Print this page
6220 memleak in l2arc on debug build
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/arc.c
+++ new/usr/src/uts/common/fs/zfs/arc.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
24 24 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
25 25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
27 27 */
28 28
29 29 /*
30 30 * DVA-based Adjustable Replacement Cache
31 31 *
32 32 * While much of the theory of operation used here is
33 33 * based on the self-tuning, low overhead replacement cache
34 34 * presented by Megiddo and Modha at FAST 2003, there are some
35 35 * significant differences:
36 36 *
37 37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 38 * Pages in its cache cannot be "locked" into memory. This makes
39 39 * the eviction algorithm simple: evict the last page in the list.
40 40 * This also make the performance characteristics easy to reason
41 41 * about. Our cache is not so simple. At any given moment, some
42 42 * subset of the blocks in the cache are un-evictable because we
43 43 * have handed out a reference to them. Blocks are only evictable
44 44 * when there are no external references active. This makes
45 45 * eviction far more problematic: we choose to evict the evictable
46 46 * blocks that are the "lowest" in the list.
47 47 *
48 48 * There are times when it is not possible to evict the requested
49 49 * space. In these circumstances we are unable to adjust the cache
50 50 * size. To prevent the cache growing unbounded at these times we
51 51 * implement a "cache throttle" that slows the flow of new data
52 52 * into the cache until we can make space available.
53 53 *
54 54 * 2. The Megiddo and Modha model assumes a fixed cache size.
55 55 * Pages are evicted when the cache is full and there is a cache
56 56 * miss. Our model has a variable sized cache. It grows with
57 57 * high use, but also tries to react to memory pressure from the
58 58 * operating system: decreasing its size when system memory is
59 59 * tight.
60 60 *
61 61 * 3. The Megiddo and Modha model assumes a fixed page size. All
62 62 * elements of the cache are therefore exactly the same size. So
63 63 * when adjusting the cache size following a cache miss, its simply
64 64 * a matter of choosing a single page to evict. In our model, we
65 65 * have variable sized cache blocks (rangeing from 512 bytes to
66 66 * 128K bytes). We therefore choose a set of blocks to evict to make
67 67 * space for a cache miss that approximates as closely as possible
68 68 * the space used by the new block.
69 69 *
70 70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
71 71 * by N. Megiddo & D. Modha, FAST 2003
72 72 */
73 73
74 74 /*
75 75 * The locking model:
76 76 *
77 77 * A new reference to a cache buffer can be obtained in two
78 78 * ways: 1) via a hash table lookup using the DVA as a key,
79 79 * or 2) via one of the ARC lists. The arc_read() interface
80 80 * uses method 1, while the internal arc algorithms for
81 81 * adjusting the cache use method 2. We therefore provide two
82 82 * types of locks: 1) the hash table lock array, and 2) the
83 83 * arc list locks.
84 84 *
85 85 * Buffers do not have their own mutexes, rather they rely on the
86 86 * hash table mutexes for the bulk of their protection (i.e. most
87 87 * fields in the arc_buf_hdr_t are protected by these mutexes).
88 88 *
89 89 * buf_hash_find() returns the appropriate mutex (held) when it
90 90 * locates the requested buffer in the hash table. It returns
91 91 * NULL for the mutex if the buffer was not in the table.
92 92 *
93 93 * buf_hash_remove() expects the appropriate hash mutex to be
94 94 * already held before it is invoked.
95 95 *
96 96 * Each arc state also has a mutex which is used to protect the
97 97 * buffer list associated with the state. When attempting to
98 98 * obtain a hash table lock while holding an arc list lock you
99 99 * must use: mutex_tryenter() to avoid deadlock. Also note that
100 100 * the active state mutex must be held before the ghost state mutex.
101 101 *
102 102 * Arc buffers may have an associated eviction callback function.
103 103 * This function will be invoked prior to removing the buffer (e.g.
104 104 * in arc_do_user_evicts()). Note however that the data associated
105 105 * with the buffer may be evicted prior to the callback. The callback
106 106 * must be made with *no locks held* (to prevent deadlock). Additionally,
107 107 * the users of callbacks must ensure that their private data is
108 108 * protected from simultaneous callbacks from arc_clear_callback()
109 109 * and arc_do_user_evicts().
110 110 *
111 111 * Note that the majority of the performance stats are manipulated
112 112 * with atomic operations.
113 113 *
114 114 * The L2ARC uses the l2ad_mtx on each vdev for the following:
115 115 *
116 116 * - L2ARC buflist creation
117 117 * - L2ARC buflist eviction
118 118 * - L2ARC write completion, which walks L2ARC buflists
119 119 * - ARC header destruction, as it removes from L2ARC buflists
120 120 * - ARC header release, as it removes from L2ARC buflists
121 121 */
122 122
123 123 #include <sys/spa.h>
124 124 #include <sys/zio.h>
125 125 #include <sys/zio_compress.h>
126 126 #include <sys/zfs_context.h>
127 127 #include <sys/arc.h>
128 128 #include <sys/refcount.h>
129 129 #include <sys/vdev.h>
130 130 #include <sys/vdev_impl.h>
131 131 #include <sys/dsl_pool.h>
132 132 #include <sys/multilist.h>
133 133 #ifdef _KERNEL
134 134 #include <sys/vmsystm.h>
135 135 #include <vm/anon.h>
136 136 #include <sys/fs/swapnode.h>
137 137 #include <sys/dnlc.h>
138 138 #endif
139 139 #include <sys/callb.h>
140 140 #include <sys/kstat.h>
141 141 #include <zfs_fletcher.h>
142 142
143 143 #ifndef _KERNEL
144 144 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
145 145 boolean_t arc_watch = B_FALSE;
146 146 int arc_procfd;
147 147 #endif
148 148
149 149 static kmutex_t arc_reclaim_lock;
150 150 static kcondvar_t arc_reclaim_thread_cv;
151 151 static boolean_t arc_reclaim_thread_exit;
152 152 static kcondvar_t arc_reclaim_waiters_cv;
153 153
154 154 static kmutex_t arc_user_evicts_lock;
155 155 static kcondvar_t arc_user_evicts_cv;
156 156 static boolean_t arc_user_evicts_thread_exit;
157 157
158 158 uint_t arc_reduce_dnlc_percent = 3;
159 159
160 160 /*
161 161 * The number of headers to evict in arc_evict_state_impl() before
162 162 * dropping the sublist lock and evicting from another sublist. A lower
163 163 * value means we're more likely to evict the "correct" header (i.e. the
164 164 * oldest header in the arc state), but comes with higher overhead
165 165 * (i.e. more invocations of arc_evict_state_impl()).
166 166 */
167 167 int zfs_arc_evict_batch_limit = 10;
168 168
169 169 /*
170 170 * The number of sublists used for each of the arc state lists. If this
171 171 * is not set to a suitable value by the user, it will be configured to
172 172 * the number of CPUs on the system in arc_init().
173 173 */
174 174 int zfs_arc_num_sublists_per_state = 0;
175 175
176 176 /* number of seconds before growing cache again */
177 177 static int arc_grow_retry = 60;
178 178
179 179 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */
180 180 int zfs_arc_overflow_shift = 8;
181 181
182 182 /* shift of arc_c for calculating both min and max arc_p */
183 183 static int arc_p_min_shift = 4;
184 184
185 185 /* log2(fraction of arc to reclaim) */
186 186 static int arc_shrink_shift = 7;
187 187
188 188 /*
189 189 * log2(fraction of ARC which must be free to allow growing).
190 190 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory,
191 191 * when reading a new block into the ARC, we will evict an equal-sized block
192 192 * from the ARC.
193 193 *
194 194 * This must be less than arc_shrink_shift, so that when we shrink the ARC,
195 195 * we will still not allow it to grow.
196 196 */
197 197 int arc_no_grow_shift = 5;
198 198
199 199
200 200 /*
201 201 * minimum lifespan of a prefetch block in clock ticks
202 202 * (initialized in arc_init())
203 203 */
204 204 static int arc_min_prefetch_lifespan;
205 205
206 206 /*
207 207 * If this percent of memory is free, don't throttle.
208 208 */
209 209 int arc_lotsfree_percent = 10;
210 210
211 211 static int arc_dead;
212 212
213 213 /*
214 214 * The arc has filled available memory and has now warmed up.
215 215 */
216 216 static boolean_t arc_warm;
217 217
218 218 /*
219 219 * These tunables are for performance analysis.
220 220 */
221 221 uint64_t zfs_arc_max;
222 222 uint64_t zfs_arc_min;
223 223 uint64_t zfs_arc_meta_limit = 0;
224 224 uint64_t zfs_arc_meta_min = 0;
225 225 int zfs_arc_grow_retry = 0;
226 226 int zfs_arc_shrink_shift = 0;
227 227 int zfs_arc_p_min_shift = 0;
228 228 int zfs_disable_dup_eviction = 0;
229 229 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
230 230
231 231 /*
232 232 * Note that buffers can be in one of 6 states:
233 233 * ARC_anon - anonymous (discussed below)
234 234 * ARC_mru - recently used, currently cached
235 235 * ARC_mru_ghost - recentely used, no longer in cache
236 236 * ARC_mfu - frequently used, currently cached
237 237 * ARC_mfu_ghost - frequently used, no longer in cache
238 238 * ARC_l2c_only - exists in L2ARC but not other states
239 239 * When there are no active references to the buffer, they are
240 240 * are linked onto a list in one of these arc states. These are
241 241 * the only buffers that can be evicted or deleted. Within each
242 242 * state there are multiple lists, one for meta-data and one for
243 243 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
244 244 * etc.) is tracked separately so that it can be managed more
245 245 * explicitly: favored over data, limited explicitly.
246 246 *
247 247 * Anonymous buffers are buffers that are not associated with
248 248 * a DVA. These are buffers that hold dirty block copies
249 249 * before they are written to stable storage. By definition,
250 250 * they are "ref'd" and are considered part of arc_mru
251 251 * that cannot be freed. Generally, they will aquire a DVA
252 252 * as they are written and migrate onto the arc_mru list.
253 253 *
254 254 * The ARC_l2c_only state is for buffers that are in the second
255 255 * level ARC but no longer in any of the ARC_m* lists. The second
256 256 * level ARC itself may also contain buffers that are in any of
257 257 * the ARC_m* states - meaning that a buffer can exist in two
258 258 * places. The reason for the ARC_l2c_only state is to keep the
259 259 * buffer header in the hash table, so that reads that hit the
260 260 * second level ARC benefit from these fast lookups.
261 261 */
262 262
263 263 typedef struct arc_state {
264 264 /*
265 265 * list of evictable buffers
266 266 */
267 267 multilist_t arcs_list[ARC_BUFC_NUMTYPES];
268 268 /*
269 269 * total amount of evictable data in this state
270 270 */
271 271 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES];
272 272 /*
273 273 * total amount of data in this state; this includes: evictable,
274 274 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
275 275 */
276 276 refcount_t arcs_size;
277 277 } arc_state_t;
278 278
279 279 /* The 6 states: */
280 280 static arc_state_t ARC_anon;
281 281 static arc_state_t ARC_mru;
282 282 static arc_state_t ARC_mru_ghost;
283 283 static arc_state_t ARC_mfu;
284 284 static arc_state_t ARC_mfu_ghost;
285 285 static arc_state_t ARC_l2c_only;
286 286
287 287 typedef struct arc_stats {
288 288 kstat_named_t arcstat_hits;
289 289 kstat_named_t arcstat_misses;
290 290 kstat_named_t arcstat_demand_data_hits;
291 291 kstat_named_t arcstat_demand_data_misses;
292 292 kstat_named_t arcstat_demand_metadata_hits;
293 293 kstat_named_t arcstat_demand_metadata_misses;
294 294 kstat_named_t arcstat_prefetch_data_hits;
295 295 kstat_named_t arcstat_prefetch_data_misses;
296 296 kstat_named_t arcstat_prefetch_metadata_hits;
297 297 kstat_named_t arcstat_prefetch_metadata_misses;
298 298 kstat_named_t arcstat_mru_hits;
299 299 kstat_named_t arcstat_mru_ghost_hits;
300 300 kstat_named_t arcstat_mfu_hits;
301 301 kstat_named_t arcstat_mfu_ghost_hits;
302 302 kstat_named_t arcstat_deleted;
303 303 /*
304 304 * Number of buffers that could not be evicted because the hash lock
305 305 * was held by another thread. The lock may not necessarily be held
306 306 * by something using the same buffer, since hash locks are shared
307 307 * by multiple buffers.
