]>
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. |
3bec585e | 25 | * Copyright 2014 Nexenta Systems, Inc. All rights reserved. |
34dc7c2f BB |
26 | */ |
27 | ||
34dc7c2f BB |
28 | /* |
29 | * DVA-based Adjustable Replacement Cache | |
30 | * | |
31 | * While much of the theory of operation used here is | |
32 | * based on the self-tuning, low overhead replacement cache | |
33 | * presented by Megiddo and Modha at FAST 2003, there are some | |
34 | * significant differences: | |
35 | * | |
36 | * 1. The Megiddo and Modha model assumes any page is evictable. | |
37 | * Pages in its cache cannot be "locked" into memory. This makes | |
38 | * the eviction algorithm simple: evict the last page in the list. | |
39 | * This also make the performance characteristics easy to reason | |
40 | * about. Our cache is not so simple. At any given moment, some | |
41 | * subset of the blocks in the cache are un-evictable because we | |
42 | * have handed out a reference to them. Blocks are only evictable | |
43 | * when there are no external references active. This makes | |
44 | * eviction far more problematic: we choose to evict the evictable | |
45 | * blocks that are the "lowest" in the list. | |
46 | * | |
47 | * There are times when it is not possible to evict the requested | |
48 | * space. In these circumstances we are unable to adjust the cache | |
49 | * size. To prevent the cache growing unbounded at these times we | |
50 | * implement a "cache throttle" that slows the flow of new data | |
51 | * into the cache until we can make space available. | |
52 | * | |
53 | * 2. The Megiddo and Modha model assumes a fixed cache size. | |
54 | * Pages are evicted when the cache is full and there is a cache | |
55 | * miss. Our model has a variable sized cache. It grows with | |
56 | * high use, but also tries to react to memory pressure from the | |
57 | * operating system: decreasing its size when system memory is | |
58 | * tight. | |
59 | * | |
60 | * 3. The Megiddo and Modha model assumes a fixed page size. All | |
d3cc8b15 | 61 | * elements of the cache are therefore exactly the same size. So |
34dc7c2f BB |
62 | * when adjusting the cache size following a cache miss, its simply |
63 | * a matter of choosing a single page to evict. In our model, we | |
64 | * have variable sized cache blocks (rangeing from 512 bytes to | |
d3cc8b15 | 65 | * 128K bytes). We therefore choose a set of blocks to evict to make |
34dc7c2f BB |
66 | * space for a cache miss that approximates as closely as possible |
67 | * the space used by the new block. | |
68 | * | |
69 | * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" | |
70 | * by N. Megiddo & D. Modha, FAST 2003 | |
71 | */ | |
72 | ||
73 | /* | |
74 | * The locking model: | |
75 | * | |
76 | * A new reference to a cache buffer can be obtained in two | |
77 | * ways: 1) via a hash table lookup using the DVA as a key, | |
78 | * or 2) via one of the ARC lists. The arc_read() interface | |
79 | * uses method 1, while the internal arc algorithms for | |
d3cc8b15 | 80 | * adjusting the cache use method 2. We therefore provide two |
34dc7c2f BB |
81 | * types of locks: 1) the hash table lock array, and 2) the |
82 | * arc list locks. | |
83 | * | |
5c839890 BC |
84 | * Buffers do not have their own mutexes, rather they rely on the |
85 | * hash table mutexes for the bulk of their protection (i.e. most | |
86 | * fields in the arc_buf_hdr_t are protected by these mutexes). | |
34dc7c2f BB |
87 | * |
88 | * buf_hash_find() returns the appropriate mutex (held) when it | |
89 | * locates the requested buffer in the hash table. It returns | |
90 | * NULL for the mutex if the buffer was not in the table. | |
91 | * | |
92 | * buf_hash_remove() expects the appropriate hash mutex to be | |
93 | * already held before it is invoked. | |
94 | * | |
95 | * Each arc state also has a mutex which is used to protect the | |
96 | * buffer list associated with the state. When attempting to | |
97 | * obtain a hash table lock while holding an arc list lock you | |
98 | * must use: mutex_tryenter() to avoid deadlock. Also note that | |
99 | * the active state mutex must be held before the ghost state mutex. | |
100 | * | |
101 | * Arc buffers may have an associated eviction callback function. | |
102 | * This function will be invoked prior to removing the buffer (e.g. | |
103 | * in arc_do_user_evicts()). Note however that the data associated | |
104 | * with the buffer may be evicted prior to the callback. The callback | |
105 | * must be made with *no locks held* (to prevent deadlock). Additionally, | |
106 | * the users of callbacks must ensure that their private data is | |
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); |
3bec585e SK |
1704 | vdev_space_update(l2hdr->b_dev->l2ad_vdev, |
1705 | -l2hdr->b_asize, 0, 0); | |
ecf3d9b8 | 1706 | kmem_cache_free(l2arc_hdr_cache, l2hdr); |
6e1d7276 | 1707 | arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS); |
428870ff BB |
1708 | if (hdr->b_state == arc_l2c_only) |
1709 | l2arc_hdr_stat_remove(); | |
1710 | hdr->b_l2hdr = NULL; | |
1711 | } | |
1712 | ||
1713 | if (!buflist_held) | |
1714 | mutex_exit(&l2arc_buflist_mtx); | |
34dc7c2f BB |
1715 | } |
1716 | ||
1717 | if (!BUF_EMPTY(hdr)) { | |
1718 | ASSERT(!HDR_IN_HASH_TABLE(hdr)); | |
428870ff | 1719 | buf_discard_identity(hdr); |
34dc7c2f BB |
1720 | } |
1721 | while (hdr->b_buf) { | |
1722 | arc_buf_t *buf = hdr->b_buf; | |
1723 | ||
1724 | if (buf->b_efunc) { | |
1725 | mutex_enter(&arc_eviction_mtx); | |
428870ff | 1726 | mutex_enter(&buf->b_evict_lock); |
34dc7c2f BB |
1727 | ASSERT(buf->b_hdr != NULL); |
1728 | arc_buf_destroy(hdr->b_buf, FALSE, FALSE); | |
1729 | hdr->b_buf = buf->b_next; | |
1730 | buf->b_hdr = &arc_eviction_hdr; | |
1731 | buf->b_next = arc_eviction_list; | |
1732 | arc_eviction_list = buf; | |
428870ff | 1733 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
1734 | mutex_exit(&arc_eviction_mtx); |
1735 | } else { | |
1736 | arc_buf_destroy(hdr->b_buf, FALSE, TRUE); | |
1737 | } | |
1738 | } | |
1739 | if (hdr->b_freeze_cksum != NULL) { | |
1740 | kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); | |
1741 | hdr->b_freeze_cksum = NULL; | |
1742 | } | |
1743 | ||
1744 | ASSERT(!list_link_active(&hdr->b_arc_node)); | |
1745 | ASSERT3P(hdr->b_hash_next, ==, NULL); | |
1746 | ASSERT3P(hdr->b_acb, ==, NULL); | |
1747 | kmem_cache_free(hdr_cache, hdr); | |
1748 | } | |
1749 | ||
1750 | void | |
1751 | arc_buf_free(arc_buf_t *buf, void *tag) | |
1752 | { | |
1753 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
1754 | int hashed = hdr->b_state != arc_anon; | |
1755 | ||
1756 | ASSERT(buf->b_efunc == NULL); | |
1757 | ASSERT(buf->b_data != NULL); | |
1758 | ||
1759 | if (hashed) { | |
1760 | kmutex_t *hash_lock = HDR_LOCK(hdr); | |
1761 | ||
1762 | mutex_enter(hash_lock); | |
428870ff BB |
1763 | hdr = buf->b_hdr; |
1764 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
1765 | ||
34dc7c2f | 1766 | (void) remove_reference(hdr, hash_lock, tag); |
428870ff | 1767 | if (hdr->b_datacnt > 1) { |
34dc7c2f | 1768 | arc_buf_destroy(buf, FALSE, TRUE); |
428870ff BB |
1769 | } else { |
1770 | ASSERT(buf == hdr->b_buf); | |
1771 | ASSERT(buf->b_efunc == NULL); | |
34dc7c2f | 1772 | hdr->b_flags |= ARC_BUF_AVAILABLE; |
428870ff | 1773 | } |
34dc7c2f BB |
1774 | mutex_exit(hash_lock); |
1775 | } else if (HDR_IO_IN_PROGRESS(hdr)) { | |
1776 | int destroy_hdr; | |
1777 | /* | |
1778 | * We are in the middle of an async write. Don't destroy | |
1779 | * this buffer unless the write completes before we finish | |
1780 | * decrementing the reference count. | |
1781 | */ | |
1782 | mutex_enter(&arc_eviction_mtx); | |
1783 | (void) remove_reference(hdr, NULL, tag); | |
1784 | ASSERT(refcount_is_zero(&hdr->b_refcnt)); | |
1785 | destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); | |
1786 | mutex_exit(&arc_eviction_mtx); | |
1787 | if (destroy_hdr) | |
1788 | arc_hdr_destroy(hdr); | |
1789 | } else { | |
428870ff | 1790 | if (remove_reference(hdr, NULL, tag) > 0) |
34dc7c2f | 1791 | arc_buf_destroy(buf, FALSE, TRUE); |
428870ff | 1792 | else |
34dc7c2f | 1793 | arc_hdr_destroy(hdr); |
34dc7c2f BB |
1794 | } |
1795 | } | |
1796 | ||
13fe0198 | 1797 | boolean_t |
34dc7c2f BB |
1798 | arc_buf_remove_ref(arc_buf_t *buf, void* tag) |
1799 | { | |
1800 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
b4f7f105 | 1801 | kmutex_t *hash_lock = NULL; |
13fe0198 | 1802 | boolean_t no_callback = (buf->b_efunc == NULL); |
34dc7c2f BB |
1803 | |
1804 | if (hdr->b_state == arc_anon) { | |
428870ff | 1805 | ASSERT(hdr->b_datacnt == 1); |
34dc7c2f BB |
1806 | arc_buf_free(buf, tag); |
1807 | return (no_callback); | |
1808 | } | |
1809 | ||
b4f7f105 | 1810 | hash_lock = HDR_LOCK(hdr); |
34dc7c2f | 1811 | mutex_enter(hash_lock); |
428870ff BB |
1812 | hdr = buf->b_hdr; |
1813 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
34dc7c2f BB |
1814 | ASSERT(hdr->b_state != arc_anon); |
1815 | ASSERT(buf->b_data != NULL); | |
1816 | ||
1817 | (void) remove_reference(hdr, hash_lock, tag); | |
1818 | if (hdr->b_datacnt > 1) { | |
1819 | if (no_callback) | |
1820 | arc_buf_destroy(buf, FALSE, TRUE); | |
1821 | } else if (no_callback) { | |
1822 | ASSERT(hdr->b_buf == buf && buf->b_next == NULL); | |
428870ff | 1823 | ASSERT(buf->b_efunc == NULL); |
34dc7c2f BB |
1824 | hdr->b_flags |= ARC_BUF_AVAILABLE; |
1825 | } | |
1826 | ASSERT(no_callback || hdr->b_datacnt > 1 || | |
1827 | refcount_is_zero(&hdr->b_refcnt)); | |
1828 | mutex_exit(hash_lock); | |
1829 | return (no_callback); | |
1830 | } | |
1831 | ||
1832 | int | |
1833 | arc_buf_size(arc_buf_t *buf) | |
1834 | { | |
1835 | return (buf->b_hdr->b_size); | |
1836 | } | |
1837 | ||
1eb5bfa3 GW |
1838 | /* |
1839 | * Called from the DMU to determine if the current buffer should be | |
1840 | * evicted. In order to ensure proper locking, the eviction must be initiated | |
1841 | * from the DMU. Return true if the buffer is associated with user data and | |
1842 | * duplicate buffers still exist. | |
1843 | */ | |
1844 | boolean_t | |
1845 | arc_buf_eviction_needed(arc_buf_t *buf) | |
1846 | { | |
1847 | arc_buf_hdr_t *hdr; | |
1848 | boolean_t evict_needed = B_FALSE; | |
1849 | ||
1850 | if (zfs_disable_dup_eviction) | |
1851 | return (B_FALSE); | |
1852 | ||
1853 | mutex_enter(&buf->b_evict_lock); | |
1854 | hdr = buf->b_hdr; | |
1855 | if (hdr == NULL) { | |
1856 | /* | |
1857 | * We are in arc_do_user_evicts(); let that function | |
1858 | * perform the eviction. | |
1859 | */ | |
1860 | ASSERT(buf->b_data == NULL); | |
1861 | mutex_exit(&buf->b_evict_lock); | |
1862 | return (B_FALSE); | |
1863 | } else if (buf->b_data == NULL) { | |
1864 | /* | |
1865 | * We have already been added to the arc eviction list; | |
1866 | * recommend eviction. | |
1867 | */ | |
1868 | ASSERT3P(hdr, ==, &arc_eviction_hdr); | |
1869 | mutex_exit(&buf->b_evict_lock); | |
1870 | return (B_TRUE); | |
1871 | } | |
1872 | ||
1873 | if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) | |
1874 | evict_needed = B_TRUE; | |
1875 | ||
1876 | mutex_exit(&buf->b_evict_lock); | |
1877 | return (evict_needed); | |
1878 | } | |
1879 | ||
34dc7c2f BB |
1880 | /* |
1881 | * Evict buffers from list until we've removed the specified number of | |
1882 | * bytes. Move the removed buffers to the appropriate evict state. | |
1883 | * If the recycle flag is set, then attempt to "recycle" a buffer: | |
1884 | * - look for a buffer to evict that is `bytes' long. | |
1885 | * - return the data block from this buffer rather than freeing it. | |
1886 | * This flag is used by callers that are trying to make space for a | |
1887 | * new buffer in a full arc cache. | |
1888 | * | |
1889 | * This function makes a "best effort". It skips over any buffers | |
1890 | * it can't get a hash_lock on, and so may not catch all candidates. | |
1891 | * It may also return without evicting as much space as requested. | |
1892 | */ | |
1893 | static void * | |
d164b209 | 1894 | arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, |
34dc7c2f BB |
1895 | arc_buf_contents_t type) |
1896 | { | |
1897 | arc_state_t *evicted_state; | |
1898 | uint64_t bytes_evicted = 0, skipped = 0, missed = 0; | |
1899 | arc_buf_hdr_t *ab, *ab_prev = NULL; | |
1900 | list_t *list = &state->arcs_list[type]; | |
1901 | kmutex_t *hash_lock; | |
1902 | boolean_t have_lock; | |
1903 | void *stolen = NULL; | |
e8b96c60 MA |
1904 | arc_buf_hdr_t marker = {{{ 0 }}}; |
1905 | int count = 0; | |
34dc7c2f BB |
1906 | |
1907 | ASSERT(state == arc_mru || state == arc_mfu); | |
1908 | ||
1909 | evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; | |
1910 | ||
da8ccd0e | 1911 | top: |
34dc7c2f BB |
1912 | mutex_enter(&state->arcs_mtx); |
1913 | mutex_enter(&evicted_state->arcs_mtx); | |
1914 | ||
1915 | for (ab = list_tail(list); ab; ab = ab_prev) { | |
1916 | ab_prev = list_prev(list, ab); | |
1917 | /* prefetch buffers have a minimum lifespan */ | |
1918 | if (HDR_IO_IN_PROGRESS(ab) || | |
1919 | (spa && ab->b_spa != spa) || | |
1920 | (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && | |
428870ff | 1921 | ddi_get_lbolt() - ab->b_arc_access < |
bce45ec9 | 1922 | zfs_arc_min_prefetch_lifespan)) { |
34dc7c2f BB |
1923 | skipped++; |
1924 | continue; | |
1925 | } | |
1926 | /* "lookahead" for better eviction candidate */ | |
1927 | if (recycle && ab->b_size != bytes && | |
1928 | ab_prev && ab_prev->b_size == bytes) | |
1929 | continue; | |
e8b96c60 MA |
1930 | |
1931 | /* ignore markers */ | |
1932 | if (ab->b_spa == 0) | |
1933 | continue; | |
1934 | ||
1935 | /* | |
1936 | * It may take a long time to evict all the bufs requested. | |
1937 | * To avoid blocking all arc activity, periodically drop | |
1938 | * the arcs_mtx and give other threads a chance to run | |
1939 | * before reacquiring the lock. | |
1940 | * | |
1941 | * If we are looking for a buffer to recycle, we are in | |
1942 | * the hot code path, so don't sleep. | |
1943 | */ | |
1944 | if (!recycle && count++ > arc_evict_iterations) { | |
1945 | list_insert_after(list, ab, &marker); | |
1946 | mutex_exit(&evicted_state->arcs_mtx); | |
1947 | mutex_exit(&state->arcs_mtx); | |
1948 | kpreempt(KPREEMPT_SYNC); | |
1949 | mutex_enter(&state->arcs_mtx); | |
1950 | mutex_enter(&evicted_state->arcs_mtx); | |
1951 | ab_prev = list_prev(list, &marker); | |
1952 | list_remove(list, &marker); | |
1953 | count = 0; | |
1954 | continue; | |
1955 | } | |
1956 | ||
34dc7c2f BB |
1957 | hash_lock = HDR_LOCK(ab); |
1958 | have_lock = MUTEX_HELD(hash_lock); | |
1959 | if (have_lock || mutex_tryenter(hash_lock)) { | |
c99c9001 | 1960 | ASSERT0(refcount_count(&ab->b_refcnt)); |
34dc7c2f BB |
1961 | ASSERT(ab->b_datacnt > 0); |
1962 | while (ab->b_buf) { | |
1963 | arc_buf_t *buf = ab->b_buf; | |
428870ff | 1964 | if (!mutex_tryenter(&buf->b_evict_lock)) { |
b128c09f BB |
1965 | missed += 1; |
1966 | break; | |
1967 | } | |
34dc7c2f BB |
1968 | if (buf->b_data) { |
1969 | bytes_evicted += ab->b_size; | |
1970 | if (recycle && ab->b_type == type && | |
1971 | ab->b_size == bytes && | |
1972 | !HDR_L2_WRITING(ab)) { | |
1973 | stolen = buf->b_data; | |
1974 | recycle = FALSE; | |
1975 | } | |
1976 | } | |
1977 | if (buf->b_efunc) { | |
1978 | mutex_enter(&arc_eviction_mtx); | |
1979 | arc_buf_destroy(buf, | |
1980 | buf->b_data == stolen, FALSE); | |
1981 | ab->b_buf = buf->b_next; | |
1982 | buf->b_hdr = &arc_eviction_hdr; | |
1983 | buf->b_next = arc_eviction_list; | |
1984 | arc_eviction_list = buf; | |
1985 | mutex_exit(&arc_eviction_mtx); | |
428870ff | 1986 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f | 1987 | } else { |
428870ff | 1988 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
1989 | arc_buf_destroy(buf, |
1990 | buf->b_data == stolen, TRUE); | |
1991 | } | |
1992 | } | |
428870ff BB |
1993 | |
1994 | if (ab->b_l2hdr) { | |
1995 | ARCSTAT_INCR(arcstat_evict_l2_cached, | |
1996 | ab->b_size); | |
1997 | } else { | |
1998 | if (l2arc_write_eligible(ab->b_spa, ab)) { | |
1999 | ARCSTAT_INCR(arcstat_evict_l2_eligible, | |
2000 | ab->b_size); | |
2001 | } else { | |
2002 | ARCSTAT_INCR( | |
2003 | arcstat_evict_l2_ineligible, | |
2004 | ab->b_size); | |
2005 | } | |
2006 | } | |
2007 | ||
b128c09f BB |
2008 | if (ab->b_datacnt == 0) { |
2009 | arc_change_state(evicted_state, ab, hash_lock); | |
2010 | ASSERT(HDR_IN_HASH_TABLE(ab)); | |
2011 | ab->b_flags |= ARC_IN_HASH_TABLE; | |
2012 | ab->b_flags &= ~ARC_BUF_AVAILABLE; | |
2013 | DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); | |
2014 | } | |
34dc7c2f BB |
2015 | if (!have_lock) |
2016 | mutex_exit(hash_lock); | |
2017 | if (bytes >= 0 && bytes_evicted >= bytes) | |
2018 | break; | |
2019 | } else { | |
2020 | missed += 1; | |
2021 | } | |
2022 | } | |
2023 | ||
2024 | mutex_exit(&evicted_state->arcs_mtx); | |
2025 | mutex_exit(&state->arcs_mtx); | |
2026 | ||
da8ccd0e PS |
2027 | if (list == &state->arcs_list[ARC_BUFC_DATA] && |
2028 | (bytes < 0 || bytes_evicted < bytes)) { | |
2029 | /* Prevent second pass from recycling metadata into data */ | |
2030 | recycle = FALSE; | |
2031 | type = ARC_BUFC_METADATA; | |
2032 | list = &state->arcs_list[type]; | |
2033 | goto top; | |
2034 | } | |
2035 | ||
34dc7c2f | 2036 | if (bytes_evicted < bytes) |
3f504482 | 2037 | dprintf("only evicted %lld bytes from %x\n", |
34dc7c2f BB |
2038 | (longlong_t)bytes_evicted, state); |
2039 | ||
2040 | if (skipped) | |
2041 | ARCSTAT_INCR(arcstat_evict_skip, skipped); | |
2042 | ||
2043 | if (missed) | |
2044 | ARCSTAT_INCR(arcstat_mutex_miss, missed); | |
2045 | ||
2046 | /* | |
e8b96c60 MA |
2047 | * Note: we have just evicted some data into the ghost state, |
2048 | * potentially putting the ghost size over the desired size. Rather | |
2049 | * that evicting from the ghost list in this hot code path, leave | |
2050 | * this chore to the arc_reclaim_thread(). | |
34dc7c2f | 2051 | */ |
34dc7c2f BB |
2052 | |
2053 | return (stolen); | |
2054 | } | |
2055 | ||
2056 | /* | |
2057 | * Remove buffers from list until we've removed the specified number of | |
2058 | * bytes. Destroy the buffers that are removed. | |
2059 | */ | |
2060 | static void | |
68121a03 BB |
2061 | arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes, |
2062 | arc_buf_contents_t type) | |
34dc7c2f BB |
2063 | { |
2064 | arc_buf_hdr_t *ab, *ab_prev; | |
2598c001 | 2065 | arc_buf_hdr_t marker; |
68121a03 | 2066 | list_t *list = &state->arcs_list[type]; |
34dc7c2f BB |
2067 | kmutex_t *hash_lock; |
2068 | uint64_t bytes_deleted = 0; | |
2069 | uint64_t bufs_skipped = 0; | |
e8b96c60 | 2070 | int count = 0; |
34dc7c2f BB |
2071 | |
2072 | ASSERT(GHOST_STATE(state)); | |
d1d7e268 | 2073 | bzero(&marker, sizeof (marker)); |
34dc7c2f BB |
2074 | top: |
2075 | mutex_enter(&state->arcs_mtx); | |
2076 | for (ab = list_tail(list); ab; ab = ab_prev) { | |
2077 | ab_prev = list_prev(list, ab); | |
e8b96c60 MA |
2078 | if (ab->b_type > ARC_BUFC_NUMTYPES) |
2079 | panic("invalid ab=%p", (void *)ab); | |
34dc7c2f BB |
2080 | if (spa && ab->b_spa != spa) |
2081 | continue; | |
572e2857 BB |
2082 | |
2083 | /* ignore markers */ | |
2084 | if (ab->b_spa == 0) | |
2085 | continue; | |
2086 | ||
34dc7c2f | 2087 | hash_lock = HDR_LOCK(ab); |
428870ff BB |
2088 | /* caller may be trying to modify this buffer, skip it */ |
2089 | if (MUTEX_HELD(hash_lock)) | |
2090 | continue; | |
e8b96c60 MA |
2091 | |
2092 | /* | |
2093 | * It may take a long time to evict all the bufs requested. | |
2094 | * To avoid blocking all arc activity, periodically drop | |
2095 | * the arcs_mtx and give other threads a chance to run | |
2096 | * before reacquiring the lock. | |
2097 | */ | |
2098 | if (count++ > arc_evict_iterations) { | |
2099 | list_insert_after(list, ab, &marker); | |
2100 | mutex_exit(&state->arcs_mtx); | |
2101 | kpreempt(KPREEMPT_SYNC); | |
2102 | mutex_enter(&state->arcs_mtx); | |
2103 | ab_prev = list_prev(list, &marker); | |
2104 | list_remove(list, &marker); | |
2105 | count = 0; | |
2106 | continue; | |
2107 | } | |
34dc7c2f BB |
2108 | if (mutex_tryenter(hash_lock)) { |
2109 | ASSERT(!HDR_IO_IN_PROGRESS(ab)); | |
2110 | ASSERT(ab->b_buf == NULL); | |
2111 | ARCSTAT_BUMP(arcstat_deleted); | |
2112 | bytes_deleted += ab->b_size; | |
2113 | ||
2114 | if (ab->b_l2hdr != NULL) { | |
2115 | /* | |
2116 | * This buffer is cached on the 2nd Level ARC; | |
2117 | * don't destroy the header. | |
2118 | */ | |
2119 | arc_change_state(arc_l2c_only, ab, hash_lock); | |
2120 | mutex_exit(hash_lock); | |
