb374 BB |
4055 | |
4056 | if (reserve + arc_tempreserve + anon_size > arc_c / 2 && | |
4057 | anon_size > arc_c / 4) { | |
34dc7c2f BB |
4058 | dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " |
4059 | "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", | |
4060 | arc_tempreserve>>10, | |
4061 | arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, | |
4062 | arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, | |
4063 | reserve>>10, arc_c>>10); | |
570827e1 | 4064 | DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle); |
2e528b49 | 4065 | return (SET_ERROR(ERESTART)); |
34dc7c2f BB |
4066 | } |
4067 | atomic_add_64(&arc_tempreserve, reserve); | |
4068 | return (0); | |
4069 | } | |
4070 | ||
13be560d BB |
4071 | static void |
4072 | arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, | |
4073 | kstat_named_t *evict_data, kstat_named_t *evict_metadata) | |
4074 | { | |
4075 | size->value.ui64 = state->arcs_size; | |
4076 | evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; | |
4077 | evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; | |
4078 | } | |
4079 | ||
4080 | static int | |
4081 | arc_kstat_update(kstat_t *ksp, int rw) | |
4082 | { | |
4083 | arc_stats_t *as = ksp->ks_data; | |
4084 | ||
4085 | if (rw == KSTAT_WRITE) { | |
2e528b49 | 4086 | return (SET_ERROR(EACCES)); |
13be560d BB |
4087 | } else { |
4088 | arc_kstat_update_state(arc_anon, | |
4089 | &as->arcstat_anon_size, | |
4090 | &as->arcstat_anon_evict_data, | |
4091 | &as->arcstat_anon_evict_metadata); | |
4092 | arc_kstat_update_state(arc_mru, | |
4093 | &as->arcstat_mru_size, | |
4094 | &as->arcstat_mru_evict_data, | |
4095 | &as->arcstat_mru_evict_metadata); | |
4096 | arc_kstat_update_state(arc_mru_ghost, | |
4097 | &as->arcstat_mru_ghost_size, | |
4098 | &as->arcstat_mru_ghost_evict_data, | |
4099 | &as->arcstat_mru_ghost_evict_metadata); | |
4100 | arc_kstat_update_state(arc_mfu, | |
4101 | &as->arcstat_mfu_size, | |
4102 | &as->arcstat_mfu_evict_data, | |
4103 | &as->arcstat_mfu_evict_metadata); | |
fc41c640 | 4104 | arc_kstat_update_state(arc_mfu_ghost, |
13be560d BB |
4105 | &as->arcstat_mfu_ghost_size, |
4106 | &as->arcstat_mfu_ghost_evict_data, | |
4107 | &as->arcstat_mfu_ghost_evict_metadata); | |
4108 | } | |
4109 | ||
4110 | return (0); | |
4111 | } | |
4112 | ||
34dc7c2f BB |
4113 | void |
4114 | arc_init(void) | |
4115 | { | |
4116 | mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); | |
4117 | cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); | |
4118 | ||
4119 | /* Convert seconds to clock ticks */ | |
bce45ec9 | 4120 | zfs_arc_min_prefetch_lifespan = 1 * hz; |
34dc7c2f BB |
4121 | |
4122 | /* Start out with 1/8 of all memory */ | |
4123 | arc_c = physmem * PAGESIZE / 8; | |
4124 | ||
4125 | #ifdef _KERNEL | |
4126 | /* | |
4127 | * On architectures where the physical memory can be larger | |
4128 | * than the addressable space (intel in 32-bit mode), we may | |
4129 | * need to limit the cache to 1/8 of VM size. | |
4130 | */ | |
4131 | arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); | |
7cb67b45 BB |
4132 | /* |
4133 | * Register a shrinker to support synchronous (direct) memory | |
4134 | * reclaim from the arc. This is done to prevent kswapd from | |
4135 | * swapping out pages when it is preferable to shrink the arc. | |
4136 | */ | |
4137 | spl_register_shrinker(&arc_shrinker); | |
34dc7c2f BB |
4138 | #endif |
4139 | ||
91415825 BB |
4140 | /* set min cache to zero */ |
4141 | arc_c_min = 4<<20; | |
518b4876 | 4142 | /* set max to 1/2 of all memory */ |
be5db977 | 4143 | arc_c_max = arc_c * 4; |
34dc7c2f BB |
4144 | |
4145 | /* | |
4146 | * Allow the tunables to override our calculations if they are | |
4147 | * reasonable (ie. over 64MB) | |
4148 | */ | |
4149 | if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE) | |
4150 | arc_c_max = zfs_arc_max; | |
91415825 | 4151 | if (zfs_arc_min > 0 && zfs_arc_min <= arc_c_max) |
34dc7c2f BB |
4152 | arc_c_min = zfs_arc_min; |
4153 | ||
4154 | arc_c = arc_c_max; | |
4155 | arc_p = (arc_c >> 1); | |
4156 | ||
2b13331d PS |
4157 | /* limit meta-data to 3/4 of the arc capacity */ |
4158 | arc_meta_limit = (3 * arc_c_max) / 4; | |
1834f2d8 | 4159 | arc_meta_max = 0; |
34dc7c2f BB |
4160 | |
4161 | /* Allow the tunable to override if it is reasonable */ | |
4162 | if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) | |
4163 | arc_meta_limit = zfs_arc_meta_limit; | |
4164 | ||
34dc7c2f BB |
4165 | /* if kmem_flags are set, lets try to use less memory */ |
4166 | if (kmem_debugging()) | |
4167 | arc_c = arc_c / 2; | |
4168 | if (arc_c < arc_c_min) | |
4169 | arc_c = arc_c_min; | |
4170 | ||
4171 | arc_anon = &ARC_anon; | |
4172 | arc_mru = &ARC_mru; | |
4173 | arc_mru_ghost = &ARC_mru_ghost; | |
4174 | arc_mfu = &ARC_mfu; | |
4175 | arc_mfu_ghost = &ARC_mfu_ghost; | |
4176 | arc_l2c_only = &ARC_l2c_only; | |
4177 | arc_size = 0; | |
4178 | ||
4179 | mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); | |
4180 | mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); | |
4181 | mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); | |
4182 | mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); | |
4183 | mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); | |
4184 | mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); | |
4185 | ||
4186 | list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], | |
4187 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4188 | list_create(&arc_mru->arcs_list[ARC_BUFC_DATA], | |
4189 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4190 | list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], | |
4191 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4192 | list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], | |
4193 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4194 | list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], | |
4195 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4196 | list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], | |
4197 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4198 | list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], | |
4199 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4200 | list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], | |
4201 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4202 | list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], | |
4203 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4204 | list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], | |
4205 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4206 | ||
e0b0ca98 BB |
4207 | arc_anon->arcs_state = ARC_STATE_ANON; |
4208 | arc_mru->arcs_state = ARC_STATE_MRU; | |
4209 | arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST; | |
4210 | arc_mfu->arcs_state = ARC_STATE_MFU; | |
4211 | arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST; | |
4212 | arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY; | |
4213 | ||
34dc7c2f BB |
4214 | buf_init(); |
4215 | ||
4216 | arc_thread_exit = 0; | |
ab26409d BB |
4217 | list_create(&arc_prune_list, sizeof (arc_prune_t), |
4218 | offsetof(arc_prune_t, p_node)); | |
34dc7c2f | 4219 | arc_eviction_list = NULL; |
ab26409d | 4220 | mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f BB |
4221 | mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); |
4222 | bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); | |
4223 | ||
4224 | arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, | |
4225 | sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); | |
4226 | ||
4227 | if (arc_ksp != NULL) { | |
4228 | arc_ksp->ks_data = &arc_stats; | |
13be560d | 4229 | arc_ksp->ks_update = arc_kstat_update; |
34dc7c2f BB |
4230 | kstat_install(arc_ksp); |
4231 | } | |
4232 | ||
302f753f | 4233 | (void) thread_create(NULL, 0, arc_adapt_thread, NULL, 0, &p0, |
34dc7c2f BB |
4234 | TS_RUN, minclsyspri); |
4235 | ||
4236 | arc_dead = FALSE; | |
b128c09f | 4237 | arc_warm = B_FALSE; |
34dc7c2f | 4238 | |
e8b96c60 MA |
4239 | /* |
4240 | * Calculate maximum amount of dirty data per pool. | |
4241 | * | |
4242 | * If it has been set by a module parameter, take that. | |
4243 | * Otherwise, use a percentage of physical memory defined by | |