308 308 */
309 309 kstat_named_t arcstat_mutex_miss;
310 310 /*
311 311 * Number of buffers skipped because they have I/O in progress, are
312 312 * indrect prefetch buffers that have not lived long enough, or are
313 313 * not from the spa we're trying to evict from.
314 314 */
315 315 kstat_named_t arcstat_evict_skip;
316 316 /*
317 317 * Number of times arc_evict_state() was unable to evict enough
318 318 * buffers to reach it's target amount.
319 319 */
320 320 kstat_named_t arcstat_evict_not_enough;
321 321 kstat_named_t arcstat_evict_l2_cached;
322 322 kstat_named_t arcstat_evict_l2_eligible;
323 323 kstat_named_t arcstat_evict_l2_ineligible;
324 324 kstat_named_t arcstat_evict_l2_skip;
325 325 kstat_named_t arcstat_hash_elements;
326 326 kstat_named_t arcstat_hash_elements_max;
327 327 kstat_named_t arcstat_hash_collisions;
328 328 kstat_named_t arcstat_hash_chains;
329 329 kstat_named_t arcstat_hash_chain_max;
330 330 kstat_named_t arcstat_p;
331 331 kstat_named_t arcstat_c;
332 332 kstat_named_t arcstat_c_min;
333 333 kstat_named_t arcstat_c_max;
334 334 kstat_named_t arcstat_size;
335 335 /*
336 336 * Number of bytes consumed by internal ARC structures necessary
337 337 * for tracking purposes; these structures are not actually
338 338 * backed by ARC buffers. This includes arc_buf_hdr_t structures
339 339 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
340 340 * caches), and arc_buf_t structures (allocated via arc_buf_t
341 341 * cache).
342 342 */
343 343 kstat_named_t arcstat_hdr_size;
344 344 /*
345 345 * Number of bytes consumed by ARC buffers of type equal to
346 346 * ARC_BUFC_DATA. This is generally consumed by buffers backing
347 347 * on disk user data (e.g. plain file contents).
348 348 */
349 349 kstat_named_t arcstat_data_size;
350 350 /*
351 351 * Number of bytes consumed by ARC buffers of type equal to
352 352 * ARC_BUFC_METADATA. This is generally consumed by buffers
353 353 * backing on disk data that is used for internal ZFS
354 354 * structures (e.g. ZAP, dnode, indirect blocks, etc).
355 355 */
356 356 kstat_named_t arcstat_metadata_size;
357 357 /*
358 358 * Number of bytes consumed by various buffers and structures
359 359 * not actually backed with ARC buffers. This includes bonus
360 360 * buffers (allocated directly via zio_buf_* functions),
361 361 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
362 362 * cache), and dnode_t structures (allocated via dnode_t cache).
363 363 */
364 364 kstat_named_t arcstat_other_size;
365 365 /*
366 366 * Total number of bytes consumed by ARC buffers residing in the
367 367 * arc_anon state. This includes *all* buffers in the arc_anon
368 368 * state; e.g. data, metadata, evictable, and unevictable buffers
369 369 * are all included in this value.
370 370 */
371 371 kstat_named_t arcstat_anon_size;
372 372 /*
373 373 * Number of bytes consumed by ARC buffers that meet the
374 374 * following criteria: backing buffers of type ARC_BUFC_DATA,
375 375 * residing in the arc_anon state, and are eligible for eviction
376 376 * (e.g. have no outstanding holds on the buffer).
377 377 */
378 378 kstat_named_t arcstat_anon_evictable_data;
379 379 /*
380 380 * Number of bytes consumed by ARC buffers that meet the
381 381 * following criteria: backing buffers of type ARC_BUFC_METADATA,
382 382 * residing in the arc_anon state, and are eligible for eviction
383 383 * (e.g. have no outstanding holds on the buffer).
384 384 */
385 385 kstat_named_t arcstat_anon_evictable_metadata;
386 386 /*
387 387 * Total number of bytes consumed by ARC buffers residing in the
388 388 * arc_mru state. This includes *all* buffers in the arc_mru
389 389 * state; e.g. data, metadata, evictable, and unevictable buffers
390 390 * are all included in this value.
391 391 */
392 392 kstat_named_t arcstat_mru_size;
393 393 /*
394 394 * Number of bytes consumed by ARC buffers that meet the
395 395 * following criteria: backing buffers of type ARC_BUFC_DATA,
396 396 * residing in the arc_mru state, and are eligible for eviction
397 397 * (e.g. have no outstanding holds on the buffer).
398 398 */
399 399 kstat_named_t arcstat_mru_evictable_data;
400 400 /*
401 401 * Number of bytes consumed by ARC buffers that meet the
402 402 * following criteria: backing buffers of type ARC_BUFC_METADATA,
403 403 * residing in the arc_mru state, and are eligible for eviction
404 404 * (e.g. have no outstanding holds on the buffer).
405 405 */
406 406 kstat_named_t arcstat_mru_evictable_metadata;
407 407 /*
408 408 * Total number of bytes that *would have been* consumed by ARC
409 409 * buffers in the arc_mru_ghost state. The key thing to note
410 410 * here, is the fact that this size doesn't actually indicate
411 411 * RAM consumption. The ghost lists only consist of headers and
412 412 * don't actually have ARC buffers linked off of these headers.
413 413 * Thus, *if* the headers had associated ARC buffers, these
414 414 * buffers *would have* consumed this number of bytes.
415 415 */
416 416 kstat_named_t arcstat_mru_ghost_size;
417 417 /*
418 418 * Number of bytes that *would have been* consumed by ARC
419 419 * buffers that are eligible for eviction, of type
420 420 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
421 421 */
422 422 kstat_named_t arcstat_mru_ghost_evictable_data;
423 423 /*
424 424 * Number of bytes that *would have been* consumed by ARC
425 425 * buffers that are eligible for eviction, of type
426 426 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
427 427 */
428 428 kstat_named_t arcstat_mru_ghost_evictable_metadata;
429 429 /*
430 430 * Total number of bytes consumed by ARC buffers residing in the
431 431 * arc_mfu state. This includes *all* buffers in the arc_mfu
432 432 * state; e.g. data, metadata, evictable, and unevictable buffers
433 433 * are all included in this value.
434 434 */
435 435 kstat_named_t arcstat_mfu_size;
436 436 /*
437 437 * Number of bytes consumed by ARC buffers that are eligible for
438 438 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
439 439 * state.
440 440 */
441 441 kstat_named_t arcstat_mfu_evictable_data;
442 442 /*
443 443 * Number of bytes consumed by ARC buffers that are eligible for
444 444 * eviction, of type ARC_BUFC_METADATA, and reside in the
445 445 * arc_mfu state.
446 446 */
447 447 kstat_named_t arcstat_mfu_evictable_metadata;
448 448 /*
449 449 * Total number of bytes that *would have been* consumed by ARC
450 450 * buffers in the arc_mfu_ghost state. See the comment above
451 451 * arcstat_mru_ghost_size for more details.
452 452 */
453 453 kstat_named_t arcstat_mfu_ghost_size;
454 454 /*
455 455 * Number of bytes that *would have been* consumed by ARC
456 456 * buffers that are eligible for eviction, of type
457 457 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
458 458 */
459 459 kstat_named_t arcstat_mfu_ghost_evictable_data;
460 460 /*
461 461 * Number of bytes that *would have been* consumed by ARC
462 462 * buffers that are eligible for eviction, of type
463 463 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
464 464 */
465 465 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
466 466 kstat_named_t arcstat_l2_hits;
467 467 kstat_named_t arcstat_l2_misses;
468 468 kstat_named_t arcstat_l2_feeds;
469 469 kstat_named_t arcstat_l2_rw_clash;
470 470 kstat_named_t arcstat_l2_read_bytes;
471 471 kstat_named_t arcstat_l2_write_bytes;
472 472 kstat_named_t arcstat_l2_writes_sent;
473 473 kstat_named_t arcstat_l2_writes_done;
474 474 kstat_named_t arcstat_l2_writes_error;
475 475 kstat_named_t arcstat_l2_writes_lock_retry;
476 476 kstat_named_t arcstat_l2_evict_lock_retry;
477 477 kstat_named_t arcstat_l2_evict_reading;
478 478 kstat_named_t arcstat_l2_evict_l1cached;
479 479 kstat_named_t arcstat_l2_free_on_write;
480 480 kstat_named_t arcstat_l2_cdata_free_on_write;
481 481 kstat_named_t arcstat_l2_abort_lowmem;
482 482 kstat_named_t arcstat_l2_cksum_bad;
483 483 kstat_named_t arcstat_l2_io_error;
484 484 kstat_named_t arcstat_l2_size;
485 485 kstat_named_t arcstat_l2_asize;
486 486 kstat_named_t arcstat_l2_hdr_size;
487 487 kstat_named_t arcstat_l2_compress_successes;
488 488 kstat_named_t arcstat_l2_compress_zeros;
489 489 kstat_named_t arcstat_l2_compress_failures;
490 490 kstat_named_t arcstat_memory_throttle_count;
491 491 kstat_named_t arcstat_duplicate_buffers;
492 492 kstat_named_t arcstat_duplicate_buffers_size;
493 493 kstat_named_t arcstat_duplicate_reads;
494 494 kstat_named_t arcstat_meta_used;
495 495 kstat_named_t arcstat_meta_limit;
496 496 kstat_named_t arcstat_meta_max;
497 497 kstat_named_t arcstat_meta_min;
498 498 } arc_stats_t;
499 499
500 500 static arc_stats_t arc_stats = {
501 501 { "hits", KSTAT_DATA_UINT64 },
502 502 { "misses", KSTAT_DATA_UINT64 },
503 503 { "demand_data_hits", KSTAT_DATA_UINT64 },
504 504 { "demand_data_misses", KSTAT_DATA_UINT64 },
505 505 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
506 506 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
507 507 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
508 508 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
509 509 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
510 510 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
511 511 { "mru_hits", KSTAT_DATA_UINT64 },
512 512 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
513 513 { "mfu_hits", KSTAT_DATA_UINT64 },
514 514 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
515 515 { "deleted", KSTAT_DATA_UINT64 },
516 516 { "mutex_miss", KSTAT_DATA_UINT64 },
517 517 { "evict_skip", KSTAT_DATA_UINT64 },
518 518 { "evict_not_enough", KSTAT_DATA_UINT64 },
519 519 { "evict_l2_cached", KSTAT_DATA_UINT64 },
520 520 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
521 521 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
522 522 { "evict_l2_skip", KSTAT_DATA_UINT64 },
523 523 { "hash_elements", KSTAT_DATA_UINT64 },
524 524 { "hash_elements_max", KSTAT_DATA_UINT64 },
525 525 { "hash_collisions", KSTAT_DATA_UINT64 },
526 526 { "hash_chains", KSTAT_DATA_UINT64 },
527 527 { "hash_chain_max", KSTAT_DATA_UINT64 },
528 528 { "p", KSTAT_DATA_UINT64 },
529 529 { "c", KSTAT_DATA_UINT64 },
530 530 { "c_min", KSTAT_DATA_UINT64 },
531 531 { "c_max", KSTAT_DATA_UINT64 },
532 532 { "size", KSTAT_DATA_UINT64 },
533 533 { "hdr_size", KSTAT_DATA_UINT64 },
534 534 { "data_size", KSTAT_DATA_UINT64 },
535 535 { "metadata_size", KSTAT_DATA_UINT64 },
536 536 { "other_size", KSTAT_DATA_UINT64 },
537 537 { "anon_size", KSTAT_DATA_UINT64 },
538 538 { "anon_evictable_data", KSTAT_DATA_UINT64 },
539 539 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
540 540 { "mru_size", KSTAT_DATA_UINT64 },
541 541 { "mru_evictable_data", KSTAT_DATA_UINT64 },
542 542 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
543 543 { "mru_ghost_size", KSTAT_DATA_UINT64 },
544 544 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
545 545 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
546 546 { "mfu_size", KSTAT_DATA_UINT64 },
547 547 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
548 548 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
549 549 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
550 550 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
551 551 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
552 552 { "l2_hits", KSTAT_DATA_UINT64 },
553 553 { "l2_misses", KSTAT_DATA_UINT64 },
554 554 { "l2_feeds", KSTAT_DATA_UINT64 },