2121 | } else { | |
2122 | arc_change_state(arc_anon, ab, hash_lock); | |
2123 | mutex_exit(hash_lock); | |
2124 | arc_hdr_destroy(ab); | |
2125 | } | |
2126 | ||
2127 | DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); | |
2128 | if (bytes >= 0 && bytes_deleted >= bytes) | |
2129 | break; | |
572e2857 BB |
2130 | } else if (bytes < 0) { |
2131 | /* | |
2132 | * Insert a list marker and then wait for the | |
2133 | * hash lock to become available. Once its | |
2134 | * available, restart from where we left off. | |
2135 | */ | |
2136 | list_insert_after(list, ab, &marker); | |
2137 | mutex_exit(&state->arcs_mtx); | |
2138 | mutex_enter(hash_lock); | |
2139 | mutex_exit(hash_lock); | |
2140 | mutex_enter(&state->arcs_mtx); | |
2141 | ab_prev = list_prev(list, &marker); | |
2142 | list_remove(list, &marker); | |
e8b96c60 | 2143 | } else { |
34dc7c2f | 2144 | bufs_skipped += 1; |
e8b96c60 | 2145 | } |
34dc7c2f BB |
2146 | } |
2147 | mutex_exit(&state->arcs_mtx); | |
2148 | ||
2149 | if (list == &state->arcs_list[ARC_BUFC_DATA] && | |
2150 | (bytes < 0 || bytes_deleted < bytes)) { | |
2151 | list = &state->arcs_list[ARC_BUFC_METADATA]; | |
2152 | goto top; | |
2153 | } | |
2154 | ||
2155 | if (bufs_skipped) { | |
2156 | ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); | |
2157 | ASSERT(bytes >= 0); | |
2158 | } | |
2159 | ||
2160 | if (bytes_deleted < bytes) | |
3f504482 | 2161 | dprintf("only deleted %lld bytes from %p\n", |
34dc7c2f BB |
2162 | (longlong_t)bytes_deleted, state); |
2163 | } | |
2164 | ||
2165 | static void | |
2166 | arc_adjust(void) | |
2167 | { | |
d164b209 BB |
2168 | int64_t adjustment, delta; |
2169 | ||
2170 | /* | |
2171 | * Adjust MRU size | |
2172 | */ | |
34dc7c2f | 2173 | |
572e2857 | 2174 | adjustment = MIN((int64_t)(arc_size - arc_c), |
77765b54 | 2175 | (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size - arc_p)); |
34dc7c2f | 2176 | |
da8ccd0e PS |
2177 | if (adjustment > 0 && arc_mru->arcs_size > 0) { |
2178 | delta = MIN(arc_mru->arcs_size, adjustment); | |
b8864a23 | 2179 | (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); |
34dc7c2f BB |
2180 | } |
2181 | ||
d164b209 BB |
2182 | /* |
2183 | * Adjust MFU size | |
2184 | */ | |
34dc7c2f | 2185 | |
d164b209 BB |
2186 | adjustment = arc_size - arc_c; |
2187 | ||
da8ccd0e PS |
2188 | if (adjustment > 0 && arc_mfu->arcs_size > 0) { |
2189 | delta = MIN(arc_mfu->arcs_size, adjustment); | |
b8864a23 | 2190 | (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); |
d164b209 | 2191 | } |
34dc7c2f | 2192 | |
d164b209 BB |
2193 | /* |
2194 | * Adjust ghost lists | |
2195 | */ | |
34dc7c2f | 2196 | |
d164b209 BB |
2197 | adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; |
2198 | ||
2199 | if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { | |
2200 | delta = MIN(arc_mru_ghost->arcs_size, adjustment); | |
68121a03 | 2201 | arc_evict_ghost(arc_mru_ghost, 0, delta, ARC_BUFC_DATA); |
d164b209 | 2202 | } |
34dc7c2f | 2203 | |
d164b209 BB |
2204 | adjustment = |
2205 | arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; | |
34dc7c2f | 2206 | |
d164b209 BB |
2207 | if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { |
2208 | delta = MIN(arc_mfu_ghost->arcs_size, adjustment); | |
68121a03 | 2209 | arc_evict_ghost(arc_mfu_ghost, 0, delta, ARC_BUFC_DATA); |
34dc7c2f BB |
2210 | } |
2211 | } | |
2212 | ||
ab26409d BB |
2213 | /* |
2214 | * Request that arc user drop references so that N bytes can be released | |
2215 | * from the cache. This provides a mechanism to ensure the arc can honor | |
2216 | * the arc_meta_limit and reclaim buffers which are pinned in the cache | |
2217 | * by higher layers. (i.e. the zpl) | |
2218 | */ | |
2219 | static void | |
2220 | arc_do_user_prune(int64_t adjustment) | |
2221 | { | |
2222 | arc_prune_func_t *func; | |
2223 | void *private; | |
2224 | arc_prune_t *cp, *np; | |
2225 | ||
2226 | mutex_enter(&arc_prune_mtx); | |
2227 | ||
2228 | cp = list_head(&arc_prune_list); | |
2229 | while (cp != NULL) { | |
2230 | func = cp->p_pfunc; | |
2231 | private = cp->p_private; | |
2232 | np = list_next(&arc_prune_list, cp); | |
2233 | refcount_add(&cp->p_refcnt, func); | |
2234 | mutex_exit(&arc_prune_mtx); | |
2235 | ||
2236 | if (func != NULL) | |
2237 | func(adjustment, private); | |
2238 | ||
2239 | mutex_enter(&arc_prune_mtx); | |
2240 | ||
2241 | /* User removed prune callback concurrently with execution */ | |
2242 | if (refcount_remove(&cp->p_refcnt, func) == 0) { | |
2243 | ASSERT(!list_link_active(&cp->p_node)); | |
2244 | refcount_destroy(&cp->p_refcnt); | |
2245 | kmem_free(cp, sizeof (*cp)); | |
2246 | } | |
2247 | ||
2248 | cp = np; | |
2249 | } | |
2250 | ||
2251 | ARCSTAT_BUMP(arcstat_prune); | |
2252 | mutex_exit(&arc_prune_mtx); | |
2253 | } | |
2254 | ||
34dc7c2f BB |
2255 | static void |
2256 | arc_do_user_evicts(void) | |
2257 | { | |
2258 | mutex_enter(&arc_eviction_mtx); | |
2259 | while (arc_eviction_list != NULL) { | |
2260 | arc_buf_t *buf = arc_eviction_list; | |
2261 | arc_eviction_list = buf->b_next; | |
428870ff | 2262 | mutex_enter(&buf->b_evict_lock); |
34dc7c2f | 2263 | buf->b_hdr = NULL; |
428870ff | 2264 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
2265 | mutex_exit(&arc_eviction_mtx); |
2266 | ||
2267 | if (buf->b_efunc != NULL) | |
2268 | VERIFY(buf->b_efunc(buf) == 0); | |
2269 | ||
2270 | buf->b_efunc = NULL; | |
2271 | buf->b_private = NULL; | |
2272 | kmem_cache_free(buf_cache, buf); | |
2273 | mutex_enter(&arc_eviction_mtx); | |
2274 | } | |
2275 | mutex_exit(&arc_eviction_mtx); | |
2276 | } | |
2277 | ||
ab26409d BB |
2278 | /* |
2279 | * Evict only meta data objects from the cache leaving the data objects. | |
2280 | * This is only used to enforce the tunable arc_meta_limit, if we are | |
2281 | * unable to evict enough buffers notify the user via the prune callback. | |
2282 | */ | |
94520ca4 PS |
2283 | static void |
2284 | arc_adjust_meta(void) | |
ab26409d | 2285 | { |
94520ca4 | 2286 | int64_t adjustmnt, delta; |
ab26409d | 2287 | |
94520ca4 PS |
2288 | /* |
2289 | * This slightly differs than the way we evict from the mru in | |
2290 | * arc_adjust because we don't have a "target" value (i.e. no | |
2291 | * "meta" arc_p). As a result, I think we can completely | |
2292 | * cannibalize the metadata in the MRU before we evict the | |
2293 | * metadata from the MFU. I think we probably need to implement a | |
2294 | * "metadata arc_p" value to do this properly. | |
2295 | */ | |
2296 | adjustmnt = arc_meta_used - arc_meta_limit; | |
2297 | ||
2298 | if (adjustmnt > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { | |
2299 | delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustmnt); | |
ab26409d | 2300 | arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_METADATA); |
94520ca4 | 2301 | adjustmnt -= delta; |
ab26409d BB |
2302 | } |
2303 | ||
94520ca4 PS |
2304 | /* |
2305 | * We can't afford to recalculate adjustmnt here. If we do, | |
2306 | * new metadata buffers can sneak into the MRU or ANON lists, | |
2307 | * thus penalize the MFU metadata. Although the fudge factor is | |
2308 | * small, it has been empirically shown to be significant for | |
2309 | * certain workloads (e.g. creating many empty directories). As | |
2310 | * such, we use the original calculation for adjustmnt, and | |
2311 | * simply decrement the amount of data evicted from the MRU. | |
2312 | */ | |
2313 | ||
2314 | if (adjustmnt > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { | |
2315 | delta = MIN(arc_mfu->arcs_lsize[ARC_BUFC_METADATA], adjustmnt); | |
ab26409d | 2316 | arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_METADATA); |
ab26409d BB |
2317 | } |
2318 | ||
94520ca4 PS |
2319 | adjustmnt = arc_mru->arcs_lsize[ARC_BUFC_METADATA] + |
2320 | arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] - arc_meta_limit; | |
2321 | ||
2322 | if (adjustmnt > 0 && arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] > 0) { | |
2323 | delta = MIN(adjustmnt, | |
2324 | arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA]); | |
2325 | arc_evict_ghost(arc_mru_ghost, 0, delta, ARC_BUFC_METADATA); | |
2326 | } | |
2327 | ||
2328 | adjustmnt = arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] + | |
2329 | arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA] - arc_meta_limit; | |
2330 | ||
2331 | if (adjustmnt > 0 && arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA] > 0) { | |
2332 | delta = MIN(adjustmnt, | |
2333 | arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA]); | |
2334 | arc_evict_ghost(arc_mfu_ghost, 0, delta, ARC_BUFC_METADATA); | |
2335 | } | |
2336 | ||
2337 | if (arc_meta_used > arc_meta_limit) | |
bce45ec9 | 2338 | arc_do_user_prune(zfs_arc_meta_prune); |
ab26409d BB |
2339 | } |
2340 | ||
34dc7c2f BB |
2341 | /* |
2342 | * Flush all *evictable* data from the cache for the given spa. | |
2343 | * NOTE: this will not touch "active" (i.e. referenced) data. | |
2344 | */ | |
2345 | void | |
2346 | arc_flush(spa_t *spa) | |
2347 | { | |
d164b209 BB |
2348 | uint64_t guid = 0; |
2349 | ||
2350 | if (spa) | |
3541dc6d | 2351 | guid = spa_load_guid(spa); |
d164b209 | 2352 | |
34dc7c2f | 2353 | while (list_head(&arc_mru->arcs_list[ARC_BUFC_DATA])) { |
d164b209 | 2354 | (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); |
34dc7c2f BB |
2355 | if (spa) |
2356 | break; | |
2357 | } | |
2358 | while (list_head(&arc_mru->arcs_list[ARC_BUFC_METADATA])) { | |
d164b209 | 2359 | (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); |
34dc7c2f BB |
2360 | if (spa) |
2361 | break; | |
2362 | } | |
2363 | while (list_head(&arc_mfu->arcs_list[ARC_BUFC_DATA])) { | |
d164b209 | 2364 | (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); |
34dc7c2f BB |
2365 | if (spa) |
2366 | break; | |
2367 | } | |
2368 | while (list_head(&arc_mfu->arcs_list[ARC_BUFC_METADATA])) { | |
d164b209 | 2369 | (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); |
34dc7c2f BB |
2370 | if (spa) |
2371 | break; | |
2372 | } | |
2373 | ||
68121a03 BB |
2374 | arc_evict_ghost(arc_mru_ghost, guid, -1, ARC_BUFC_DATA); |
2375 | arc_evict_ghost(arc_mfu_ghost, guid, -1, ARC_BUFC_DATA); | |
34dc7c2f BB |
2376 | |
2377 | mutex_enter(&arc_reclaim_thr_lock); | |
2378 | arc_do_user_evicts(); | |
2379 | mutex_exit(&arc_reclaim_thr_lock); | |
2380 | ASSERT(spa || arc_eviction_list == NULL); | |
2381 | } | |
2382 | ||
34dc7c2f | 2383 | void |
302f753f | 2384 | arc_shrink(uint64_t bytes) |
34dc7c2f BB |
2385 | { |
2386 | if (arc_c > arc_c_min) { | |
2387 | uint64_t to_free; | |
2388 | ||
bce45ec9 | 2389 | to_free = bytes ? bytes : arc_c >> zfs_arc_shrink_shift; |
302f753f | 2390 | |
34dc7c2f BB |
2391 | if (arc_c > arc_c_min + to_free) |
2392 | atomic_add_64(&arc_c, -to_free); | |
2393 | else | |
2394 | arc_c = arc_c_min; | |
2395 | ||
39e055c4 PS |
2396 | to_free = bytes ? bytes : arc_p >> zfs_arc_shrink_shift; |
2397 | ||
f521ce1b | 2398 | if (arc_p > to_free) |
39e055c4 PS |
2399 | atomic_add_64(&arc_p, -to_free); |
2400 | else | |
f521ce1b | 2401 | arc_p = 0; |
39e055c4 | 2402 | |
34dc7c2f BB |
2403 | if (arc_c > arc_size) |
2404 | arc_c = MAX(arc_size, arc_c_min); | |
2405 | if (arc_p > arc_c) | |
2406 | arc_p = (arc_c >> 1); | |
2407 | ASSERT(arc_c >= arc_c_min); | |
2408 | ASSERT((int64_t)arc_p >= 0); | |
2409 | } | |
2410 | ||
2411 | if (arc_size > arc_c) | |
2412 | arc_adjust(); | |
2413 | } | |
2414 | ||
34dc7c2f | 2415 | static void |
302f753f | 2416 | arc_kmem_reap_now(arc_reclaim_strategy_t strat, uint64_t bytes) |
34dc7c2f BB |
2417 | { |
2418 | size_t i; | |
2419 | kmem_cache_t *prev_cache = NULL; | |
2420 | kmem_cache_t *prev_data_cache = NULL; | |
2421 | extern kmem_cache_t *zio_buf_cache[]; | |
2422 | extern kmem_cache_t *zio_data_buf_cache[]; | |
34dc7c2f BB |
2423 | |
2424 | /* | |
2425 | * An aggressive reclamation will shrink the cache size as well as | |
2426 | * reap free buffers from the arc kmem caches. | |
2427 | */ | |
2428 | if (strat == ARC_RECLAIM_AGGR) | |
302f753f | 2429 | arc_shrink(bytes); |
34dc7c2f BB |
2430 | |
2431 | for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { | |
2432 | if (zio_buf_cache[i] != prev_cache) { | |
2433 | prev_cache = zio_buf_cache[i]; | |
2434 | kmem_cache_reap_now(zio_buf_cache[i]); | |
2435 | } | |
2436 | if (zio_data_buf_cache[i] != prev_data_cache) { | |
2437 | prev_data_cache = zio_data_buf_cache[i]; | |
2438 | kmem_cache_reap_now(zio_data_buf_cache[i]); | |
2439 | } | |
2440 | } | |
ab26409d | 2441 | |
34dc7c2f BB |
2442 | kmem_cache_reap_now(buf_cache); |
2443 | kmem_cache_reap_now(hdr_cache); | |
2444 | } | |
2445 | ||
302f753f BB |
2446 | /* |
2447 | * Unlike other ZFS implementations this thread is only responsible for | |
2448 | * adapting the target ARC size on Linux. The responsibility for memory | |
2449 | * reclamation has been entirely delegated to the arc_shrinker_func() | |
2450 | * which is registered with the VM. To reflect this change in behavior | |
2451 | * the arc_reclaim thread has been renamed to arc_adapt. | |
2452 | */ | |
34dc7c2f | 2453 | static void |
302f753f | 2454 | arc_adapt_thread(void) |
34dc7c2f | 2455 | { |
34dc7c2f BB |
2456 | callb_cpr_t cpr; |
2457 | ||
2458 | CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); | |
2459 | ||
2460 | mutex_enter(&arc_reclaim_thr_lock); | |
2461 | while (arc_thread_exit == 0) { | |
302f753f BB |
2462 | #ifndef _KERNEL |
2463 | arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; | |
2464 | ||
2465 | if (spa_get_random(100) == 0) { | |
34dc7c2f BB |
2466 | |
2467 | if (arc_no_grow) { | |
2468 | if (last_reclaim == ARC_RECLAIM_CONS) { | |
2469 | last_reclaim = ARC_RECLAIM_AGGR; | |
2470 | } else { | |
2471 | last_reclaim = ARC_RECLAIM_CONS; | |
2472 | } | |
2473 | } else { | |
2474 | arc_no_grow = TRUE; | |
2475 | last_reclaim = ARC_RECLAIM_AGGR; | |
2476 | membar_producer(); | |
2477 | } | |
2478 | ||
2479 | /* reset the growth delay for every reclaim */ | |
d1d7e268 MK |
2480 | arc_grow_time = ddi_get_lbolt() + |
2481 | (zfs_arc_grow_retry * hz); | |
34dc7c2f | 2482 | |
302f753f | 2483 | arc_kmem_reap_now(last_reclaim, 0); |
b128c09f | 2484 | arc_warm = B_TRUE; |
302f753f BB |
2485 | } |
2486 | #endif /* !_KERNEL */ | |
34dc7c2f | 2487 | |
302f753f | 2488 | /* No recent memory pressure allow the ARC to grow. */ |
0b75bdb3 CC |
2489 | if (arc_no_grow && |
2490 | ddi_time_after_eq(ddi_get_lbolt(), arc_grow_time)) | |
34dc7c2f | 2491 | arc_no_grow = FALSE; |
34dc7c2f | 2492 | |
94520ca4 | 2493 | arc_adjust_meta(); |
6a8f9b6b | 2494 | |
572e2857 | 2495 | arc_adjust(); |
34dc7c2f BB |
2496 | |
2497 | if (arc_eviction_list != NULL) | |
2498 | arc_do_user_evicts(); | |
2499 | ||
2500 | /* block until needed, or one second, whichever is shorter */ | |
2501 | CALLB_CPR_SAFE_BEGIN(&cpr); | |
5b63b3eb | 2502 | (void) cv_timedwait_interruptible(&arc_reclaim_thr_cv, |
428870ff | 2503 | &arc_reclaim_thr_lock, (ddi_get_lbolt() + hz)); |
34dc7c2f | 2504 | CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); |
bce45ec9 BB |
2505 | |
2506 | ||
2507 | /* Allow the module options to be changed */ | |
2508 | if (zfs_arc_max > 64 << 20 && | |
2509 | zfs_arc_max < physmem * PAGESIZE && | |
2510 | zfs_arc_max != arc_c_max) | |
2511 | arc_c_max = zfs_arc_max; | |
2512 | ||
2513 | if (zfs_arc_min > 0 && | |
2514 | zfs_arc_min < arc_c_max && | |
2515 | zfs_arc_min != arc_c_min) | |
2516 | arc_c_min = zfs_arc_min; | |
2517 | ||
2518 | if (zfs_arc_meta_limit > 0 && | |
2519 | zfs_arc_meta_limit <= arc_c_max && | |
2520 | zfs_arc_meta_limit != arc_meta_limit) | |
2521 | arc_meta_limit = zfs_arc_meta_limit; | |
2522 | ||
2523 | ||
2524 | ||
34dc7c2f BB |
2525 | } |
2526 | ||
2527 | arc_thread_exit = 0; | |
2528 | cv_broadcast(&arc_reclaim_thr_cv); | |
2529 | CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ | |
2530 | thread_exit(); | |
2531 | } | |
2532 | ||
7cb67b45 BB |
2533 | #ifdef _KERNEL |
2534 | /* | |
302f753f BB |
2535 | * Determine the amount of memory eligible for eviction contained in the |
2536 | * ARC. All clean data reported by the ghost lists can always be safely | |
2537 | * evicted. Due to arc_c_min, the same does not hold for all clean data | |
2538 | * contained by the regular mru and mfu lists. | |
2539 | * | |
2540 | * In the case of the regular mru and mfu lists, we need to report as | |
2541 | * much clean data as possible, such that evicting that same reported | |
2542 | * data will not bring arc_size below arc_c_min. Thus, in certain | |
2543 | * circumstances, the total amount of clean data in the mru and mfu | |
2544 | * lists might not actually be evictable. | |
2545 | * | |
2546 | * The following two distinct cases are accounted for: | |
2547 | * | |
2548 | * 1. The sum of the amount of dirty data contained by both the mru and | |
2549 | * mfu lists, plus the ARC's other accounting (e.g. the anon list), | |
2550 | * is greater than or equal to arc_c_min. | |
2551 | * (i.e. amount of dirty data >= arc_c_min) | |
2552 | * | |
2553 | * This is the easy case; all clean data contained by the mru and mfu | |
2554 | * lists is evictable. Evicting all clean data can only drop arc_size | |
2555 | * to the amount of dirty data, which is greater than arc_c_min. | |
2556 | * | |
2557 | * 2. The sum of the amount of dirty data contained by both the mru and | |
2558 | * mfu lists, plus the ARC's other accounting (e.g. the anon list), | |
2559 | * is less than arc_c_min. | |
2560 | * (i.e. arc_c_min > amount of dirty data) | |
2561 | * | |
2562 | * 2.1. arc_size is greater than or equal arc_c_min. | |
2563 | * (i.e. arc_size >= arc_c_min > amount of dirty data) | |
2564 | * | |
2565 | * In this case, not all clean data from the regular mru and mfu | |
2566 | * lists is actually evictable; we must leave enough clean data | |
2567 | * to keep arc_size above arc_c_min. Thus, the maximum amount of | |
2568 | * evictable data from the two lists combined, is exactly the | |
2569 | * difference between arc_size and arc_c_min. | |
2570 | * | |
2571 | * 2.2. arc_size is less than arc_c_min | |
2572 | * (i.e. arc_c_min > arc_size > amount of dirty data) | |
2573 | * | |
2574 | * In this case, none of the data contained in the mru and mfu | |
2575 | * lists is evictable, even if it's clean. Since arc_size is | |
2576 | * already below arc_c_min, evicting any more would only | |
2577 | * increase this negative difference. | |
7cb67b45 | 2578 | */ |
302f753f BB |
2579 | static uint64_t |
2580 | arc_evictable_memory(void) { | |
2581 | uint64_t arc_clean = | |
2582 | arc_mru->arcs_lsize[ARC_BUFC_DATA] + | |
2583 | arc_mru->arcs_lsize[ARC_BUFC_METADATA] + | |
2584 | arc_mfu->arcs_lsize[ARC_BUFC_DATA] + | |
2585 | arc_mfu->arcs_lsize[ARC_BUFC_METADATA]; | |
2586 | uint64_t ghost_clean = | |
2587 | arc_mru_ghost->arcs_lsize[ARC_BUFC_DATA] + | |
2588 | arc_mru_ghost->arcs_lsize[ARC_BUFC_METADATA] + | |
2589 | arc_mfu_ghost->arcs_lsize[ARC_BUFC_DATA] + | |
2590 | arc_mfu_ghost->arcs_lsize[ARC_BUFC_METADATA]; | |
2591 | uint64_t arc_dirty = MAX((int64_t)arc_size - (int64_t)arc_clean, 0); | |
2592 | ||
2593 | if (arc_dirty >= arc_c_min) | |
2594 | return (ghost_clean + arc_clean); | |
2595 | ||
2596 | return (ghost_clean + MAX((int64_t)arc_size - (int64_t)arc_c_min, 0)); | |
2597 | } | |
2598 | ||
7e7baeca BB |
2599 | static int |
2600 | __arc_shrinker_func(struct shrinker *shrink, struct shrink_control *sc) | |
7cb67b45 | 2601 | { |
302f753f | 2602 | uint64_t pages; |
7cb67b45 | 2603 | |
302f753f BB |
2604 | /* The arc is considered warm once reclaim has occurred */ |
2605 | if (unlikely(arc_warm == B_FALSE)) | |
2606 | arc_warm = B_TRUE; | |
7cb67b45 | 2607 | |
302f753f BB |
2608 | /* Return the potential number of reclaimable pages */ |
2609 | pages = btop(arc_evictable_memory()); | |
2610 | if (sc->nr_to_scan == 0) | |
2611 | return (pages); | |
3fd70ee6 BB |
2612 | |
2613 | /* Not allowed to perform filesystem reclaim */ | |
7e7baeca | 2614 | if (!(sc->gfp_mask & __GFP_FS)) |
3fd70ee6 BB |
2615 | return (-1); |
2616 | ||
7cb67b45 BB |
2617 | /* Reclaim in progress */ |
2618 | if (mutex_tryenter(&arc_reclaim_thr_lock) == 0) | |
2619 | return (-1); | |
2620 | ||
302f753f BB |
2621 | /* |
2622 | * Evict the requested number of pages by shrinking arc_c the | |
2623 | * requested amount. If there is nothing left to evict just | |
2624 | * reap whatever we can from the various arc slabs. | |
2625 | */ | |
2626 | if (pages > 0) { | |
2627 | arc_kmem_reap_now(ARC_RECLAIM_AGGR, ptob(sc->nr_to_scan)); | |
1e3cb67b | 2628 | pages = btop(arc_evictable_memory()); |