4244 | * zfs_dirty_data_max_percent (default 10%) with a cap at | |
4245 | * zfs_dirty_data_max_max (default 25% of physical memory). | |
4246 | */ | |
4247 | if (zfs_dirty_data_max_max == 0) | |
4248 | zfs_dirty_data_max_max = physmem * PAGESIZE * | |
4249 | zfs_dirty_data_max_max_percent / 100; | |
4250 | ||
4251 | if (zfs_dirty_data_max == 0) { | |
4252 | zfs_dirty_data_max = physmem * PAGESIZE * | |
4253 | zfs_dirty_data_max_percent / 100; | |
4254 | zfs_dirty_data_max = MIN(zfs_dirty_data_max, | |
4255 | zfs_dirty_data_max_max); | |
4256 | } | |
34dc7c2f BB |
4257 | } |
4258 | ||
4259 | void | |
4260 | arc_fini(void) | |
4261 | { | |
ab26409d BB |
4262 | arc_prune_t *p; |
4263 | ||
34dc7c2f | 4264 | mutex_enter(&arc_reclaim_thr_lock); |
7cb67b45 BB |
4265 | #ifdef _KERNEL |
4266 | spl_unregister_shrinker(&arc_shrinker); | |
4267 | #endif /* _KERNEL */ | |
4268 | ||
34dc7c2f BB |
4269 | arc_thread_exit = 1; |
4270 | while (arc_thread_exit != 0) | |
4271 | cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); | |
4272 | mutex_exit(&arc_reclaim_thr_lock); | |
4273 | ||
4274 | arc_flush(NULL); | |
4275 | ||
4276 | arc_dead = TRUE; | |
4277 | ||
4278 | if (arc_ksp != NULL) { | |
4279 | kstat_delete(arc_ksp); | |
4280 | arc_ksp = NULL; | |
4281 | } | |
4282 | ||
ab26409d BB |
4283 | mutex_enter(&arc_prune_mtx); |
4284 | while ((p = list_head(&arc_prune_list)) != NULL) { | |
4285 | list_remove(&arc_prune_list, p); | |
4286 | refcount_remove(&p->p_refcnt, &arc_prune_list); | |
4287 | refcount_destroy(&p->p_refcnt); | |
4288 | kmem_free(p, sizeof (*p)); | |
4289 | } | |
4290 | mutex_exit(&arc_prune_mtx); | |
4291 | ||
4292 | list_destroy(&arc_prune_list); | |
4293 | mutex_destroy(&arc_prune_mtx); | |
34dc7c2f BB |
4294 | mutex_destroy(&arc_eviction_mtx); |
4295 | mutex_destroy(&arc_reclaim_thr_lock); | |
4296 | cv_destroy(&arc_reclaim_thr_cv); | |
4297 | ||
4298 | list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); | |
4299 | list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); | |
4300 | list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); | |
4301 | list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); | |
4302 | list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); | |
4303 | list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); | |
4304 | list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); | |
4305 | list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); | |
4306 | ||
4307 | mutex_destroy(&arc_anon->arcs_mtx); | |
4308 | mutex_destroy(&arc_mru->arcs_mtx); | |
4309 | mutex_destroy(&arc_mru_ghost->arcs_mtx); | |
4310 | mutex_destroy(&arc_mfu->arcs_mtx); | |
4311 | mutex_destroy(&arc_mfu_ghost->arcs_mtx); | |
fb5f0bc8 | 4312 | mutex_destroy(&arc_l2c_only->arcs_mtx); |
34dc7c2f BB |
4313 | |
4314 | buf_fini(); | |
9babb374 BB |
4315 | |
4316 | ASSERT(arc_loaned_bytes == 0); | |
34dc7c2f BB |
4317 | } |
4318 | ||
4319 | /* | |
4320 | * Level 2 ARC | |
4321 | * | |
4322 | * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. | |
4323 | * It uses dedicated storage devices to hold cached data, which are populated | |
4324 | * using large infrequent writes. The main role of this cache is to boost | |
4325 | * the performance of random read workloads. The intended L2ARC devices | |
4326 | * include short-stroked disks, solid state disks, and other media with | |
4327 | * substantially faster read latency than disk. | |
4328 | * | |
4329 | * +-----------------------+ | |
4330 | * | ARC | | |
4331 | * +-----------------------+ | |
4332 | * | ^ ^ | |
4333 | * | | | | |
4334 | * l2arc_feed_thread() arc_read() | |
4335 | * | | | | |
4336 | * | l2arc read | | |
4337 | * V | | | |
4338 | * +---------------+ | | |
4339 | * | L2ARC | | | |
4340 | * +---------------+ | | |
4341 | * | ^ | | |
4342 | * l2arc_write() | | | |
4343 | * | | | | |
4344 | * V | | | |
4345 | * +-------+ +-------+ | |
4346 | * | vdev | | vdev | | |
4347 | * | cache | | cache | | |
4348 | * +-------+ +-------+ | |
4349 | * +=========+ .-----. | |
4350 | * : L2ARC : |-_____-| | |
4351 | * : devices : | Disks | | |
4352 | * +=========+ `-_____-' | |
4353 | * | |
4354 | * Read requests are satisfied from the following sources, in order: | |
4355 | * | |
4356 | * 1) ARC | |
4357 | * 2) vdev cache of L2ARC devices | |
4358 | * 3) L2ARC devices | |
4359 | * 4) vdev cache of disks | |
4360 | * 5) disks | |
4361 | * | |
4362 | * Some L2ARC device types exhibit extremely slow write performance. | |
4363 | * To accommodate for this there are some significant differences between | |
4364 | * the L2ARC and traditional cache design: | |
4365 | * | |
4366 | * 1. There is no eviction path from the ARC to the L2ARC. Evictions from | |
4367 | * the ARC behave as usual, freeing buffers and placing headers on ghost | |
4368 | * lists. The ARC does not send buffers to the L2ARC during eviction as | |
4369 | * this would add inflated write latencies for all ARC memory pressure. | |
4370 | * | |
4371 | * 2. The L2ARC attempts to cache data from the ARC before it is evicted. | |
4372 | * It does this by periodically scanning buffers from the eviction-end of | |
4373 | * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are | |
3a17a7a9 SK |
4374 | * not already there. It scans until a headroom of buffers is satisfied, |
4375 | * which itself is a buffer for ARC eviction. If a compressible buffer is | |
4376 | * found during scanning and selected for writing to an L2ARC device, we | |
4377 | * temporarily boost scanning headroom during the next scan cycle to make | |
4378 | * sure we adapt to compression effects (which might significantly reduce | |
4379 | * the data volume we write to L2ARC). The thread that does this is | |
34dc7c2f BB |
4380 | * l2arc_feed_thread(), illustrated below; example sizes are included to |
4381 | * provide a better sense of ratio than this diagram: | |
4382 | * | |
4383 | * head --> tail | |
4384 | * +---------------------+----------+ | |
4385 | * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC | |
4386 | * +---------------------+----------+ | o L2ARC eligible | |
4387 | * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer | |
4388 | * +---------------------+----------+ | | |
4389 | * 15.9 Gbytes ^ 32 Mbytes | | |
4390 | * headroom | | |
4391 | * l2arc_feed_thread() | |
4392 | * | | |
4393 | * l2arc write hand <--[oooo]--' | |
4394 | * | 8 Mbyte | |
4395 | * | write max | |
4396 | * V | |
4397 | * +==============================+ | |
4398 | * L2ARC dev |####|#|###|###| |####| ... | | |
4399 | * +==============================+ | |
4400 | * 32 Gbytes | |
4401 | * | |
4402 | * 3. If an ARC buffer is copied to the L2ARC but then hit instead of | |
4403 | * evicted, then the L2ARC has cached a buffer much sooner than it probably | |
4404 | * needed to, potentially wasting L2ARC device bandwidth and storage. It is | |
4405 | * safe to say that this is an uncommon case, since buffers at the end of | |
4406 | * the ARC lists have moved there due to inactivity. | |
4407 | * | |
4408 | * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, | |
4409 | * then the L2ARC simply misses copying some buffers. This serves as a | |
4410 | * pressure valve to prevent heavy read workloads from both stalling the ARC | |
4411 | * with waits and clogging the L2ARC with writes. This also helps prevent | |
4412 | * the potential for the L2ARC to churn if it attempts to cache content too | |
4413 | * quickly, such as during backups of the entire pool. | |
4414 | * | |
b128c09f BB |
4415 | * 5. After system boot and before the ARC has filled main memory, there are |
4416 | * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru | |
4417 | * lists can remain mostly static. Instead of searching from tail of these | |
4418 | * lists as pictured, the l2arc_feed_thread() will search from the list heads | |
4419 | * for eligible buffers, greatly increasing its chance of finding them. | |
4420 | * | |
4421 | * The L2ARC device write speed is also boosted during this time so that | |
4422 | * the L2ARC warms up faster. Since there have been no ARC evictions yet, | |
4423 | * there are no L2ARC reads, and no fear of degrading read performance | |
4424 | * through increased writes. | |