555 555 { "l2_rw_clash", KSTAT_DATA_UINT64 },
556 556 { "l2_read_bytes", KSTAT_DATA_UINT64 },
557 557 { "l2_write_bytes", KSTAT_DATA_UINT64 },
558 558 { "l2_writes_sent", KSTAT_DATA_UINT64 },
559 559 { "l2_writes_done", KSTAT_DATA_UINT64 },
560 560 { "l2_writes_error", KSTAT_DATA_UINT64 },
561 561 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
562 562 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
563 563 { "l2_evict_reading", KSTAT_DATA_UINT64 },
564 564 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
565 565 { "l2_free_on_write", KSTAT_DATA_UINT64 },
566 566 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 },
567 567 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
568 568 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
569 569 { "l2_io_error", KSTAT_DATA_UINT64 },
570 570 { "l2_size", KSTAT_DATA_UINT64 },
571 571 { "l2_asize", KSTAT_DATA_UINT64 },
572 572 { "l2_hdr_size", KSTAT_DATA_UINT64 },
573 573 { "l2_compress_successes", KSTAT_DATA_UINT64 },
574 574 { "l2_compress_zeros", KSTAT_DATA_UINT64 },
575 575 { "l2_compress_failures", KSTAT_DATA_UINT64 },
576 576 { "memory_throttle_count", KSTAT_DATA_UINT64 },
577 577 { "duplicate_buffers", KSTAT_DATA_UINT64 },
578 578 { "duplicate_buffers_size", KSTAT_DATA_UINT64 },
579 579 { "duplicate_reads", KSTAT_DATA_UINT64 },
580 580 { "arc_meta_used", KSTAT_DATA_UINT64 },
581 581 { "arc_meta_limit", KSTAT_DATA_UINT64 },
582 582 { "arc_meta_max", KSTAT_DATA_UINT64 },
583 583 { "arc_meta_min", KSTAT_DATA_UINT64 }
584 584 };
585 585
586 586 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
587 587
588 588 #define ARCSTAT_INCR(stat, val) \
589 589 atomic_add_64(&arc_stats.stat.value.ui64, (val))
590 590
591 591 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
592 592 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
593 593
594 594 #define ARCSTAT_MAX(stat, val) { \
595 595 uint64_t m; \
596 596 while ((val) > (m = arc_stats.stat.value.ui64) && \
597 597 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
598 598 continue; \
599 599 }
600 600
601 601 #define ARCSTAT_MAXSTAT(stat) \
602 602 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
603 603
604 604 /*
605 605 * We define a macro to allow ARC hits/misses to be easily broken down by
606 606 * two separate conditions, giving a total of four different subtypes for
607 607 * each of hits and misses (so eight statistics total).
608 608 */
609 609 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
610 610 if (cond1) { \
611 611 if (cond2) { \
612 612 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
613 613 } else { \
614 614 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
615 615 } \
616 616 } else { \
617 617 if (cond2) { \
618 618 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
619 619 } else { \
620 620 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
621 621 } \
622 622 }
623 623
624 624 kstat_t *arc_ksp;
625 625 static arc_state_t *arc_anon;
626 626 static arc_state_t *arc_mru;
627 627 static arc_state_t *arc_mru_ghost;
628 628 static arc_state_t *arc_mfu;
629 629 static arc_state_t *arc_mfu_ghost;
630 630 static arc_state_t *arc_l2c_only;
631 631
632 632 /*
633 633 * There are several ARC variables that are critical to export as kstats --
634 634 * but we don't want to have to grovel around in the kstat whenever we wish to
635 635 * manipulate them. For these variables, we therefore define them to be in
636 636 * terms of the statistic variable. This assures that we are not introducing
637 637 * the possibility of inconsistency by having shadow copies of the variables,
638 638 * while still allowing the code to be readable.
639 639 */
640 640 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
641 641 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
642 642 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
643 643 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
644 644 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
645 645 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
646 646 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
647 647 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
648 648 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
649 649
650 650 #define L2ARC_IS_VALID_COMPRESS(_c_) \
651 651 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY)
652 652
653 653 static int arc_no_grow; /* Don't try to grow cache size */
654 654 static uint64_t arc_tempreserve;
655 655 static uint64_t arc_loaned_bytes;
656 656
657 657 typedef struct arc_callback arc_callback_t;
658 658
659 659 struct arc_callback {
660 660 void *acb_private;
661 661 arc_done_func_t *acb_done;
662 662 arc_buf_t *acb_buf;
663 663 zio_t *acb_zio_dummy;
664 664 arc_callback_t *acb_next;
665 665 };
666 666
667 667 typedef struct arc_write_callback arc_write_callback_t;
668 668
669 669 struct arc_write_callback {
670 670 void *awcb_private;
671 671 arc_done_func_t *awcb_ready;
672 672 arc_done_func_t *awcb_physdone;
673 673 arc_done_func_t *awcb_done;
674 674 arc_buf_t *awcb_buf;
675 675 };
676 676
677 677 /*
678 678 * ARC buffers are separated into multiple structs as a memory saving measure:
679 679 * - Common fields struct, always defined, and embedded within it:
680 680 * - L2-only fields, always allocated but undefined when not in L2ARC
681 681 * - L1-only fields, only allocated when in L1ARC
682 682 *
683 683 * Buffer in L1 Buffer only in L2
684 684 * +------------------------+ +------------------------+
685 685 * | arc_buf_hdr_t | | arc_buf_hdr_t |
686 686 * | | | |
687 687 * | | | |
688 688 * | | | |
689 689 * +------------------------+ +------------------------+
690 690 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
691 691 * | (undefined if L1-only) | | |
692 692 * +------------------------+ +------------------------+
693 693 * | l1arc_buf_hdr_t |
694 694 * | |
695 695 * | |
696 696 * | |
697 697 * | |
698 698 * +------------------------+
699 699 *
700 700 * Because it's possible for the L2ARC to become extremely large, we can wind
701 701 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
702 702 * is minimized by only allocating the fields necessary for an L1-cached buffer
703 703 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
704 704 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
705 705 * words in pointers. arc_hdr_realloc() is used to switch a header between
706 706 * these two allocation states.
707 707 */
708 708 typedef struct l1arc_buf_hdr {
709 709 kmutex_t b_freeze_lock;
710 710 #ifdef ZFS_DEBUG
711 711 /*
712 712 * used for debugging wtih kmem_flags - by allocating and freeing
713 713 * b_thawed when the buffer is thawed, we get a record of the stack
714 714 * trace that thawed it.
715 715 */
716 716 void *b_thawed;
717 717 #endif
718 718
719 719 arc_buf_t *b_buf;
720 720 uint32_t b_datacnt;
721 721 /* for waiting on writes to complete */
722 722 kcondvar_t b_cv;
723 723
724 724 /* protected by arc state mutex */
725 725 arc_state_t *b_state;
726 726 multilist_node_t b_arc_node;
727 727
728 728 /* updated atomically */
729 729 clock_t b_arc_access;
730 730
731 731 /* self protecting */
732 732 refcount_t b_refcnt;
733 733
734 734 arc_callback_t *b_acb;
735 735 /* temporary buffer holder for in-flight compressed data */
736 736 void *b_tmp_cdata;
737 737 } l1arc_buf_hdr_t;
738 738
739 739 typedef struct l2arc_dev l2arc_dev_t;
740 740
741 741 typedef struct l2arc_buf_hdr {
742 742 /* protected by arc_buf_hdr mutex */
743 743 l2arc_dev_t *b_dev; /* L2ARC device */
744 744 uint64_t b_daddr; /* disk address, offset byte */
745 745 /* real alloc'd buffer size depending on b_compress applied */
746 746 int32_t b_asize;
747 747 uint8_t b_compress;
748 748
749 749 list_node_t b_l2node;
750 750 } l2arc_buf_hdr_t;
751 751
752 752 struct arc_buf_hdr {
753 753 /* protected by hash lock */
754 754 dva_t b_dva;
755 755 uint64_t b_birth;
756 756 /*
757 757 * Even though this checksum is only set/verified when a buffer is in
758 758 * the L1 cache, it needs to be in the set of common fields because it
759 759 * must be preserved from the time before a buffer is written out to
760 760 * L2ARC until after it is read back in.
761 761 */
762 762 zio_cksum_t *b_freeze_cksum;
763 763
764 764 arc_buf_hdr_t *b_hash_next;
765 765 arc_flags_t b_flags;
766 766
767 767 /* immutable */
768 768 int32_t b_size;
769 769 uint64_t b_spa;
770 770
771 771 /* L2ARC fields. Undefined when not in L2ARC. */
772 772 l2arc_buf_hdr_t b_l2hdr;
773 773 /* L1ARC fields. Undefined when in l2arc_only state */
774 774 l1arc_buf_hdr_t b_l1hdr;
775 775 };
776 776
777 777 static arc_buf_t *arc_eviction_list;
778 778 static arc_buf_hdr_t arc_eviction_hdr;
779 779
780 780 #define GHOST_STATE(state) \
781 781 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
782 782 (state) == arc_l2c_only)
783 783
784 784 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
785 785 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
786 786 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
787 787 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
788 788 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ)
789 789 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE)
790 790
791 791 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
792 792 #define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS)
793 793 #define HDR_L2_READING(hdr) \
794 794 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
795 795 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
796 796 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
797 797 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
798 798 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
799 799
800 800 #define HDR_ISTYPE_METADATA(hdr) \
801 801 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
802 802 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
803 803
804 804 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
805 805 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
806 806
807 807 /*
808 808 * Other sizes
809 809 */
810 810
811 811 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
812 812 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
813 813
814 814 /*
815 815 * Hash table routines
816 816 */
817 817
818 818 #define HT_LOCK_PAD 64
819 819
820 820 struct ht_lock {
821 821 kmutex_t ht_lock;
822 822 #ifdef _KERNEL
823 823 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
824 824 #endif
825 825 };
826 826
827 827 #define BUF_LOCKS 256
828 828 typedef struct buf_hash_table {
829 829 uint64_t ht_mask;
830 830 arc_buf_hdr_t **ht_table;
831 831 struct ht_lock ht_locks[BUF_LOCKS];
832 832 } buf_hash_table_t;
833 833
834 834 static buf_hash_table_t buf_hash_table;
835 835
836 836 #define BUF_HASH_INDEX(spa, dva, birth) \
837 837 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
838 838 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
839 839 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
840 840 #define HDR_LOCK(hdr) \
841 841 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
842 842
843 843 uint64_t zfs_crc64_table[256];
844 844
845 845 /*
846 846 * Level 2 ARC
847 847 */
848 848
849 849 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
850 850 #define L2ARC_HEADROOM 2 /* num of writes */
851 851 /*
852 852 * If we discover during ARC scan any buffers to be compressed, we boost
853 853 * our headroom for the next scanning cycle by this percentage multiple.