302f753f BB |
2629 | } else { |
2630 | arc_kmem_reap_now(ARC_RECLAIM_CONS, ptob(sc->nr_to_scan)); | |
1e3cb67b | 2631 | pages = -1; |
302f753f BB |
2632 | } |
2633 | ||
2634 | /* | |
2635 | * When direct reclaim is observed it usually indicates a rapid | |
2636 | * increase in memory pressure. This occurs because the kswapd | |
2637 | * threads were unable to asynchronously keep enough free memory | |
2638 | * available. In this case set arc_no_grow to briefly pause arc | |
2639 | * growth to avoid compounding the memory pressure. | |
2640 | */ | |
7cb67b45 | 2641 | if (current_is_kswapd()) { |
302f753f | 2642 | ARCSTAT_BUMP(arcstat_memory_indirect_count); |
7cb67b45 | 2643 | } else { |
302f753f | 2644 | arc_no_grow = B_TRUE; |
bce45ec9 | 2645 | arc_grow_time = ddi_get_lbolt() + (zfs_arc_grow_retry * hz); |
302f753f | 2646 | ARCSTAT_BUMP(arcstat_memory_direct_count); |
7cb67b45 BB |
2647 | } |
2648 | ||
7cb67b45 BB |
2649 | mutex_exit(&arc_reclaim_thr_lock); |
2650 | ||
1e3cb67b | 2651 | return (pages); |
7cb67b45 | 2652 | } |
7e7baeca | 2653 | SPL_SHRINKER_CALLBACK_WRAPPER(arc_shrinker_func); |
7cb67b45 BB |
2654 | |
2655 | SPL_SHRINKER_DECLARE(arc_shrinker, arc_shrinker_func, DEFAULT_SEEKS); | |
2656 | #endif /* _KERNEL */ | |
2657 | ||
34dc7c2f BB |
2658 | /* |
2659 | * Adapt arc info given the number of bytes we are trying to add and | |
2660 | * the state that we are comming from. This function is only called | |
2661 | * when we are adding new content to the cache. | |
2662 | */ | |
2663 | static void | |
2664 | arc_adapt(int bytes, arc_state_t *state) | |
2665 | { | |
2666 | int mult; | |
2667 | ||
2668 | if (state == arc_l2c_only) | |
2669 | return; | |
2670 | ||
2671 | ASSERT(bytes > 0); | |
2672 | /* | |
2673 | * Adapt the target size of the MRU list: | |
2674 | * - if we just hit in the MRU ghost list, then increase | |
2675 | * the target size of the MRU list. | |
2676 | * - if we just hit in the MFU ghost list, then increase | |
2677 | * the target size of the MFU list by decreasing the | |
2678 | * target size of the MRU list. | |
2679 | */ | |
2680 | if (state == arc_mru_ghost) { | |
2681 | mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? | |
2682 | 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); | |
62422785 PS |
2683 | |
2684 | if (!zfs_arc_p_dampener_disable) | |
2685 | mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ | |
34dc7c2f | 2686 | |
f521ce1b | 2687 | arc_p = MIN(arc_c, arc_p + bytes * mult); |
34dc7c2f | 2688 | } else if (state == arc_mfu_ghost) { |
d164b209 BB |
2689 | uint64_t delta; |
2690 | ||
34dc7c2f BB |
2691 | mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? |
2692 | 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); | |
62422785 PS |
2693 | |
2694 | if (!zfs_arc_p_dampener_disable) | |
2695 | mult = MIN(mult, 10); | |
34dc7c2f | 2696 | |
d164b209 | 2697 | delta = MIN(bytes * mult, arc_p); |
f521ce1b | 2698 | arc_p = MAX(0, arc_p - delta); |
34dc7c2f BB |
2699 | } |
2700 | ASSERT((int64_t)arc_p >= 0); | |
2701 | ||
34dc7c2f BB |
2702 | if (arc_no_grow) |
2703 | return; | |
2704 | ||
2705 | if (arc_c >= arc_c_max) | |
2706 | return; | |
2707 | ||
2708 | /* | |
2709 | * If we're within (2 * maxblocksize) bytes of the target | |
2710 | * cache size, increment the target cache size | |
2711 | */ | |
2712 | if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { | |
2713 | atomic_add_64(&arc_c, (int64_t)bytes); | |
2714 | if (arc_c > arc_c_max) | |
2715 | arc_c = arc_c_max; | |
2716 | else if (state == arc_anon) | |
2717 | atomic_add_64(&arc_p, (int64_t)bytes); | |
2718 | if (arc_p > arc_c) | |
2719 | arc_p = arc_c; | |
2720 | } | |
2721 | ASSERT((int64_t)arc_p >= 0); | |
2722 | } | |
2723 | ||
2724 | /* | |
2725 | * Check if the cache has reached its limits and eviction is required | |
2726 | * prior to insert. | |
2727 | */ | |
2728 | static int | |
2729 | arc_evict_needed(arc_buf_contents_t type) | |
2730 | { | |
2731 | if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) | |
2732 | return (1); | |
2733 | ||
302f753f | 2734 | if (arc_no_grow) |
34dc7c2f BB |
2735 | return (1); |
2736 | ||
2737 | return (arc_size > arc_c); | |
2738 | } | |
2739 | ||
2740 | /* | |
2741 | * The buffer, supplied as the first argument, needs a data block. | |
2742 | * So, if we are at cache max, determine which cache should be victimized. | |
2743 | * We have the following cases: | |
2744 | * | |
2745 | * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> | |
2746 | * In this situation if we're out of space, but the resident size of the MFU is | |
2747 | * under the limit, victimize the MFU cache to satisfy this insertion request. | |
2748 | * | |
2749 | * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> | |
2750 | * Here, we've used up all of the available space for the MRU, so we need to | |
2751 | * evict from our own cache instead. Evict from the set of resident MRU | |
2752 | * entries. | |
2753 | * | |
2754 | * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> | |
2755 | * c minus p represents the MFU space in the cache, since p is the size of the | |
2756 | * cache that is dedicated to the MRU. In this situation there's still space on | |
2757 | * the MFU side, so the MRU side needs to be victimized. | |
2758 | * | |
2759 | * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> | |
2760 | * MFU's resident set is consuming more space than it has been allotted. In | |
2761 | * this situation, we must victimize our own cache, the MFU, for this insertion. | |
2762 | */ | |
2763 | static void | |
2764 | arc_get_data_buf(arc_buf_t *buf) | |
2765 | { | |
2766 | arc_state_t *state = buf->b_hdr->b_state; | |
2767 | uint64_t size = buf->b_hdr->b_size; | |
2768 | arc_buf_contents_t type = buf->b_hdr->b_type; | |
da8ccd0e PS |
2769 | arc_buf_contents_t evict = ARC_BUFC_DATA; |
2770 | boolean_t recycle = TRUE; | |
34dc7c2f BB |
2771 | |
2772 | arc_adapt(size, state); | |
2773 | ||
2774 | /* | |
2775 | * We have not yet reached cache maximum size, | |
2776 | * just allocate a new buffer. | |
2777 | */ | |
2778 | if (!arc_evict_needed(type)) { | |
2779 | if (type == ARC_BUFC_METADATA) { | |
2780 | buf->b_data = zio_buf_alloc(size); | |
cc7f677c | 2781 | arc_space_consume(size, ARC_SPACE_META); |
34dc7c2f BB |
2782 | } else { |
2783 | ASSERT(type == ARC_BUFC_DATA); | |
2784 | buf->b_data = zio_data_buf_alloc(size); | |
cc7f677c | 2785 | arc_space_consume(size, ARC_SPACE_DATA); |
34dc7c2f BB |
2786 | } |
2787 | goto out; | |
2788 | } | |
2789 | ||
2790 | /* | |
2791 | * If we are prefetching from the mfu ghost list, this buffer | |
2792 | * will end up on the mru list; so steal space from there. | |
2793 | */ | |
2794 | if (state == arc_mfu_ghost) | |
2795 | state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; | |
2796 | else if (state == arc_mru_ghost) | |
2797 | state = arc_mru; | |
2798 | ||
2799 | if (state == arc_mru || state == arc_anon) { | |
2800 | uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; | |
d164b209 | 2801 | state = (arc_mfu->arcs_lsize[type] >= size && |
34dc7c2f BB |
2802 | arc_p > mru_used) ? arc_mfu : arc_mru; |
2803 | } else { | |
2804 | /* MFU cases */ | |
2805 | uint64_t mfu_space = arc_c - arc_p; | |
d164b209 | 2806 | state = (arc_mru->arcs_lsize[type] >= size && |
34dc7c2f BB |
2807 | mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; |
2808 | } | |
ab26409d | 2809 | |
da8ccd0e PS |
2810 | /* |
2811 | * Evict data buffers prior to metadata buffers, unless we're | |
2812 | * over the metadata limit and adding a metadata buffer. | |
2813 | */ | |
2814 | if (type == ARC_BUFC_METADATA) { | |
2815 | if (arc_meta_used >= arc_meta_limit) | |
2816 | evict = ARC_BUFC_METADATA; | |
2817 | else | |
2818 | /* | |
2819 | * In this case, we're evicting data while | |
2820 | * adding metadata. Thus, to prevent recycling a | |
2821 | * data buffer into a metadata buffer, recycling | |
2822 | * is disabled in the following arc_evict call. | |
2823 | */ | |
2824 | recycle = FALSE; | |
2825 | } | |
2826 | ||
2827 | if ((buf->b_data = arc_evict(state, 0, size, recycle, evict)) == NULL) { | |
34dc7c2f BB |
2828 | if (type == ARC_BUFC_METADATA) { |
2829 | buf->b_data = zio_buf_alloc(size); | |
cc7f677c | 2830 | arc_space_consume(size, ARC_SPACE_META); |
ab26409d BB |
2831 | |
2832 | /* | |
2833 | * If we are unable to recycle an existing meta buffer | |
2834 | * signal the reclaim thread. It will notify users | |
2835 | * via the prune callback to drop references. The | |
2836 | * prune callback in run in the context of the reclaim | |
2837 | * thread to avoid deadlocking on the hash_lock. | |
da8ccd0e | 2838 | * Of course, only do this when recycle is true. |
ab26409d | 2839 | */ |
da8ccd0e PS |
2840 | if (recycle) |
2841 | cv_signal(&arc_reclaim_thr_cv); | |
34dc7c2f BB |
2842 | } else { |
2843 | ASSERT(type == ARC_BUFC_DATA); | |
2844 | buf->b_data = zio_data_buf_alloc(size); | |
cc7f677c | 2845 | arc_space_consume(size, ARC_SPACE_DATA); |
34dc7c2f | 2846 | } |
ab26409d | 2847 | |
da8ccd0e PS |
2848 | /* Only bump this if we tried to recycle and failed */ |
2849 | if (recycle) | |
2850 | ARCSTAT_BUMP(arcstat_recycle_miss); | |
34dc7c2f BB |
2851 | } |
2852 | ASSERT(buf->b_data != NULL); | |
2853 | out: | |
2854 | /* | |
2855 | * Update the state size. Note that ghost states have a | |
2856 | * "ghost size" and so don't need to be updated. | |
2857 | */ | |
2858 | if (!GHOST_STATE(buf->b_hdr->b_state)) { | |
2859 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
2860 | ||
2861 | atomic_add_64(&hdr->b_state->arcs_size, size); | |
2862 | if (list_link_active(&hdr->b_arc_node)) { | |
2863 | ASSERT(refcount_is_zero(&hdr->b_refcnt)); | |
2864 | atomic_add_64(&hdr->b_state->arcs_lsize[type], size); | |
2865 | } | |
2866 | /* | |
2867 | * If we are growing the cache, and we are adding anonymous | |
2868 | * data, and we have outgrown arc_p, update arc_p | |
2869 | */ | |
89c8cac4 PS |
2870 | if (!zfs_arc_p_aggressive_disable && |
2871 | arc_size < arc_c && hdr->b_state == arc_anon && | |
34dc7c2f BB |
2872 | arc_anon->arcs_size + arc_mru->arcs_size > arc_p) |
2873 | arc_p = MIN(arc_c, arc_p + size); | |
2874 | } | |
2875 | } | |
2876 | ||
2877 | /* | |
2878 | * This routine is called whenever a buffer is accessed. | |
2879 | * NOTE: the hash lock is dropped in this function. | |
2880 | */ | |
2881 | static void | |
2882 | arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) | |
2883 | { | |
428870ff BB |
2884 | clock_t now; |
2885 | ||
34dc7c2f BB |
2886 | ASSERT(MUTEX_HELD(hash_lock)); |
2887 | ||
2888 | if (buf->b_state == arc_anon) { | |
2889 | /* | |
2890 | * This buffer is not in the cache, and does not | |
2891 | * appear in our "ghost" list. Add the new buffer | |
2892 | * to the MRU state. | |
2893 | */ | |
2894 | ||
2895 | ASSERT(buf->b_arc_access == 0); | |
428870ff | 2896 | buf->b_arc_access = ddi_get_lbolt(); |
34dc7c2f BB |
2897 | DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); |
2898 | arc_change_state(arc_mru, buf, hash_lock); | |
2899 | ||
2900 | } else if (buf->b_state == arc_mru) { | |
428870ff BB |
2901 | now = ddi_get_lbolt(); |
2902 | ||
34dc7c2f BB |
2903 | /* |
2904 | * If this buffer is here because of a prefetch, then either: | |
2905 | * - clear the flag if this is a "referencing" read | |
2906 | * (any subsequent access will bump this into the MFU state). | |
2907 | * or | |
2908 | * - move the buffer to the head of the list if this is | |
2909 | * another prefetch (to make it less likely to be evicted). | |
2910 | */ | |
2911 | if ((buf->b_flags & ARC_PREFETCH) != 0) { | |
2912 | if (refcount_count(&buf->b_refcnt) == 0) { | |
2913 | ASSERT(list_link_active(&buf->b_arc_node)); | |
2914 | } else { | |
2915 | buf->b_flags &= ~ARC_PREFETCH; | |
e0b0ca98 | 2916 | atomic_inc_32(&buf->b_mru_hits); |
34dc7c2f BB |
2917 | ARCSTAT_BUMP(arcstat_mru_hits); |
2918 | } | |
428870ff | 2919 | buf->b_arc_access = now; |
34dc7c2f BB |
2920 | return; |
2921 | } | |
2922 | ||
2923 | /* | |
2924 | * This buffer has been "accessed" only once so far, | |
2925 | * but it is still in the cache. Move it to the MFU | |
2926 | * state. | |
2927 | */ | |
0b75bdb3 | 2928 | if (ddi_time_after(now, buf->b_arc_access + ARC_MINTIME)) { |
34dc7c2f BB |
2929 | /* |
2930 | * More than 125ms have passed since we | |
2931 | * instantiated this buffer. Move it to the | |
2932 | * most frequently used state. | |
2933 | */ | |
428870ff | 2934 | buf->b_arc_access = now; |
34dc7c2f BB |
2935 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); |
2936 | arc_change_state(arc_mfu, buf, hash_lock); | |
2937 | } | |
e0b0ca98 | 2938 | atomic_inc_32(&buf->b_mru_hits); |
34dc7c2f BB |
2939 | ARCSTAT_BUMP(arcstat_mru_hits); |
2940 | } else if (buf->b_state == arc_mru_ghost) { | |
2941 | arc_state_t *new_state; | |
2942 | /* | |
2943 | * This buffer has been "accessed" recently, but | |
2944 | * was evicted from the cache. Move it to the | |
2945 | * MFU state. | |
2946 | */ | |
2947 | ||
2948 | if (buf->b_flags & ARC_PREFETCH) { | |
2949 | new_state = arc_mru; | |
2950 | if (refcount_count(&buf->b_refcnt) > 0) | |
2951 | buf->b_flags &= ~ARC_PREFETCH; | |
2952 | DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); | |
2953 | } else { | |
2954 | new_state = arc_mfu; | |
2955 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); | |
2956 | } | |
2957 | ||
428870ff | 2958 | buf->b_arc_access = ddi_get_lbolt(); |
34dc7c2f BB |
2959 | arc_change_state(new_state, buf, hash_lock); |
2960 | ||
e0b0ca98 | 2961 | atomic_inc_32(&buf->b_mru_ghost_hits); |
34dc7c2f BB |
2962 | ARCSTAT_BUMP(arcstat_mru_ghost_hits); |
2963 | } else if (buf->b_state == arc_mfu) { | |
2964 | /* | |
2965 | * This buffer has been accessed more than once and is | |
2966 | * still in the cache. Keep it in the MFU state. | |
2967 | * | |
2968 | * NOTE: an add_reference() that occurred when we did | |
2969 | * the arc_read() will have kicked this off the list. | |
2970 | * If it was a prefetch, we will explicitly move it to | |
2971 | * the head of the list now. | |
2972 | */ | |
2973 | if ((buf->b_flags & ARC_PREFETCH) != 0) { | |
2974 | ASSERT(refcount_count(&buf->b_refcnt) == 0); | |
2975 | ASSERT(list_link_active(&buf->b_arc_node)); | |
2976 | } | |
e0b0ca98 | 2977 | atomic_inc_32(&buf->b_mfu_hits); |
34dc7c2f | 2978 | ARCSTAT_BUMP(arcstat_mfu_hits); |
428870ff | 2979 | buf->b_arc_access = ddi_get_lbolt(); |
34dc7c2f BB |
2980 | } else if (buf->b_state == arc_mfu_ghost) { |
2981 | arc_state_t *new_state = arc_mfu; | |
2982 | /* | |
2983 | * This buffer has been accessed more than once but has | |
2984 | * been evicted from the cache. Move it back to the | |
2985 | * MFU state. | |
2986 | */ | |
2987 | ||
2988 | if (buf->b_flags & ARC_PREFETCH) { | |
2989 | /* | |
2990 | * This is a prefetch access... | |
2991 | * move this block back to the MRU state. | |
2992 | */ | |
c99c9001 | 2993 | ASSERT0(refcount_count(&buf->b_refcnt)); |
34dc7c2f BB |
2994 | new_state = arc_mru; |
2995 | } | |
2996 | ||
428870ff | 2997 | buf->b_arc_access = ddi_get_lbolt(); |
34dc7c2f BB |
2998 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); |
2999 | arc_change_state(new_state, buf, hash_lock); | |
3000 | ||
e0b0ca98 | 3001 | atomic_inc_32(&buf->b_mfu_ghost_hits); |
34dc7c2f BB |
3002 | ARCSTAT_BUMP(arcstat_mfu_ghost_hits); |
3003 | } else if (buf->b_state == arc_l2c_only) { | |
3004 | /* | |
3005 | * This buffer is on the 2nd Level ARC. | |
3006 | */ | |
3007 | ||
428870ff | 3008 | buf->b_arc_access = ddi_get_lbolt(); |
34dc7c2f BB |
3009 | DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); |
3010 | arc_change_state(arc_mfu, buf, hash_lock); | |
3011 | } else { | |
3012 | ASSERT(!"invalid arc state"); | |
3013 | } | |
3014 | } | |
3015 | ||
3016 | /* a generic arc_done_func_t which you can use */ | |
3017 | /* ARGSUSED */ | |
3018 | void | |
3019 | arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) | |
3020 | { | |
428870ff BB |
3021 | if (zio == NULL || zio->io_error == 0) |
3022 | bcopy(buf->b_data, arg, buf->b_hdr->b_size); | |
13fe0198 | 3023 | VERIFY(arc_buf_remove_ref(buf, arg)); |
34dc7c2f BB |
3024 | } |
3025 | ||
3026 | /* a generic arc_done_func_t */ | |
3027 | void | |
3028 | arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) | |
3029 | { | |
3030 | arc_buf_t **bufp = arg; | |
3031 | if (zio && zio->io_error) { | |
13fe0198 | 3032 | VERIFY(arc_buf_remove_ref(buf, arg)); |
34dc7c2f BB |
3033 | *bufp = NULL; |
3034 | } else { | |
3035 | *bufp = buf; | |
428870ff | 3036 | ASSERT(buf->b_data); |
34dc7c2f BB |
3037 | } |
3038 | } | |
3039 | ||
3040 | static void | |
3041 | arc_read_done(zio_t *zio) | |
3042 | { | |
9b67f605 | 3043 | arc_buf_hdr_t *hdr; |
34dc7c2f BB |
3044 | arc_buf_t *buf; |
3045 | arc_buf_t *abuf; /* buffer we're assigning to callback */ | |
9b67f605 | 3046 | kmutex_t *hash_lock = NULL; |
34dc7c2f BB |
3047 | arc_callback_t *callback_list, *acb; |
3048 | int freeable = FALSE; | |
3049 | ||
3050 | buf = zio->io_private; | |
3051 | hdr = buf->b_hdr; | |
3052 | ||
3053 | /* | |
3054 | * The hdr was inserted into hash-table and removed from lists | |
3055 | * prior to starting I/O. We should find this header, since | |
3056 | * it's in the hash table, and it should be legit since it's | |
3057 | * not possible to evict it during the I/O. The only possible | |
3058 | * reason for it not to be found is if we were freed during the | |
3059 | * read. | |
3060 | */ | |
9b67f605 MA |
3061 | if (HDR_IN_HASH_TABLE(hdr)) { |
3062 | arc_buf_hdr_t *found; | |
3063 | ||
3064 | ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); | |
3065 | ASSERT3U(hdr->b_dva.dva_word[0], ==, | |
3066 | BP_IDENTITY(zio->io_bp)->dva_word[0]); | |
3067 | ASSERT3U(hdr->b_dva.dva_word[1], ==, | |
3068 | BP_IDENTITY(zio->io_bp)->dva_word[1]); | |
3069 | ||
3070 | found = buf_hash_find(hdr->b_spa, zio->io_bp, | |
3071 | &hash_lock); | |
3072 | ||
3073 | ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && | |
3074 | hash_lock == NULL) || | |
3075 | (found == hdr && | |
3076 | DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || | |
3077 | (found == hdr && HDR_L2_READING(hdr))); | |
3078 | } | |
34dc7c2f | 3079 | |
b128c09f | 3080 | hdr->b_flags &= ~ARC_L2_EVICTED; |
34dc7c2f | 3081 | if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) |
b128c09f | 3082 | hdr->b_flags &= ~ARC_L2CACHE; |
34dc7c2f BB |
3083 | |
3084 | /* byteswap if necessary */ | |
3085 | callback_list = hdr->b_acb; | |
3086 | ASSERT(callback_list != NULL); | |
428870ff | 3087 | if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { |
9ae529ec CS |
3088 | dmu_object_byteswap_t bswap = |
3089 | DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); | |
b01615d5 RY |
3090 | if (BP_GET_LEVEL(zio->io_bp) > 0) |
3091 | byteswap_uint64_array(buf->b_data, hdr->b_size); | |
3092 | else | |
3093 | dmu_ot_byteswap[bswap].ob_func(buf->b_data, hdr->b_size); | |
b128c09f | 3094 | } |
34dc7c2f BB |
3095 | |
3096 | arc_cksum_compute(buf, B_FALSE); | |
498877ba | 3097 | arc_buf_watch(buf); |
34dc7c2f | 3098 | |
428870ff BB |
3099 | if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { |
3100 | /* | |
3101 | * Only call arc_access on anonymous buffers. This is because | |
3102 | * if we've issued an I/O for an evicted buffer, we've already | |
3103 | * called arc_access (to prevent any simultaneous readers from | |
3104 | * getting confused). | |
3105 | */ | |
3106 | arc_access(hdr, hash_lock); | |
3107 | } | |
3108 | ||
34dc7c2f BB |
3109 | /* create copies of the data buffer for the callers */ |
3110 | abuf = buf; | |
3111 | for (acb = callback_list; acb; acb = acb->acb_next) { | |
3112 | if (acb->acb_done) { | |
1eb5bfa3 GW |
3113 | if (abuf == NULL) { |
3114 | ARCSTAT_BUMP(arcstat_duplicate_reads); | |
34dc7c2f | 3115 | abuf = arc_buf_clone(buf); |
1eb5bfa3 | 3116 | } |
34dc7c2f BB |
3117 | acb->acb_buf = abuf; |
3118 | abuf = NULL; | |
3119 | } | |
3120 | } | |
3121 | hdr->b_acb = NULL; | |
3122 | hdr->b_flags &= ~ARC_IO_IN_PROGRESS; | |
3123 | ASSERT(!HDR_BUF_AVAILABLE(hdr)); | |
428870ff BB |
3124 | if (abuf == buf) { |
3125 | ASSERT(buf->b_efunc == NULL); | |
3126 | ASSERT(hdr->b_datacnt == 1); | |
34dc7c2f | 3127 | hdr->b_flags |= ARC_BUF_AVAILABLE; |
428870ff | 3128 | } |
34dc7c2f BB |
3129 | |
3130 | ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); | |
3131 | ||
3132 | if (zio->io_error != 0) { | |
3133 | hdr->b_flags |= ARC_IO_ERROR; | |
3134 | if (hdr->b_state != arc_anon) | |
3135 | arc_change_state(arc_anon, hdr, hash_lock); | |