4425 | * | |
4426 | * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that | |
34dc7c2f BB |
4427 | * the vdev queue can aggregate them into larger and fewer writes. Each |
4428 | * device is written to in a rotor fashion, sweeping writes through | |
4429 | * available space then repeating. | |
4430 | * | |
b128c09f | 4431 | * 7. The L2ARC does not store dirty content. It never needs to flush |
34dc7c2f BB |
4432 | * write buffers back to disk based storage. |
4433 | * | |
b128c09f | 4434 | * 8. If an ARC buffer is written (and dirtied) which also exists in the |
34dc7c2f BB |
4435 | * L2ARC, the now stale L2ARC buffer is immediately dropped. |
4436 | * | |
4437 | * The performance of the L2ARC can be tweaked by a number of tunables, which | |
4438 | * may be necessary for different workloads: | |
4439 | * | |
4440 | * l2arc_write_max max write bytes per interval | |
b128c09f | 4441 | * l2arc_write_boost extra write bytes during device warmup |
34dc7c2f | 4442 | * l2arc_noprefetch skip caching prefetched buffers |
3a17a7a9 | 4443 | * l2arc_nocompress skip compressing buffers |
34dc7c2f | 4444 | * l2arc_headroom number of max device writes to precache |
3a17a7a9 SK |
4445 | * l2arc_headroom_boost when we find compressed buffers during ARC |
4446 | * scanning, we multiply headroom by this | |
4447 | * percentage factor for the next scan cycle, | |
4448 | * since more compressed buffers are likely to | |
4449 | * be present | |
34dc7c2f BB |
4450 | * l2arc_feed_secs seconds between L2ARC writing |
4451 | * | |
4452 | * Tunables may be removed or added as future performance improvements are | |
4453 | * integrated, and also may become zpool properties. | |
d164b209 BB |
4454 | * |
4455 | * There are three key functions that control how the L2ARC warms up: | |
4456 | * | |
4457 | * l2arc_write_eligible() check if a buffer is eligible to cache | |
4458 | * l2arc_write_size() calculate how much to write | |
4459 | * l2arc_write_interval() calculate sleep delay between writes | |
4460 | * | |
4461 | * These three functions determine what to write, how much, and how quickly | |
4462 | * to send writes. | |
34dc7c2f BB |
4463 | */ |
4464 | ||
d164b209 BB |
4465 | static boolean_t |
4466 | l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) | |
4467 | { | |
4468 | /* | |
4469 | * A buffer is *not* eligible for the L2ARC if it: | |
4470 | * 1. belongs to a different spa. | |
428870ff BB |
4471 | * 2. is already cached on the L2ARC. |
4472 | * 3. has an I/O in progress (it may be an incomplete read). | |
4473 | * 4. is flagged not eligible (zfs property). | |
d164b209 | 4474 | */ |
428870ff | 4475 | if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL || |
d164b209 BB |
4476 | HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab)) |
4477 | return (B_FALSE); | |
4478 | ||
4479 | return (B_TRUE); | |
4480 | } | |
4481 | ||
4482 | static uint64_t | |
3a17a7a9 | 4483 | l2arc_write_size(void) |
d164b209 BB |
4484 | { |
4485 | uint64_t size; | |
4486 | ||
3a17a7a9 SK |
4487 | /* |
4488 | * Make sure our globals have meaningful values in case the user | |
4489 | * altered them. | |
4490 | */ | |
4491 | size = l2arc_write_max; | |
4492 | if (size == 0) { | |
4493 | cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " | |
4494 | "be greater than zero, resetting it to the default (%d)", | |
4495 | L2ARC_WRITE_SIZE); | |
4496 | size = l2arc_write_max = L2ARC_WRITE_SIZE; | |
4497 | } | |
d164b209 BB |
4498 | |
4499 | if (arc_warm == B_FALSE) | |
3a17a7a9 | 4500 | size += l2arc_write_boost; |
d164b209 BB |
4501 | |
4502 | return (size); | |
4503 | ||
4504 | } | |
4505 | ||
4506 | static clock_t | |
4507 | l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) | |
4508 | { | |
428870ff | 4509 | clock_t interval, next, now; |
d164b209 BB |
4510 | |
4511 | /* | |
4512 | * If the ARC lists are busy, increase our write rate; if the | |
4513 | * lists are stale, idle back. This is achieved by checking | |
4514 | * how much we previously wrote - if it was more than half of | |
4515 | * what we wanted, schedule the next write much sooner. | |
4516 | */ | |
4517 | if (l2arc_feed_again && wrote > (wanted / 2)) | |
4518 | interval = (hz * l2arc_feed_min_ms) / 1000; | |
4519 | else | |
4520 | interval = hz * l2arc_feed_secs; | |
4521 | ||
428870ff BB |
4522 | now = ddi_get_lbolt(); |
4523 | next = MAX(now, MIN(now + interval, began + interval)); | |
d164b209 BB |
4524 | |
4525 | return (next); | |
4526 | } | |
4527 | ||
34dc7c2f BB |
4528 | static void |
4529 | l2arc_hdr_stat_add(void) | |
4530 | { | |
6e1d7276 | 4531 | ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE); |
34dc7c2f BB |
4532 | ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); |
4533 | } | |
4534 | ||
4535 | static void | |
4536 | l2arc_hdr_stat_remove(void) | |
4537 | { | |
6e1d7276 | 4538 | ARCSTAT_INCR(arcstat_l2_hdr_size, -HDR_SIZE); |
34dc7c2f BB |
4539 | ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); |
4540 | } | |
4541 | ||
4542 | /* | |
4543 | * Cycle through L2ARC devices. This is how L2ARC load balances. | |
b128c09f | 4544 | * If a device is returned, this also returns holding the spa config lock. |
34dc7c2f BB |
4545 | */ |
4546 | static l2arc_dev_t * | |
4547 | l2arc_dev_get_next(void) | |
4548 | { | |
b128c09f | 4549 | l2arc_dev_t *first, *next = NULL; |
34dc7c2f | 4550 | |
b128c09f BB |
4551 | /* |
4552 | * Lock out the removal of spas (spa_namespace_lock), then removal | |
4553 | * of cache devices (l2arc_dev_mtx). Once a device has been selected, | |
4554 | * both locks will be dropped and a spa config lock held instead. | |
4555 | */ | |
4556 | mutex_enter(&spa_namespace_lock); | |
4557 | mutex_enter(&l2arc_dev_mtx); | |
4558 | ||
4559 | /* if there are no vdevs, there is nothing to do */ | |
4560 | if (l2arc_ndev == 0) | |
4561 | goto out; | |
4562 | ||
4563 | first = NULL; | |
4564 | next = l2arc_dev_last; | |
4565 | do { | |
4566 | /* loop around the list looking for a non-faulted vdev */ | |
4567 | if (next == NULL) { | |
34dc7c2f | 4568 | next = list_head(l2arc_dev_list); |
b128c09f BB |
4569 | } else { |
4570 | next = list_next(l2arc_dev_list, next); | |
4571 | if (next == NULL) | |
4572 | next = list_head(l2arc_dev_list); | |
4573 | } | |
4574 | ||
4575 | /* if we have come back to the start, bail out */ | |
4576 | if (first == NULL) | |
4577 | first = next; | |
4578 | else if (next == first) | |
4579 | break; | |
4580 | ||
4581 | } while (vdev_is_dead(next->l2ad_vdev)); | |
4582 | ||
4583 | /* if we were unable to find any usable vdevs, return NULL */ | |
4584 | if (vdev_is_dead(next->l2ad_vdev)) | |
4585 | next = NULL; | |
34dc7c2f BB |
4586 | |
4587 | l2arc_dev_last = next; | |
4588 | ||
b128c09f BB |
4589 | out: |
4590 | mutex_exit(&l2arc_dev_mtx); | |
4591 | ||
4592 | /* | |
4593 | * Grab the config lock to prevent the 'next' device from being | |
4594 | * removed while we are writing to it. | |
4595 | */ | |
4596 | if (next != NULL) | |
4597 | spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); | |
4598 | mutex_exit(&spa_namespace_lock); | |
4599 | ||
34dc7c2f BB |
4600 | return (next); |
4601 | } | |
4602 | ||
b128c09f BB |
4603 | /* |
4604 | * Free buffers that were tagged for destruction. | |
4605 | */ | |
4606 | static void | |
0bc8fd78 | 4607 | l2arc_do_free_on_write(void) |
b128c09f BB |
4608 | { |
4609 | list_t *buflist; | |
4610 | l2arc_data_free_t *df, *df_prev; | |
4611 | ||
4612 | mutex_enter(&l2arc_free_on_write_mtx); | |
4613 | buflist = l2arc_free_on_write; | |
4614 | ||
4615 | for (df = list_tail(buflist); df; df = df_prev) { | |
4616 | df_prev = list_prev(buflist, df); | |
4617 | ASSERT(df->l2df_data != NULL); | |
4618 | ASSERT(df->l2df_func != NULL); | |
4619 | df->l2df_func(df->l2df_data, df->l2df_size); | |
4620 | list_remove(buflist, df); | |
4621 | kmem_free(df, sizeof (l2arc_data_free_t)); | |
4622 | } | |
4623 | ||
4624 | mutex_exit(&l2arc_free_on_write_mtx); | |
4625 | } | |
4626 | ||
34dc7c2f BB |
4627 | /* |
4628 | * A write to a cache device has completed. Update all headers to allow | |
4629 | * reads from these buffers to begin. | |
4630 | */ | |
4631 | static void | |
4632 | l2arc_write_done(zio_t *zio) | |
4633 | { | |
4634 | l2arc_write_callback_t *cb; | |