854 854 */
855 855 #define L2ARC_HEADROOM_BOOST 200
856 856 #define L2ARC_FEED_SECS 1 /* caching interval secs */
857 857 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
858 858
859 859 /*
860 860 * Used to distinguish headers that are being process by
861 861 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk
862 862 * address. This can happen when the header is added to the l2arc's list
863 863 * of buffers to write in the first stage of l2arc_write_buffers(), but
864 864 * has not yet been written out which happens in the second stage of
865 865 * l2arc_write_buffers().
866 866 */
867 867 #define L2ARC_ADDR_UNSET ((uint64_t)(-1))
868 868
869 869 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
870 870 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
871 871
872 872 /* L2ARC Performance Tunables */
873 873 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
874 874 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
875 875 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
876 876 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
877 877 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
878 878 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
879 879 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
880 880 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
881 881 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
882 882
883 883 /*
884 884 * L2ARC Internals
885 885 */
886 886 struct l2arc_dev {
887 887 vdev_t *l2ad_vdev; /* vdev */
888 888 spa_t *l2ad_spa; /* spa */
889 889 uint64_t l2ad_hand; /* next write location */
890 890 uint64_t l2ad_start; /* first addr on device */
891 891 uint64_t l2ad_end; /* last addr on device */
892 892 boolean_t l2ad_first; /* first sweep through */
893 893 boolean_t l2ad_writing; /* currently writing */
894 894 kmutex_t l2ad_mtx; /* lock for buffer list */
895 895 list_t l2ad_buflist; /* buffer list */
896 896 list_node_t l2ad_node; /* device list node */
897 897 refcount_t l2ad_alloc; /* allocated bytes */
898 898 };
899 899
900 900 static list_t L2ARC_dev_list; /* device list */
901 901 static list_t *l2arc_dev_list; /* device list pointer */
902 902 static kmutex_t l2arc_dev_mtx; /* device list mutex */
903 903 static l2arc_dev_t *l2arc_dev_last; /* last device used */
904 904 static list_t L2ARC_free_on_write; /* free after write buf list */
905 905 static list_t *l2arc_free_on_write; /* free after write list ptr */
906 906 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
907 907 static uint64_t l2arc_ndev; /* number of devices */
908 908
909 909 typedef struct l2arc_read_callback {
910 910 arc_buf_t *l2rcb_buf; /* read buffer */
911 911 spa_t *l2rcb_spa; /* spa */
912 912 blkptr_t l2rcb_bp; /* original blkptr */
913 913 zbookmark_phys_t l2rcb_zb; /* original bookmark */
914 914 int l2rcb_flags; /* original flags */
915 915 enum zio_compress l2rcb_compress; /* applied compress */
916 916 } l2arc_read_callback_t;
917 917
918 918 typedef struct l2arc_write_callback {
919 919 l2arc_dev_t *l2wcb_dev; /* device info */
920 920 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
921 921 } l2arc_write_callback_t;
922 922
923 923 typedef struct l2arc_data_free {
924 924 /* protected by l2arc_free_on_write_mtx */
925 925 void *l2df_data;
926 926 size_t l2df_size;
927 927 void (*l2df_func)(void *, size_t);
928 928 list_node_t l2df_list_node;
929 929 } l2arc_data_free_t;
930 930
931 931 static kmutex_t l2arc_feed_thr_lock;
932 932 static kcondvar_t l2arc_feed_thr_cv;
933 933 static uint8_t l2arc_thread_exit;
934 934
935 935 static void arc_get_data_buf(arc_buf_t *);
936 936 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
937 937 static boolean_t arc_is_overflowing();
938 938 static void arc_buf_watch(arc_buf_t *);
939 939
940 940 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
941 941 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
942 942
943 943 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
944 944 static void l2arc_read_done(zio_t *);
945 945
946 946 static boolean_t l2arc_compress_buf(arc_buf_hdr_t *);
947 947 static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress);
948 948 static void l2arc_release_cdata_buf(arc_buf_hdr_t *);
949 949
950 950 static uint64_t
951 951 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
952 952 {
953 953 uint8_t *vdva = (uint8_t *)dva;
954 954 uint64_t crc = -1ULL;
955 955 int i;
956 956
957 957 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
958 958
959 959 for (i = 0; i < sizeof (dva_t); i++)
960 960 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
961 961
962 962 crc ^= (spa>>8) ^ birth;
963 963
964 964 return (crc);
965 965 }
966 966
967 967 #define BUF_EMPTY(buf) \
968 968 ((buf)->b_dva.dva_word[0] == 0 && \
969 969 (buf)->b_dva.dva_word[1] == 0)
970 970
971 971 #define BUF_EQUAL(spa, dva, birth, buf) \
972 972 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
973 973 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
974 974 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
975 975
976 976 static void
977 977 buf_discard_identity(arc_buf_hdr_t *hdr)
978 978 {
979 979 hdr->b_dva.dva_word[0] = 0;
980 980 hdr->b_dva.dva_word[1] = 0;
981 981 hdr->b_birth = 0;
982 982 }
983 983
984 984 static arc_buf_hdr_t *
985 985 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
986 986 {
987 987 const dva_t *dva = BP_IDENTITY(bp);
988 988 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
989 989 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
990 990 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
991 991 arc_buf_hdr_t *hdr;
992 992
993 993 mutex_enter(hash_lock);
994 994 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
995 995 hdr = hdr->b_hash_next) {
996 996 if (BUF_EQUAL(spa, dva, birth, hdr)) {
997 997 *lockp = hash_lock;
998 998 return (hdr);
999 999 }
1000 1000 }
1001 1001 mutex_exit(hash_lock);
1002 1002 *lockp = NULL;
1003 1003 return (NULL);
1004 1004 }
1005 1005
1006 1006 /*
1007 1007 * Insert an entry into the hash table. If there is already an element
1008 1008 * equal to elem in the hash table, then the already existing element
1009 1009 * will be returned and the new element will not be inserted.
1010 1010 * Otherwise returns NULL.
1011 1011 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1012 1012 */
1013 1013 static arc_buf_hdr_t *
1014 1014 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1015 1015 {
1016 1016 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1017 1017 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1018 1018 arc_buf_hdr_t *fhdr;
1019 1019 uint32_t i;
1020 1020
1021 1021 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1022 1022 ASSERT(hdr->b_birth != 0);
1023 1023 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1024 1024
1025 1025 if (lockp != NULL) {
1026 1026 *lockp = hash_lock;
1027 1027 mutex_enter(hash_lock);
1028 1028 } else {
1029 1029 ASSERT(MUTEX_HELD(hash_lock));
1030 1030 }
1031 1031
1032 1032 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1033 1033 fhdr = fhdr->b_hash_next, i++) {
1034 1034 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1035 1035 return (fhdr);
1036 1036 }
1037 1037
1038 1038 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1039 1039 buf_hash_table.ht_table[idx] = hdr;
1040 1040 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
1041 1041
1042 1042 /* collect some hash table performance data */
1043 1043 if (i > 0) {
1044 1044 ARCSTAT_BUMP(arcstat_hash_collisions);
1045 1045 if (i == 1)
1046 1046 ARCSTAT_BUMP(arcstat_hash_chains);
1047 1047
1048 1048 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1049 1049 }
1050 1050
1051 1051 ARCSTAT_BUMP(arcstat_hash_elements);
1052 1052 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1053 1053
1054 1054 return (NULL);
1055 1055 }
1056 1056
1057 1057 static void
1058 1058 buf_hash_remove(arc_buf_hdr_t *hdr)
1059 1059 {
1060 1060 arc_buf_hdr_t *fhdr, **hdrp;
1061 1061 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1062 1062
1063 1063 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1064 1064 ASSERT(HDR_IN_HASH_TABLE(hdr));
1065 1065
1066 1066 hdrp = &buf_hash_table.ht_table[idx];
1067 1067 while ((fhdr = *hdrp) != hdr) {
1068 1068 ASSERT(fhdr != NULL);
1069 1069 hdrp = &fhdr->b_hash_next;
1070 1070 }
1071 1071 *hdrp = hdr->b_hash_next;
1072 1072 hdr->b_hash_next = NULL;
1073 1073 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE;
1074 1074
1075 1075 /* collect some hash table performance data */
1076 1076 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1077 1077
1078 1078 if (buf_hash_table.ht_table[idx] &&
1079 1079 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1080 1080 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1081 1081 }
1082 1082
1083 1083 /*
1084 1084 * Global data structures and functions for the buf kmem cache.
1085 1085 */
1086 1086 static kmem_cache_t *hdr_full_cache;
1087 1087 static kmem_cache_t *hdr_l2only_cache;
1088 1088 static kmem_cache_t *buf_cache;
1089 1089
1090 1090 static void
1091 1091 buf_fini(void)
1092 1092 {
1093 1093 int i;
1094 1094
1095 1095 kmem_free(buf_hash_table.ht_table,
1096 1096 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1097 1097 for (i = 0; i < BUF_LOCKS; i++)
1098 1098 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1099 1099 kmem_cache_destroy(hdr_full_cache);
1100 1100 kmem_cache_destroy(hdr_l2only_cache);
1101 1101 kmem_cache_destroy(buf_cache);
1102 1102 }
1103 1103
1104 1104 /*
1105 1105 * Constructor callback - called when the cache is empty
1106 1106 * and a new buf is requested.
1107 1107 */
1108 1108 /* ARGSUSED */
1109 1109 static int
1110 1110 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1111 1111 {
1112 1112 arc_buf_hdr_t *hdr = vbuf;
1113 1113
1114 1114 bzero(hdr, HDR_FULL_SIZE);
1115 1115 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1116 1116 refcount_create(&hdr->b_l1hdr.b_refcnt);
1117 1117 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1118 1118 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1119 1119 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1120 1120
1121 1121 return (0);
1122 1122 }
1123 1123
1124 1124 /* ARGSUSED */
1125 1125 static int
1126 1126 hdr_l2only_cons(void *vbuf, void *unused, int kmflag)
1127 1127 {
1128 1128 arc_buf_hdr_t *hdr = vbuf;
1129 1129
1130 1130 bzero(hdr, HDR_L2ONLY_SIZE);
1131 1131 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1132 1132
1133 1133 return (0);
1134 1134 }
1135 1135
1136 1136 /* ARGSUSED */
1137 1137 static int
1138 1138 buf_cons(void *vbuf, void *unused, int kmflag)
1139 1139 {
1140 1140 arc_buf_t *buf = vbuf;
1141 1141
1142 1142 bzero(buf, sizeof (arc_buf_t));
1143 1143 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL);
1144 1144 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1145 1145
1146 1146 return (0);
1147 1147 }
1148 1148
1149 1149 /*
1150 1150 * Destructor callback - called when a cached buf is
1151 1151 * no longer required.