3136 | if (HDR_IN_HASH_TABLE(hdr)) | |
3137 | buf_hash_remove(hdr); | |
3138 | freeable = refcount_is_zero(&hdr->b_refcnt); | |
34dc7c2f BB |
3139 | } |
3140 | ||
3141 | /* | |
3142 | * Broadcast before we drop the hash_lock to avoid the possibility | |
3143 | * that the hdr (and hence the cv) might be freed before we get to | |
3144 | * the cv_broadcast(). | |
3145 | */ | |
3146 | cv_broadcast(&hdr->b_cv); | |
3147 | ||
3148 | if (hash_lock) { | |
34dc7c2f BB |
3149 | mutex_exit(hash_lock); |
3150 | } else { | |
3151 | /* | |
3152 | * This block was freed while we waited for the read to | |
3153 | * complete. It has been removed from the hash table and | |
3154 | * moved to the anonymous state (so that it won't show up | |
3155 | * in the cache). | |
3156 | */ | |
3157 | ASSERT3P(hdr->b_state, ==, arc_anon); | |
3158 | freeable = refcount_is_zero(&hdr->b_refcnt); | |
3159 | } | |
3160 | ||
3161 | /* execute each callback and free its structure */ | |
3162 | while ((acb = callback_list) != NULL) { | |
3163 | if (acb->acb_done) | |
3164 | acb->acb_done(zio, acb->acb_buf, acb->acb_private); | |
3165 | ||
3166 | if (acb->acb_zio_dummy != NULL) { | |
3167 | acb->acb_zio_dummy->io_error = zio->io_error; | |
3168 | zio_nowait(acb->acb_zio_dummy); | |
3169 | } | |
3170 | ||
3171 | callback_list = acb->acb_next; | |
3172 | kmem_free(acb, sizeof (arc_callback_t)); | |
3173 | } | |
3174 | ||
3175 | if (freeable) | |
3176 | arc_hdr_destroy(hdr); | |
3177 | } | |
3178 | ||
3179 | /* | |
5c839890 | 3180 | * "Read" the block at the specified DVA (in bp) via the |
34dc7c2f BB |
3181 | * cache. If the block is found in the cache, invoke the provided |
3182 | * callback immediately and return. Note that the `zio' parameter | |
3183 | * in the callback will be NULL in this case, since no IO was | |
3184 | * required. If the block is not in the cache pass the read request | |
3185 | * on to the spa with a substitute callback function, so that the | |
3186 | * requested block will be added to the cache. | |
3187 | * | |
3188 | * If a read request arrives for a block that has a read in-progress, | |
3189 | * either wait for the in-progress read to complete (and return the | |
3190 | * results); or, if this is a read with a "done" func, add a record | |
3191 | * to the read to invoke the "done" func when the read completes, | |
3192 | * and return; or just return. | |
3193 | * | |
3194 | * arc_read_done() will invoke all the requested "done" functions | |
3195 | * for readers of this block. | |
3196 | */ | |
3197 | int | |
294f6806 | 3198 | arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, |
e8b96c60 | 3199 | void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, |
294f6806 | 3200 | const zbookmark_t *zb) |
34dc7c2f | 3201 | { |
9b67f605 | 3202 | arc_buf_hdr_t *hdr = NULL; |
d4ed6673 | 3203 | arc_buf_t *buf = NULL; |
9b67f605 | 3204 | kmutex_t *hash_lock = NULL; |
34dc7c2f | 3205 | zio_t *rzio; |
3541dc6d | 3206 | uint64_t guid = spa_load_guid(spa); |
1421c891 | 3207 | int rc = 0; |
34dc7c2f | 3208 | |
9b67f605 MA |
3209 | ASSERT(!BP_IS_EMBEDDED(bp) || |
3210 | BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); | |
3211 | ||
34dc7c2f | 3212 | top: |
9b67f605 MA |
3213 | if (!BP_IS_EMBEDDED(bp)) { |
3214 | /* | |
3215 | * Embedded BP's have no DVA and require no I/O to "read". | |
3216 | * Create an anonymous arc buf to back it. | |
3217 | */ | |
3218 | hdr = buf_hash_find(guid, bp, &hash_lock); | |
3219 | } | |
3220 | ||
3221 | if (hdr != NULL && hdr->b_datacnt > 0) { | |
34dc7c2f BB |
3222 | |
3223 | *arc_flags |= ARC_CACHED; | |
3224 | ||
3225 | if (HDR_IO_IN_PROGRESS(hdr)) { | |
3226 | ||
3227 | if (*arc_flags & ARC_WAIT) { | |
3228 | cv_wait(&hdr->b_cv, hash_lock); | |
3229 | mutex_exit(hash_lock); | |
3230 | goto top; | |
3231 | } | |
3232 | ASSERT(*arc_flags & ARC_NOWAIT); | |
3233 | ||
3234 | if (done) { | |
3235 | arc_callback_t *acb = NULL; | |
3236 | ||
3237 | acb = kmem_zalloc(sizeof (arc_callback_t), | |
691f6ac4 | 3238 | KM_PUSHPAGE); |
34dc7c2f BB |
3239 | acb->acb_done = done; |
3240 | acb->acb_private = private; | |
34dc7c2f BB |
3241 | if (pio != NULL) |
3242 | acb->acb_zio_dummy = zio_null(pio, | |
d164b209 | 3243 | spa, NULL, NULL, NULL, zio_flags); |
34dc7c2f BB |
3244 | |
3245 | ASSERT(acb->acb_done != NULL); | |
3246 | acb->acb_next = hdr->b_acb; | |
3247 | hdr->b_acb = acb; | |
3248 | add_reference(hdr, hash_lock, private); | |
3249 | mutex_exit(hash_lock); | |
1421c891 | 3250 | goto out; |
34dc7c2f BB |
3251 | } |
3252 | mutex_exit(hash_lock); | |
1421c891 | 3253 | goto out; |
34dc7c2f BB |
3254 | } |
3255 | ||
3256 | ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); | |
3257 | ||
3258 | if (done) { | |
3259 | add_reference(hdr, hash_lock, private); | |
3260 | /* | |
3261 | * If this block is already in use, create a new | |
3262 | * copy of the data so that we will be guaranteed | |
3263 | * that arc_release() will always succeed. | |
3264 | */ | |
3265 | buf = hdr->b_buf; | |
3266 | ASSERT(buf); | |
3267 | ASSERT(buf->b_data); | |
3268 | if (HDR_BUF_AVAILABLE(hdr)) { | |
3269 | ASSERT(buf->b_efunc == NULL); | |
3270 | hdr->b_flags &= ~ARC_BUF_AVAILABLE; | |
3271 | } else { | |
3272 | buf = arc_buf_clone(buf); | |
3273 | } | |
428870ff | 3274 | |
34dc7c2f BB |
3275 | } else if (*arc_flags & ARC_PREFETCH && |
3276 | refcount_count(&hdr->b_refcnt) == 0) { | |
3277 | hdr->b_flags |= ARC_PREFETCH; | |
3278 | } | |
3279 | DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); | |
3280 | arc_access(hdr, hash_lock); | |
b128c09f BB |
3281 | if (*arc_flags & ARC_L2CACHE) |
3282 | hdr->b_flags |= ARC_L2CACHE; | |
3a17a7a9 SK |
3283 | if (*arc_flags & ARC_L2COMPRESS) |
3284 | hdr->b_flags |= ARC_L2COMPRESS; | |
34dc7c2f BB |
3285 | mutex_exit(hash_lock); |
3286 | ARCSTAT_BUMP(arcstat_hits); | |
3287 | ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), | |
3288 | demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, | |
3289 | data, metadata, hits); | |
3290 | ||
3291 | if (done) | |
3292 | done(NULL, buf, private); | |
3293 | } else { | |
3294 | uint64_t size = BP_GET_LSIZE(bp); | |
9b67f605 | 3295 | arc_callback_t *acb; |
b128c09f | 3296 | vdev_t *vd = NULL; |
a117a6d6 | 3297 | uint64_t addr = 0; |
d164b209 | 3298 | boolean_t devw = B_FALSE; |
0ed212dc BP |
3299 | enum zio_compress b_compress = ZIO_COMPRESS_OFF; |
3300 | uint64_t b_asize = 0; | |
34dc7c2f BB |
3301 | |
3302 | if (hdr == NULL) { | |
3303 | /* this block is not in the cache */ | |
9b67f605 | 3304 | arc_buf_hdr_t *exists = NULL; |
34dc7c2f BB |
3305 | arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); |
3306 | buf = arc_buf_alloc(spa, size, private, type); | |
3307 | hdr = buf->b_hdr; | |
9b67f605 MA |
3308 | if (!BP_IS_EMBEDDED(bp)) { |
3309 | hdr->b_dva = *BP_IDENTITY(bp); | |
3310 | hdr->b_birth = BP_PHYSICAL_BIRTH(bp); | |
3311 | hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; | |
3312 | exists = buf_hash_insert(hdr, &hash_lock); | |
3313 | } | |
3314 | if (exists != NULL) { | |
34dc7c2f BB |
3315 | /* somebody beat us to the hash insert */ |
3316 | mutex_exit(hash_lock); | |
428870ff | 3317 | buf_discard_identity(hdr); |
34dc7c2f BB |
3318 | (void) arc_buf_remove_ref(buf, private); |
3319 | goto top; /* restart the IO request */ | |
3320 | } | |
3321 | /* if this is a prefetch, we don't have a reference */ | |
3322 | if (*arc_flags & ARC_PREFETCH) { | |
3323 | (void) remove_reference(hdr, hash_lock, | |
3324 | private); | |
3325 | hdr->b_flags |= ARC_PREFETCH; | |
3326 | } | |
b128c09f BB |
3327 | if (*arc_flags & ARC_L2CACHE) |
3328 | hdr->b_flags |= ARC_L2CACHE; | |
3a17a7a9 SK |
3329 | if (*arc_flags & ARC_L2COMPRESS) |
3330 | hdr->b_flags |= ARC_L2COMPRESS; | |
34dc7c2f BB |
3331 | if (BP_GET_LEVEL(bp) > 0) |
3332 | hdr->b_flags |= ARC_INDIRECT; | |
3333 | } else { | |
3334 | /* this block is in the ghost cache */ | |
3335 | ASSERT(GHOST_STATE(hdr->b_state)); | |
3336 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
c99c9001 | 3337 | ASSERT0(refcount_count(&hdr->b_refcnt)); |
34dc7c2f BB |
3338 | ASSERT(hdr->b_buf == NULL); |
3339 | ||
3340 | /* if this is a prefetch, we don't have a reference */ | |
3341 | if (*arc_flags & ARC_PREFETCH) | |
3342 | hdr->b_flags |= ARC_PREFETCH; | |
3343 | else | |
3344 | add_reference(hdr, hash_lock, private); | |
b128c09f BB |
3345 | if (*arc_flags & ARC_L2CACHE) |
3346 | hdr->b_flags |= ARC_L2CACHE; | |
3a17a7a9 SK |
3347 | if (*arc_flags & ARC_L2COMPRESS) |
3348 | hdr->b_flags |= ARC_L2COMPRESS; | |
34dc7c2f BB |
3349 | buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); |
3350 | buf->b_hdr = hdr; | |
3351 | buf->b_data = NULL; | |
3352 | buf->b_efunc = NULL; | |
3353 | buf->b_private = NULL; | |
3354 | buf->b_next = NULL; | |
3355 | hdr->b_buf = buf; | |
34dc7c2f BB |
3356 | ASSERT(hdr->b_datacnt == 0); |
3357 | hdr->b_datacnt = 1; | |
428870ff BB |
3358 | arc_get_data_buf(buf); |
3359 | arc_access(hdr, hash_lock); | |
34dc7c2f BB |
3360 | } |
3361 | ||
428870ff BB |
3362 | ASSERT(!GHOST_STATE(hdr->b_state)); |
3363 | ||
691f6ac4 | 3364 | acb = kmem_zalloc(sizeof (arc_callback_t), KM_PUSHPAGE); |
34dc7c2f BB |
3365 | acb->acb_done = done; |
3366 | acb->acb_private = private; | |
34dc7c2f BB |
3367 | |
3368 | ASSERT(hdr->b_acb == NULL); | |
3369 | hdr->b_acb = acb; | |
3370 | hdr->b_flags |= ARC_IO_IN_PROGRESS; | |
3371 | ||
0ed212dc | 3372 | if (hdr->b_l2hdr != NULL && |
b128c09f | 3373 | (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { |
d164b209 | 3374 | devw = hdr->b_l2hdr->b_dev->l2ad_writing; |
b128c09f | 3375 | addr = hdr->b_l2hdr->b_daddr; |
0ed212dc BP |
3376 | b_compress = hdr->b_l2hdr->b_compress; |
3377 | b_asize = hdr->b_l2hdr->b_asize; | |
b128c09f BB |
3378 | /* |
3379 | * Lock out device removal. | |
3380 | */ | |
3381 | if (vdev_is_dead(vd) || | |
3382 | !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) | |
3383 | vd = NULL; | |
3384 | } | |
3385 | ||
9b67f605 MA |
3386 | if (hash_lock != NULL) |
3387 | mutex_exit(hash_lock); | |
b128c09f | 3388 | |
e49f1e20 WA |
3389 | /* |
3390 | * At this point, we have a level 1 cache miss. Try again in | |
3391 | * L2ARC if possible. | |
3392 | */ | |
34dc7c2f | 3393 | ASSERT3U(hdr->b_size, ==, size); |
428870ff BB |
3394 | DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, |
3395 | uint64_t, size, zbookmark_t *, zb); | |
34dc7c2f BB |
3396 | ARCSTAT_BUMP(arcstat_misses); |
3397 | ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), | |
3398 | demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, | |
3399 | data, metadata, misses); | |
3400 | ||
d164b209 | 3401 | if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { |
34dc7c2f BB |
3402 | /* |
3403 | * Read from the L2ARC if the following are true: | |
b128c09f BB |
3404 | * 1. The L2ARC vdev was previously cached. |
3405 | * 2. This buffer still has L2ARC metadata. | |
3406 | * 3. This buffer isn't currently writing to the L2ARC. | |
3407 | * 4. The L2ARC entry wasn't evicted, which may | |
3408 | * also have invalidated the vdev. | |
d164b209 | 3409 | * 5. This isn't prefetch and l2arc_noprefetch is set. |
34dc7c2f | 3410 | */ |
b128c09f | 3411 | if (hdr->b_l2hdr != NULL && |
d164b209 BB |
3412 | !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && |
3413 | !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { | |
34dc7c2f BB |
3414 | l2arc_read_callback_t *cb; |
3415 | ||
3416 | DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); | |
3417 | ARCSTAT_BUMP(arcstat_l2_hits); | |
e0b0ca98 | 3418 | atomic_inc_32(&hdr->b_l2hdr->b_hits); |
34dc7c2f | 3419 | |
34dc7c2f | 3420 | cb = kmem_zalloc(sizeof (l2arc_read_callback_t), |
691f6ac4 | 3421 | KM_PUSHPAGE); |
34dc7c2f BB |
3422 | cb->l2rcb_buf = buf; |
3423 | cb->l2rcb_spa = spa; | |
3424 | cb->l2rcb_bp = *bp; | |
3425 | cb->l2rcb_zb = *zb; | |
b128c09f | 3426 | cb->l2rcb_flags = zio_flags; |
0ed212dc | 3427 | cb->l2rcb_compress = b_compress; |
34dc7c2f | 3428 | |
a117a6d6 GW |
3429 | ASSERT(addr >= VDEV_LABEL_START_SIZE && |
3430 | addr + size < vd->vdev_psize - | |
3431 | VDEV_LABEL_END_SIZE); | |
3432 | ||
34dc7c2f | 3433 | /* |
b128c09f BB |
3434 | * l2arc read. The SCL_L2ARC lock will be |
3435 | * released by l2arc_read_done(). | |
3a17a7a9 SK |
3436 | * Issue a null zio if the underlying buffer |
3437 | * was squashed to zero size by compression. | |
34dc7c2f | 3438 | */ |
0ed212dc | 3439 | if (b_compress == ZIO_COMPRESS_EMPTY) { |
3a17a7a9 SK |
3440 | rzio = zio_null(pio, spa, vd, |
3441 | l2arc_read_done, cb, | |
3442 | zio_flags | ZIO_FLAG_DONT_CACHE | | |
3443 | ZIO_FLAG_CANFAIL | | |
3444 | ZIO_FLAG_DONT_PROPAGATE | | |
3445 | ZIO_FLAG_DONT_RETRY); | |
3446 | } else { | |
3447 | rzio = zio_read_phys(pio, vd, addr, | |
0ed212dc BP |
3448 | b_asize, buf->b_data, |
3449 | ZIO_CHECKSUM_OFF, | |
3a17a7a9 SK |
3450 | l2arc_read_done, cb, priority, |
3451 | zio_flags | ZIO_FLAG_DONT_CACHE | | |
3452 | ZIO_FLAG_CANFAIL | | |
3453 | ZIO_FLAG_DONT_PROPAGATE | | |
3454 | ZIO_FLAG_DONT_RETRY, B_FALSE); | |
3455 | } | |
34dc7c2f BB |
3456 | DTRACE_PROBE2(l2arc__read, vdev_t *, vd, |
3457 | zio_t *, rzio); | |
0ed212dc | 3458 | ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); |
34dc7c2f | 3459 | |
b128c09f BB |
3460 | if (*arc_flags & ARC_NOWAIT) { |
3461 | zio_nowait(rzio); | |
1421c891 | 3462 | goto out; |
b128c09f | 3463 | } |
34dc7c2f | 3464 | |
b128c09f BB |
3465 | ASSERT(*arc_flags & ARC_WAIT); |
3466 | if (zio_wait(rzio) == 0) | |
1421c891 | 3467 | goto out; |
b128c09f BB |
3468 | |
3469 | /* l2arc read error; goto zio_read() */ | |
34dc7c2f BB |
3470 | } else { |
3471 | DTRACE_PROBE1(l2arc__miss, | |
3472 | arc_buf_hdr_t *, hdr); | |
3473 | ARCSTAT_BUMP(arcstat_l2_misses); | |
3474 | if (HDR_L2_WRITING(hdr)) | |
3475 | ARCSTAT_BUMP(arcstat_l2_rw_clash); | |
b128c09f | 3476 | spa_config_exit(spa, SCL_L2ARC, vd); |
34dc7c2f | 3477 | } |
d164b209 BB |
3478 | } else { |
3479 | if (vd != NULL) | |
3480 | spa_config_exit(spa, SCL_L2ARC, vd); | |
3481 | if (l2arc_ndev != 0) { | |
3482 | DTRACE_PROBE1(l2arc__miss, | |
3483 | arc_buf_hdr_t *, hdr); | |
3484 | ARCSTAT_BUMP(arcstat_l2_misses); | |
3485 | } | |
34dc7c2f | 3486 | } |
34dc7c2f BB |
3487 | |
3488 | rzio = zio_read(pio, spa, bp, buf->b_data, size, | |
b128c09f | 3489 | arc_read_done, buf, priority, zio_flags, zb); |
34dc7c2f | 3490 | |
1421c891 PS |
3491 | if (*arc_flags & ARC_WAIT) { |
3492 | rc = zio_wait(rzio); | |
3493 | goto out; | |
3494 | } | |
34dc7c2f BB |
3495 | |
3496 | ASSERT(*arc_flags & ARC_NOWAIT); | |
3497 | zio_nowait(rzio); | |
3498 | } | |
1421c891 PS |
3499 | |
3500 | out: | |
3501 | spa_read_history_add(spa, zb, *arc_flags); | |
3502 | return (rc); | |
34dc7c2f BB |
3503 | } |
3504 | ||
ab26409d BB |
3505 | arc_prune_t * |
3506 | arc_add_prune_callback(arc_prune_func_t *func, void *private) | |
3507 | { | |
3508 | arc_prune_t *p; | |
3509 | ||
d1d7e268 | 3510 | p = kmem_alloc(sizeof (*p), KM_SLEEP); |
ab26409d BB |
3511 | p->p_pfunc = func; |
3512 | p->p_private = private; | |
3513 | list_link_init(&p->p_node); | |
3514 | refcount_create(&p->p_refcnt); | |
3515 | ||
3516 | mutex_enter(&arc_prune_mtx); | |
3517 | refcount_add(&p->p_refcnt, &arc_prune_list); | |
3518 | list_insert_head(&arc_prune_list, p); | |
3519 | mutex_exit(&arc_prune_mtx); | |
3520 | ||
3521 | return (p); | |
3522 | } | |
3523 | ||
3524 | void | |
3525 | arc_remove_prune_callback(arc_prune_t *p) | |
3526 | { | |
3527 | mutex_enter(&arc_prune_mtx); | |
3528 | list_remove(&arc_prune_list, p); | |
3529 | if (refcount_remove(&p->p_refcnt, &arc_prune_list) == 0) { | |
3530 | refcount_destroy(&p->p_refcnt); | |
3531 | kmem_free(p, sizeof (*p)); | |
3532 | } | |
3533 | mutex_exit(&arc_prune_mtx); | |
3534 | } | |
3535 | ||
34dc7c2f BB |
3536 | void |
3537 | arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) | |
3538 | { | |
3539 | ASSERT(buf->b_hdr != NULL); | |
3540 | ASSERT(buf->b_hdr->b_state != arc_anon); | |
3541 | ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); | |
428870ff BB |
3542 | ASSERT(buf->b_efunc == NULL); |
3543 | ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); | |
3544 | ||
34dc7c2f BB |
3545 | buf->b_efunc = func; |
3546 | buf->b_private = private; | |
3547 | } | |
3548 | ||
df4474f9 MA |
3549 | /* |
3550 | * Notify the arc that a block was freed, and thus will never be used again. | |
3551 | */ | |
3552 | void | |
3553 | arc_freed(spa_t *spa, const blkptr_t *bp) | |
3554 | { | |
3555 | arc_buf_hdr_t *hdr; | |
3556 | kmutex_t *hash_lock; | |
3557 | uint64_t guid = spa_load_guid(spa); | |
3558 | ||
9b67f605 MA |
3559 | ASSERT(!BP_IS_EMBEDDED(bp)); |
3560 | ||
3561 | hdr = buf_hash_find(guid, bp, &hash_lock); | |
df4474f9 MA |
3562 | if (hdr == NULL) |
3563 | return; | |
3564 | if (HDR_BUF_AVAILABLE(hdr)) { | |
3565 | arc_buf_t *buf = hdr->b_buf; | |
3566 | add_reference(hdr, hash_lock, FTAG); | |
3567 | hdr->b_flags &= ~ARC_BUF_AVAILABLE; | |
3568 | mutex_exit(hash_lock); | |
3569 | ||
3570 | arc_release(buf, FTAG); | |
3571 | (void) arc_buf_remove_ref(buf, FTAG); | |
3572 | } else { | |
3573 | mutex_exit(hash_lock); | |
3574 | } | |
3575 | ||
3576 | } | |
3577 | ||
34dc7c2f BB |
3578 | /* |
3579 | * This is used by the DMU to let the ARC know that a buffer is | |
3580 | * being evicted, so the ARC should clean up. If this arc buf | |
3581 | * is not yet in the evicted state, it will be put there. | |
3582 | */ | |
3583 | int | |
3584 | arc_buf_evict(arc_buf_t *buf) | |
3585 | { | |
3586 | arc_buf_hdr_t *hdr; | |
3587 | kmutex_t *hash_lock; | |
3588 | arc_buf_t **bufp; | |
3589 | ||
428870ff | 3590 | mutex_enter(&buf->b_evict_lock); |
34dc7c2f BB |
3591 | hdr = buf->b_hdr; |
3592 | if (hdr == NULL) { | |
3593 | /* | |
3594 | * We are in arc_do_user_evicts(). | |
3595 | */ | |
3596 | ASSERT(buf->b_data == NULL); | |
428870ff | 3597 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f | 3598 | return (0); |
b128c09f BB |
3599 | } else if (buf->b_data == NULL) { |
3600 | arc_buf_t copy = *buf; /* structure assignment */ | |
34dc7c2f | 3601 | /* |
b128c09f BB |
3602 | * We are on the eviction list; process this buffer now |
3603 | * but let arc_do_user_evicts() do the reaping. | |
34dc7c2f | 3604 | */ |
b128c09f | 3605 | buf->b_efunc = NULL; |
428870ff | 3606 | mutex_exit(&buf->b_evict_lock); |
b128c09f BB |
3607 | VERIFY(copy.b_efunc(©) == 0); |
3608 | return (1); | |
34dc7c2f | 3609 | } |
b128c09f BB |
3610 | hash_lock = HDR_LOCK(hdr); |
3611 | mutex_enter(hash_lock); | |
428870ff BB |
3612 | hdr = buf->b_hdr; |
3613 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
34dc7c2f | 3614 | |
34dc7c2f BB |
3615 | ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); |
3616 | ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); | |
3617 | ||
3618 | /* | |
3619 | * Pull this buffer off of the hdr | |
3620 | */ | |
3621 | bufp = &hdr->b_buf; | |
3622 | while (*bufp != buf) | |
3623 | bufp = &(*bufp)->b_next; | |
3624 | *bufp = buf->b_next; | |
3625 | ||
3626 | ASSERT(buf->b_data != NULL); | |
3627 | arc_buf_destroy(buf, FALSE, FALSE); | |
3628 | ||
3629 | if (hdr->b_datacnt == 0) { | |
3630 | arc_state_t *old_state = hdr->b_state; | |
3631 | arc_state_t *evicted_state; | |
3632 | ||
428870ff | 3633 | ASSERT(hdr->b_buf == NULL); |
34dc7c2f BB |
3634 | ASSERT(refcount_is_zero(&hdr->b_refcnt)); |
3635 | ||
3636 | evicted_state = | |
3637 | (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; | |
3638 | ||
3639 | mutex_enter(&old_state->arcs_mtx); | |
3640 | mutex_enter(&evicted_state->arcs_mtx); | |
3641 | ||
3642 | arc_change_state(evicted_state, hdr, hash_lock); | |
3643 | ASSERT(HDR_IN_HASH_TABLE(hdr)); | |
3644 | hdr->b_flags |= ARC_IN_HASH_TABLE; | |
3645 | hdr->b_flags &= ~ARC_BUF_AVAILABLE; | |
3646 | ||
3647 | mutex_exit(&evicted_state->arcs_mtx); | |
3648 | mutex_exit(&old_state->arcs_mtx); | |
3649 | } | |
3650 | mutex_exit(hash_lock); | |
428870ff | 3651 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
3652 | |
3653 | VERIFY(buf->b_efunc(buf) == 0); | |
3654 | buf->b_efunc = NULL; | |
3655 | buf->b_private = NULL; | |
3656 | buf->b_hdr = NULL; | |
428870ff | 3657 | buf->b_next = NULL; |
34dc7c2f BB |
3658 | kmem_cache_free(buf_cache, buf); |
3659 | return (1); | |
3660 | } | |
3661 | ||
3662 | /* | |
e49f1e20 WA |
3663 | * Release this buffer from the cache, making it an anonymous buffer. This |
3664 | * must be done after a read and prior to modifying the buffer contents. | |
34dc7c2f | 3665 | * If the buffer has more than one reference, we must make |
b128c09f | 3666 | * a new hdr for the buffer. |
34dc7c2f BB |