4635 | l2arc_dev_t *dev; | |
4636 | list_t *buflist; | |
34dc7c2f | 4637 | arc_buf_hdr_t *head, *ab, *ab_prev; |
b128c09f | 4638 | l2arc_buf_hdr_t *abl2; |
34dc7c2f BB |
4639 | kmutex_t *hash_lock; |
4640 | ||
4641 | cb = zio->io_private; | |
4642 | ASSERT(cb != NULL); | |
4643 | dev = cb->l2wcb_dev; | |
4644 | ASSERT(dev != NULL); | |
4645 | head = cb->l2wcb_head; | |
4646 | ASSERT(head != NULL); | |
4647 | buflist = dev->l2ad_buflist; | |
4648 | ASSERT(buflist != NULL); | |
4649 | DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, | |
4650 | l2arc_write_callback_t *, cb); | |
4651 | ||
4652 | if (zio->io_error != 0) | |
4653 | ARCSTAT_BUMP(arcstat_l2_writes_error); | |
4654 | ||
4655 | mutex_enter(&l2arc_buflist_mtx); | |
4656 | ||
4657 | /* | |
4658 | * All writes completed, or an error was hit. | |
4659 | */ | |
4660 | for (ab = list_prev(buflist, head); ab; ab = ab_prev) { | |
4661 | ab_prev = list_prev(buflist, ab); | |
1ca546b3 SK |
4662 | abl2 = ab->b_l2hdr; |
4663 | ||
4664 | /* | |
4665 | * Release the temporary compressed buffer as soon as possible. | |
4666 | */ | |
4667 | if (abl2->b_compress != ZIO_COMPRESS_OFF) | |
4668 | l2arc_release_cdata_buf(ab); | |
34dc7c2f BB |
4669 | |
4670 | hash_lock = HDR_LOCK(ab); | |
4671 | if (!mutex_tryenter(hash_lock)) { | |
4672 | /* | |
4673 | * This buffer misses out. It may be in a stage | |
4674 | * of eviction. Its ARC_L2_WRITING flag will be | |
4675 | * left set, denying reads to this buffer. | |
4676 | */ | |
4677 | ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); | |
4678 | continue; | |
4679 | } | |
4680 | ||
4681 | if (zio->io_error != 0) { | |
4682 | /* | |
b128c09f | 4683 | * Error - drop L2ARC entry. |
34dc7c2f | 4684 | */ |
b128c09f | 4685 | list_remove(buflist, ab); |
3a17a7a9 | 4686 | ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); |
34dc7c2f | 4687 | ab->b_l2hdr = NULL; |
ecf3d9b8 | 4688 | kmem_cache_free(l2arc_hdr_cache, abl2); |
6e1d7276 | 4689 | arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS); |
b128c09f | 4690 | ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); |
34dc7c2f BB |
4691 | } |
4692 | ||
4693 | /* | |
4694 | * Allow ARC to begin reads to this L2ARC entry. | |
4695 | */ | |
4696 | ab->b_flags &= ~ARC_L2_WRITING; | |
4697 | ||
4698 | mutex_exit(hash_lock); | |
4699 | } | |
4700 | ||
4701 | atomic_inc_64(&l2arc_writes_done); | |
4702 | list_remove(buflist, head); | |
4703 | kmem_cache_free(hdr_cache, head); | |
4704 | mutex_exit(&l2arc_buflist_mtx); | |
4705 | ||
b128c09f | 4706 | l2arc_do_free_on_write(); |
34dc7c2f BB |
4707 | |
4708 | kmem_free(cb, sizeof (l2arc_write_callback_t)); | |
4709 | } | |
4710 | ||
4711 | /* | |
4712 | * A read to a cache device completed. Validate buffer contents before | |
4713 | * handing over to the regular ARC routines. | |
4714 | */ | |
4715 | static void | |
4716 | l2arc_read_done(zio_t *zio) | |
4717 | { | |
4718 | l2arc_read_callback_t *cb; | |
4719 | arc_buf_hdr_t *hdr; | |
4720 | arc_buf_t *buf; | |
34dc7c2f | 4721 | kmutex_t *hash_lock; |
b128c09f BB |
4722 | int equal; |
4723 | ||
4724 | ASSERT(zio->io_vd != NULL); | |
4725 | ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); | |
4726 | ||
4727 | spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); | |
34dc7c2f BB |
4728 | |
4729 | cb = zio->io_private; | |
4730 | ASSERT(cb != NULL); | |
4731 | buf = cb->l2rcb_buf; | |
4732 | ASSERT(buf != NULL); | |
34dc7c2f | 4733 | |
428870ff | 4734 | hash_lock = HDR_LOCK(buf->b_hdr); |
34dc7c2f | 4735 | mutex_enter(hash_lock); |
428870ff BB |
4736 | hdr = buf->b_hdr; |
4737 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
34dc7c2f | 4738 | |
3a17a7a9 SK |
4739 | /* |
4740 | * If the buffer was compressed, decompress it first. | |
4741 | */ | |
4742 | if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) | |
4743 | l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); | |
4744 | ASSERT(zio->io_data != NULL); | |
4745 | ||
34dc7c2f BB |
4746 | /* |
4747 | * Check this survived the L2ARC journey. | |
4748 | */ | |
4749 | equal = arc_cksum_equal(buf); | |
4750 | if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { | |
4751 | mutex_exit(hash_lock); | |
4752 | zio->io_private = buf; | |
b128c09f BB |
4753 | zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ |
4754 | zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ | |
34dc7c2f BB |
4755 | arc_read_done(zio); |
4756 | } else { | |
4757 | mutex_exit(hash_lock); | |
4758 | /* | |
4759 | * Buffer didn't survive caching. Increment stats and | |
4760 | * reissue to the original storage device. | |
4761 | */ | |
b128c09f | 4762 | if (zio->io_error != 0) { |
34dc7c2f | 4763 | ARCSTAT_BUMP(arcstat_l2_io_error); |
b128c09f | 4764 | } else { |
2e528b49 | 4765 | zio->io_error = SET_ERROR(EIO); |
b128c09f | 4766 | } |
34dc7c2f BB |
4767 | if (!equal) |
4768 | ARCSTAT_BUMP(arcstat_l2_cksum_bad); | |
4769 | ||
34dc7c2f | 4770 | /* |
b128c09f BB |
4771 | * If there's no waiter, issue an async i/o to the primary |
4772 | * storage now. If there *is* a waiter, the caller must | |
4773 | * issue the i/o in a context where it's OK to block. | |
34dc7c2f | 4774 | */ |
d164b209 BB |
4775 | if (zio->io_waiter == NULL) { |
4776 | zio_t *pio = zio_unique_parent(zio); | |
4777 | ||
4778 | ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); | |
4779 | ||
4780 | zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, | |
b128c09f BB |
4781 | buf->b_data, zio->io_size, arc_read_done, buf, |
4782 | zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); | |
d164b209 | 4783 | } |
34dc7c2f BB |
4784 | } |
4785 | ||
4786 | kmem_free(cb, sizeof (l2arc_read_callback_t)); | |
4787 | } | |
4788 | ||
4789 | /* | |
4790 | * This is the list priority from which the L2ARC will search for pages to | |
4791 | * cache. This is used within loops (0..3) to cycle through lists in the | |
4792 | * desired order. This order can have a significant effect on cache | |
4793 | * performance. | |
4794 | * | |
4795 | * Currently the metadata lists are hit first, MFU then MRU, followed by | |
4796 | * the data lists. This function returns a locked list, and also returns | |
4797 | * the lock pointer. | |
4798 | */ | |
4799 | static list_t * | |
4800 | l2arc_list_locked(int list_num, kmutex_t **lock) | |
4801 | { | |
d4ed6673 | 4802 | list_t *list = NULL; |
34dc7c2f BB |
4803 | |
4804 | ASSERT(list_num >= 0 && list_num <= 3); | |
4805 | ||
4806 | switch (list_num) { | |
4807 | case 0: | |
4808 | list = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; | |
4809 | *lock = &arc_mfu->arcs_mtx; | |
4810 | break; | |
4811 | case 1: | |
4812 | list = &arc_mru->arcs_list[ARC_BUFC_METADATA]; | |
4813 | *lock = &arc_mru->arcs_mtx; | |
4814 | break; | |
4815 | case 2: | |
4816 | list = &arc_mfu->arcs_list[ARC_BUFC_DATA]; | |
4817 | *lock = &arc_mfu->arcs_mtx; | |
4818 | break; | |
4819 | case 3: | |
4820 | list = &arc_mru->arcs_list[ARC_BUFC_DATA]; | |
4821 | *lock = &arc_mru->arcs_mtx; | |
4822 | break; | |
4823 | } | |
4824 | ||
4825 | ASSERT(!(MUTEX_HELD(*lock))); | |
4826 | mutex_enter(*lock); | |
4827 | return (list); | |
4828 | } | |
4829 | ||
4830 | /* | |
4831 | * Evict buffers from the device write hand to the distance specified in | |
4832 | * bytes. This distance may span populated buffers, it may span nothing. | |
4833 | * This is clearing a region on the L2ARC device ready for writing. | |
4834 | * If the 'all' boolean is set, every buffer is evicted. | |
4835 | */ | |
4836 | static void | |
4837 | l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) | |
4838 | { | |
4839 | list_t *buflist; | |
4840 | l2arc_buf_hdr_t *abl2; | |
4841 | arc_buf_hdr_t *ab, *ab_prev; | |
4842 | kmutex_t *hash_lock; | |
4843 | uint64_t taddr; | |
4844 | ||
34dc7c2f BB |
4845 | buflist = dev->l2ad_buflist; |
4846 | ||
4847 | if (buflist == NULL) | |
4848 | return; | |
4849 | ||
4850 | if (!all && dev->l2ad_first) { | |
4851 | /* | |
4852 | * This is the first sweep through the device. There is | |
4853 | * nothing to evict. | |
4854 | */ | |
4855 | return; | |
4856 | } | |
4857 | ||
b128c09f | 4858 | if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { |
34dc7c2f BB |
4859 | /* |
4860 | * When nearing the end of the device, evict to the end | |