1152 1152 */
1153 1153 /* ARGSUSED */
1154 1154 static void
1155 1155 hdr_full_dest(void *vbuf, void *unused)
1156 1156 {
1157 1157 arc_buf_hdr_t *hdr = vbuf;
1158 1158
1159 1159 ASSERT(BUF_EMPTY(hdr));
1160 1160 cv_destroy(&hdr->b_l1hdr.b_cv);
1161 1161 refcount_destroy(&hdr->b_l1hdr.b_refcnt);
1162 1162 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock);
1163 1163 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
1164 1164 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS);
1165 1165 }
1166 1166
1167 1167 /* ARGSUSED */
1168 1168 static void
1169 1169 hdr_l2only_dest(void *vbuf, void *unused)
1170 1170 {
1171 1171 arc_buf_hdr_t *hdr = vbuf;
1172 1172
1173 1173 ASSERT(BUF_EMPTY(hdr));
1174 1174 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS);
1175 1175 }
1176 1176
1177 1177 /* ARGSUSED */
1178 1178 static void
1179 1179 buf_dest(void *vbuf, void *unused)
1180 1180 {
1181 1181 arc_buf_t *buf = vbuf;
1182 1182
1183 1183 mutex_destroy(&buf->b_evict_lock);
1184 1184 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
1185 1185 }
1186 1186
1187 1187 /*
1188 1188 * Reclaim callback -- invoked when memory is low.
1189 1189 */
1190 1190 /* ARGSUSED */
1191 1191 static void
1192 1192 hdr_recl(void *unused)
1193 1193 {
1194 1194 dprintf("hdr_recl called\n");
1195 1195 /*
1196 1196 * umem calls the reclaim func when we destroy the buf cache,
1197 1197 * which is after we do arc_fini().
1198 1198 */
1199 1199 if (!arc_dead)
1200 1200 cv_signal(&arc_reclaim_thread_cv);
1201 1201 }
1202 1202
1203 1203 static void
1204 1204 buf_init(void)
1205 1205 {
1206 1206 uint64_t *ct;
1207 1207 uint64_t hsize = 1ULL << 12;
1208 1208 int i, j;
1209 1209
1210 1210 /*
1211 1211 * The hash table is big enough to fill all of physical memory
1212 1212 * with an average block size of zfs_arc_average_blocksize (default 8K).
1213 1213 * By default, the table will take up
1214 1214 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1215 1215 */
1216 1216 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1217 1217 hsize <<= 1;
1218 1218 retry:
1219 1219 buf_hash_table.ht_mask = hsize - 1;
1220 1220 buf_hash_table.ht_table =
1221 1221 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1222 1222 if (buf_hash_table.ht_table == NULL) {
1223 1223 ASSERT(hsize > (1ULL << 8));
1224 1224 hsize >>= 1;
1225 1225 goto retry;
1226 1226 }
1227 1227
1228 1228 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1229 1229 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1230 1230 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1231 1231 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1232 1232 NULL, NULL, 0);
1233 1233 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1234 1234 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1235 1235
1236 1236 for (i = 0; i < 256; i++)
1237 1237 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1238 1238 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1239 1239
1240 1240 for (i = 0; i < BUF_LOCKS; i++) {
1241 1241 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1242 1242 NULL, MUTEX_DEFAULT, NULL);
1243 1243 }
1244 1244 }
1245 1245
1246 1246 /*
1247 1247 * Transition between the two allocation states for the arc_buf_hdr struct.
1248 1248 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without
1249 1249 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller
1250 1250 * version is used when a cache buffer is only in the L2ARC in order to reduce
1251 1251 * memory usage.
1252 1252 */
1253 1253 static arc_buf_hdr_t *
1254 1254 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new)
1255 1255 {
1256 1256 ASSERT(HDR_HAS_L2HDR(hdr));
1257 1257
1258 1258 arc_buf_hdr_t *nhdr;
1259 1259 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
1260 1260
1261 1261 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) ||
1262 1262 (old == hdr_l2only_cache && new == hdr_full_cache));
1263 1263
1264 1264 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE);
1265 1265
1266 1266 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
1267 1267 buf_hash_remove(hdr);
1268 1268
1269 1269 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
1270 1270
1271 1271 if (new == hdr_full_cache) {
1272 1272 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1273 1273 /*
1274 1274 * arc_access and arc_change_state need to be aware that a
1275 1275 * header has just come out of L2ARC, so we set its state to
1276 1276 * l2c_only even though it's about to change.
1277 1277 */
1278 1278 nhdr->b_l1hdr.b_state = arc_l2c_only;
1279 1279
1280 1280 /* Verify previous threads set to NULL before freeing */
1281 1281 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1282 1282 } else {
1283 1283 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1284 1284 ASSERT0(hdr->b_l1hdr.b_datacnt);
1285 1285
1286 1286 /*
1287 1287 * If we've reached here, We must have been called from
1288 1288 * arc_evict_hdr(), as such we should have already been
1289 1289 * removed from any ghost list we were previously on
1290 1290 * (which protects us from racing with arc_evict_state),
1291 1291 * thus no locking is needed during this check.
1292 1292 */
1293 1293 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
|
↓ open down ↓ |
1293 lines elided |
↑ open up ↑ |
1294 1294
1295 1295 /*
1296 1296 * A buffer must not be moved into the arc_l2c_only
1297 1297 * state if it's not finished being written out to the
1298 1298 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field
1299 1299 * might try to be accessed, even though it was removed.
1300 1300 */
1301 1301 VERIFY(!HDR_L2_WRITING(hdr));
1302 1302 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
1303 1303
1304 +#ifdef ZFS_DEBUG
1305 + if (hdr->b_l1hdr.b_thawed != NULL) {
1306 + kmem_free(hdr->b_l1hdr.b_thawed, 1);
1307 + hdr->b_l1hdr.b_thawed = NULL;
1308 + }
1309 +#endif
1310 +
1304 1311 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR;
1305 1312 }
1306 1313 /*
1307 1314 * The header has been reallocated so we need to re-insert it into any
1308 1315 * lists it was on.
1309 1316 */
1310 1317 (void) buf_hash_insert(nhdr, NULL);
1311 1318
1312 1319 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node));
1313 1320
1314 1321 mutex_enter(&dev->l2ad_mtx);
1315 1322
1316 1323 /*
1317 1324 * We must place the realloc'ed header back into the list at
1318 1325 * the same spot. Otherwise, if it's placed earlier in the list,
1319 1326 * l2arc_write_buffers() could find it during the function's
1320 1327 * write phase, and try to write it out to the l2arc.
1321 1328 */
1322 1329 list_insert_after(&dev->l2ad_buflist, hdr, nhdr);
1323 1330 list_remove(&dev->l2ad_buflist, hdr);
1324 1331
1325 1332 mutex_exit(&dev->l2ad_mtx);
1326 1333
1327 1334 /*
1328 1335 * Since we're using the pointer address as the tag when
1329 1336 * incrementing and decrementing the l2ad_alloc refcount, we
1330 1337 * must remove the old pointer (that we're about to destroy) and
1331 1338 * add the new pointer to the refcount. Otherwise we'd remove
1332 1339 * the wrong pointer address when calling arc_hdr_destroy() later.
1333 1340 */
1334 1341
1335 1342 (void) refcount_remove_many(&dev->l2ad_alloc,
1336 1343 hdr->b_l2hdr.b_asize, hdr);
1337 1344
1338 1345 (void) refcount_add_many(&dev->l2ad_alloc,
1339 1346 nhdr->b_l2hdr.b_asize, nhdr);
1340 1347
1341 1348 buf_discard_identity(hdr);
1342 1349 hdr->b_freeze_cksum = NULL;
1343 1350 kmem_cache_free(old, hdr);
1344 1351
1345 1352 return (nhdr);
1346 1353 }
1347 1354
1348 1355
1349 1356 #define ARC_MINTIME (hz>>4) /* 62 ms */
1350 1357
1351 1358 static void
1352 1359 arc_cksum_verify(arc_buf_t *buf)
1353 1360 {
1354 1361 zio_cksum_t zc;
1355 1362
1356 1363 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1357 1364 return;
1358 1365
1359 1366 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1360 1367 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) {
1361 1368 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1362 1369 return;
1363 1370 }
1364 1371 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1365 1372 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1366 1373 panic("buffer modified while frozen!");
1367 1374 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1368 1375 }
1369 1376
1370 1377 static int
1371 1378 arc_cksum_equal(arc_buf_t *buf)
1372 1379 {
1373 1380 zio_cksum_t zc;
1374 1381 int equal;
1375 1382
1376 1383 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1377 1384 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1378 1385 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1379 1386 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1380 1387
1381 1388 return (equal);
1382 1389 }
1383 1390
1384 1391 static void
1385 1392 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1386 1393 {
1387 1394 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1388 1395 return;
1389 1396
1390 1397 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1391 1398 if (buf->b_hdr->b_freeze_cksum != NULL) {
1392 1399 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1393 1400 return;
1394 1401 }
1395 1402 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1396 1403 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1397 1404 buf->b_hdr->b_freeze_cksum);
1398 1405 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1399 1406 arc_buf_watch(buf);
1400 1407 }
1401 1408
1402 1409 #ifndef _KERNEL
1403 1410 typedef struct procctl {
1404 1411 long cmd;
1405 1412 prwatch_t prwatch;
1406 1413 } procctl_t;
1407 1414 #endif
1408 1415
1409 1416 /* ARGSUSED */
1410 1417 static void
1411 1418 arc_buf_unwatch(arc_buf_t *buf)
1412 1419 {
1413 1420 #ifndef _KERNEL
1414 1421 if (arc_watch) {
1415 1422 int result;
1416 1423 procctl_t ctl;
1417 1424 ctl.cmd = PCWATCH;
1418 1425 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1419 1426 ctl.prwatch.pr_size = 0;
1420 1427 ctl.prwatch.pr_wflags = 0;
1421 1428 result = write(arc_procfd, &ctl, sizeof (ctl));
1422 1429 ASSERT3U(result, ==, sizeof (ctl));
1423 1430 }
1424 1431 #endif
1425 1432 }
1426 1433
1427 1434 /* ARGSUSED */
1428 1435 static void
1429 1436 arc_buf_watch(arc_buf_t *buf)
1430 1437 {
1431 1438 #ifndef _KERNEL
1432 1439 if (arc_watch) {
1433 1440 int result;
1434 1441 procctl_t ctl;
1435 1442 ctl.cmd = PCWATCH;
1436 1443 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1437 1444 ctl.prwatch.pr_size = buf->b_hdr->b_size;
1438 1445 ctl.prwatch.pr_wflags = WA_WRITE;
1439 1446 result = write(arc_procfd, &ctl, sizeof (ctl));