3667 | */ |
3668 | void | |
3669 | arc_release(arc_buf_t *buf, void *tag) | |
3670 | { | |
b128c09f | 3671 | arc_buf_hdr_t *hdr; |
428870ff | 3672 | kmutex_t *hash_lock = NULL; |
b128c09f | 3673 | l2arc_buf_hdr_t *l2hdr; |
d4ed6673 | 3674 | uint64_t buf_size = 0; |
34dc7c2f | 3675 | |
428870ff BB |
3676 | /* |
3677 | * It would be nice to assert that if it's DMU metadata (level > | |
3678 | * 0 || it's the dnode file), then it must be syncing context. | |
3679 | * But we don't know that information at this level. | |
3680 | */ | |
3681 | ||
3682 | mutex_enter(&buf->b_evict_lock); | |
b128c09f BB |
3683 | hdr = buf->b_hdr; |
3684 | ||
34dc7c2f BB |
3685 | /* this buffer is not on any list */ |
3686 | ASSERT(refcount_count(&hdr->b_refcnt) > 0); | |
3687 | ||
3688 | if (hdr->b_state == arc_anon) { | |
3689 | /* this buffer is already released */ | |
34dc7c2f | 3690 | ASSERT(buf->b_efunc == NULL); |
9babb374 BB |
3691 | } else { |
3692 | hash_lock = HDR_LOCK(hdr); | |
3693 | mutex_enter(hash_lock); | |
428870ff BB |
3694 | hdr = buf->b_hdr; |
3695 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
34dc7c2f BB |
3696 | } |
3697 | ||
b128c09f BB |
3698 | l2hdr = hdr->b_l2hdr; |
3699 | if (l2hdr) { | |
3700 | mutex_enter(&l2arc_buflist_mtx); | |
3701 | hdr->b_l2hdr = NULL; | |
47fe91b5 | 3702 | list_remove(l2hdr->b_dev->l2ad_buflist, hdr); |
b128c09f | 3703 | } |
a117a6d6 | 3704 | buf_size = hdr->b_size; |
b128c09f | 3705 | |
34dc7c2f BB |
3706 | /* |
3707 | * Do we have more than one buf? | |
3708 | */ | |
b128c09f | 3709 | if (hdr->b_datacnt > 1) { |
34dc7c2f BB |
3710 | arc_buf_hdr_t *nhdr; |
3711 | arc_buf_t **bufp; | |
3712 | uint64_t blksz = hdr->b_size; | |
d164b209 | 3713 | uint64_t spa = hdr->b_spa; |
34dc7c2f BB |
3714 | arc_buf_contents_t type = hdr->b_type; |
3715 | uint32_t flags = hdr->b_flags; | |
3716 | ||
b128c09f | 3717 | ASSERT(hdr->b_buf != buf || buf->b_next != NULL); |
34dc7c2f | 3718 | /* |
428870ff BB |
3719 | * Pull the data off of this hdr and attach it to |
3720 | * a new anonymous hdr. | |
34dc7c2f BB |
3721 | */ |
3722 | (void) remove_reference(hdr, hash_lock, tag); | |
3723 | bufp = &hdr->b_buf; | |
3724 | while (*bufp != buf) | |
3725 | bufp = &(*bufp)->b_next; | |
428870ff | 3726 | *bufp = buf->b_next; |
34dc7c2f BB |
3727 | buf->b_next = NULL; |
3728 | ||
3729 | ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); | |
3730 | atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); | |
3731 | if (refcount_is_zero(&hdr->b_refcnt)) { | |
3732 | uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; | |
3733 | ASSERT3U(*size, >=, hdr->b_size); | |
3734 | atomic_add_64(size, -hdr->b_size); | |
3735 | } | |
1eb5bfa3 GW |
3736 | |
3737 | /* | |
3738 | * We're releasing a duplicate user data buffer, update | |
3739 | * our statistics accordingly. | |
3740 | */ | |
3741 | if (hdr->b_type == ARC_BUFC_DATA) { | |
3742 | ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); | |
3743 | ARCSTAT_INCR(arcstat_duplicate_buffers_size, | |
3744 | -hdr->b_size); | |
3745 | } | |
34dc7c2f | 3746 | hdr->b_datacnt -= 1; |
34dc7c2f | 3747 | arc_cksum_verify(buf); |
498877ba | 3748 | arc_buf_unwatch(buf); |
34dc7c2f BB |
3749 | |
3750 | mutex_exit(hash_lock); | |
3751 | ||
3752 | nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); | |
3753 | nhdr->b_size = blksz; | |
3754 | nhdr->b_spa = spa; | |
3755 | nhdr->b_type = type; | |
3756 | nhdr->b_buf = buf; | |
3757 | nhdr->b_state = arc_anon; | |
3758 | nhdr->b_arc_access = 0; | |
e0b0ca98 BB |
3759 | nhdr->b_mru_hits = 0; |
3760 | nhdr->b_mru_ghost_hits = 0; | |
3761 | nhdr->b_mfu_hits = 0; | |
3762 | nhdr->b_mfu_ghost_hits = 0; | |
3763 | nhdr->b_l2_hits = 0; | |
34dc7c2f BB |
3764 | nhdr->b_flags = flags & ARC_L2_WRITING; |
3765 | nhdr->b_l2hdr = NULL; | |
3766 | nhdr->b_datacnt = 1; | |
3767 | nhdr->b_freeze_cksum = NULL; | |
3768 | (void) refcount_add(&nhdr->b_refcnt, tag); | |
3769 | buf->b_hdr = nhdr; | |
428870ff | 3770 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
3771 | atomic_add_64(&arc_anon->arcs_size, blksz); |
3772 | } else { | |
428870ff | 3773 | mutex_exit(&buf->b_evict_lock); |
34dc7c2f BB |
3774 | ASSERT(refcount_count(&hdr->b_refcnt) == 1); |
3775 | ASSERT(!list_link_active(&hdr->b_arc_node)); | |
3776 | ASSERT(!HDR_IO_IN_PROGRESS(hdr)); | |
428870ff BB |
3777 | if (hdr->b_state != arc_anon) |
3778 | arc_change_state(arc_anon, hdr, hash_lock); | |
34dc7c2f | 3779 | hdr->b_arc_access = 0; |
e0b0ca98 BB |
3780 | hdr->b_mru_hits = 0; |
3781 | hdr->b_mru_ghost_hits = 0; | |
3782 | hdr->b_mfu_hits = 0; | |
3783 | hdr->b_mfu_ghost_hits = 0; | |
3784 | hdr->b_l2_hits = 0; | |
428870ff BB |
3785 | if (hash_lock) |
3786 | mutex_exit(hash_lock); | |
34dc7c2f | 3787 | |
428870ff | 3788 | buf_discard_identity(hdr); |
34dc7c2f BB |
3789 | arc_buf_thaw(buf); |
3790 | } | |
3791 | buf->b_efunc = NULL; | |
3792 | buf->b_private = NULL; | |
3793 | ||
3794 | if (l2hdr) { | |
3a17a7a9 | 3795 | ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); |
3bec585e SK |
3796 | vdev_space_update(l2hdr->b_dev->l2ad_vdev, |
3797 | -l2hdr->b_asize, 0, 0); | |
ecf3d9b8 | 3798 | kmem_cache_free(l2arc_hdr_cache, l2hdr); |
6e1d7276 | 3799 | arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS); |
34dc7c2f | 3800 | ARCSTAT_INCR(arcstat_l2_size, -buf_size); |
34dc7c2f | 3801 | mutex_exit(&l2arc_buflist_mtx); |
b128c09f | 3802 | } |
34dc7c2f BB |
3803 | } |
3804 | ||
3805 | int | |
3806 | arc_released(arc_buf_t *buf) | |
3807 | { | |
b128c09f BB |
3808 | int released; |
3809 | ||
428870ff | 3810 | mutex_enter(&buf->b_evict_lock); |
b128c09f | 3811 | released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); |
428870ff | 3812 | mutex_exit(&buf->b_evict_lock); |
b128c09f | 3813 | return (released); |
34dc7c2f BB |
3814 | } |
3815 | ||
3816 | int | |
3817 | arc_has_callback(arc_buf_t *buf) | |
3818 | { | |
b128c09f BB |
3819 | int callback; |
3820 | ||
428870ff | 3821 | mutex_enter(&buf->b_evict_lock); |
b128c09f | 3822 | callback = (buf->b_efunc != NULL); |
428870ff | 3823 | mutex_exit(&buf->b_evict_lock); |
b128c09f | 3824 | return (callback); |
34dc7c2f BB |
3825 | } |
3826 | ||
3827 | #ifdef ZFS_DEBUG | |
3828 | int | |
3829 | arc_referenced(arc_buf_t *buf) | |
3830 | { | |
b128c09f BB |
3831 | int referenced; |
3832 | ||
428870ff | 3833 | mutex_enter(&buf->b_evict_lock); |
b128c09f | 3834 | referenced = (refcount_count(&buf->b_hdr->b_refcnt)); |
428870ff | 3835 | mutex_exit(&buf->b_evict_lock); |
b128c09f | 3836 | return (referenced); |
34dc7c2f BB |
3837 | } |
3838 | #endif | |
3839 | ||
3840 | static void | |
3841 | arc_write_ready(zio_t *zio) | |
3842 | { | |
3843 | arc_write_callback_t *callback = zio->io_private; | |
3844 | arc_buf_t *buf = callback->awcb_buf; | |
3845 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
3846 | ||
b128c09f BB |
3847 | ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); |
3848 | callback->awcb_ready(zio, buf, callback->awcb_private); | |
3849 | ||
34dc7c2f BB |
3850 | /* |
3851 | * If the IO is already in progress, then this is a re-write | |
b128c09f BB |
3852 | * attempt, so we need to thaw and re-compute the cksum. |
3853 | * It is the responsibility of the callback to handle the | |
3854 | * accounting for any re-write attempt. | |
34dc7c2f BB |
3855 | */ |
3856 | if (HDR_IO_IN_PROGRESS(hdr)) { | |
34dc7c2f BB |
3857 | mutex_enter(&hdr->b_freeze_lock); |
3858 | if (hdr->b_freeze_cksum != NULL) { | |
3859 | kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); | |
3860 | hdr->b_freeze_cksum = NULL; | |
3861 | } | |
3862 | mutex_exit(&hdr->b_freeze_lock); | |
3863 | } | |
3864 | arc_cksum_compute(buf, B_FALSE); | |
3865 | hdr->b_flags |= ARC_IO_IN_PROGRESS; | |
3866 | } | |
3867 | ||
e8b96c60 MA |
3868 | /* |
3869 | * The SPA calls this callback for each physical write that happens on behalf | |
3870 | * of a logical write. See the comment in dbuf_write_physdone() for details. | |
3871 | */ | |
3872 | static void | |
3873 | arc_write_physdone(zio_t *zio) | |
3874 | { | |
3875 | arc_write_callback_t *cb = zio->io_private; | |
3876 | if (cb->awcb_physdone != NULL) | |
3877 | cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); | |
3878 | } | |
3879 | ||
34dc7c2f BB |
3880 | static void |
3881 | arc_write_done(zio_t *zio) | |
3882 | { | |
3883 | arc_write_callback_t *callback = zio->io_private; | |
3884 | arc_buf_t *buf = callback->awcb_buf; | |
3885 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
3886 | ||
428870ff BB |
3887 | ASSERT(hdr->b_acb == NULL); |
3888 | ||
3889 | if (zio->io_error == 0) { | |
9b67f605 | 3890 | if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { |
b0bc7a84 MG |
3891 | buf_discard_identity(hdr); |
3892 | } else { | |
3893 | hdr->b_dva = *BP_IDENTITY(zio->io_bp); | |
3894 | hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); | |
3895 | hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; | |
3896 | } | |
428870ff BB |
3897 | } else { |
3898 | ASSERT(BUF_EMPTY(hdr)); | |
3899 | } | |
34dc7c2f | 3900 | |
34dc7c2f | 3901 | /* |
9b67f605 MA |
3902 | * If the block to be written was all-zero or compressed enough to be |
3903 | * embedded in the BP, no write was performed so there will be no | |
3904 | * dva/birth/checksum. The buffer must therefore remain anonymous | |
3905 | * (and uncached). | |
34dc7c2f BB |
3906 | */ |
3907 | if (!BUF_EMPTY(hdr)) { | |
3908 | arc_buf_hdr_t *exists; | |
3909 | kmutex_t *hash_lock; | |
3910 | ||
428870ff BB |
3911 | ASSERT(zio->io_error == 0); |
3912 | ||
34dc7c2f BB |
3913 | arc_cksum_verify(buf); |
3914 | ||
3915 | exists = buf_hash_insert(hdr, &hash_lock); | |
3916 | if (exists) { | |
3917 | /* | |
3918 | * This can only happen if we overwrite for | |
3919 | * sync-to-convergence, because we remove | |
3920 | * buffers from the hash table when we arc_free(). | |
3921 | */ | |
428870ff BB |
3922 | if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { |
3923 | if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) | |
3924 | panic("bad overwrite, hdr=%p exists=%p", | |
3925 | (void *)hdr, (void *)exists); | |
3926 | ASSERT(refcount_is_zero(&exists->b_refcnt)); | |
3927 | arc_change_state(arc_anon, exists, hash_lock); | |
3928 | mutex_exit(hash_lock); | |
3929 | arc_hdr_destroy(exists); | |
3930 | exists = buf_hash_insert(hdr, &hash_lock); | |
3931 | ASSERT3P(exists, ==, NULL); | |
03c6040b GW |
3932 | } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { |
3933 | /* nopwrite */ | |
3934 | ASSERT(zio->io_prop.zp_nopwrite); | |
3935 | if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) | |
3936 | panic("bad nopwrite, hdr=%p exists=%p", | |
3937 | (void *)hdr, (void *)exists); | |
428870ff BB |
3938 | } else { |
3939 | /* Dedup */ | |
3940 | ASSERT(hdr->b_datacnt == 1); | |
3941 | ASSERT(hdr->b_state == arc_anon); | |
3942 | ASSERT(BP_GET_DEDUP(zio->io_bp)); | |
3943 | ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); | |
3944 | } | |
34dc7c2f BB |
3945 | } |
3946 | hdr->b_flags &= ~ARC_IO_IN_PROGRESS; | |
b128c09f | 3947 | /* if it's not anon, we are doing a scrub */ |
428870ff | 3948 | if (!exists && hdr->b_state == arc_anon) |
b128c09f | 3949 | arc_access(hdr, hash_lock); |
34dc7c2f | 3950 | mutex_exit(hash_lock); |
34dc7c2f BB |
3951 | } else { |
3952 | hdr->b_flags &= ~ARC_IO_IN_PROGRESS; | |
3953 | } | |
3954 | ||
428870ff BB |
3955 | ASSERT(!refcount_is_zero(&hdr->b_refcnt)); |
3956 | callback->awcb_done(zio, buf, callback->awcb_private); | |
34dc7c2f BB |
3957 | |
3958 | kmem_free(callback, sizeof (arc_write_callback_t)); | |
3959 | } | |
3960 | ||
3961 | zio_t * | |
428870ff | 3962 | arc_write(zio_t *pio, spa_t *spa, uint64_t txg, |
3a17a7a9 | 3963 | blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, |
e8b96c60 MA |
3964 | const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, |
3965 | arc_done_func_t *done, void *private, zio_priority_t priority, | |
3966 | int zio_flags, const zbookmark_t *zb) | |
34dc7c2f BB |
3967 | { |
3968 | arc_buf_hdr_t *hdr = buf->b_hdr; | |
3969 | arc_write_callback_t *callback; | |
b128c09f | 3970 | zio_t *zio; |
34dc7c2f | 3971 | |
b128c09f | 3972 | ASSERT(ready != NULL); |
428870ff | 3973 | ASSERT(done != NULL); |
34dc7c2f BB |
3974 | ASSERT(!HDR_IO_ERROR(hdr)); |
3975 | ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); | |
428870ff | 3976 | ASSERT(hdr->b_acb == NULL); |
b128c09f BB |
3977 | if (l2arc) |
3978 | hdr->b_flags |= ARC_L2CACHE; | |
3a17a7a9 SK |
3979 | if (l2arc_compress) |
3980 | hdr->b_flags |= ARC_L2COMPRESS; | |
b8d06fca | 3981 | callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_PUSHPAGE); |
34dc7c2f | 3982 | callback->awcb_ready = ready; |
e8b96c60 | 3983 | callback->awcb_physdone = physdone; |
34dc7c2f BB |
3984 | callback->awcb_done = done; |
3985 | callback->awcb_private = private; | |
3986 | callback->awcb_buf = buf; | |
b128c09f | 3987 | |
428870ff | 3988 | zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, |
e8b96c60 MA |
3989 | arc_write_ready, arc_write_physdone, arc_write_done, callback, |
3990 | priority, zio_flags, zb); | |
34dc7c2f BB |
3991 | |
3992 | return (zio); | |
3993 | } | |
3994 | ||
34dc7c2f | 3995 | static int |
e8b96c60 | 3996 | arc_memory_throttle(uint64_t reserve, uint64_t txg) |
34dc7c2f BB |
3997 | { |
3998 | #ifdef _KERNEL | |
0c5493d4 BB |
3999 | if (zfs_arc_memory_throttle_disable) |
4000 | return (0); | |
4001 | ||
e8b96c60 | 4002 | if (freemem <= physmem * arc_lotsfree_percent / 100) { |
34dc7c2f | 4003 | ARCSTAT_INCR(arcstat_memory_throttle_count, 1); |
570827e1 | 4004 | DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim); |
2e528b49 | 4005 | return (SET_ERROR(EAGAIN)); |
34dc7c2f | 4006 | } |
34dc7c2f BB |
4007 | #endif |
4008 | return (0); | |
4009 | } | |
4010 | ||
4011 | void | |
4012 | arc_tempreserve_clear(uint64_t reserve) | |
4013 | { | |
4014 | atomic_add_64(&arc_tempreserve, -reserve); | |
4015 | ASSERT((int64_t)arc_tempreserve >= 0); | |
4016 | } | |
4017 | ||
4018 | int | |
4019 | arc_tempreserve_space(uint64_t reserve, uint64_t txg) | |
4020 | { | |
4021 | int error; | |
9babb374 | 4022 | uint64_t anon_size; |
34dc7c2f | 4023 | |
34dc7c2f BB |
4024 | if (reserve > arc_c/4 && !arc_no_grow) |
4025 | arc_c = MIN(arc_c_max, reserve * 4); | |
12f9a6a3 BB |
4026 | |
4027 | /* | |
4028 | * Throttle when the calculated memory footprint for the TXG | |
4029 | * exceeds the target ARC size. | |
4030 | */ | |
570827e1 BB |
4031 | if (reserve > arc_c) { |
4032 | DMU_TX_STAT_BUMP(dmu_tx_memory_reserve); | |
12f9a6a3 | 4033 | return (SET_ERROR(ERESTART)); |
570827e1 | 4034 | } |
34dc7c2f | 4035 | |
9babb374 BB |
4036 | /* |
4037 | * Don't count loaned bufs as in flight dirty data to prevent long | |
4038 | * network delays from blocking transactions that are ready to be | |
4039 | * assigned to a txg. | |
4040 | */ | |
4041 | anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); | |
4042 | ||
34dc7c2f BB |
4043 | /* |
4044 | * Writes will, almost always, require additional memory allocations | |
d3cc8b15 | 4045 | * in order to compress/encrypt/etc the data. We therefore need to |
34dc7c2f BB |
4046 | * make sure that there is sufficient available memory for this. |
4047 | */ | |
e8b96c60 MA |
4048 | error = arc_memory_throttle(reserve, txg); |
4049 | if (error != 0) | |
34dc7c2f BB |
4050 | return (error); |
4051 | ||
4052 | /* | |
4053 | * Throttle writes when the amount of dirty data in the cache | |
4054 | * gets too large. We try to keep the cache less than half full | |
4055 | * of dirty blocks so that our sync times don't grow too large. | |
4056 | * Note: if two requests come in concurrently, we might let them | |
4057 | * both succeed, when one of them should fail. Not a huge deal. | |
4058 | */ | |
9babb374 BB |
4059 | |
4060 | if (reserve + arc_tempreserve + anon_size > arc_c / 2 && | |
4061 | anon_size > arc_c / 4) { | |
34dc7c2f BB |
4062 | dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " |
4063 | "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", | |
4064 | arc_tempreserve>>10, | |
4065 | arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, | |
4066 | arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, | |
4067 | reserve>>10, arc_c>>10); | |
570827e1 | 4068 | DMU_TX_STAT_BUMP(dmu_tx_dirty_throttle); |
2e528b49 | 4069 | return (SET_ERROR(ERESTART)); |
34dc7c2f BB |
4070 | } |
4071 | atomic_add_64(&arc_tempreserve, reserve); | |
4072 | return (0); | |
4073 | } | |
4074 | ||
13be560d BB |
4075 | static void |
4076 | arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, | |
4077 | kstat_named_t *evict_data, kstat_named_t *evict_metadata) | |
4078 | { | |
4079 | size->value.ui64 = state->arcs_size; | |
4080 | evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; | |
4081 | evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; | |
4082 | } | |
4083 | ||
4084 | static int | |
4085 | arc_kstat_update(kstat_t *ksp, int rw) | |
4086 | { | |
4087 | arc_stats_t *as = ksp->ks_data; | |
4088 | ||
4089 | if (rw == KSTAT_WRITE) { | |
2e528b49 | 4090 | return (SET_ERROR(EACCES)); |
13be560d BB |
4091 | } else { |
4092 | arc_kstat_update_state(arc_anon, | |
4093 | &as->arcstat_anon_size, | |
4094 | &as->arcstat_anon_evict_data, | |
4095 | &as->arcstat_anon_evict_metadata); | |
4096 | arc_kstat_update_state(arc_mru, | |
4097 | &as->arcstat_mru_size, | |
4098 | &as->arcstat_mru_evict_data, | |
4099 | &as->arcstat_mru_evict_metadata); | |
4100 | arc_kstat_update_state(arc_mru_ghost, | |
4101 | &as->arcstat_mru_ghost_size, | |
4102 | &as->arcstat_mru_ghost_evict_data, | |
4103 | &as->arcstat_mru_ghost_evict_metadata); | |
4104 | arc_kstat_update_state(arc_mfu, | |
4105 | &as->arcstat_mfu_size, | |
4106 | &as->arcstat_mfu_evict_data, | |
4107 | &as->arcstat_mfu_evict_metadata); | |
fc41c640 | 4108 | arc_kstat_update_state(arc_mfu_ghost, |
13be560d BB |
4109 | &as->arcstat_mfu_ghost_size, |
4110 | &as->arcstat_mfu_ghost_evict_data, | |
4111 | &as->arcstat_mfu_ghost_evict_metadata); | |
4112 | } | |
4113 | ||
4114 | return (0); | |
4115 | } | |
4116 | ||
34dc7c2f BB |
4117 | void |
4118 | arc_init(void) | |
4119 | { | |
4120 | mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); | |
4121 | cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); | |
4122 | ||
4123 | /* Convert seconds to clock ticks */ | |
bce45ec9 | 4124 | zfs_arc_min_prefetch_lifespan = 1 * hz; |
34dc7c2f BB |
4125 | |
4126 | /* Start out with 1/8 of all memory */ | |
4127 | arc_c = physmem * PAGESIZE / 8; | |
4128 | ||
4129 | #ifdef _KERNEL | |
4130 | /* | |
4131 | * On architectures where the physical memory can be larger | |
4132 | * than the addressable space (intel in 32-bit mode), we may | |
4133 | * need to limit the cache to 1/8 of VM size. | |
4134 | */ | |
4135 | arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); | |
7cb67b45 BB |
4136 | /* |
4137 | * Register a shrinker to support synchronous (direct) memory | |
4138 | * reclaim from the arc. This is done to prevent kswapd from | |
4139 | * swapping out pages when it is preferable to shrink the arc. | |
4140 | */ | |
4141 | spl_register_shrinker(&arc_shrinker); | |
34dc7c2f BB |
4142 | #endif |
4143 | ||
91415825 BB |
4144 | /* set min cache to zero */ |
4145 | arc_c_min = 4<<20; | |
518b4876 | 4146 | /* set max to 1/2 of all memory */ |
be5db977 | 4147 | arc_c_max = arc_c * 4; |
34dc7c2f BB |
4148 | |
4149 | /* | |
4150 | * Allow the tunables to override our calculations if they are | |
4151 | * reasonable (ie. over 64MB) | |
4152 | */ | |
4153 | if (zfs_arc_max > 64<<20 && zfs_arc_max < physmem * PAGESIZE) | |
4154 | arc_c_max = zfs_arc_max; | |
91415825 | 4155 | if (zfs_arc_min > 0 && zfs_arc_min <= arc_c_max) |
34dc7c2f BB |
4156 | arc_c_min = zfs_arc_min; |
4157 | ||
4158 | arc_c = arc_c_max; | |
4159 | arc_p = (arc_c >> 1); | |
4160 | ||
2b13331d PS |
4161 | /* limit meta-data to 3/4 of the arc capacity */ |
4162 | arc_meta_limit = (3 * arc_c_max) / 4; | |
1834f2d8 | 4163 | arc_meta_max = 0; |
34dc7c2f BB |
4164 | |
4165 | /* Allow the tunable to override if it is reasonable */ | |
4166 | if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) | |
4167 | arc_meta_limit = zfs_arc_meta_limit; | |
4168 | ||
34dc7c2f BB |
4169 | /* if kmem_flags are set, lets try to use less memory */ |
4170 | if (kmem_debugging()) | |
4171 | arc_c = arc_c / 2; | |
4172 | if (arc_c < arc_c_min) | |
4173 | arc_c = arc_c_min; | |
4174 | ||
4175 | arc_anon = &ARC_anon; | |
4176 | arc_mru = &ARC_mru; | |
4177 | arc_mru_ghost = &ARC_mru_ghost; | |
4178 | arc_mfu = &ARC_mfu; | |