4861 | * before the device write hand jumps to the start. | |
4862 | */ | |
4863 | taddr = dev->l2ad_end; | |
4864 | } else { | |
4865 | taddr = dev->l2ad_hand + distance; | |
4866 | } | |
4867 | DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, | |
4868 | uint64_t, taddr, boolean_t, all); | |
4869 | ||
4870 | top: | |
4871 | mutex_enter(&l2arc_buflist_mtx); | |
4872 | for (ab = list_tail(buflist); ab; ab = ab_prev) { | |
4873 | ab_prev = list_prev(buflist, ab); | |
4874 | ||
4875 | hash_lock = HDR_LOCK(ab); | |
4876 | if (!mutex_tryenter(hash_lock)) { | |
4877 | /* | |
4878 | * Missed the hash lock. Retry. | |
4879 | */ | |
4880 | ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); | |
4881 | mutex_exit(&l2arc_buflist_mtx); | |
4882 | mutex_enter(hash_lock); | |
4883 | mutex_exit(hash_lock); | |
4884 | goto top; | |
4885 | } | |
4886 | ||
4887 | if (HDR_L2_WRITE_HEAD(ab)) { | |
4888 | /* | |
4889 | * We hit a write head node. Leave it for | |
4890 | * l2arc_write_done(). | |
4891 | */ | |
4892 | list_remove(buflist, ab); | |
4893 | mutex_exit(hash_lock); | |
4894 | continue; | |
4895 | } | |
4896 | ||
4897 | if (!all && ab->b_l2hdr != NULL && | |
4898 | (ab->b_l2hdr->b_daddr > taddr || | |
4899 | ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { | |
4900 | /* | |
4901 | * We've evicted to the target address, | |
4902 | * or the end of the device. | |
4903 | */ | |
4904 | mutex_exit(hash_lock); | |
4905 | break; | |
4906 | } | |
4907 | ||
4908 | if (HDR_FREE_IN_PROGRESS(ab)) { | |
4909 | /* | |
4910 | * Already on the path to destruction. | |
4911 | */ | |
4912 | mutex_exit(hash_lock); | |
4913 | continue; | |
4914 | } | |
4915 | ||
4916 | if (ab->b_state == arc_l2c_only) { | |
4917 | ASSERT(!HDR_L2_READING(ab)); | |
4918 | /* | |
4919 | * This doesn't exist in the ARC. Destroy. | |
4920 | * arc_hdr_destroy() will call list_remove() | |
4921 | * and decrement arcstat_l2_size. | |
4922 | */ | |
4923 | arc_change_state(arc_anon, ab, hash_lock); | |
4924 | arc_hdr_destroy(ab); | |
4925 | } else { | |
b128c09f BB |
4926 | /* |
4927 | * Invalidate issued or about to be issued | |
4928 | * reads, since we may be about to write | |
4929 | * over this location. | |
4930 | */ | |
4931 | if (HDR_L2_READING(ab)) { | |
4932 | ARCSTAT_BUMP(arcstat_l2_evict_reading); | |
4933 | ab->b_flags |= ARC_L2_EVICTED; | |
4934 | } | |
4935 | ||
34dc7c2f BB |
4936 | /* |
4937 | * Tell ARC this no longer exists in L2ARC. | |
4938 | */ | |
4939 | if (ab->b_l2hdr != NULL) { | |
4940 | abl2 = ab->b_l2hdr; | |
3a17a7a9 | 4941 | ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); |
34dc7c2f | 4942 | ab->b_l2hdr = NULL; |
ecf3d9b8 | 4943 | kmem_cache_free(l2arc_hdr_cache, abl2); |
6e1d7276 | 4944 | arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS); |
34dc7c2f BB |
4945 | ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); |
4946 | } | |
4947 | list_remove(buflist, ab); | |
4948 | ||
4949 | /* | |
4950 | * This may have been leftover after a | |
4951 | * failed write. | |
4952 | */ | |
4953 | ab->b_flags &= ~ARC_L2_WRITING; | |
34dc7c2f BB |
4954 | } |
4955 | mutex_exit(hash_lock); | |
4956 | } | |
4957 | mutex_exit(&l2arc_buflist_mtx); | |
4958 | ||
428870ff | 4959 | vdev_space_update(dev->l2ad_vdev, -(taddr - dev->l2ad_evict), 0, 0); |
34dc7c2f BB |
4960 | dev->l2ad_evict = taddr; |
4961 | } | |
4962 | ||
4963 | /* | |
4964 | * Find and write ARC buffers to the L2ARC device. | |
4965 | * | |
4966 | * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid | |
4967 | * for reading until they have completed writing. | |
3a17a7a9 SK |
4968 | * The headroom_boost is an in-out parameter used to maintain headroom boost |
4969 | * state between calls to this function. | |
4970 | * | |
4971 | * Returns the number of bytes actually written (which may be smaller than | |
4972 | * the delta by which the device hand has changed due to alignment). | |
34dc7c2f | 4973 | */ |
d164b209 | 4974 | static uint64_t |
3a17a7a9 SK |
4975 | l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, |
4976 | boolean_t *headroom_boost) | |
34dc7c2f BB |
4977 | { |
4978 | arc_buf_hdr_t *ab, *ab_prev, *head; | |
34dc7c2f | 4979 | list_t *list; |
3a17a7a9 SK |
4980 | uint64_t write_asize, write_psize, write_sz, headroom, |
4981 | buf_compress_minsz; | |
34dc7c2f | 4982 | void *buf_data; |
3a17a7a9 SK |
4983 | kmutex_t *list_lock = NULL; |
4984 | boolean_t full; | |
34dc7c2f BB |
4985 | l2arc_write_callback_t *cb; |
4986 | zio_t *pio, *wzio; | |
3541dc6d | 4987 | uint64_t guid = spa_load_guid(spa); |
d6320ddb | 4988 | int try; |
3a17a7a9 | 4989 | const boolean_t do_headroom_boost = *headroom_boost; |
34dc7c2f | 4990 | |
34dc7c2f BB |
4991 | ASSERT(dev->l2ad_vdev != NULL); |
4992 | ||
3a17a7a9 SK |
4993 | /* Lower the flag now, we might want to raise it again later. */ |
4994 | *headroom_boost = B_FALSE; | |
4995 | ||
34dc7c2f | 4996 | pio = NULL; |
3a17a7a9 | 4997 | write_sz = write_asize = write_psize = 0; |
34dc7c2f BB |
4998 | full = B_FALSE; |
4999 | head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); | |
5000 | head->b_flags |= ARC_L2_WRITE_HEAD; | |
5001 | ||
3a17a7a9 SK |
5002 | /* |
5003 | * We will want to try to compress buffers that are at least 2x the | |
5004 | * device sector size. | |
5005 | */ | |
5006 | buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; | |
5007 | ||
34dc7c2f BB |
5008 | /* |
5009 | * Copy buffers for L2ARC writing. | |
5010 | */ | |
5011 | mutex_enter(&l2arc_buflist_mtx); | |
d6320ddb | 5012 | for (try = 0; try <= 3; try++) { |
3a17a7a9 SK |
5013 | uint64_t passed_sz = 0; |
5014 | ||
34dc7c2f | 5015 | list = l2arc_list_locked(try, &list_lock); |
34dc7c2f | 5016 | |
b128c09f BB |
5017 | /* |
5018 | * L2ARC fast warmup. | |
5019 | * | |
5020 | * Until the ARC is warm and starts to evict, read from the | |
5021 | * head of the ARC lists rather than the tail. | |
5022 | */ | |
b128c09f BB |
5023 | if (arc_warm == B_FALSE) |
5024 | ab = list_head(list); | |
5025 | else | |
5026 | ab = list_tail(list); | |
5027 | ||
3a17a7a9 SK |
5028 | headroom = target_sz * l2arc_headroom; |
5029 | if (do_headroom_boost) | |
5030 | headroom = (headroom * l2arc_headroom_boost) / 100; | |
5031 | ||
b128c09f | 5032 | for (; ab; ab = ab_prev) { |
3a17a7a9 SK |
5033 | l2arc_buf_hdr_t *l2hdr; |
5034 | kmutex_t *hash_lock; | |
5035 | uint64_t buf_sz; | |
5036 | ||
b128c09f BB |
5037 | if (arc_warm == B_FALSE) |
5038 | ab_prev = list_next(list, ab); | |
5039 | else | |
5040 | ab_prev = list_prev(list, ab); | |
34dc7c2f BB |
5041 | |
5042 | hash_lock = HDR_LOCK(ab); | |
3a17a7a9 | 5043 | if (!mutex_tryenter(hash_lock)) { |
34dc7c2f BB |
5044 | /* |
5045 | * Skip this buffer rather than waiting. | |
5046 | */ | |
5047 | continue; | |
5048 | } | |
5049 | ||
5050 | passed_sz += ab->b_size; | |
5051 | if (passed_sz > headroom) { | |
5052 | /* | |
5053 | * Searched too far. | |
5054 | */ | |
5055 | mutex_exit(hash_lock); | |
5056 | break; | |
5057 | } | |
5058 | ||
d164b209 | 5059 | if (!l2arc_write_eligible(guid, ab)) { |
34dc7c2f BB |
5060 | mutex_exit(hash_lock); |
5061 | continue; | |
5062 | } | |
5063 | ||
5064 | if ((write_sz + ab->b_size) > target_sz) { | |
5065 | full = B_TRUE; | |
5066 | mutex_exit(hash_lock); | |
5067 | break; | |
5068 | } | |
5069 | ||
34dc7c2f BB |
5070 | if (pio == NULL) { |
5071 | /* | |
5072 | * Insert a dummy header on the buflist so | |
5073 | * l2arc_write_done() can find where the | |
5074 | * write buffers begin without searching. | |
5075 | */ | |
5076 | list_insert_head(dev->l2ad_buflist, head); | |
5077 | ||
409dc1a5 | 5078 | cb = kmem_alloc(sizeof (l2arc_write_callback_t), |
d1d7e268 | 5079 | KM_PUSHPAGE); |
34dc7c2f BB |
5080 | cb->l2wcb_dev = dev; |
5081 | cb->l2wcb_head = head; | |
5082 | pio = zio_root(spa, l2arc_write_done, cb, | |
5083 | ZIO_FLAG_CANFAIL); | |
5084 | } | |
5085 | ||
5086 | /* | |
5087 | * Create and add a new L2ARC header. | |
5088 | */ | |
ecf3d9b8 | 5089 | l2hdr = kmem_cache_alloc(l2arc_hdr_cache, KM_PUSHPAGE); |
3a17a7a9 | 5090 | l2hdr->b_dev = dev; |
ecf3d9b8 | 5091 | l2hdr->b_daddr = 0; |
6e1d7276 | 5092 | arc_space_consume(L2HDR_SIZE, ARC_SPACE_L2HDRS); |
34dc7c2f BB |
5093 | |
5094 | ab->b_flags |= ARC_L2_WRITING; | |
3a17a7a9 SK |
5095 | |
5096 | /* | |