1440 1447 ASSERT3U(result, ==, sizeof (ctl));
1441 1448 }
1442 1449 #endif
1443 1450 }
1444 1451
1445 1452 static arc_buf_contents_t
1446 1453 arc_buf_type(arc_buf_hdr_t *hdr)
1447 1454 {
1448 1455 if (HDR_ISTYPE_METADATA(hdr)) {
1449 1456 return (ARC_BUFC_METADATA);
1450 1457 } else {
1451 1458 return (ARC_BUFC_DATA);
1452 1459 }
1453 1460 }
1454 1461
1455 1462 static uint32_t
1456 1463 arc_bufc_to_flags(arc_buf_contents_t type)
1457 1464 {
1458 1465 switch (type) {
1459 1466 case ARC_BUFC_DATA:
1460 1467 /* metadata field is 0 if buffer contains normal data */
1461 1468 return (0);
1462 1469 case ARC_BUFC_METADATA:
1463 1470 return (ARC_FLAG_BUFC_METADATA);
1464 1471 default:
1465 1472 break;
1466 1473 }
1467 1474 panic("undefined ARC buffer type!");
1468 1475 return ((uint32_t)-1);
1469 1476 }
1470 1477
1471 1478 void
1472 1479 arc_buf_thaw(arc_buf_t *buf)
1473 1480 {
1474 1481 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1475 1482 if (buf->b_hdr->b_l1hdr.b_state != arc_anon)
1476 1483 panic("modifying non-anon buffer!");
1477 1484 if (HDR_IO_IN_PROGRESS(buf->b_hdr))
1478 1485 panic("modifying buffer while i/o in progress!");
1479 1486 arc_cksum_verify(buf);
1480 1487 }
1481 1488
1482 1489 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1483 1490 if (buf->b_hdr->b_freeze_cksum != NULL) {
1484 1491 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1485 1492 buf->b_hdr->b_freeze_cksum = NULL;
1486 1493 }
1487 1494
1488 1495 #ifdef ZFS_DEBUG
1489 1496 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1490 1497 if (buf->b_hdr->b_l1hdr.b_thawed != NULL)
1491 1498 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1);
1492 1499 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1493 1500 }
1494 1501 #endif
1495 1502
1496 1503 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1497 1504
1498 1505 arc_buf_unwatch(buf);
1499 1506 }
1500 1507
1501 1508 void
1502 1509 arc_buf_freeze(arc_buf_t *buf)
1503 1510 {
1504 1511 kmutex_t *hash_lock;
1505 1512
1506 1513 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1507 1514 return;
1508 1515
1509 1516 hash_lock = HDR_LOCK(buf->b_hdr);
1510 1517 mutex_enter(hash_lock);
1511 1518
1512 1519 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1513 1520 buf->b_hdr->b_l1hdr.b_state == arc_anon);
1514 1521 arc_cksum_compute(buf, B_FALSE);
1515 1522 mutex_exit(hash_lock);
1516 1523
1517 1524 }
1518 1525
1519 1526 static void
1520 1527 add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1521 1528 {
1522 1529 ASSERT(HDR_HAS_L1HDR(hdr));
1523 1530 ASSERT(MUTEX_HELD(hash_lock));
1524 1531 arc_state_t *state = hdr->b_l1hdr.b_state;
1525 1532
1526 1533 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) &&
1527 1534 (state != arc_anon)) {
1528 1535 /* We don't use the L2-only state list. */
1529 1536 if (state != arc_l2c_only) {
1530 1537 arc_buf_contents_t type = arc_buf_type(hdr);
1531 1538 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt;
1532 1539 multilist_t *list = &state->arcs_list[type];
1533 1540 uint64_t *size = &state->arcs_lsize[type];
1534 1541
1535 1542 multilist_remove(list, hdr);
1536 1543
1537 1544 if (GHOST_STATE(state)) {
1538 1545 ASSERT0(hdr->b_l1hdr.b_datacnt);
1539 1546 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
1540 1547 delta = hdr->b_size;
1541 1548 }
1542 1549 ASSERT(delta > 0);
1543 1550 ASSERT3U(*size, >=, delta);
1544 1551 atomic_add_64(size, -delta);
1545 1552 }
1546 1553 /* remove the prefetch flag if we get a reference */
1547 1554 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
1548 1555 }
1549 1556 }
1550 1557
1551 1558 static int
1552 1559 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag)
1553 1560 {
1554 1561 int cnt;
1555 1562 arc_state_t *state = hdr->b_l1hdr.b_state;
1556 1563
1557 1564 ASSERT(HDR_HAS_L1HDR(hdr));
1558 1565 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1559 1566 ASSERT(!GHOST_STATE(state));
1560 1567
1561 1568 /*
1562 1569 * arc_l2c_only counts as a ghost state so we don't need to explicitly
1563 1570 * check to prevent usage of the arc_l2c_only list.
1564 1571 */
1565 1572 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
1566 1573 (state != arc_anon)) {
1567 1574 arc_buf_contents_t type = arc_buf_type(hdr);
1568 1575 multilist_t *list = &state->arcs_list[type];
1569 1576 uint64_t *size = &state->arcs_lsize[type];
1570 1577
1571 1578 multilist_insert(list, hdr);
1572 1579
1573 1580 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
1574 1581 atomic_add_64(size, hdr->b_size *
1575 1582 hdr->b_l1hdr.b_datacnt);
1576 1583 }
1577 1584 return (cnt);
1578 1585 }
1579 1586
1580 1587 /*
1581 1588 * Move the supplied buffer to the indicated state. The hash lock
1582 1589 * for the buffer must be held by the caller.
1583 1590 */
1584 1591 static void
1585 1592 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr,
1586 1593 kmutex_t *hash_lock)
1587 1594 {
1588 1595 arc_state_t *old_state;
1589 1596 int64_t refcnt;
1590 1597 uint32_t datacnt;
1591 1598 uint64_t from_delta, to_delta;
1592 1599 arc_buf_contents_t buftype = arc_buf_type(hdr);
1593 1600
1594 1601 /*
1595 1602 * We almost always have an L1 hdr here, since we call arc_hdr_realloc()
1596 1603 * in arc_read() when bringing a buffer out of the L2ARC. However, the
1597 1604 * L1 hdr doesn't always exist when we change state to arc_anon before
1598 1605 * destroying a header, in which case reallocating to add the L1 hdr is
1599 1606 * pointless.
1600 1607 */
1601 1608 if (HDR_HAS_L1HDR(hdr)) {
1602 1609 old_state = hdr->b_l1hdr.b_state;
1603 1610 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt);
1604 1611 datacnt = hdr->b_l1hdr.b_datacnt;
1605 1612 } else {
1606 1613 old_state = arc_l2c_only;
1607 1614 refcnt = 0;
1608 1615 datacnt = 0;
1609 1616 }
1610 1617
1611 1618 ASSERT(MUTEX_HELD(hash_lock));
1612 1619 ASSERT3P(new_state, !=, old_state);
1613 1620 ASSERT(refcnt == 0 || datacnt > 0);
1614 1621 ASSERT(!GHOST_STATE(new_state) || datacnt == 0);
1615 1622 ASSERT(old_state != arc_anon || datacnt <= 1);
1616 1623
1617 1624 from_delta = to_delta = datacnt * hdr->b_size;
1618 1625
1619 1626 /*
1620 1627 * If this buffer is evictable, transfer it from the
1621 1628 * old state list to the new state list.
1622 1629 */
1623 1630 if (refcnt == 0) {
1624 1631 if (old_state != arc_anon && old_state != arc_l2c_only) {
1625 1632 uint64_t *size = &old_state->arcs_lsize[buftype];
1626 1633
1627 1634 ASSERT(HDR_HAS_L1HDR(hdr));
1628 1635 multilist_remove(&old_state->arcs_list[buftype], hdr);
1629 1636
1630 1637 /*
1631 1638 * If prefetching out of the ghost cache,
1632 1639 * we will have a non-zero datacnt.
1633 1640 */
1634 1641 if (GHOST_STATE(old_state) && datacnt == 0) {
1635 1642 /* ghost elements have a ghost size */
1636 1643 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1637 1644 from_delta = hdr->b_size;
1638 1645 }
1639 1646 ASSERT3U(*size, >=, from_delta);
1640 1647 atomic_add_64(size, -from_delta);
1641 1648 }
1642 1649 if (new_state != arc_anon && new_state != arc_l2c_only) {
1643 1650 uint64_t *size = &new_state->arcs_lsize[buftype];
1644 1651
1645 1652 /*
1646 1653 * An L1 header always exists here, since if we're
1647 1654 * moving to some L1-cached state (i.e. not l2c_only or
1648 1655 * anonymous), we realloc the header to add an L1hdr
1649 1656 * beforehand.
1650 1657 */
1651 1658 ASSERT(HDR_HAS_L1HDR(hdr));
1652 1659 multilist_insert(&new_state->arcs_list[buftype], hdr);
1653 1660
1654 1661 /* ghost elements have a ghost size */
1655 1662 if (GHOST_STATE(new_state)) {
1656 1663 ASSERT0(datacnt);
1657 1664 ASSERT(hdr->b_l1hdr.b_buf == NULL);
1658 1665 to_delta = hdr->b_size;
1659 1666 }
1660 1667 atomic_add_64(size, to_delta);
1661 1668 }
1662 1669 }
1663 1670
1664 1671 ASSERT(!BUF_EMPTY(hdr));
1665 1672 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
1666 1673 buf_hash_remove(hdr);
1667 1674
1668 1675 /* adjust state sizes (ignore arc_l2c_only) */
1669 1676
1670 1677 if (to_delta && new_state != arc_l2c_only) {
1671 1678 ASSERT(HDR_HAS_L1HDR(hdr));
1672 1679 if (GHOST_STATE(new_state)) {
1673 1680 ASSERT0(datacnt);
1674 1681
1675 1682 /*
1676 1683 * We moving a header to a ghost state, we first
1677 1684 * remove all arc buffers. Thus, we'll have a
1678 1685 * datacnt of zero, and no arc buffer to use for
1679 1686 * the reference. As a result, we use the arc
1680 1687 * header pointer for the reference.
1681 1688 */
1682 1689 (void) refcount_add_many(&new_state->arcs_size,
1683 1690 hdr->b_size, hdr);
1684 1691 } else {
1685 1692 ASSERT3U(datacnt, !=, 0);
1686 1693
1687 1694 /*
1688 1695 * Each individual buffer holds a unique reference,
1689 1696 * thus we must remove each of these references one
1690 1697 * at a time.
1691 1698 */
1692 1699 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1693 1700 buf = buf->b_next) {
1694 1701 (void) refcount_add_many(&new_state->arcs_size,
1695 1702 hdr->b_size, buf);
1696 1703 }
1697 1704 }
1698 1705 }
1699 1706
1700 1707 if (from_delta && old_state != arc_l2c_only) {
1701 1708 ASSERT(HDR_HAS_L1HDR(hdr));
1702 1709 if (GHOST_STATE(old_state)) {
1703 1710 /*
1704 1711 * When moving a header off of a ghost state,
1705 1712 * there's the possibility for datacnt to be
1706 1713 * non-zero. This is because we first add the
1707 1714 * arc buffer to the header prior to changing
1708 1715 * the header's state. Since we used the header
1709 1716 * for the reference when putting the header on
1710 1717 * the ghost state, we must balance that and use
1711 1718 * the header when removing off the ghost state
1712 1719 * (even though datacnt is non zero).
1713 1720 */
1714 1721
1715 1722 IMPLY(datacnt == 0, new_state == arc_anon ||
1716 1723 new_state == arc_l2c_only);
1717 1724
1718 1725 (void) refcount_remove_many(&old_state->arcs_size,
1719 1726 hdr->b_size, hdr);
1720 1727 } else {
1721 1728 ASSERT3P(datacnt, !=, 0);
1722 1729
1723 1730 /*
1724 1731 * Each individual buffer holds a unique reference,
1725 1732 * thus we must remove each of these references one
1726 1733 * at a time.
1727 1734 */
1728 1735 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL;
1729 1736 buf = buf->b_next) {
1730 1737 (void) refcount_remove_many(
1731 1738 &old_state->arcs_size, hdr->b_size, buf);
1732 1739 }
1733 1740 }
1734 1741 }
1735 1742
1736 1743 if (HDR_HAS_L1HDR(hdr))
1737 1744 hdr->b_l1hdr.b_state = new_state;
1738 1745
1739 1746 /*
1740 1747 * L2 headers should never be on the L2 state list since they don't
1741 1748 * have L1 headers allocated.