4179 | arc_mfu_ghost = &ARC_mfu_ghost; | |
4180 | arc_l2c_only = &ARC_l2c_only; | |
4181 | arc_size = 0; | |
4182 | ||
4183 | mutex_init(&arc_anon->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); | |
4184 | mutex_init(&arc_mru->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); | |
4185 | mutex_init(&arc_mru_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); | |
4186 | mutex_init(&arc_mfu->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); | |
4187 | mutex_init(&arc_mfu_ghost->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); | |
4188 | mutex_init(&arc_l2c_only->arcs_mtx, NULL, MUTEX_DEFAULT, NULL); | |
4189 | ||
4190 | list_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], | |
4191 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4192 | list_create(&arc_mru->arcs_list[ARC_BUFC_DATA], | |
4193 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4194 | list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], | |
4195 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4196 | list_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], | |
4197 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4198 | list_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], | |
4199 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4200 | list_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], | |
4201 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4202 | list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], | |
4203 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4204 | list_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], | |
4205 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4206 | list_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], | |
4207 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4208 | list_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], | |
4209 | sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); | |
4210 | ||
e0b0ca98 BB |
4211 | arc_anon->arcs_state = ARC_STATE_ANON; |
4212 | arc_mru->arcs_state = ARC_STATE_MRU; | |
4213 | arc_mru_ghost->arcs_state = ARC_STATE_MRU_GHOST; | |
4214 | arc_mfu->arcs_state = ARC_STATE_MFU; | |
4215 | arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST; | |
4216 | arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY; | |
4217 | ||
34dc7c2f BB |
4218 | buf_init(); |
4219 | ||
4220 | arc_thread_exit = 0; | |
ab26409d BB |
4221 | list_create(&arc_prune_list, sizeof (arc_prune_t), |
4222 | offsetof(arc_prune_t, p_node)); | |
34dc7c2f | 4223 | arc_eviction_list = NULL; |
ab26409d | 4224 | mutex_init(&arc_prune_mtx, NULL, MUTEX_DEFAULT, NULL); |
34dc7c2f BB |
4225 | mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); |
4226 | bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); | |
4227 | ||
4228 | arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, | |
4229 | sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); | |
4230 | ||
4231 | if (arc_ksp != NULL) { | |
4232 | arc_ksp->ks_data = &arc_stats; | |
13be560d | 4233 | arc_ksp->ks_update = arc_kstat_update; |
34dc7c2f BB |
4234 | kstat_install(arc_ksp); |
4235 | } | |
4236 | ||
302f753f | 4237 | (void) thread_create(NULL, 0, arc_adapt_thread, NULL, 0, &p0, |
34dc7c2f BB |
4238 | TS_RUN, minclsyspri); |
4239 | ||
4240 | arc_dead = FALSE; | |
b128c09f | 4241 | arc_warm = B_FALSE; |
34dc7c2f | 4242 | |
e8b96c60 MA |
4243 | /* |
4244 | * Calculate maximum amount of dirty data per pool. | |
4245 | * | |
4246 | * If it has been set by a module parameter, take that. | |
4247 | * Otherwise, use a percentage of physical memory defined by | |
4248 | * zfs_dirty_data_max_percent (default 10%) with a cap at | |
4249 | * zfs_dirty_data_max_max (default 25% of physical memory). | |
4250 | */ | |
4251 | if (zfs_dirty_data_max_max == 0) | |
4252 | zfs_dirty_data_max_max = physmem * PAGESIZE * | |
4253 | zfs_dirty_data_max_max_percent / 100; | |
4254 | ||
4255 | if (zfs_dirty_data_max == 0) { | |
4256 | zfs_dirty_data_max = physmem * PAGESIZE * | |
4257 | zfs_dirty_data_max_percent / 100; | |
4258 | zfs_dirty_data_max = MIN(zfs_dirty_data_max, | |
4259 | zfs_dirty_data_max_max); | |
4260 | } | |
34dc7c2f BB |
4261 | } |
4262 | ||
4263 | void | |
4264 | arc_fini(void) | |
4265 | { | |
ab26409d BB |
4266 | arc_prune_t *p; |
4267 | ||
34dc7c2f | 4268 | mutex_enter(&arc_reclaim_thr_lock); |
7cb67b45 BB |
4269 | #ifdef _KERNEL |
4270 | spl_unregister_shrinker(&arc_shrinker); | |
4271 | #endif /* _KERNEL */ | |
4272 | ||
34dc7c2f BB |
4273 | arc_thread_exit = 1; |
4274 | while (arc_thread_exit != 0) | |
4275 | cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); | |
4276 | mutex_exit(&arc_reclaim_thr_lock); | |
4277 | ||
4278 | arc_flush(NULL); | |
4279 | ||
4280 | arc_dead = TRUE; | |
4281 | ||
4282 | if (arc_ksp != NULL) { | |
4283 | kstat_delete(arc_ksp); | |
4284 | arc_ksp = NULL; | |
4285 | } | |
4286 | ||
ab26409d BB |
4287 | mutex_enter(&arc_prune_mtx); |
4288 | while ((p = list_head(&arc_prune_list)) != NULL) { | |
4289 | list_remove(&arc_prune_list, p); | |
4290 | refcount_remove(&p->p_refcnt, &arc_prune_list); | |
4291 | refcount_destroy(&p->p_refcnt); | |
4292 | kmem_free(p, sizeof (*p)); | |
4293 | } | |
4294 | mutex_exit(&arc_prune_mtx); | |
4295 | ||
4296 | list_destroy(&arc_prune_list); | |
4297 | mutex_destroy(&arc_prune_mtx); | |
34dc7c2f BB |
4298 | mutex_destroy(&arc_eviction_mtx); |
4299 | mutex_destroy(&arc_reclaim_thr_lock); | |
4300 | cv_destroy(&arc_reclaim_thr_cv); | |
4301 | ||
4302 | list_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); | |
4303 | list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); | |
4304 | list_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); | |
4305 | list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); | |
4306 | list_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); | |
4307 | list_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); | |
4308 | list_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); | |
4309 | list_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); | |
4310 | ||
4311 | mutex_destroy(&arc_anon->arcs_mtx); | |
4312 | mutex_destroy(&arc_mru->arcs_mtx); | |
4313 | mutex_destroy(&arc_mru_ghost->arcs_mtx); | |
4314 | mutex_destroy(&arc_mfu->arcs_mtx); | |
4315 | mutex_destroy(&arc_mfu_ghost->arcs_mtx); | |
fb5f0bc8 | 4316 | mutex_destroy(&arc_l2c_only->arcs_mtx); |
34dc7c2f BB |
4317 | |
4318 | buf_fini(); | |
9babb374 BB |
4319 | |
4320 | ASSERT(arc_loaned_bytes == 0); | |
34dc7c2f BB |
4321 | } |
4322 | ||
4323 | /* | |
4324 | * Level 2 ARC | |
4325 | * | |
4326 | * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. | |
4327 | * It uses dedicated storage devices to hold cached data, which are populated | |
4328 | * using large infrequent writes. The main role of this cache is to boost | |
4329 | * the performance of random read workloads. The intended L2ARC devices | |
4330 | * include short-stroked disks, solid state disks, and other media with | |
4331 | * substantially faster read latency than disk. | |
4332 | * | |
4333 | * +-----------------------+ | |
4334 | * | ARC | | |
4335 | * +-----------------------+ | |
4336 | * | ^ ^ | |
4337 | * | | | | |
4338 | * l2arc_feed_thread() arc_read() | |
4339 | * | | | | |
4340 | * | l2arc read | | |
4341 | * V | | | |
4342 | * +---------------+ | | |
4343 | * | L2ARC | | | |
4344 | * +---------------+ | | |
4345 | * | ^ | | |
4346 | * l2arc_write() | | | |
4347 | * | | | | |
4348 | * V | | | |
4349 | * +-------+ +-------+ | |
4350 | * | vdev | | vdev | | |
4351 | * | cache | | cache | | |
4352 | * +-------+ +-------+ | |
4353 | * +=========+ .-----. | |
4354 | * : L2ARC : |-_____-| | |
4355 | * : devices : | Disks | | |
4356 | * +=========+ `-_____-' | |
4357 | * | |
4358 | * Read requests are satisfied from the following sources, in order: | |
4359 | * | |
4360 | * 1) ARC | |
4361 | * 2) vdev cache of L2ARC devices | |
4362 | * 3) L2ARC devices | |
4363 | * 4) vdev cache of disks | |
4364 | * 5) disks | |
4365 | * | |
4366 | * Some L2ARC device types exhibit extremely slow write performance. | |
4367 | * To accommodate for this there are some significant differences between | |
4368 | * the L2ARC and traditional cache design: | |
4369 | * | |
4370 | * 1. There is no eviction path from the ARC to the L2ARC. Evictions from | |
4371 | * the ARC behave as usual, freeing buffers and placing headers on ghost | |
4372 | * lists. The ARC does not send buffers to the L2ARC during eviction as | |
4373 | * this would add inflated write latencies for all ARC memory pressure. | |
4374 | * | |
4375 | * 2. The L2ARC attempts to cache data from the ARC before it is evicted. | |
4376 | * It does this by periodically scanning buffers from the eviction-end of | |
4377 | * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are | |
3a17a7a9 SK |
4378 | * not already there. It scans until a headroom of buffers is satisfied, |
4379 | * which itself is a buffer for ARC eviction. If a compressible buffer is | |
4380 | * found during scanning and selected for writing to an L2ARC device, we | |
4381 | * temporarily boost scanning headroom during the next scan cycle to make | |
4382 | * sure we adapt to compression effects (which might significantly reduce | |
4383 | * the data volume we write to L2ARC). The thread that does this is | |
34dc7c2f BB |
4384 | * l2arc_feed_thread(), illustrated below; example sizes are included to |
4385 | * provide a better sense of ratio than this diagram: | |
4386 | * | |
4387 | * head --> tail | |
4388 | * +---------------------+----------+ | |
4389 | * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC | |
4390 | * +---------------------+----------+ | o L2ARC eligible | |
4391 | * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer | |
4392 | * +---------------------+----------+ | | |
4393 | * 15.9 Gbytes ^ 32 Mbytes | | |
4394 | * headroom | | |
4395 | * l2arc_feed_thread() | |
4396 | * | | |
4397 | * l2arc write hand <--[oooo]--' | |
4398 | * | 8 Mbyte | |
4399 | * | write max | |
4400 | * V | |
4401 | * +==============================+ | |
4402 | * L2ARC dev |####|#|###|###| |####| ... | | |
4403 | * +==============================+ | |
4404 | * 32 Gbytes | |
4405 | * | |
4406 | * 3. If an ARC buffer is copied to the L2ARC but then hit instead of | |
4407 | * evicted, then the L2ARC has cached a buffer much sooner than it probably | |
4408 | * needed to, potentially wasting L2ARC device bandwidth and storage. It is | |
4409 | * safe to say that this is an uncommon case, since buffers at the end of | |
4410 | * the ARC lists have moved there due to inactivity. | |
4411 | * | |
4412 | * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, | |
4413 | * then the L2ARC simply misses copying some buffers. This serves as a | |
4414 | * pressure valve to prevent heavy read workloads from both stalling the ARC | |
4415 | * with waits and clogging the L2ARC with writes. This also helps prevent | |
4416 | * the potential for the L2ARC to churn if it attempts to cache content too | |
4417 | * quickly, such as during backups of the entire pool. | |
4418 | * | |
b128c09f BB |
4419 | * 5. After system boot and before the ARC has filled main memory, there are |
4420 | * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru | |
4421 | * lists can remain mostly static. Instead of searching from tail of these | |
4422 | * lists as pictured, the l2arc_feed_thread() will search from the list heads | |
4423 | * for eligible buffers, greatly increasing its chance of finding them. | |
4424 | * | |
4425 | * The L2ARC device write speed is also boosted during this time so that | |
4426 | * the L2ARC warms up faster. Since there have been no ARC evictions yet, | |
4427 | * there are no L2ARC reads, and no fear of degrading read performance | |
4428 | * through increased writes. | |
4429 | * | |
4430 | * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that | |
34dc7c2f BB |
4431 | * the vdev queue can aggregate them into larger and fewer writes. Each |
4432 | * device is written to in a rotor fashion, sweeping writes through | |
4433 | * available space then repeating. | |
4434 | * | |
b128c09f | 4435 | * 7. The L2ARC does not store dirty content. It never needs to flush |
34dc7c2f BB |
4436 | * write buffers back to disk based storage. |
4437 | * | |
b128c09f | 4438 | * 8. If an ARC buffer is written (and dirtied) which also exists in the |
34dc7c2f BB |
4439 | * L2ARC, the now stale L2ARC buffer is immediately dropped. |
4440 | * | |
4441 | * The performance of the L2ARC can be tweaked by a number of tunables, which | |
4442 | * may be necessary for different workloads: | |
4443 | * | |
4444 | * l2arc_write_max max write bytes per interval | |
b128c09f | 4445 | * l2arc_write_boost extra write bytes during device warmup |
34dc7c2f | 4446 | * l2arc_noprefetch skip caching prefetched buffers |
3a17a7a9 | 4447 | * l2arc_nocompress skip compressing buffers |
34dc7c2f | 4448 | * l2arc_headroom number of max device writes to precache |
3a17a7a9 SK |
4449 | * l2arc_headroom_boost when we find compressed buffers during ARC |
4450 | * scanning, we multiply headroom by this | |
4451 | * percentage factor for the next scan cycle, | |
4452 | * since more compressed buffers are likely to | |
4453 | * be present | |
34dc7c2f BB |
4454 | * l2arc_feed_secs seconds between L2ARC writing |
4455 | * | |
4456 | * Tunables may be removed or added as future performance improvements are | |
4457 | * integrated, and also may become zpool properties. | |
d164b209 BB |
4458 | * |
4459 | * There are three key functions that control how the L2ARC warms up: | |
4460 | * | |
4461 | * l2arc_write_eligible() check if a buffer is eligible to cache | |
4462 | * l2arc_write_size() calculate how much to write | |
4463 | * l2arc_write_interval() calculate sleep delay between writes | |
4464 | * | |
4465 | * These three functions determine what to write, how much, and how quickly | |
4466 | * to send writes. | |
34dc7c2f BB |
4467 | */ |
4468 | ||
d164b209 BB |
4469 | static boolean_t |
4470 | l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) | |
4471 | { | |
4472 | /* | |
4473 | * A buffer is *not* eligible for the L2ARC if it: | |
4474 | * 1. belongs to a different spa. | |
428870ff BB |
4475 | * 2. is already cached on the L2ARC. |
4476 | * 3. has an I/O in progress (it may be an incomplete read). | |
4477 | * 4. is flagged not eligible (zfs property). | |
d164b209 | 4478 | */ |
428870ff | 4479 | if (ab->b_spa != spa_guid || ab->b_l2hdr != NULL || |
d164b209 BB |
4480 | HDR_IO_IN_PROGRESS(ab) || !HDR_L2CACHE(ab)) |
4481 | return (B_FALSE); | |
4482 | ||
4483 | return (B_TRUE); | |
4484 | } | |
4485 | ||
4486 | static uint64_t | |
3a17a7a9 | 4487 | l2arc_write_size(void) |
d164b209 BB |
4488 | { |
4489 | uint64_t size; | |
4490 | ||
3a17a7a9 SK |
4491 | /* |
4492 | * Make sure our globals have meaningful values in case the user | |
4493 | * altered them. | |
4494 | */ | |
4495 | size = l2arc_write_max; | |
4496 | if (size == 0) { | |
4497 | cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " | |
4498 | "be greater than zero, resetting it to the default (%d)", | |
4499 | L2ARC_WRITE_SIZE); | |
4500 | size = l2arc_write_max = L2ARC_WRITE_SIZE; | |
4501 | } | |
d164b209 BB |
4502 | |
4503 | if (arc_warm == B_FALSE) | |
3a17a7a9 | 4504 | size += l2arc_write_boost; |
d164b209 BB |
4505 | |
4506 | return (size); | |
4507 | ||
4508 | } | |
4509 | ||
4510 | static clock_t | |
4511 | l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) | |
4512 | { | |
428870ff | 4513 | clock_t interval, next, now; |
d164b209 BB |
4514 | |
4515 | /* | |
4516 | * If the ARC lists are busy, increase our write rate; if the | |
4517 | * lists are stale, idle back. This is achieved by checking | |
4518 | * how much we previously wrote - if it was more than half of | |
4519 | * what we wanted, schedule the next write much sooner. | |
4520 | */ | |
4521 | if (l2arc_feed_again && wrote > (wanted / 2)) | |
4522 | interval = (hz * l2arc_feed_min_ms) / 1000; | |
4523 | else | |
4524 | interval = hz * l2arc_feed_secs; | |
4525 | ||
428870ff BB |
4526 | now = ddi_get_lbolt(); |
4527 | next = MAX(now, MIN(now + interval, began + interval)); | |
d164b209 BB |
4528 | |
4529 | return (next); | |
4530 | } | |
4531 | ||
34dc7c2f BB |
4532 | static void |
4533 | l2arc_hdr_stat_add(void) | |
4534 | { | |
6e1d7276 | 4535 | ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE); |
34dc7c2f BB |
4536 | ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); |
4537 | } | |
4538 | ||
4539 | static void | |
4540 | l2arc_hdr_stat_remove(void) | |
4541 | { | |
6e1d7276 | 4542 | ARCSTAT_INCR(arcstat_l2_hdr_size, -HDR_SIZE); |
34dc7c2f BB |
4543 | ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); |
4544 | } | |
4545 | ||
4546 | /* | |
4547 | * Cycle through L2ARC devices. This is how L2ARC load balances. | |
b128c09f | 4548 | * If a device is returned, this also returns holding the spa config lock. |
34dc7c2f BB |
4549 | */ |
4550 | static l2arc_dev_t * | |
4551 | l2arc_dev_get_next(void) | |
4552 | { | |
b128c09f | 4553 | l2arc_dev_t *first, *next = NULL; |
34dc7c2f | 4554 | |
b128c09f BB |
4555 | /* |
4556 | * Lock out the removal of spas (spa_namespace_lock), then removal | |
4557 | * of cache devices (l2arc_dev_mtx). Once a device has been selected, | |
4558 | * both locks will be dropped and a spa config lock held instead. | |
4559 | */ | |
4560 | mutex_enter(&spa_namespace_lock); | |
4561 | mutex_enter(&l2arc_dev_mtx); | |
4562 | ||
4563 | /* if there are no vdevs, there is nothing to do */ | |
4564 | if (l2arc_ndev == 0) | |
4565 | goto out; | |
4566 | ||
4567 | first = NULL; | |
4568 | next = l2arc_dev_last; | |
4569 | do { | |
4570 | /* loop around the list looking for a non-faulted vdev */ | |
4571 | if (next == NULL) { | |
34dc7c2f | 4572 | next = list_head(l2arc_dev_list); |
b128c09f BB |
4573 | } else { |
4574 | next = list_next(l2arc_dev_list, next); | |
4575 | if (next == NULL) | |
4576 | next = list_head(l2arc_dev_list); | |
4577 | } | |
4578 | ||
4579 | /* if we have come back to the start, bail out */ | |
4580 | if (first == NULL) | |
4581 | first = next; | |
4582 | else if (next == first) | |
4583 | break; | |
4584 | ||
4585 | } while (vdev_is_dead(next->l2ad_vdev)); | |
4586 | ||
4587 | /* if we were unable to find any usable vdevs, return NULL */ | |
4588 | if (vdev_is_dead(next->l2ad_vdev)) | |
4589 | next = NULL; | |
34dc7c2f BB |
4590 | |
4591 | l2arc_dev_last = next; | |
4592 | ||
b128c09f BB |
4593 | out: |
4594 | mutex_exit(&l2arc_dev_mtx); | |
4595 | ||
4596 | /* | |
4597 | * Grab the config lock to prevent the 'next' device from being | |
4598 | * removed while we are writing to it. | |
4599 | */ | |
4600 | if (next != NULL) | |
4601 | spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); | |
4602 | mutex_exit(&spa_namespace_lock); | |
4603 | ||
34dc7c2f BB |
4604 | return (next); |
4605 | } | |
4606 | ||
b128c09f BB |
4607 | /* |
4608 | * Free buffers that were tagged for destruction. | |
4609 | */ | |
4610 | static void | |
0bc8fd78 | 4611 | l2arc_do_free_on_write(void) |
b128c09f BB |
4612 | { |
4613 | list_t *buflist; | |
4614 | l2arc_data_free_t *df, *df_prev; | |
4615 | ||
4616 | mutex_enter(&l2arc_free_on_write_mtx); | |
4617 | buflist = l2arc_free_on_write; | |
4618 | ||
4619 | for (df = list_tail(buflist); df; df = df_prev) { | |
4620 | df_prev = list_prev(buflist, df); | |
4621 | ASSERT(df->l2df_data != NULL); | |
4622 | ASSERT(df->l2df_func != NULL); | |
4623 | df->l2df_func(df->l2df_data, df->l2df_size); | |
4624 | list_remove(buflist, df); | |
4625 | kmem_free(df, sizeof (l2arc_data_free_t)); | |
4626 | } | |
4627 | ||
4628 | mutex_exit(&l2arc_free_on_write_mtx); | |
4629 | } | |
4630 | ||
34dc7c2f BB |
4631 | /* |
4632 | * A write to a cache device has completed. Update all headers to allow | |
4633 | * reads from these buffers to begin. | |
4634 | */ | |
4635 | static void | |
4636 | l2arc_write_done(zio_t *zio) | |
4637 | { | |
4638 | l2arc_write_callback_t *cb; | |
4639 | l2arc_dev_t *dev; | |
4640 | list_t *buflist; | |
34dc7c2f | 4641 | arc_buf_hdr_t *head, *ab, *ab_prev; |
b128c09f | 4642 | l2arc_buf_hdr_t *abl2; |
34dc7c2f | 4643 | kmutex_t *hash_lock; |
3bec585e | 4644 | int64_t bytes_dropped = 0; |
34dc7c2f BB |
4645 | |
4646 | cb = zio->io_private; | |
4647 | ASSERT(cb != NULL); | |
4648 | dev = cb->l2wcb_dev; | |
4649 | ASSERT(dev != NULL); | |
4650 | head = cb->l2wcb_head; | |
4651 | ASSERT(head != NULL); | |
4652 | buflist = dev->l2ad_buflist; | |