5097 | * Temporarily stash the data buffer in b_tmp_cdata. | |
5098 | * The subsequent write step will pick it up from | |
5099 | * there. This is because can't access ab->b_buf | |
5100 | * without holding the hash_lock, which we in turn | |
5101 | * can't access without holding the ARC list locks | |
5102 | * (which we want to avoid during compression/writing) | |
5103 | */ | |
5104 | l2hdr->b_compress = ZIO_COMPRESS_OFF; | |
5105 | l2hdr->b_asize = ab->b_size; | |
5106 | l2hdr->b_tmp_cdata = ab->b_buf->b_data; | |
e0b0ca98 | 5107 | l2hdr->b_hits = 0; |
3a17a7a9 | 5108 | |
34dc7c2f | 5109 | buf_sz = ab->b_size; |
3a17a7a9 SK |
5110 | ab->b_l2hdr = l2hdr; |
5111 | ||
5112 | list_insert_head(dev->l2ad_buflist, ab); | |
34dc7c2f BB |
5113 | |
5114 | /* | |
5115 | * Compute and store the buffer cksum before | |
5116 | * writing. On debug the cksum is verified first. | |
5117 | */ | |
5118 | arc_cksum_verify(ab->b_buf); | |
5119 | arc_cksum_compute(ab->b_buf, B_TRUE); | |
5120 | ||
5121 | mutex_exit(hash_lock); | |
5122 | ||
3a17a7a9 SK |
5123 | write_sz += buf_sz; |
5124 | } | |
5125 | ||
5126 | mutex_exit(list_lock); | |
5127 | ||
5128 | if (full == B_TRUE) | |
5129 | break; | |
5130 | } | |
5131 | ||
5132 | /* No buffers selected for writing? */ | |
5133 | if (pio == NULL) { | |
5134 | ASSERT0(write_sz); | |
5135 | mutex_exit(&l2arc_buflist_mtx); | |
5136 | kmem_cache_free(hdr_cache, head); | |
5137 | return (0); | |
5138 | } | |
5139 | ||
5140 | /* | |
5141 | * Now start writing the buffers. We're starting at the write head | |
5142 | * and work backwards, retracing the course of the buffer selector | |
5143 | * loop above. | |
5144 | */ | |
5145 | for (ab = list_prev(dev->l2ad_buflist, head); ab; | |
5146 | ab = list_prev(dev->l2ad_buflist, ab)) { | |
5147 | l2arc_buf_hdr_t *l2hdr; | |
5148 | uint64_t buf_sz; | |
5149 | ||
5150 | /* | |
5151 | * We shouldn't need to lock the buffer here, since we flagged | |
5152 | * it as ARC_L2_WRITING in the previous step, but we must take | |
5153 | * care to only access its L2 cache parameters. In particular, | |
5154 | * ab->b_buf may be invalid by now due to ARC eviction. | |
5155 | */ | |
5156 | l2hdr = ab->b_l2hdr; | |
5157 | l2hdr->b_daddr = dev->l2ad_hand; | |
5158 | ||
5159 | if (!l2arc_nocompress && (ab->b_flags & ARC_L2COMPRESS) && | |
5160 | l2hdr->b_asize >= buf_compress_minsz) { | |
5161 | if (l2arc_compress_buf(l2hdr)) { | |
5162 | /* | |
5163 | * If compression succeeded, enable headroom | |
5164 | * boost on the next scan cycle. | |
5165 | */ | |
5166 | *headroom_boost = B_TRUE; | |
5167 | } | |
5168 | } | |
5169 | ||
5170 | /* | |
5171 | * Pick up the buffer data we had previously stashed away | |
5172 | * (and now potentially also compressed). | |
5173 | */ | |
5174 | buf_data = l2hdr->b_tmp_cdata; | |
5175 | buf_sz = l2hdr->b_asize; | |
5176 | ||
5177 | /* Compression may have squashed the buffer to zero length. */ | |
5178 | if (buf_sz != 0) { | |
5179 | uint64_t buf_p_sz; | |
5180 | ||
34dc7c2f BB |
5181 | wzio = zio_write_phys(pio, dev->l2ad_vdev, |
5182 | dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, | |
5183 | NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, | |
5184 | ZIO_FLAG_CANFAIL, B_FALSE); | |
5185 | ||
5186 | DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, | |
5187 | zio_t *, wzio); | |
5188 | (void) zio_nowait(wzio); | |
5189 | ||
3a17a7a9 | 5190 | write_asize += buf_sz; |
b128c09f BB |
5191 | /* |
5192 | * Keep the clock hand suitably device-aligned. | |
5193 | */ | |
3a17a7a9 SK |
5194 | buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); |
5195 | write_psize += buf_p_sz; | |
5196 | dev->l2ad_hand += buf_p_sz; | |
34dc7c2f | 5197 | } |
34dc7c2f | 5198 | } |
34dc7c2f | 5199 | |
3a17a7a9 | 5200 | mutex_exit(&l2arc_buflist_mtx); |
34dc7c2f | 5201 | |
3a17a7a9 | 5202 | ASSERT3U(write_asize, <=, target_sz); |
34dc7c2f | 5203 | ARCSTAT_BUMP(arcstat_l2_writes_sent); |
3a17a7a9 | 5204 | ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); |
34dc7c2f | 5205 | ARCSTAT_INCR(arcstat_l2_size, write_sz); |
3a17a7a9 SK |
5206 | ARCSTAT_INCR(arcstat_l2_asize, write_asize); |
5207 | vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); | |
34dc7c2f BB |
5208 | |
5209 | /* | |
5210 | * Bump device hand to the device start if it is approaching the end. | |
5211 | * l2arc_evict() will already have evicted ahead for this case. | |
5212 | */ | |
b128c09f | 5213 | if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { |
428870ff BB |
5214 | vdev_space_update(dev->l2ad_vdev, |
5215 | dev->l2ad_end - dev->l2ad_hand, 0, 0); | |
34dc7c2f BB |
5216 | dev->l2ad_hand = dev->l2ad_start; |
5217 | dev->l2ad_evict = dev->l2ad_start; | |
5218 | dev->l2ad_first = B_FALSE; | |
5219 | } | |
5220 | ||
d164b209 | 5221 | dev->l2ad_writing = B_TRUE; |
34dc7c2f | 5222 | (void) zio_wait(pio); |
d164b209 BB |
5223 | dev->l2ad_writing = B_FALSE; |
5224 | ||
3a17a7a9 SK |
5225 | return (write_asize); |
5226 | } | |
5227 | ||
5228 | /* | |
5229 | * Compresses an L2ARC buffer. | |
5230 | * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its | |
5231 | * size in l2hdr->b_asize. This routine tries to compress the data and | |
5232 | * depending on the compression result there are three possible outcomes: | |
5233 | * *) The buffer was incompressible. The original l2hdr contents were left | |
5234 | * untouched and are ready for writing to an L2 device. | |
5235 | * *) The buffer was all-zeros, so there is no need to write it to an L2 | |
5236 | * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is | |
5237 | * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. | |
5238 | * *) Compression succeeded and b_tmp_cdata was replaced with a temporary | |
5239 | * data buffer which holds the compressed data to be written, and b_asize | |
5240 | * tells us how much data there is. b_compress is set to the appropriate | |
5241 | * compression algorithm. Once writing is done, invoke | |
5242 | * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. | |
5243 | * | |
5244 | * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the | |
5245 | * buffer was incompressible). | |
5246 | */ | |
5247 | static boolean_t | |
5248 | l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) | |
5249 | { | |
5250 | void *cdata; | |
9b67f605 | 5251 | size_t csize, len, rounded; |
3a17a7a9 SK |
5252 | |
5253 | ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); | |
5254 | ASSERT(l2hdr->b_tmp_cdata != NULL); | |
5255 | ||
5256 | len = l2hdr->b_asize; | |
5257 | cdata = zio_data_buf_alloc(len); | |
5258 | csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, | |
5259 | cdata, l2hdr->b_asize); | |
5260 | ||
9b67f605 MA |
5261 | rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); |
5262 | if (rounded > csize) { | |
5263 | bzero((char *)cdata + csize, rounded - csize); | |
5264 | csize = rounded; | |
5265 | } | |
5266 | ||
3a17a7a9 SK |
5267 | if (csize == 0) { |
5268 | /* zero block, indicate that there's nothing to write */ | |
5269 | zio_data_buf_free(cdata, len); | |
5270 | l2hdr->b_compress = ZIO_COMPRESS_EMPTY; | |
5271 | l2hdr->b_asize = 0; | |
5272 | l2hdr->b_tmp_cdata = NULL; | |
5273 | ARCSTAT_BUMP(arcstat_l2_compress_zeros); | |
5274 | return (B_TRUE); | |
5275 | } else if (csize > 0 && csize < len) { | |
5276 | /* | |
5277 | * Compression succeeded, we'll keep the cdata around for | |
5278 | * writing and release it afterwards. | |
5279 | */ | |
5280 | l2hdr->b_compress = ZIO_COMPRESS_LZ4; | |
5281 | l2hdr->b_asize = csize; | |
5282 | l2hdr->b_tmp_cdata = cdata; | |
5283 | ARCSTAT_BUMP(arcstat_l2_compress_successes); | |
5284 | return (B_TRUE); | |
5285 | } else { | |
5286 | /* | |
5287 | * Compression failed, release the compressed buffer. | |
5288 | * l2hdr will be left unmodified. | |
5289 | */ | |
5290 | zio_data_buf_free(cdata, len); | |
5291 | ARCSTAT_BUMP(arcstat_l2_compress_failures); | |
5292 | return (B_FALSE); | |
5293 | } | |
5294 | } | |
5295 | ||
5296 | /* | |
5297 | * Decompresses a zio read back from an l2arc device. On success, the | |
5298 | * underlying zio's io_data buffer is overwritten by the uncompressed | |
5299 | * version. On decompression error (corrupt compressed stream), the | |
5300 | * zio->io_error value is set to signal an I/O error. | |