1742 1749 */
1743 1750 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
1744 1751 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
1745 1752 }
1746 1753
1747 1754 void
1748 1755 arc_space_consume(uint64_t space, arc_space_type_t type)
1749 1756 {
1750 1757 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1751 1758
1752 1759 switch (type) {
1753 1760 case ARC_SPACE_DATA:
1754 1761 ARCSTAT_INCR(arcstat_data_size, space);
1755 1762 break;
1756 1763 case ARC_SPACE_META:
1757 1764 ARCSTAT_INCR(arcstat_metadata_size, space);
1758 1765 break;
1759 1766 case ARC_SPACE_OTHER:
1760 1767 ARCSTAT_INCR(arcstat_other_size, space);
1761 1768 break;
1762 1769 case ARC_SPACE_HDRS:
1763 1770 ARCSTAT_INCR(arcstat_hdr_size, space);
1764 1771 break;
1765 1772 case ARC_SPACE_L2HDRS:
1766 1773 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1767 1774 break;
1768 1775 }
1769 1776
1770 1777 if (type != ARC_SPACE_DATA)
1771 1778 ARCSTAT_INCR(arcstat_meta_used, space);
1772 1779
1773 1780 atomic_add_64(&arc_size, space);
1774 1781 }
1775 1782
1776 1783 void
1777 1784 arc_space_return(uint64_t space, arc_space_type_t type)
1778 1785 {
1779 1786 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1780 1787
1781 1788 switch (type) {
1782 1789 case ARC_SPACE_DATA:
1783 1790 ARCSTAT_INCR(arcstat_data_size, -space);
1784 1791 break;
1785 1792 case ARC_SPACE_META:
1786 1793 ARCSTAT_INCR(arcstat_metadata_size, -space);
1787 1794 break;
1788 1795 case ARC_SPACE_OTHER:
1789 1796 ARCSTAT_INCR(arcstat_other_size, -space);
1790 1797 break;
1791 1798 case ARC_SPACE_HDRS:
1792 1799 ARCSTAT_INCR(arcstat_hdr_size, -space);
1793 1800 break;
1794 1801 case ARC_SPACE_L2HDRS:
1795 1802 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1796 1803 break;
1797 1804 }
1798 1805
1799 1806 if (type != ARC_SPACE_DATA) {
1800 1807 ASSERT(arc_meta_used >= space);
1801 1808 if (arc_meta_max < arc_meta_used)
1802 1809 arc_meta_max = arc_meta_used;
1803 1810 ARCSTAT_INCR(arcstat_meta_used, -space);
1804 1811 }
1805 1812
1806 1813 ASSERT(arc_size >= space);
1807 1814 atomic_add_64(&arc_size, -space);
1808 1815 }
1809 1816
1810 1817 arc_buf_t *
1811 1818 arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type)
1812 1819 {
1813 1820 arc_buf_hdr_t *hdr;
1814 1821 arc_buf_t *buf;
1815 1822
1816 1823 ASSERT3U(size, >, 0);
1817 1824 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
1818 1825 ASSERT(BUF_EMPTY(hdr));
1819 1826 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
1820 1827 hdr->b_size = size;
1821 1828 hdr->b_spa = spa_load_guid(spa);
1822 1829
1823 1830 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1824 1831 buf->b_hdr = hdr;
1825 1832 buf->b_data = NULL;
1826 1833 buf->b_efunc = NULL;
1827 1834 buf->b_private = NULL;
1828 1835 buf->b_next = NULL;
1829 1836
1830 1837 hdr->b_flags = arc_bufc_to_flags(type);
1831 1838 hdr->b_flags |= ARC_FLAG_HAS_L1HDR;
1832 1839
1833 1840 hdr->b_l1hdr.b_buf = buf;
1834 1841 hdr->b_l1hdr.b_state = arc_anon;
1835 1842 hdr->b_l1hdr.b_arc_access = 0;
1836 1843 hdr->b_l1hdr.b_datacnt = 1;
1837 1844 hdr->b_l1hdr.b_tmp_cdata = NULL;
1838 1845
1839 1846 arc_get_data_buf(buf);
1840 1847 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
1841 1848 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
1842 1849
1843 1850 return (buf);
1844 1851 }
1845 1852
1846 1853 static char *arc_onloan_tag = "onloan";
1847 1854
1848 1855 /*
1849 1856 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
1850 1857 * flight data by arc_tempreserve_space() until they are "returned". Loaned
1851 1858 * buffers must be returned to the arc before they can be used by the DMU or
1852 1859 * freed.
1853 1860 */
1854 1861 arc_buf_t *
1855 1862 arc_loan_buf(spa_t *spa, int size)
1856 1863 {
1857 1864 arc_buf_t *buf;
1858 1865
1859 1866 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA);
1860 1867
1861 1868 atomic_add_64(&arc_loaned_bytes, size);
1862 1869 return (buf);
1863 1870 }
1864 1871
1865 1872 /*
1866 1873 * Return a loaned arc buffer to the arc.
1867 1874 */
1868 1875 void
1869 1876 arc_return_buf(arc_buf_t *buf, void *tag)
1870 1877 {
1871 1878 arc_buf_hdr_t *hdr = buf->b_hdr;
1872 1879
1873 1880 ASSERT(buf->b_data != NULL);
1874 1881 ASSERT(HDR_HAS_L1HDR(hdr));
1875 1882 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag);
1876 1883 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
1877 1884
1878 1885 atomic_add_64(&arc_loaned_bytes, -hdr->b_size);
1879 1886 }
1880 1887
1881 1888 /* Detach an arc_buf from a dbuf (tag) */
1882 1889 void
1883 1890 arc_loan_inuse_buf(arc_buf_t *buf, void *tag)
1884 1891 {
1885 1892 arc_buf_hdr_t *hdr = buf->b_hdr;
1886 1893
1887 1894 ASSERT(buf->b_data != NULL);
1888 1895 ASSERT(HDR_HAS_L1HDR(hdr));
1889 1896 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag);
1890 1897 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag);
1891 1898 buf->b_efunc = NULL;
1892 1899 buf->b_private = NULL;
1893 1900
1894 1901 atomic_add_64(&arc_loaned_bytes, hdr->b_size);
1895 1902 }
1896 1903
1897 1904 static arc_buf_t *
1898 1905 arc_buf_clone(arc_buf_t *from)
1899 1906 {
1900 1907 arc_buf_t *buf;
1901 1908 arc_buf_hdr_t *hdr = from->b_hdr;
1902 1909 uint64_t size = hdr->b_size;
1903 1910
1904 1911 ASSERT(HDR_HAS_L1HDR(hdr));
1905 1912 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
1906 1913
1907 1914 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1908 1915 buf->b_hdr = hdr;
1909 1916 buf->b_data = NULL;
1910 1917 buf->b_efunc = NULL;
1911 1918 buf->b_private = NULL;
1912 1919 buf->b_next = hdr->b_l1hdr.b_buf;
1913 1920 hdr->b_l1hdr.b_buf = buf;
1914 1921 arc_get_data_buf(buf);
1915 1922 bcopy(from->b_data, buf->b_data, size);
1916 1923
1917 1924 /*
1918 1925 * This buffer already exists in the arc so create a duplicate
1919 1926 * copy for the caller. If the buffer is associated with user data
1920 1927 * then track the size and number of duplicates. These stats will be
1921 1928 * updated as duplicate buffers are created and destroyed.
1922 1929 */
1923 1930 if (HDR_ISTYPE_DATA(hdr)) {
1924 1931 ARCSTAT_BUMP(arcstat_duplicate_buffers);
1925 1932 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size);
1926 1933 }
1927 1934 hdr->b_l1hdr.b_datacnt += 1;
1928 1935 return (buf);
1929 1936 }
1930 1937
1931 1938 void
1932 1939 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1933 1940 {
1934 1941 arc_buf_hdr_t *hdr;
1935 1942 kmutex_t *hash_lock;
1936 1943
1937 1944 /*
1938 1945 * Check to see if this buffer is evicted. Callers
1939 1946 * must verify b_data != NULL to know if the add_ref
1940 1947 * was successful.
1941 1948 */
1942 1949 mutex_enter(&buf->b_evict_lock);
1943 1950 if (buf->b_data == NULL) {
1944 1951 mutex_exit(&buf->b_evict_lock);
1945 1952 return;
1946 1953 }
1947 1954 hash_lock = HDR_LOCK(buf->b_hdr);
1948 1955 mutex_enter(hash_lock);
1949 1956 hdr = buf->b_hdr;
1950 1957 ASSERT(HDR_HAS_L1HDR(hdr));
1951 1958 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
1952 1959 mutex_exit(&buf->b_evict_lock);
1953 1960
1954 1961 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
1955 1962 hdr->b_l1hdr.b_state == arc_mfu);
1956 1963
1957 1964 add_reference(hdr, hash_lock, tag);
1958 1965 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1959 1966 arc_access(hdr, hash_lock);
1960 1967 mutex_exit(hash_lock);
1961 1968 ARCSTAT_BUMP(arcstat_hits);
1962 1969 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
1963 1970 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
1964 1971 data, metadata, hits);
1965 1972 }
1966 1973
1967 1974 static void
1968 1975 arc_buf_free_on_write(void *data, size_t size,
1969 1976 void (*free_func)(void *, size_t))
1970 1977 {
1971 1978 l2arc_data_free_t *df;
1972 1979
1973 1980 df = kmem_alloc(sizeof (*df), KM_SLEEP);
1974 1981 df->l2df_data = data;
1975 1982 df->l2df_size = size;
1976 1983 df->l2df_func = free_func;
1977 1984 mutex_enter(&l2arc_free_on_write_mtx);
1978 1985 list_insert_head(l2arc_free_on_write, df);
1979 1986 mutex_exit(&l2arc_free_on_write_mtx);
1980 1987 }
1981 1988
1982 1989 /*
1983 1990 * Free the arc data buffer. If it is an l2arc write in progress,
1984 1991 * the buffer is placed on l2arc_free_on_write to be freed later.
1985 1992 */
1986 1993 static void
1987 1994 arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t))
1988 1995 {
1989 1996 arc_buf_hdr_t *hdr = buf->b_hdr;
1990 1997
1991 1998 if (HDR_L2_WRITING(hdr)) {
1992 1999 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func);
1993 2000 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1994 2001 } else {
1995 2002 free_func(buf->b_data, hdr->b_size);
1996 2003 }
1997 2004 }
1998 2005
1999 2006 static void
2000 2007 arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr)
2001 2008 {
2002 2009 ASSERT(HDR_HAS_L2HDR(hdr));
2003 2010 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx));
2004 2011
2005 2012 /*
2006 2013 * The b_tmp_cdata field is linked off of the b_l1hdr, so if
2007 2014 * that doesn't exist, the header is in the arc_l2c_only state,
2008 2015 * and there isn't anything to free (it's already been freed).
2009 2016 */
2010 2017 if (!HDR_HAS_L1HDR(hdr))
2011 2018 return;
2012 2019
2013 2020 /*
2014 2021 * The header isn't being written to the l2arc device, thus it
2015 2022 * shouldn't have a b_tmp_cdata to free.