4653 | ASSERT(buflist != NULL); | |
4654 | DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, | |
4655 | l2arc_write_callback_t *, cb); | |
4656 | ||
4657 | if (zio->io_error != 0) | |
4658 | ARCSTAT_BUMP(arcstat_l2_writes_error); | |
4659 | ||
4660 | mutex_enter(&l2arc_buflist_mtx); | |
4661 | ||
4662 | /* | |
4663 | * All writes completed, or an error was hit. | |
4664 | */ | |
4665 | for (ab = list_prev(buflist, head); ab; ab = ab_prev) { | |
4666 | ab_prev = list_prev(buflist, ab); | |
1ca546b3 SK |
4667 | abl2 = ab->b_l2hdr; |
4668 | ||
4669 | /* | |
4670 | * Release the temporary compressed buffer as soon as possible. | |
4671 | */ | |
4672 | if (abl2->b_compress != ZIO_COMPRESS_OFF) | |
4673 | l2arc_release_cdata_buf(ab); | |
34dc7c2f BB |
4674 | |
4675 | hash_lock = HDR_LOCK(ab); | |
4676 | if (!mutex_tryenter(hash_lock)) { | |
4677 | /* | |
4678 | * This buffer misses out. It may be in a stage | |
4679 | * of eviction. Its ARC_L2_WRITING flag will be | |
4680 | * left set, denying reads to this buffer. | |
4681 | */ | |
4682 | ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); | |
4683 | continue; | |
4684 | } | |
4685 | ||
4686 | if (zio->io_error != 0) { | |
4687 | /* | |
b128c09f | 4688 | * Error - drop L2ARC entry. |
34dc7c2f | 4689 | */ |
b128c09f | 4690 | list_remove(buflist, ab); |
3a17a7a9 | 4691 | ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); |
3bec585e | 4692 | bytes_dropped += abl2->b_asize; |
34dc7c2f | 4693 | ab->b_l2hdr = NULL; |
ecf3d9b8 | 4694 | kmem_cache_free(l2arc_hdr_cache, abl2); |
6e1d7276 | 4695 | arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS); |
b128c09f | 4696 | ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); |
34dc7c2f BB |
4697 | } |
4698 | ||
4699 | /* | |
4700 | * Allow ARC to begin reads to this L2ARC entry. | |
4701 | */ | |
4702 | ab->b_flags &= ~ARC_L2_WRITING; | |
4703 | ||
4704 | mutex_exit(hash_lock); | |
4705 | } | |
4706 | ||
4707 | atomic_inc_64(&l2arc_writes_done); | |
4708 | list_remove(buflist, head); | |
4709 | kmem_cache_free(hdr_cache, head); | |
4710 | mutex_exit(&l2arc_buflist_mtx); | |
4711 | ||
3bec585e SK |
4712 | vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); |
4713 | ||
b128c09f | 4714 | l2arc_do_free_on_write(); |
34dc7c2f BB |
4715 | |
4716 | kmem_free(cb, sizeof (l2arc_write_callback_t)); | |
4717 | } | |
4718 | ||
4719 | /* | |
4720 | * A read to a cache device completed. Validate buffer contents before | |
4721 | * handing over to the regular ARC routines. | |
4722 | */ | |
4723 | static void | |
4724 | l2arc_read_done(zio_t *zio) | |
4725 | { | |
4726 | l2arc_read_callback_t *cb; | |
4727 | arc_buf_hdr_t *hdr; | |
4728 | arc_buf_t *buf; | |
34dc7c2f | 4729 | kmutex_t *hash_lock; |
b128c09f BB |
4730 | int equal; |
4731 | ||
4732 | ASSERT(zio->io_vd != NULL); | |
4733 | ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); | |
4734 | ||
4735 | spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); | |
34dc7c2f BB |
4736 | |
4737 | cb = zio->io_private; | |
4738 | ASSERT(cb != NULL); | |
4739 | buf = cb->l2rcb_buf; | |
4740 | ASSERT(buf != NULL); | |
34dc7c2f | 4741 | |
428870ff | 4742 | hash_lock = HDR_LOCK(buf->b_hdr); |
34dc7c2f | 4743 | mutex_enter(hash_lock); |
428870ff BB |
4744 | hdr = buf->b_hdr; |
4745 | ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); | |
34dc7c2f | 4746 | |
3a17a7a9 SK |
4747 | /* |
4748 | * If the buffer was compressed, decompress it first. | |
4749 | */ | |
4750 | if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) | |
4751 | l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); | |
4752 | ASSERT(zio->io_data != NULL); | |
4753 | ||
34dc7c2f BB |
4754 | /* |
4755 | * Check this survived the L2ARC journey. | |
4756 | */ | |
4757 | equal = arc_cksum_equal(buf); | |
4758 | if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { | |
4759 | mutex_exit(hash_lock); | |
4760 | zio->io_private = buf; | |
b128c09f BB |
4761 | zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ |
4762 | zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ | |
34dc7c2f BB |
4763 | arc_read_done(zio); |
4764 | } else { | |
4765 | mutex_exit(hash_lock); | |
4766 | /* | |
4767 | * Buffer didn't survive caching. Increment stats and | |
4768 | * reissue to the original storage device. | |
4769 | */ | |
b128c09f | 4770 | if (zio->io_error != 0) { |
34dc7c2f | 4771 | ARCSTAT_BUMP(arcstat_l2_io_error); |
b128c09f | 4772 | } else { |
2e528b49 | 4773 | zio->io_error = SET_ERROR(EIO); |
b128c09f | 4774 | } |
34dc7c2f BB |
4775 | if (!equal) |
4776 | ARCSTAT_BUMP(arcstat_l2_cksum_bad); | |
4777 | ||
34dc7c2f | 4778 | /* |
b128c09f BB |
4779 | * If there's no waiter, issue an async i/o to the primary |
4780 | * storage now. If there *is* a waiter, the caller must | |
4781 | * issue the i/o in a context where it's OK to block. | |
34dc7c2f | 4782 | */ |
d164b209 BB |
4783 | if (zio->io_waiter == NULL) { |
4784 | zio_t *pio = zio_unique_parent(zio); | |
4785 | ||
4786 | ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); | |
4787 | ||
4788 | zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, | |
b128c09f BB |
4789 | buf->b_data, zio->io_size, arc_read_done, buf, |
4790 | zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); | |
d164b209 | 4791 | } |
34dc7c2f BB |
4792 | } |
4793 | ||
4794 | kmem_free(cb, sizeof (l2arc_read_callback_t)); | |
4795 | } | |
4796 | ||
4797 | /* | |
4798 | * This is the list priority from which the L2ARC will search for pages to | |
4799 | * cache. This is used within loops (0..3) to cycle through lists in the | |
4800 | * desired order. This order can have a significant effect on cache | |
4801 | * performance. | |
4802 | * | |
4803 | * Currently the metadata lists are hit first, MFU then MRU, followed by | |
4804 | * the data lists. This function returns a locked list, and also returns | |
4805 | * the lock pointer. | |
4806 | */ | |
4807 | static list_t * | |
4808 | l2arc_list_locked(int list_num, kmutex_t **lock) | |
4809 | { | |
d4ed6673 | 4810 | list_t *list = NULL; |
34dc7c2f BB |
4811 | |
4812 | ASSERT(list_num >= 0 && list_num <= 3); | |
4813 | ||
4814 | switch (list_num) { | |
4815 | case 0: | |
4816 | list = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; | |
4817 | *lock = &arc_mfu->arcs_mtx; | |
4818 | break; | |
4819 | case 1: | |
4820 | list = &arc_mru->arcs_list[ARC_BUFC_METADATA]; | |
4821 | *lock = &arc_mru->arcs_mtx; | |
4822 | break; | |
4823 | case 2: | |
4824 | list = &arc_mfu->arcs_list[ARC_BUFC_DATA]; | |
4825 | *lock = &arc_mfu->arcs_mtx; | |
4826 | break; | |
4827 | case 3: | |
4828 | list = &arc_mru->arcs_list[ARC_BUFC_DATA]; | |
4829 | *lock = &arc_mru->arcs_mtx; | |
4830 | break; | |
4831 | } | |
4832 | ||
4833 | ASSERT(!(MUTEX_HELD(*lock))); | |
4834 | mutex_enter(*lock); | |
4835 | return (list); | |
4836 | } | |
4837 | ||
4838 | /* | |
4839 | * Evict buffers from the device write hand to the distance specified in | |
4840 | * bytes. This distance may span populated buffers, it may span nothing. | |
4841 | * This is clearing a region on the L2ARC device ready for writing. | |
4842 | * If the 'all' boolean is set, every buffer is evicted. | |
4843 | */ | |
4844 | static void | |
4845 | l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) | |
4846 | { | |
4847 | list_t *buflist; | |
4848 | l2arc_buf_hdr_t *abl2; | |
4849 | arc_buf_hdr_t *ab, *ab_prev; | |
4850 | kmutex_t *hash_lock; | |
4851 | uint64_t taddr; | |
3bec585e | 4852 | int64_t bytes_evicted = 0; |
34dc7c2f | 4853 | |
34dc7c2f BB |
4854 | buflist = dev->l2ad_buflist; |
4855 | ||
4856 | if (buflist == NULL) | |
4857 | return; | |
4858 | ||
4859 | if (!all && dev->l2ad_first) { | |
4860 | /* | |
4861 | * This is the first sweep through the device. There is | |
4862 | * nothing to evict. | |
4863 | */ | |
4864 | return; | |
4865 | } | |
4866 | ||
b128c09f | 4867 | if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { |
34dc7c2f BB |
4868 | /* |
4869 | * When nearing the end of the device, evict to the end | |
4870 | * before the device write hand jumps to the start. | |
4871 | */ | |
4872 | taddr = dev->l2ad_end; | |
4873 | } else { | |
4874 | taddr = dev->l2ad_hand + distance; | |
4875 | } | |
4876 | DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, | |
4877 | uint64_t, taddr, boolean_t, all); | |
4878 | ||
4879 | top: | |
4880 | mutex_enter(&l2arc_buflist_mtx); | |
4881 | for (ab = list_tail(buflist); ab; ab = ab_prev) { | |
4882 | ab_prev = list_prev(buflist, ab); | |
4883 | ||
4884 | hash_lock = HDR_LOCK(ab); | |
4885 | if (!mutex_tryenter(hash_lock)) { | |
4886 | /* | |
4887 | * Missed the hash lock. Retry. | |
4888 | */ | |
4889 | ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); | |
4890 | mutex_exit(&l2arc_buflist_mtx); | |
4891 | mutex_enter(hash_lock); | |
4892 | mutex_exit(hash_lock); | |
4893 | goto top; | |
4894 | } | |
4895 | ||
4896 | if (HDR_L2_WRITE_HEAD(ab)) { | |
4897 | /* | |
4898 | * We hit a write head node. Leave it for | |
4899 | * l2arc_write_done(). | |
4900 | */ | |
4901 | list_remove(buflist, ab); | |
4902 | mutex_exit(hash_lock); | |
4903 | continue; | |
4904 | } | |
4905 | ||
4906 | if (!all && ab->b_l2hdr != NULL && | |
4907 | (ab->b_l2hdr->b_daddr > taddr || | |
4908 | ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { | |
4909 | /* | |
4910 | * We've evicted to the target address, | |
4911 | * or the end of the device. | |
4912 | */ | |
4913 | mutex_exit(hash_lock); | |
4914 | break; | |
4915 | } | |
4916 | ||
4917 | if (HDR_FREE_IN_PROGRESS(ab)) { | |
4918 | /* | |
4919 | * Already on the path to destruction. | |
4920 | */ | |
4921 | mutex_exit(hash_lock); | |
4922 | continue; | |
4923 | } | |
4924 | ||
4925 | if (ab->b_state == arc_l2c_only) { | |
4926 | ASSERT(!HDR_L2_READING(ab)); | |
4927 | /* | |
4928 | * This doesn't exist in the ARC. Destroy. | |
4929 | * arc_hdr_destroy() will call list_remove() | |
4930 | * and decrement arcstat_l2_size. | |
4931 | */ | |
4932 | arc_change_state(arc_anon, ab, hash_lock); | |
4933 | arc_hdr_destroy(ab); | |
4934 | } else { | |
b128c09f BB |
4935 | /* |
4936 | * Invalidate issued or about to be issued | |
4937 | * reads, since we may be about to write | |
4938 | * over this location. | |
4939 | */ | |
4940 | if (HDR_L2_READING(ab)) { | |
4941 | ARCSTAT_BUMP(arcstat_l2_evict_reading); | |
4942 | ab->b_flags |= ARC_L2_EVICTED; | |
4943 | } | |
4944 | ||
34dc7c2f BB |
4945 | /* |
4946 | * Tell ARC this no longer exists in L2ARC. | |
4947 | */ | |
4948 | if (ab->b_l2hdr != NULL) { | |
4949 | abl2 = ab->b_l2hdr; | |
3a17a7a9 | 4950 | ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); |
3bec585e | 4951 | bytes_evicted += abl2->b_asize; |
34dc7c2f | 4952 | ab->b_l2hdr = NULL; |
ecf3d9b8 | 4953 | kmem_cache_free(l2arc_hdr_cache, abl2); |
6e1d7276 | 4954 | arc_space_return(L2HDR_SIZE, ARC_SPACE_L2HDRS); |
34dc7c2f BB |
4955 | ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); |
4956 | } | |
4957 | list_remove(buflist, ab); | |
4958 | ||
4959 | /* | |
4960 | * This may have been leftover after a | |
4961 | * failed write. | |
4962 | */ | |
4963 | ab->b_flags &= ~ARC_L2_WRITING; | |
34dc7c2f BB |
4964 | } |
4965 | mutex_exit(hash_lock); | |
4966 | } | |
4967 | mutex_exit(&l2arc_buflist_mtx); | |
4968 | ||
3bec585e | 4969 | vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); |
34dc7c2f BB |
4970 | dev->l2ad_evict = taddr; |
4971 | } | |
4972 | ||
4973 | /* | |
4974 | * Find and write ARC buffers to the L2ARC device. | |
4975 | * | |
4976 | * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid | |
4977 | * for reading until they have completed writing. | |
3a17a7a9 SK |
4978 | * The headroom_boost is an in-out parameter used to maintain headroom boost |
4979 | * state between calls to this function. | |
4980 | * | |
4981 | * Returns the number of bytes actually written (which may be smaller than | |
4982 | * the delta by which the device hand has changed due to alignment). | |
34dc7c2f | 4983 | */ |
d164b209 | 4984 | static uint64_t |
3a17a7a9 SK |
4985 | l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, |
4986 | boolean_t *headroom_boost) | |
34dc7c2f BB |
4987 | { |
4988 | arc_buf_hdr_t *ab, *ab_prev, *head; | |
34dc7c2f | 4989 | list_t *list; |
3a17a7a9 SK |
4990 | uint64_t write_asize, write_psize, write_sz, headroom, |
4991 | buf_compress_minsz; | |
34dc7c2f | 4992 | void *buf_data; |
3a17a7a9 SK |
4993 | kmutex_t *list_lock = NULL; |
4994 | boolean_t full; | |
34dc7c2f BB |
4995 | l2arc_write_callback_t *cb; |
4996 | zio_t *pio, *wzio; | |
3541dc6d | 4997 | uint64_t guid = spa_load_guid(spa); |
d6320ddb | 4998 | int try; |
3a17a7a9 | 4999 | const boolean_t do_headroom_boost = *headroom_boost; |
34dc7c2f | 5000 | |
34dc7c2f BB |
5001 | ASSERT(dev->l2ad_vdev != NULL); |
5002 | ||
3a17a7a9 SK |
5003 | /* Lower the flag now, we might want to raise it again later. */ |
5004 | *headroom_boost = B_FALSE; | |
5005 | ||
34dc7c2f | 5006 | pio = NULL; |
3a17a7a9 | 5007 | write_sz = write_asize = write_psize = 0; |
34dc7c2f BB |
5008 | full = B_FALSE; |
5009 | head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); | |
5010 | head->b_flags |= ARC_L2_WRITE_HEAD; | |
5011 | ||
3a17a7a9 SK |
5012 | /* |
5013 | * We will want to try to compress buffers that are at least 2x the | |
5014 | * device sector size. | |
5015 | */ | |
5016 | buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; | |
5017 | ||
34dc7c2f BB |
5018 | /* |
5019 | * Copy buffers for L2ARC writing. | |
5020 | */ | |
5021 | mutex_enter(&l2arc_buflist_mtx); | |
d6320ddb | 5022 | for (try = 0; try <= 3; try++) { |
3a17a7a9 SK |
5023 | uint64_t passed_sz = 0; |
5024 | ||
34dc7c2f | 5025 | list = l2arc_list_locked(try, &list_lock); |
34dc7c2f | 5026 | |
b128c09f BB |
5027 | /* |
5028 | * L2ARC fast warmup. | |
5029 | * | |
5030 | * Until the ARC is warm and starts to evict, read from the | |
5031 | * head of the ARC lists rather than the tail. | |
5032 | */ | |
b128c09f BB |
5033 | if (arc_warm == B_FALSE) |
5034 | ab = list_head(list); | |
5035 | else | |
5036 | ab = list_tail(list); | |
5037 | ||
3a17a7a9 SK |
5038 | headroom = target_sz * l2arc_headroom; |
5039 | if (do_headroom_boost) | |
5040 | headroom = (headroom * l2arc_headroom_boost) / 100; | |
5041 | ||
b128c09f | 5042 | for (; ab; ab = ab_prev) { |
3a17a7a9 SK |
5043 | l2arc_buf_hdr_t *l2hdr; |
5044 | kmutex_t *hash_lock; | |
5045 | uint64_t buf_sz; | |
5046 | ||
b128c09f BB |
5047 | if (arc_warm == B_FALSE) |
5048 | ab_prev = list_next(list, ab); | |
5049 | else | |
5050 | ab_prev = list_prev(list, ab); | |
34dc7c2f BB |
5051 | |
5052 | hash_lock = HDR_LOCK(ab); | |
3a17a7a9 | 5053 | if (!mutex_tryenter(hash_lock)) { |
34dc7c2f BB |
5054 | /* |
5055 | * Skip this buffer rather than waiting. | |
5056 | */ | |
5057 | continue; | |
5058 | } | |
5059 | ||
5060 | passed_sz += ab->b_size; | |
5061 | if (passed_sz > headroom) { | |
5062 | /* | |
5063 | * Searched too far. | |
5064 | */ | |
5065 | mutex_exit(hash_lock); | |
5066 | break; | |
5067 | } | |
5068 | ||
d164b209 | 5069 | if (!l2arc_write_eligible(guid, ab)) { |
34dc7c2f BB |
5070 | mutex_exit(hash_lock); |
5071 | continue; | |
5072 | } | |
5073 | ||
5074 | if ((write_sz + ab->b_size) > target_sz) { | |
5075 | full = B_TRUE; | |
5076 | mutex_exit(hash_lock); | |
5077 | break; | |
5078 | } | |
5079 | ||
34dc7c2f BB |
5080 | if (pio == NULL) { |
5081 | /* | |
5082 | * Insert a dummy header on the buflist so | |
5083 | * l2arc_write_done() can find where the | |
5084 | * write buffers begin without searching. | |
5085 | */ | |
5086 | list_insert_head(dev->l2ad_buflist, head); | |
5087 | ||
409dc1a5 | 5088 | cb = kmem_alloc(sizeof (l2arc_write_callback_t), |
d1d7e268 | 5089 | KM_PUSHPAGE); |
34dc7c2f BB |
5090 | cb->l2wcb_dev = dev; |
5091 | cb->l2wcb_head = head; | |
5092 | pio = zio_root(spa, l2arc_write_done, cb, | |
5093 | ZIO_FLAG_CANFAIL); | |
5094 | } | |
5095 | ||
5096 | /* | |
5097 | * Create and add a new L2ARC header. | |
5098 | */ | |
ecf3d9b8 | 5099 | l2hdr = kmem_cache_alloc(l2arc_hdr_cache, KM_PUSHPAGE); |
3a17a7a9 | 5100 | l2hdr->b_dev = dev; |
ecf3d9b8 | 5101 | l2hdr->b_daddr = 0; |
6e1d7276 | 5102 | arc_space_consume(L2HDR_SIZE, ARC_SPACE_L2HDRS); |
34dc7c2f BB |
5103 | |
5104 | ab->b_flags |= ARC_L2_WRITING; | |
3a17a7a9 SK |
5105 | |
5106 | /* | |
5107 | * Temporarily stash the data buffer in b_tmp_cdata. | |
5108 | * The subsequent write step will pick it up from | |
5109 | * there. This is because can't access ab->b_buf | |
5110 | * without holding the hash_lock, which we in turn | |
5111 | * can't access without holding the ARC list locks | |
5112 | * (which we want to avoid during compression/writing) | |
5113 | */ | |
5114 | l2hdr->b_compress = ZIO_COMPRESS_OFF; | |
5115 | l2hdr->b_asize = ab->b_size; | |
5116 | l2hdr->b_tmp_cdata = ab->b_buf->b_data; | |
e0b0ca98 | 5117 | l2hdr->b_hits = 0; |
3a17a7a9 | 5118 | |
34dc7c2f | 5119 | buf_sz = ab->b_size; |
3a17a7a9 SK |
5120 | ab->b_l2hdr = l2hdr; |
5121 | ||
5122 | list_insert_head(dev->l2ad_buflist, ab); | |
34dc7c2f BB |
5123 | |
5124 | /* | |
5125 | * Compute and store the buffer cksum before | |
5126 | * writing. On debug the cksum is verified first. | |
5127 | */ | |
5128 | arc_cksum_verify(ab->b_buf); | |
5129 | arc_cksum_compute(ab->b_buf, B_TRUE); | |
5130 | ||
5131 | mutex_exit(hash_lock); | |
5132 | ||
3a17a7a9 SK |
5133 | write_sz += buf_sz; |
5134 | } | |
5135 | ||
5136 | mutex_exit(list_lock); | |
5137 | ||
5138 | if (full == B_TRUE) | |
5139 | break; | |
5140 | } | |
5141 | ||
5142 | /* No buffers selected for writing? */ | |
5143 | if (pio == NULL) { | |
5144 | ASSERT0(write_sz); | |
5145 | mutex_exit(&l2arc_buflist_mtx); | |
5146 | kmem_cache_free(hdr_cache, head); | |
5147 | return (0); | |
5148 | } | |
5149 | ||
5150 | /* | |
5151 | * Now start writing the buffers. We're starting at the write head | |
5152 | * and work backwards, retracing the course of the buffer selector | |
5153 | * loop above. | |
5154 | */ | |
5155 | for (ab = list_prev(dev->l2ad_buflist, head); ab; | |
5156 | ab = list_prev(dev->l2ad_buflist, ab)) { | |
5157 | l2arc_buf_hdr_t *l2hdr; | |
5158 | uint64_t buf_sz; | |
5159 | ||
5160 | /* | |
5161 | * We shouldn't need to lock the buffer here, since we flagged | |
5162 | * it as ARC_L2_WRITING in the previous step, but we must take | |
5163 | * care to only access its L2 cache parameters. In particular, | |
5164 | * ab->b_buf may be invalid by now due to ARC eviction. | |
5165 | */ | |
5166 | l2hdr = ab->b_l2hdr; | |
5167 | l2hdr->b_daddr = dev->l2ad_hand; | |
5168 | ||
5169 | if (!l2arc_nocompress && (ab->b_flags & ARC_L2COMPRESS) && | |
5170 | l2hdr->b_asize >= buf_compress_minsz) { | |
5171 | if (l2arc_compress_buf(l2hdr)) { | |
5172 | /* | |
5173 | * If compression succeeded, enable headroom | |
5174 | * boost on the next scan cycle. | |
5175 | */ | |
5176 | *headroom_boost = B_TRUE; | |
5177 | } | |
5178 | } | |
5179 | ||
5180 | /* | |
5181 | * Pick up the buffer data we had previously stashed away | |
5182 | * (and now potentially also compressed). | |
5183 | */ | |
5184 | buf_data = l2hdr->b_tmp_cdata; | |
5185 | buf_sz = l2hdr->b_asize; | |
5186 | ||
5187 | /* Compression may have squashed the buffer to zero length. */ | |
5188 | if (buf_sz != 0) { | |
5189 | uint64_t buf_p_sz; | |
5190 | ||
34dc7c2f BB |
5191 | wzio = zio_write_phys(pio, dev->l2ad_vdev, |
5192 | dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, | |
5193 | NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, | |
5194 | ZIO_FLAG_CANFAIL, B_FALSE); | |
5195 | ||
5196 | DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, | |
5197 | zio_t *, wzio); | |
5198 | (void) zio_nowait(wzio); | |
5199 | ||
3a17a7a9 | 5200 | write_asize += buf_sz; |
b128c09f BB |
5201 | /* |
5202 | * Keep the clock hand suitably device-aligned. | |
5203 | */ | |
3a17a7a9 SK |
5204 | buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); |
5205 | write_psize += buf_p_sz; | |
5206 | dev->l2ad_hand += buf_p_sz; | |
34dc7c2f | 5207 | } |
34dc7c2f | 5208 | } |
34dc7c2f | 5209 | |
3a17a7a9 | 5210 | mutex_exit(&l2arc_buflist_mtx); |
34dc7c2f | 5211 | |
3a17a7a9 | 5212 | ASSERT3U(write_asize, <=, target_sz); |
34dc7c2f | 5213 | ARCSTAT_BUMP(arcstat_l2_writes_sent); |
3a17a7a9 | 5214 | ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); |
34dc7c2f | 5215 | ARCSTAT_INCR(arcstat_l2_size, write_sz); |
3a17a7a9 | 5216 | ARCSTAT_INCR(arcstat_l2_asize, write_asize); |
3bec585e | 5217 | vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); |
34dc7c2f BB |
5218 | |