5301 | * | |
5302 | * Please note that the compressed data stream is not checksummed, so | |
5303 | * if the underlying device is experiencing data corruption, we may feed | |
5304 | * corrupt data to the decompressor, so the decompressor needs to be | |
5305 | * able to handle this situation (LZ4 does). | |
5306 | */ | |
5307 | static void | |
5308 | l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) | |
5309 | { | |
5310 | uint64_t csize; | |
5311 | void *cdata; | |
5312 | ||
5313 | ASSERT(L2ARC_IS_VALID_COMPRESS(c)); | |
5314 | ||
5315 | if (zio->io_error != 0) { | |
5316 | /* | |
5317 | * An io error has occured, just restore the original io | |
5318 | * size in preparation for a main pool read. | |
5319 | */ | |
5320 | zio->io_orig_size = zio->io_size = hdr->b_size; | |
5321 | return; | |
5322 | } | |
5323 | ||
5324 | if (c == ZIO_COMPRESS_EMPTY) { | |
5325 | /* | |
5326 | * An empty buffer results in a null zio, which means we | |
5327 | * need to fill its io_data after we're done restoring the | |
5328 | * buffer's contents. | |
5329 | */ | |
5330 | ASSERT(hdr->b_buf != NULL); | |
5331 | bzero(hdr->b_buf->b_data, hdr->b_size); | |
5332 | zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; | |
5333 | } else { | |
5334 | ASSERT(zio->io_data != NULL); | |
5335 | /* | |
5336 | * We copy the compressed data from the start of the arc buffer | |
5337 | * (the zio_read will have pulled in only what we need, the | |
5338 | * rest is garbage which we will overwrite at decompression) | |
5339 | * and then decompress back to the ARC data buffer. This way we | |
5340 | * can minimize copying by simply decompressing back over the | |
5341 | * original compressed data (rather than decompressing to an | |
5342 | * aux buffer and then copying back the uncompressed buffer, | |
5343 | * which is likely to be much larger). | |
5344 | */ | |
5345 | csize = zio->io_size; | |
5346 | cdata = zio_data_buf_alloc(csize); | |
5347 | bcopy(zio->io_data, cdata, csize); | |
5348 | if (zio_decompress_data(c, cdata, zio->io_data, csize, | |
5349 | hdr->b_size) != 0) | |
2e528b49 | 5350 | zio->io_error = SET_ERROR(EIO); |
3a17a7a9 SK |
5351 | zio_data_buf_free(cdata, csize); |
5352 | } | |
5353 | ||
5354 | /* Restore the expected uncompressed IO size. */ | |
5355 | zio->io_orig_size = zio->io_size = hdr->b_size; | |
5356 | } | |
5357 | ||
5358 | /* | |
5359 | * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. | |
5360 | * This buffer serves as a temporary holder of compressed data while | |
5361 | * the buffer entry is being written to an l2arc device. Once that is | |
5362 | * done, we can dispose of it. | |
5363 | */ | |
5364 | static void | |
5365 | l2arc_release_cdata_buf(arc_buf_hdr_t *ab) | |
5366 | { | |
5367 | l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; | |
5368 | ||
5369 | if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { | |
5370 | /* | |
5371 | * If the data was compressed, then we've allocated a | |
5372 | * temporary buffer for it, so now we need to release it. | |
5373 | */ | |
5374 | ASSERT(l2hdr->b_tmp_cdata != NULL); | |
5375 | zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); | |
5376 | } | |
5377 | l2hdr->b_tmp_cdata = NULL; | |
34dc7c2f BB |
5378 | } |
5379 | ||
5380 | /* | |
5381 | * This thread feeds the L2ARC at regular intervals. This is the beating | |
5382 | * heart of the L2ARC. | |
5383 | */ | |
5384 | static void | |
5385 | l2arc_feed_thread(void) | |
5386 | { | |
5387 | callb_cpr_t cpr; | |
5388 | l2arc_dev_t *dev; | |
5389 | spa_t *spa; | |
d164b209 | 5390 | uint64_t size, wrote; |
428870ff | 5391 | clock_t begin, next = ddi_get_lbolt(); |
3a17a7a9 | 5392 | boolean_t headroom_boost = B_FALSE; |
34dc7c2f BB |
5393 | |
5394 | CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); | |
5395 | ||
5396 | mutex_enter(&l2arc_feed_thr_lock); | |
5397 | ||
5398 | while (l2arc_thread_exit == 0) { | |
34dc7c2f | 5399 | CALLB_CPR_SAFE_BEGIN(&cpr); |
5b63b3eb BB |
5400 | (void) cv_timedwait_interruptible(&l2arc_feed_thr_cv, |
5401 | &l2arc_feed_thr_lock, next); | |
34dc7c2f | 5402 | CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); |
428870ff | 5403 | next = ddi_get_lbolt() + hz; |
34dc7c2f BB |
5404 | |
5405 | /* | |
b128c09f | 5406 | * Quick check for L2ARC devices. |
34dc7c2f BB |
5407 | */ |
5408 | mutex_enter(&l2arc_dev_mtx); | |
5409 | if (l2arc_ndev == 0) { | |
5410 | mutex_exit(&l2arc_dev_mtx); | |
5411 | continue; | |
5412 | } | |
b128c09f | 5413 | mutex_exit(&l2arc_dev_mtx); |
428870ff | 5414 | begin = ddi_get_lbolt(); |
34dc7c2f BB |
5415 | |
5416 | /* | |
b128c09f BB |
5417 | * This selects the next l2arc device to write to, and in |
5418 | * doing so the next spa to feed from: dev->l2ad_spa. This | |
5419 | * will return NULL if there are now no l2arc devices or if | |
5420 | * they are all faulted. | |
5421 | * | |
5422 | * If a device is returned, its spa's config lock is also | |
5423 | * held to prevent device removal. l2arc_dev_get_next() | |
5424 | * will grab and release l2arc_dev_mtx. | |
34dc7c2f | 5425 | */ |
b128c09f | 5426 | if ((dev = l2arc_dev_get_next()) == NULL) |
34dc7c2f | 5427 | continue; |
b128c09f BB |
5428 | |
5429 | spa = dev->l2ad_spa; | |
5430 | ASSERT(spa != NULL); | |
34dc7c2f | 5431 | |
572e2857 BB |
5432 | /* |
5433 | * If the pool is read-only then force the feed thread to | |
5434 | * sleep a little longer. | |
5435 | */ | |
5436 | if (!spa_writeable(spa)) { | |
5437 | next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; | |
5438 | spa_config_exit(spa, SCL_L2ARC, dev); | |
5439 | continue; | |
5440 | } | |
5441 | ||
34dc7c2f | 5442 | /* |
b128c09f | 5443 | * Avoid contributing to memory pressure. |
34dc7c2f | 5444 | */ |
302f753f | 5445 | if (arc_no_grow) { |
b128c09f BB |
5446 | ARCSTAT_BUMP(arcstat_l2_abort_lowmem); |
5447 | spa_config_exit(spa, SCL_L2ARC, dev); | |
34dc7c2f BB |
5448 | continue; |
5449 | } | |
b128c09f | 5450 | |
34dc7c2f BB |
5451 | ARCSTAT_BUMP(arcstat_l2_feeds); |
5452 | ||
3a17a7a9 | 5453 | size = l2arc_write_size(); |
b128c09f | 5454 | |
34dc7c2f BB |
5455 | /* |
5456 | * Evict L2ARC buffers that will be overwritten. | |
5457 | */ | |
b128c09f | 5458 | l2arc_evict(dev, size, B_FALSE); |
34dc7c2f BB |
5459 | |
5460 | /* | |
5461 | * Write ARC buffers. | |
5462 | */ | |
3a17a7a9 | 5463 | wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); |
d164b209 BB |
5464 | |
5465 | /* | |
5466 | * Calculate interval between writes. | |
5467 | */ | |
5468 | next = l2arc_write_interval(begin, size, wrote); | |
b128c09f | 5469 | spa_config_exit(spa, SCL_L2ARC, dev); |
34dc7c2f BB |
5470 | } |
5471 | ||
5472 | l2arc_thread_exit = 0; | |
5473 | cv_broadcast(&l2arc_feed_thr_cv); | |
5474 | CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ | |
5475 | thread_exit(); | |
5476 | } | |
5477 | ||
b128c09f BB |
5478 | boolean_t |
5479 | l2arc_vdev_present(vdev_t *vd) | |
5480 | { | |
5481 | l2arc_dev_t *dev; | |
5482 | ||
5483 | mutex_enter(&l2arc_dev_mtx); | |
5484 | for (dev = list_head(l2arc_dev_list); dev != NULL; | |
5485 | dev = list_next(l2arc_dev_list, dev)) { | |
5486 | if (dev->l2ad_vdev == vd) | |
5487 | break; | |
5488 | } | |
5489 | mutex_exit(&l2arc_dev_mtx); | |
5490 | ||
5491 | return (dev != NULL); | |
5492 | } | |
5493 | ||
34dc7c2f BB |
5494 | /* |
5495 | * Add a vdev for use by the L2ARC. By this point the spa has already | |
5496 | * validated the vdev and opened it. | |
5497 | */ | |
5498 | void | |
9babb374 | 5499 | l2arc_add_vdev(spa_t *spa, vdev_t *vd) |
34dc7c2f BB |
5500 | { |
5501 | l2arc_dev_t *adddev; | |
5502 | ||
b128c09f BB |
5503 | ASSERT(!l2arc_vdev_present(vd)); |
5504 | ||
34dc7c2f BB |
5505 | /* |
5506 | * Create a new l2arc device entry. | |
5507 | */ | |
5508 | adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); | |
5509 | adddev->l2ad_spa = spa; | |
5510 | adddev->l2ad_vdev = vd; | |
9babb374 BB |
5511 | adddev->l2ad_start = VDEV_LABEL_START_SIZE; |
5512 | adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); | |
34dc7c2f BB |
5513 | adddev->l2ad_hand = adddev->l2ad_start; |
5514 | adddev->l2ad_evict = adddev->l2ad_start; | |
5515 | adddev->l2ad_first = B_TRUE; | |
d164b209 | 5516 | adddev->l2ad_writing = B_FALSE; |
98f72a53 | 5517 | list_link_init(&adddev->l2ad_node); |
34dc7c2f BB |
5518 | |
5519 | /* | |
5520 | * This is a list of all ARC buffers that are still valid on the | |
5521 | * device. | |
5522 | */ | |
5523 | adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); | |
5524 | list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), | |
5525 | offsetof(arc_buf_hdr_t, b_l2node)); | |
5526 | ||
428870ff | 5527 | vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); |
34dc7c2f BB |
5528 | |
5529 | /* | |
5530 | * Add device to global list | |
5531 | */ | |
5532 | mutex_enter(&l2arc_dev_mtx); | |
5533 | list_insert_head(l2arc_dev_list, adddev); | |
5534 | atomic_inc_64(&l2arc_ndev); | |
5535 | mutex_exit(&l2arc_dev_mtx); | |
5536 | } | |
5537 | ||
5538 | /* | |
5539 | * Remove a vdev from the L2ARC. | |
5540 | */ | |
5541 | void | |
5542 | l2arc_remove_vdev(vdev_t *vd) | |
5543 | { | |
5544 | l2arc_dev_t *dev, *nextdev, *remdev = NULL; | |
5545 | ||
34dc7c2f BB |
5546 | /* |
5547 | * Find the device by vdev | |
5548 | */ | |
5549 | mutex_enter(&l2arc_dev_mtx); | |
5550 | for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { | |
5551 | nextdev = list_next(l2arc_dev_list, dev); | |
5552 | if (vd == dev->l2ad_vdev) { | |
5553 | remdev = dev; | |
5554 | break; | |
5555 | } | |
5556 | } | |
5557 | ASSERT(remdev != NULL); | |
5558 | ||
5559 | /* | |
5560 | * Remove device from global list | |
5561 | */ | |
5562 | list_remove(l2arc_dev_list, remdev); | |
5563 | l2arc_dev_last = NULL; /* may have been invalidated */ | |
b128c09f BB |
5564 | atomic_dec_64(&l2arc_ndev); |
5565 | mutex_exit(&l2arc_dev_mtx); | |
34dc7c2f BB |
5566 | |
5567 | /* | |
5568 | * Clear all buflists and ARC references. L2ARC device flush. | |
5569 | */ | |
5570 | l2arc_evict(remdev, 0, B_TRUE); | |
5571 | list_destroy(remdev->l2ad_buflist); | |
5572 | kmem_free(remdev->l2ad_buflist, sizeof (list_t)); | |
5573 | kmem_free(remdev, sizeof (l2arc_dev_t)); | |
34dc7c2f BB |
5574 | } |
5575 | ||
5576 | void | |
b128c09f | 5577 | l2arc_init(void) |
34dc7c2f BB |
5578 | { |
5579 | l2arc_thread_exit = 0; | |
5580 | l2arc_ndev = 0; | |
5581 | l2arc_writes_sent = 0; | |
5582 | l2arc_writes_done = 0; | |
5583 | ||
5584 | mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); | |
5585 | cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); | |
5586 | mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); | |
5587 | mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); | |
5588 | mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); | |
5589 | ||
5590 | l2arc_dev_list = &L2ARC_dev_list; | |
5591 | l2arc_free_on_write = &L2ARC_free_on_write; | |
5592 | list_create(l2arc_dev_list, sizeof (l2arc_dev_t), | |
5593 | offsetof(l2arc_dev_t, l2ad_node)); | |
5594 | list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), | |
5595 | offsetof(l2arc_data_free_t, l2df_list_node)); | |
34dc7c2f BB |
5596 | } |
5597 | ||
5598 | void | |
b128c09f | 5599 | l2arc_fini(void) |
34dc7c2f | 5600 | { |
b128c09f BB |
5601 | /* |
5602 | * This is called from dmu_fini(), which is called from spa_fini(); | |
5603 | * Because of this, we can assume that all l2arc devices have | |
5604 | * already been removed when the pools themselves were removed. | |
5605 | */ | |
5606 | ||
5607 | l2arc_do_free_on_write(); | |
34dc7c2f BB |
5608 | |
5609 | mutex_destroy(&l2arc_feed_thr_lock); | |
5610 | cv_destroy(&l2arc_feed_thr_cv); | |
5611 | mutex_destroy(&l2arc_dev_mtx); | |
5612 | mutex_destroy(&l2arc_buflist_mtx); | |
5613 | mutex_destroy(&l2arc_free_on_write_mtx); | |
5614 | ||
5615 | list_destroy(l2arc_dev_list); | |
5616 | list_destroy(l2arc_free_on_write); | |
5617 | } | |
b128c09f BB |
5618 | |
5619 | void | |
5620 | l2arc_start(void) | |
5621 | { | |
fb5f0bc8 | 5622 | if (!(spa_mode_global & FWRITE)) |
b128c09f BB |
5623 | return; |
5624 | ||
5625 | (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, | |
5626 | TS_RUN, minclsyspri); | |
5627 | } | |
5628 | ||
5629 | void | |
5630 | l2arc_stop(void) | |
5631 | { | |
fb5f0bc8 | 5632 | if (!(spa_mode_global & FWRITE)) |
b128c09f BB |
5633 | return; |
5634 | ||
5635 | mutex_enter(&l2arc_feed_thr_lock); | |
5636 | cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ | |
5637 | l2arc_thread_exit = 1; | |
5638 | while (l2arc_thread_exit != 0) | |
5639 | cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); | |
5640 | mutex_exit(&l2arc_feed_thr_lock); | |
5641 | } | |
c28b2279 BB |
5642 | |
5643 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
5644 | EXPORT_SYMBOL(arc_read); | |
5645 | EXPORT_SYMBOL(arc_buf_remove_ref); | |
e0b0ca98 | 5646 | EXPORT_SYMBOL(arc_buf_info); |
c28b2279 | 5647 | EXPORT_SYMBOL(arc_getbuf_func); |
ab26409d BB |
5648 | EXPORT_SYMBOL(arc_add_prune_callback); |
5649 | EXPORT_SYMBOL(arc_remove_prune_callback); | |
c28b2279 | 5650 | |
bce45ec9 | 5651 | module_param(zfs_arc_min, ulong, 0644); |
c409e464 | 5652 | MODULE_PARM_DESC(zfs_arc_min, "Min arc size"); |
c28b2279 | 5653 | |
bce45ec9 | 5654 | module_param(zfs_arc_max, ulong, 0644); |
c409e464 | 5655 | MODULE_PARM_DESC(zfs_arc_max, "Max arc size"); |
c28b2279 | 5656 | |
bce45ec9 | 5657 | module_param(zfs_arc_meta_limit, ulong, 0644); |
c28b2279 | 5658 | MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size"); |
6a8f9b6b | 5659 | |
bce45ec9 | 5660 | module_param(zfs_arc_meta_prune, int, 0644); |
ab26409d | 5661 | MODULE_PARM_DESC(zfs_arc_meta_prune, "Bytes of meta data to prune"); |
c409e464 | 5662 | |
bce45ec9 | 5663 | module_param(zfs_arc_grow_retry, int, 0644); |
c409e464 BB |
5664 | MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size"); |
5665 | ||
89c8cac4 PS |
5666 | module_param(zfs_arc_p_aggressive_disable, int, 0644); |
5667 | MODULE_PARM_DESC(zfs_arc_p_aggressive_disable, "disable aggressive arc_p grow"); | |
5668 | ||
62422785 PS |
5669 | module_param(zfs_arc_p_dampener_disable, int, 0644); |
5670 | MODULE_PARM_DESC(zfs_arc_p_dampener_disable, "disable arc_p adapt dampener"); | |
5671 | ||
bce45ec9 | 5672 | module_param(zfs_arc_shrink_shift, int, 0644); |
c409e464 BB |
5673 | MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)"); |
5674 | ||
1f7c30df BB |
5675 | module_param(zfs_disable_dup_eviction, int, 0644); |
5676 | MODULE_PARM_DESC(zfs_disable_dup_eviction, "disable duplicate buffer eviction"); | |
5677 | ||
0c5493d4 BB |
5678 | module_param(zfs_arc_memory_throttle_disable, int, 0644); |
5679 | MODULE_PARM_DESC(zfs_arc_memory_throttle_disable, "disable memory throttle"); | |
5680 | ||
bce45ec9 BB |
5681 | module_param(zfs_arc_min_prefetch_lifespan, int, 0644); |
5682 | MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan, "Min life of prefetch block"); | |
5683 | ||
5684 | module_param(l2arc_write_max, ulong, 0644); | |
abd8610c BB |
5685 | MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval"); |
5686 | ||
bce45ec9 | 5687 | module_param(l2arc_write_boost, ulong, 0644); |
abd8610c BB |
5688 | MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup"); |
5689 | ||
bce45ec9 | 5690 | module_param(l2arc_headroom, ulong, 0644); |
abd8610c BB |
5691 | MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache"); |
5692 | ||
3a17a7a9 SK |
5693 | module_param(l2arc_headroom_boost, ulong, 0644); |
5694 | MODULE_PARM_DESC(l2arc_headroom_boost, "Compressed l2arc_headroom multiplier"); | |
5695 | ||
bce45ec9 | 5696 | module_param(l2arc_feed_secs, ulong, 0644); |
abd8610c BB |
5697 | MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing"); |
5698 | ||
bce45ec9 | 5699 | module_param(l2arc_feed_min_ms, ulong, 0644); |
abd8610c BB |
5700 | MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds"); |
5701 | ||
bce45ec9 | 5702 | module_param(l2arc_noprefetch, int, 0644); |
abd8610c BB |
5703 | MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers"); |
5704 | ||
3a17a7a9 SK |
5705 | module_param(l2arc_nocompress, int, 0644); |
5706 | MODULE_PARM_DESC(l2arc_nocompress, "Skip compressing L2ARC buffers"); | |
5707 | ||
bce45ec9 | 5708 | module_param(l2arc_feed_again, int, 0644); |
abd8610c BB |
5709 | MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup"); |
5710 | ||
bce45ec9 | 5711 | module_param(l2arc_norw, int, 0644); |
abd8610c BB |
5712 | MODULE_PARM_DESC(l2arc_norw, "No reads during writes"); |
5713 | ||
c28b2279 | 5714 | #endif |