5219 | /* | |
5220 | * Bump device hand to the device start if it is approaching the end. | |
5221 | * l2arc_evict() will already have evicted ahead for this case. | |
5222 | */ | |
b128c09f | 5223 | if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { |
34dc7c2f BB |
5224 | dev->l2ad_hand = dev->l2ad_start; |
5225 | dev->l2ad_evict = dev->l2ad_start; | |
5226 | dev->l2ad_first = B_FALSE; | |
5227 | } | |
5228 | ||
d164b209 | 5229 | dev->l2ad_writing = B_TRUE; |
34dc7c2f | 5230 | (void) zio_wait(pio); |
d164b209 BB |
5231 | dev->l2ad_writing = B_FALSE; |
5232 | ||
3a17a7a9 SK |
5233 | return (write_asize); |
5234 | } | |
5235 | ||
5236 | /* | |
5237 | * Compresses an L2ARC buffer. | |
5238 | * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its | |
5239 | * size in l2hdr->b_asize. This routine tries to compress the data and | |
5240 | * depending on the compression result there are three possible outcomes: | |
5241 | * *) The buffer was incompressible. The original l2hdr contents were left | |
5242 | * untouched and are ready for writing to an L2 device. | |
5243 | * *) The buffer was all-zeros, so there is no need to write it to an L2 | |
5244 | * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is | |
5245 | * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. | |
5246 | * *) Compression succeeded and b_tmp_cdata was replaced with a temporary | |
5247 | * data buffer which holds the compressed data to be written, and b_asize | |
5248 | * tells us how much data there is. b_compress is set to the appropriate | |
5249 | * compression algorithm. Once writing is done, invoke | |
5250 | * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. | |
5251 | * | |
5252 | * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the | |
5253 | * buffer was incompressible). | |
5254 | */ | |
5255 | static boolean_t | |
5256 | l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) | |
5257 | { | |
5258 | void *cdata; | |
9b67f605 | 5259 | size_t csize, len, rounded; |
3a17a7a9 SK |
5260 | |
5261 | ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); | |
5262 | ASSERT(l2hdr->b_tmp_cdata != NULL); | |
5263 | ||
5264 | len = l2hdr->b_asize; | |
5265 | cdata = zio_data_buf_alloc(len); | |
5266 | csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, | |
5267 | cdata, l2hdr->b_asize); | |
5268 | ||
9b67f605 MA |
5269 | rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); |
5270 | if (rounded > csize) { | |
5271 | bzero((char *)cdata + csize, rounded - csize); | |
5272 | csize = rounded; | |
5273 | } | |
5274 | ||
3a17a7a9 SK |
5275 | if (csize == 0) { |
5276 | /* zero block, indicate that there's nothing to write */ | |
5277 | zio_data_buf_free(cdata, len); | |
5278 | l2hdr->b_compress = ZIO_COMPRESS_EMPTY; | |
5279 | l2hdr->b_asize = 0; | |
5280 | l2hdr->b_tmp_cdata = NULL; | |
5281 | ARCSTAT_BUMP(arcstat_l2_compress_zeros); | |
5282 | return (B_TRUE); | |
5283 | } else if (csize > 0 && csize < len) { | |
5284 | /* | |
5285 | * Compression succeeded, we'll keep the cdata around for | |
5286 | * writing and release it afterwards. | |
5287 | */ | |
5288 | l2hdr->b_compress = ZIO_COMPRESS_LZ4; | |
5289 | l2hdr->b_asize = csize; | |
5290 | l2hdr->b_tmp_cdata = cdata; | |
5291 | ARCSTAT_BUMP(arcstat_l2_compress_successes); | |
5292 | return (B_TRUE); | |
5293 | } else { | |
5294 | /* | |
5295 | * Compression failed, release the compressed buffer. | |
5296 | * l2hdr will be left unmodified. | |
5297 | */ | |
5298 | zio_data_buf_free(cdata, len); | |
5299 | ARCSTAT_BUMP(arcstat_l2_compress_failures); | |
5300 | return (B_FALSE); | |
5301 | } | |
5302 | } | |
5303 | ||
5304 | /* | |
5305 | * Decompresses a zio read back from an l2arc device. On success, the | |
5306 | * underlying zio's io_data buffer is overwritten by the uncompressed | |
5307 | * version. On decompression error (corrupt compressed stream), the | |
5308 | * zio->io_error value is set to signal an I/O error. | |
5309 | * | |
5310 | * Please note that the compressed data stream is not checksummed, so | |
5311 | * if the underlying device is experiencing data corruption, we may feed | |
5312 | * corrupt data to the decompressor, so the decompressor needs to be | |
5313 | * able to handle this situation (LZ4 does). | |
5314 | */ | |
5315 | static void | |
5316 | l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) | |
5317 | { | |
5318 | uint64_t csize; | |
5319 | void *cdata; | |
5320 | ||
5321 | ASSERT(L2ARC_IS_VALID_COMPRESS(c)); | |
5322 | ||
5323 | if (zio->io_error != 0) { | |
5324 | /* | |
5325 | * An io error has occured, just restore the original io | |
5326 | * size in preparation for a main pool read. | |
5327 | */ | |
5328 | zio->io_orig_size = zio->io_size = hdr->b_size; | |
5329 | return; | |
5330 | } | |
5331 | ||
5332 | if (c == ZIO_COMPRESS_EMPTY) { | |
5333 | /* | |
5334 | * An empty buffer results in a null zio, which means we | |
5335 | * need to fill its io_data after we're done restoring the | |
5336 | * buffer's contents. | |
5337 | */ | |
5338 | ASSERT(hdr->b_buf != NULL); | |
5339 | bzero(hdr->b_buf->b_data, hdr->b_size); | |
5340 | zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; | |
5341 | } else { | |
5342 | ASSERT(zio->io_data != NULL); | |
5343 | /* | |
5344 | * We copy the compressed data from the start of the arc buffer | |
5345 | * (the zio_read will have pulled in only what we need, the | |
5346 | * rest is garbage which we will overwrite at decompression) | |
5347 | * and then decompress back to the ARC data buffer. This way we | |
5348 | * can minimize copying by simply decompressing back over the | |
5349 | * original compressed data (rather than decompressing to an | |
5350 | * aux buffer and then copying back the uncompressed buffer, | |
5351 | * which is likely to be much larger). | |
5352 | */ | |
5353 | csize = zio->io_size; | |
5354 | cdata = zio_data_buf_alloc(csize); | |
5355 | bcopy(zio->io_data, cdata, csize); | |
5356 | if (zio_decompress_data(c, cdata, zio->io_data, csize, | |
5357 | hdr->b_size) != 0) | |
2e528b49 | 5358 | zio->io_error = SET_ERROR(EIO); |
3a17a7a9 SK |
5359 | zio_data_buf_free(cdata, csize); |
5360 | } | |
5361 | ||
5362 | /* Restore the expected uncompressed IO size. */ | |
5363 | zio->io_orig_size = zio->io_size = hdr->b_size; | |
5364 | } | |
5365 | ||
5366 | /* | |
5367 | * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. | |
5368 | * This buffer serves as a temporary holder of compressed data while | |
5369 | * the buffer entry is being written to an l2arc device. Once that is | |
5370 | * done, we can dispose of it. | |
5371 | */ | |
5372 | static void | |
5373 | l2arc_release_cdata_buf(arc_buf_hdr_t *ab) | |
5374 | { | |
5375 | l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; | |
5376 | ||
5377 | if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { | |
5378 | /* | |
5379 | * If the data was compressed, then we've allocated a | |
5380 | * temporary buffer for it, so now we need to release it. | |
5381 | */ | |
5382 | ASSERT(l2hdr->b_tmp_cdata != NULL); | |
5383 | zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); | |
5384 | } | |
5385 | l2hdr->b_tmp_cdata = NULL; | |
34dc7c2f BB |
5386 | } |
5387 | ||
5388 | /* | |
5389 | * This thread feeds the L2ARC at regular intervals. This is the beating | |
5390 | * heart of the L2ARC. | |
5391 | */ | |
5392 | static void | |
5393 | l2arc_feed_thread(void) | |
5394 | { | |
5395 | callb_cpr_t cpr; | |
5396 | l2arc_dev_t *dev; | |
5397 | spa_t *spa; | |
d164b209 | 5398 | uint64_t size, wrote; |
428870ff | 5399 | clock_t begin, next = ddi_get_lbolt(); |
3a17a7a9 | 5400 | boolean_t headroom_boost = B_FALSE; |
34dc7c2f BB |
5401 | |
5402 | CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); | |
5403 | ||
5404 | mutex_enter(&l2arc_feed_thr_lock); | |
5405 | ||
5406 | while (l2arc_thread_exit == 0) { | |
34dc7c2f | 5407 | CALLB_CPR_SAFE_BEGIN(&cpr); |
5b63b3eb BB |
5408 | (void) cv_timedwait_interruptible(&l2arc_feed_thr_cv, |
5409 | &l2arc_feed_thr_lock, next); | |
34dc7c2f | 5410 | CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); |
428870ff | 5411 | next = ddi_get_lbolt() + hz; |
34dc7c2f BB |
5412 | |
5413 | /* | |
b128c09f | 5414 | * Quick check for L2ARC devices. |
34dc7c2f BB |
5415 | */ |
5416 | mutex_enter(&l2arc_dev_mtx); | |
5417 | if (l2arc_ndev == 0) { | |
5418 | mutex_exit(&l2arc_dev_mtx); | |
5419 | continue; | |
5420 | } | |
b128c09f | 5421 | mutex_exit(&l2arc_dev_mtx); |
428870ff | 5422 | begin = ddi_get_lbolt(); |
34dc7c2f BB |
5423 | |
5424 | /* | |
b128c09f BB |
5425 | * This selects the next l2arc device to write to, and in |
5426 | * doing so the next spa to feed from: dev->l2ad_spa. This | |
5427 | * will return NULL if there are now no l2arc devices or if | |
5428 | * they are all faulted. | |
5429 | * | |
5430 | * If a device is returned, its spa's config lock is also | |
5431 | * held to prevent device removal. l2arc_dev_get_next() | |
5432 | * will grab and release l2arc_dev_mtx. | |
34dc7c2f | 5433 | */ |
b128c09f | 5434 | if ((dev = l2arc_dev_get_next()) == NULL) |
34dc7c2f | 5435 | continue; |
b128c09f BB |
5436 | |
5437 | spa = dev->l2ad_spa; | |
5438 | ASSERT(spa != NULL); | |
34dc7c2f | 5439 | |
572e2857 BB |
5440 | /* |
5441 | * If the pool is read-only then force the feed thread to | |
5442 | * sleep a little longer. | |
5443 | */ | |
5444 | if (!spa_writeable(spa)) { | |
5445 | next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; | |
5446 | spa_config_exit(spa, SCL_L2ARC, dev); | |
5447 | continue; | |
5448 | } | |
5449 | ||
34dc7c2f | 5450 | /* |
b128c09f | 5451 | * Avoid contributing to memory pressure. |
34dc7c2f | 5452 | */ |
302f753f | 5453 | if (arc_no_grow) { |
b128c09f BB |
5454 | ARCSTAT_BUMP(arcstat_l2_abort_lowmem); |
5455 | spa_config_exit(spa, SCL_L2ARC, dev); | |
34dc7c2f BB |
5456 | continue; |
5457 | } | |
b128c09f | 5458 | |
34dc7c2f BB |
5459 | ARCSTAT_BUMP(arcstat_l2_feeds); |
5460 | ||
3a17a7a9 | 5461 | size = l2arc_write_size(); |
b128c09f | 5462 | |
34dc7c2f BB |
5463 | /* |
5464 | * Evict L2ARC buffers that will be overwritten. | |
5465 | */ | |
b128c09f | 5466 | l2arc_evict(dev, size, B_FALSE); |
34dc7c2f BB |
5467 | |
5468 | /* | |
5469 | * Write ARC buffers. | |
5470 | */ | |
3a17a7a9 | 5471 | wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); |
d164b209 BB |
5472 | |
5473 | /* | |
5474 | * Calculate interval between writes. | |
5475 | */ | |
5476 | next = l2arc_write_interval(begin, size, wrote); | |
b128c09f | 5477 | spa_config_exit(spa, SCL_L2ARC, dev); |
34dc7c2f BB |
5478 | } |
5479 | ||
5480 | l2arc_thread_exit = 0; | |
5481 | cv_broadcast(&l2arc_feed_thr_cv); | |
5482 | CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ | |
5483 | thread_exit(); | |
5484 | } | |
5485 | ||
b128c09f BB |
5486 | boolean_t |
5487 | l2arc_vdev_present(vdev_t *vd) | |
5488 | { | |
5489 | l2arc_dev_t *dev; | |
5490 | ||
5491 | mutex_enter(&l2arc_dev_mtx); | |
5492 | for (dev = list_head(l2arc_dev_list); dev != NULL; | |
5493 | dev = list_next(l2arc_dev_list, dev)) { | |
5494 | if (dev->l2ad_vdev == vd) | |
5495 | break; | |
5496 | } | |
5497 | mutex_exit(&l2arc_dev_mtx); | |
5498 | ||
5499 | return (dev != NULL); | |
5500 | } | |
5501 | ||
34dc7c2f BB |
5502 | /* |
5503 | * Add a vdev for use by the L2ARC. By this point the spa has already | |
5504 | * validated the vdev and opened it. | |
5505 | */ | |
5506 | void | |
9babb374 | 5507 | l2arc_add_vdev(spa_t *spa, vdev_t *vd) |
34dc7c2f BB |
5508 | { |
5509 | l2arc_dev_t *adddev; | |
5510 | ||
b128c09f BB |
5511 | ASSERT(!l2arc_vdev_present(vd)); |
5512 | ||
34dc7c2f BB |
5513 | /* |
5514 | * Create a new l2arc device entry. | |
5515 | */ | |
5516 | adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); | |
5517 | adddev->l2ad_spa = spa; | |
5518 | adddev->l2ad_vdev = vd; | |
9babb374 BB |
5519 | adddev->l2ad_start = VDEV_LABEL_START_SIZE; |
5520 | adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); | |
34dc7c2f BB |
5521 | adddev->l2ad_hand = adddev->l2ad_start; |
5522 | adddev->l2ad_evict = adddev->l2ad_start; | |
5523 | adddev->l2ad_first = B_TRUE; | |
d164b209 | 5524 | adddev->l2ad_writing = B_FALSE; |
98f72a53 | 5525 | list_link_init(&adddev->l2ad_node); |
34dc7c2f BB |
5526 | |
5527 | /* | |
5528 | * This is a list of all ARC buffers that are still valid on the | |
5529 | * device. | |
5530 | */ | |
5531 | adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); | |
5532 | list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), | |
5533 | offsetof(arc_buf_hdr_t, b_l2node)); | |
5534 | ||
428870ff | 5535 | vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); |
34dc7c2f BB |
5536 | |
5537 | /* | |
5538 | * Add device to global list | |
5539 | */ | |
5540 | mutex_enter(&l2arc_dev_mtx); | |
5541 | list_insert_head(l2arc_dev_list, adddev); | |
5542 | atomic_inc_64(&l2arc_ndev); | |
5543 | mutex_exit(&l2arc_dev_mtx); | |
5544 | } | |
5545 | ||
5546 | /* | |
5547 | * Remove a vdev from the L2ARC. | |
5548 | */ | |
5549 | void | |
5550 | l2arc_remove_vdev(vdev_t *vd) | |
5551 | { | |
5552 | l2arc_dev_t *dev, *nextdev, *remdev = NULL; | |
5553 | ||
34dc7c2f BB |
5554 | /* |
5555 | * Find the device by vdev | |
5556 | */ | |
5557 | mutex_enter(&l2arc_dev_mtx); | |
5558 | for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { | |
5559 | nextdev = list_next(l2arc_dev_list, dev); | |
5560 | if (vd == dev->l2ad_vdev) { | |
5561 | remdev = dev; | |
5562 | break; | |
5563 | } | |
5564 | } | |
5565 | ASSERT(remdev != NULL); | |
5566 | ||
5567 | /* | |
5568 | * Remove device from global list | |
5569 | */ | |
5570 | list_remove(l2arc_dev_list, remdev); | |
5571 | l2arc_dev_last = NULL; /* may have been invalidated */ | |
b128c09f BB |
5572 | atomic_dec_64(&l2arc_ndev); |
5573 | mutex_exit(&l2arc_dev_mtx); | |
34dc7c2f BB |
5574 | |
5575 | /* | |
5576 | * Clear all buflists and ARC references. L2ARC device flush. | |
5577 | */ | |
5578 | l2arc_evict(remdev, 0, B_TRUE); | |
5579 | list_destroy(remdev->l2ad_buflist); | |
5580 | kmem_free(remdev->l2ad_buflist, sizeof (list_t)); | |
5581 | kmem_free(remdev, sizeof (l2arc_dev_t)); | |
34dc7c2f BB |
5582 | } |
5583 | ||
5584 | void | |
b128c09f | 5585 | l2arc_init(void) |
34dc7c2f BB |
5586 | { |
5587 | l2arc_thread_exit = 0; | |
5588 | l2arc_ndev = 0; | |
5589 | l2arc_writes_sent = 0; | |
5590 | l2arc_writes_done = 0; | |
5591 | ||
5592 | mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); | |
5593 | cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); | |
5594 | mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); | |
5595 | mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); | |
5596 | mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); | |
5597 | ||
5598 | l2arc_dev_list = &L2ARC_dev_list; | |
5599 | l2arc_free_on_write = &L2ARC_free_on_write; | |
5600 | list_create(l2arc_dev_list, sizeof (l2arc_dev_t), | |
5601 | offsetof(l2arc_dev_t, l2ad_node)); | |
5602 | list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), | |
5603 | offsetof(l2arc_data_free_t, l2df_list_node)); | |
34dc7c2f BB |
5604 | } |
5605 | ||
5606 | void | |
b128c09f | 5607 | l2arc_fini(void) |
34dc7c2f | 5608 | { |
b128c09f BB |
5609 | /* |
5610 | * This is called from dmu_fini(), which is called from spa_fini(); | |
5611 | * Because of this, we can assume that all l2arc devices have | |
5612 | * already been removed when the pools themselves were removed. | |
5613 | */ | |
5614 | ||
5615 | l2arc_do_free_on_write(); | |
34dc7c2f BB |
5616 | |
5617 | mutex_destroy(&l2arc_feed_thr_lock); | |
5618 | cv_destroy(&l2arc_feed_thr_cv); | |
5619 | mutex_destroy(&l2arc_dev_mtx); | |
5620 | mutex_destroy(&l2arc_buflist_mtx); | |
5621 | mutex_destroy(&l2arc_free_on_write_mtx); | |
5622 | ||
5623 | list_destroy(l2arc_dev_list); | |
5624 | list_destroy(l2arc_free_on_write); | |
5625 | } | |
b128c09f BB |
5626 | |
5627 | void | |
5628 | l2arc_start(void) | |
5629 | { | |
fb5f0bc8 | 5630 | if (!(spa_mode_global & FWRITE)) |
b128c09f BB |
5631 | return; |
5632 | ||
5633 | (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, | |
5634 | TS_RUN, minclsyspri); | |
5635 | } | |
5636 | ||
5637 | void | |
5638 | l2arc_stop(void) | |
5639 | { | |
fb5f0bc8 | 5640 | if (!(spa_mode_global & FWRITE)) |
b128c09f BB |
5641 | return; |
5642 | ||
5643 | mutex_enter(&l2arc_feed_thr_lock); | |
5644 | cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ | |
5645 | l2arc_thread_exit = 1; | |
5646 | while (l2arc_thread_exit != 0) | |
5647 | cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); | |
5648 | mutex_exit(&l2arc_feed_thr_lock); | |
5649 | } | |
c28b2279 BB |
5650 | |
5651 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
5652 | EXPORT_SYMBOL(arc_read); | |
5653 | EXPORT_SYMBOL(arc_buf_remove_ref); | |
e0b0ca98 | 5654 | EXPORT_SYMBOL(arc_buf_info); |
c28b2279 | 5655 | EXPORT_SYMBOL(arc_getbuf_func); |
ab26409d BB |
5656 | EXPORT_SYMBOL(arc_add_prune_callback); |
5657 | EXPORT_SYMBOL(arc_remove_prune_callback); | |
c28b2279 | 5658 | |
bce45ec9 | 5659 | module_param(zfs_arc_min, ulong, 0644); |
c409e464 | 5660 | MODULE_PARM_DESC(zfs_arc_min, "Min arc size"); |
c28b2279 | 5661 | |
bce45ec9 | 5662 | module_param(zfs_arc_max, ulong, 0644); |
c409e464 | 5663 | MODULE_PARM_DESC(zfs_arc_max, "Max arc size"); |
c28b2279 | 5664 | |
bce45ec9 | 5665 | module_param(zfs_arc_meta_limit, ulong, 0644); |
c28b2279 | 5666 | MODULE_PARM_DESC(zfs_arc_meta_limit, "Meta limit for arc size"); |
6a8f9b6b | 5667 | |
bce45ec9 | 5668 | module_param(zfs_arc_meta_prune, int, 0644); |
ab26409d | 5669 | MODULE_PARM_DESC(zfs_arc_meta_prune, "Bytes of meta data to prune"); |
c409e464 | 5670 | |
bce45ec9 | 5671 | module_param(zfs_arc_grow_retry, int, 0644); |
c409e464 BB |
5672 | MODULE_PARM_DESC(zfs_arc_grow_retry, "Seconds before growing arc size"); |
5673 | ||
89c8cac4 PS |
5674 | module_param(zfs_arc_p_aggressive_disable, int, 0644); |
5675 | MODULE_PARM_DESC(zfs_arc_p_aggressive_disable, "disable aggressive arc_p grow"); | |
5676 | ||
62422785 PS |
5677 | module_param(zfs_arc_p_dampener_disable, int, 0644); |
5678 | MODULE_PARM_DESC(zfs_arc_p_dampener_disable, "disable arc_p adapt dampener"); | |
5679 | ||
bce45ec9 | 5680 | module_param(zfs_arc_shrink_shift, int, 0644); |
c409e464 BB |
5681 | MODULE_PARM_DESC(zfs_arc_shrink_shift, "log2(fraction of arc to reclaim)"); |
5682 | ||
1f7c30df BB |
5683 | module_param(zfs_disable_dup_eviction, int, 0644); |
5684 | MODULE_PARM_DESC(zfs_disable_dup_eviction, "disable duplicate buffer eviction"); | |
5685 | ||
0c5493d4 BB |
5686 | module_param(zfs_arc_memory_throttle_disable, int, 0644); |
5687 | MODULE_PARM_DESC(zfs_arc_memory_throttle_disable, "disable memory throttle"); | |
5688 | ||
bce45ec9 BB |
5689 | module_param(zfs_arc_min_prefetch_lifespan, int, 0644); |
5690 | MODULE_PARM_DESC(zfs_arc_min_prefetch_lifespan, "Min life of prefetch block"); | |
5691 | ||
5692 | module_param(l2arc_write_max, ulong, 0644); | |
abd8610c BB |
5693 | MODULE_PARM_DESC(l2arc_write_max, "Max write bytes per interval"); |
5694 | ||
bce45ec9 | 5695 | module_param(l2arc_write_boost, ulong, 0644); |
abd8610c BB |
5696 | MODULE_PARM_DESC(l2arc_write_boost, "Extra write bytes during device warmup"); |
5697 | ||
bce45ec9 | 5698 | module_param(l2arc_headroom, ulong, 0644); |
abd8610c BB |
5699 | MODULE_PARM_DESC(l2arc_headroom, "Number of max device writes to precache"); |
5700 | ||
3a17a7a9 SK |
5701 | module_param(l2arc_headroom_boost, ulong, 0644); |
5702 | MODULE_PARM_DESC(l2arc_headroom_boost, "Compressed l2arc_headroom multiplier"); | |
5703 | ||
bce45ec9 | 5704 | module_param(l2arc_feed_secs, ulong, 0644); |
abd8610c BB |
5705 | MODULE_PARM_DESC(l2arc_feed_secs, "Seconds between L2ARC writing"); |
5706 | ||
bce45ec9 | 5707 | module_param(l2arc_feed_min_ms, ulong, 0644); |
abd8610c BB |
5708 | MODULE_PARM_DESC(l2arc_feed_min_ms, "Min feed interval in milliseconds"); |
5709 | ||
bce45ec9 | 5710 | module_param(l2arc_noprefetch, int, 0644); |
abd8610c BB |
5711 | MODULE_PARM_DESC(l2arc_noprefetch, "Skip caching prefetched buffers"); |
5712 | ||
3a17a7a9 SK |
5713 | module_param(l2arc_nocompress, int, 0644); |
5714 | MODULE_PARM_DESC(l2arc_nocompress, "Skip compressing L2ARC buffers"); | |
5715 | ||
bce45ec9 | 5716 | module_param(l2arc_feed_again, int, 0644); |
abd8610c BB |
5717 | MODULE_PARM_DESC(l2arc_feed_again, "Turbo L2ARC warmup"); |
5718 | ||
bce45ec9 | 5719 | module_param(l2arc_norw, int, 0644); |
abd8610c BB |
5720 | MODULE_PARM_DESC(l2arc_norw, "No reads during writes"); |
5721 | ||
c28b2279 | 5722 | #endif |