2016 2023 */
2017 2024 if (!HDR_L2_WRITING(hdr)) {
2018 2025 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2019 2026 return;
2020 2027 }
2021 2028
2022 2029 /*
2023 2030 * The header does not have compression enabled. This can be due
2024 2031 * to the buffer not being compressible, or because we're
2025 2032 * freeing the buffer before the second phase of
2026 2033 * l2arc_write_buffer() has started (which does the compression
2027 2034 * step). In either case, b_tmp_cdata does not point to a
2028 2035 * separately compressed buffer, so there's nothing to free (it
2029 2036 * points to the same buffer as the arc_buf_t's b_data field).
2030 2037 */
2031 2038 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_OFF) {
2032 2039 hdr->b_l1hdr.b_tmp_cdata = NULL;
2033 2040 return;
2034 2041 }
2035 2042
2036 2043 /*
2037 2044 * There's nothing to free since the buffer was all zero's and
2038 2045 * compressed to a zero length buffer.
2039 2046 */
2040 2047 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_EMPTY) {
2041 2048 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
2042 2049 return;
2043 2050 }
2044 2051
2045 2052 ASSERT(L2ARC_IS_VALID_COMPRESS(hdr->b_l2hdr.b_compress));
2046 2053
2047 2054 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata,
2048 2055 hdr->b_size, zio_data_buf_free);
2049 2056
2050 2057 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write);
2051 2058 hdr->b_l1hdr.b_tmp_cdata = NULL;
2052 2059 }
2053 2060
2054 2061 /*
2055 2062 * Free up buf->b_data and if 'remove' is set, then pull the
2056 2063 * arc_buf_t off of the the arc_buf_hdr_t's list and free it.
2057 2064 */
2058 2065 static void
2059 2066 arc_buf_destroy(arc_buf_t *buf, boolean_t remove)
2060 2067 {
2061 2068 arc_buf_t **bufp;
2062 2069
2063 2070 /* free up data associated with the buf */
2064 2071 if (buf->b_data != NULL) {
2065 2072 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
2066 2073 uint64_t size = buf->b_hdr->b_size;
2067 2074 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
2068 2075
2069 2076 arc_cksum_verify(buf);
2070 2077 arc_buf_unwatch(buf);
2071 2078
2072 2079 if (type == ARC_BUFC_METADATA) {
2073 2080 arc_buf_data_free(buf, zio_buf_free);
2074 2081 arc_space_return(size, ARC_SPACE_META);
2075 2082 } else {
2076 2083 ASSERT(type == ARC_BUFC_DATA);
2077 2084 arc_buf_data_free(buf, zio_data_buf_free);
2078 2085 arc_space_return(size, ARC_SPACE_DATA);
2079 2086 }
2080 2087
2081 2088 /* protected by hash lock, if in the hash table */
2082 2089 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) {
2083 2090 uint64_t *cnt = &state->arcs_lsize[type];
2084 2091
2085 2092 ASSERT(refcount_is_zero(
2086 2093 &buf->b_hdr->b_l1hdr.b_refcnt));
2087 2094 ASSERT(state != arc_anon && state != arc_l2c_only);
2088 2095
2089 2096 ASSERT3U(*cnt, >=, size);
2090 2097 atomic_add_64(cnt, -size);
2091 2098 }
2092 2099
2093 2100 (void) refcount_remove_many(&state->arcs_size, size, buf);
2094 2101 buf->b_data = NULL;
2095 2102
2096 2103 /*
2097 2104 * If we're destroying a duplicate buffer make sure
2098 2105 * that the appropriate statistics are updated.
2099 2106 */
2100 2107 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 &&
2101 2108 HDR_ISTYPE_DATA(buf->b_hdr)) {
2102 2109 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
2103 2110 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size);
2104 2111 }
2105 2112 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0);
2106 2113 buf->b_hdr->b_l1hdr.b_datacnt -= 1;
2107 2114 }
2108 2115
2109 2116 /* only remove the buf if requested */
2110 2117 if (!remove)
2111 2118 return;
2112 2119
2113 2120 /* remove the buf from the hdr list */
2114 2121 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf;
2115 2122 bufp = &(*bufp)->b_next)
2116 2123 continue;
2117 2124 *bufp = buf->b_next;
2118 2125 buf->b_next = NULL;
2119 2126
2120 2127 ASSERT(buf->b_efunc == NULL);
2121 2128
2122 2129 /* clean up the buf */
2123 2130 buf->b_hdr = NULL;
2124 2131 kmem_cache_free(buf_cache, buf);
2125 2132 }
2126 2133
2127 2134 static void
2128 2135 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
2129 2136 {
2130 2137 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
2131 2138 l2arc_dev_t *dev = l2hdr->b_dev;
2132 2139
2133 2140 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
2134 2141 ASSERT(HDR_HAS_L2HDR(hdr));
2135 2142
2136 2143 list_remove(&dev->l2ad_buflist, hdr);
2137 2144
2138 2145 /*
2139 2146 * We don't want to leak the b_tmp_cdata buffer that was
2140 2147 * allocated in l2arc_write_buffers()
2141 2148 */
2142 2149 arc_buf_l2_cdata_free(hdr);
2143 2150
2144 2151 /*
2145 2152 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then
2146 2153 * this header is being processed by l2arc_write_buffers() (i.e.
2147 2154 * it's in the first stage of l2arc_write_buffers()).
2148 2155 * Re-affirming that truth here, just to serve as a reminder. If
2149 2156 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or
2150 2157 * may not have its HDR_L2_WRITING flag set. (the write may have
2151 2158 * completed, in which case HDR_L2_WRITING will be false and the
2152 2159 * b_daddr field will point to the address of the buffer on disk).
2153 2160 */
2154 2161 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr));
2155 2162
2156 2163 /*
2157 2164 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with
2158 2165 * l2arc_write_buffers(). Since we've just removed this header
2159 2166 * from the l2arc buffer list, this header will never reach the
2160 2167 * second stage of l2arc_write_buffers(), which increments the
2161 2168 * accounting stats for this header. Thus, we must be careful
2162 2169 * not to decrement them for this header either.
2163 2170 */
2164 2171 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) {
2165 2172 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize);
2166 2173 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
2167 2174
2168 2175 vdev_space_update(dev->l2ad_vdev,
2169 2176 -l2hdr->b_asize, 0, 0);
2170 2177
2171 2178 (void) refcount_remove_many(&dev->l2ad_alloc,
2172 2179 l2hdr->b_asize, hdr);
2173 2180 }
2174 2181
2175 2182 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
2176 2183 }
2177 2184
2178 2185 static void
2179 2186 arc_hdr_destroy(arc_buf_hdr_t *hdr)
2180 2187 {
2181 2188 if (HDR_HAS_L1HDR(hdr)) {
2182 2189 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
2183 2190 hdr->b_l1hdr.b_datacnt > 0);
2184 2191 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2185 2192 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2186 2193 }
2187 2194 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2188 2195 ASSERT(!HDR_IN_HASH_TABLE(hdr));
2189 2196
2190 2197 if (HDR_HAS_L2HDR(hdr)) {
2191 2198 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
2192 2199 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
2193 2200
2194 2201 if (!buflist_held)
2195 2202 mutex_enter(&dev->l2ad_mtx);
2196 2203
2197 2204 /*
2198 2205 * Even though we checked this conditional above, we
2199 2206 * need to check this again now that we have the
2200 2207 * l2ad_mtx. This is because we could be racing with
2201 2208 * another thread calling l2arc_evict() which might have
2202 2209 * destroyed this header's L2 portion as we were waiting
2203 2210 * to acquire the l2ad_mtx. If that happens, we don't
2204 2211 * want to re-destroy the header's L2 portion.
2205 2212 */
2206 2213 if (HDR_HAS_L2HDR(hdr))
2207 2214 arc_hdr_l2hdr_destroy(hdr);
2208 2215
2209 2216 if (!buflist_held)
2210 2217 mutex_exit(&dev->l2ad_mtx);
2211 2218 }
2212 2219
2213 2220 if (!BUF_EMPTY(hdr))
2214 2221 buf_discard_identity(hdr);
2215 2222
2216 2223 if (hdr->b_freeze_cksum != NULL) {
2217 2224 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
2218 2225 hdr->b_freeze_cksum = NULL;
2219 2226 }
2220 2227
2221 2228 if (HDR_HAS_L1HDR(hdr)) {
2222 2229 while (hdr->b_l1hdr.b_buf) {
2223 2230 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2224 2231
2225 2232 if (buf->b_efunc != NULL) {
2226 2233 mutex_enter(&arc_user_evicts_lock);
2227 2234 mutex_enter(&buf->b_evict_lock);
2228 2235 ASSERT(buf->b_hdr != NULL);
2229 2236 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE);
2230 2237 hdr->b_l1hdr.b_buf = buf->b_next;
2231 2238 buf->b_hdr = &arc_eviction_hdr;
2232 2239 buf->b_next = arc_eviction_list;
2233 2240 arc_eviction_list = buf;
2234 2241 mutex_exit(&buf->b_evict_lock);
2235 2242 cv_signal(&arc_user_evicts_cv);
2236 2243 mutex_exit(&arc_user_evicts_lock);
2237 2244 } else {
2238 2245 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE);
2239 2246 }
2240 2247 }
2241 2248 #ifdef ZFS_DEBUG
2242 2249 if (hdr->b_l1hdr.b_thawed != NULL) {
2243 2250 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2244 2251 hdr->b_l1hdr.b_thawed = NULL;
2245 2252 }
2246 2253 #endif
2247 2254 }
2248 2255
2249 2256 ASSERT3P(hdr->b_hash_next, ==, NULL);
2250 2257 if (HDR_HAS_L1HDR(hdr)) {
2251 2258 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2252 2259 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
2253 2260 kmem_cache_free(hdr_full_cache, hdr);
2254 2261 } else {
2255 2262 kmem_cache_free(hdr_l2only_cache, hdr);
2256 2263 }
2257 2264 }
2258 2265
2259 2266 void
2260 2267 arc_buf_free(arc_buf_t *buf, void *tag)
2261 2268 {
2262 2269 arc_buf_hdr_t *hdr = buf->b_hdr;
2263 2270 int hashed = hdr->b_l1hdr.b_state != arc_anon;
2264 2271
2265 2272 ASSERT(buf->b_efunc == NULL);
2266 2273 ASSERT(buf->b_data != NULL);
2267 2274
2268 2275 if (hashed) {
2269 2276 kmutex_t *hash_lock = HDR_LOCK(hdr);
2270 2277
2271 2278 mutex_enter(hash_lock);
2272 2279 hdr = buf->b_hdr;
2273 2280 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2274 2281
2275 2282 (void) remove_reference(hdr, hash_lock, tag);
2276 2283 if (hdr->b_l1hdr.b_datacnt > 1) {
2277 2284 arc_buf_destroy(buf, TRUE);
2278 2285 } else {
2279 2286 ASSERT(buf == hdr->b_l1hdr.b_buf);
2280 2287 ASSERT(buf->b_efunc == NULL);
2281 2288 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2282 2289 }
2283 2290 mutex_exit(hash_lock);
2284 2291 } else if (HDR_IO_IN_PROGRESS(hdr)) {
2285 2292 int destroy_hdr;
2286 2293 /*
2287 2294 * We are in the middle of an async write. Don't destroy
2288 2295 * this buffer unless the write completes before we finish
2289 2296 * decrementing the reference count.
2290 2297 */
2291 2298 mutex_enter(&arc_user_evicts_lock);
2292 2299 (void) remove_reference(hdr, NULL, tag);
2293 2300 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2294 2301 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
2295 2302 mutex_exit(&arc_user_evicts_lock);
2296 2303 if (destroy_hdr)
2297 2304 arc_hdr_destroy(hdr);
2298 2305 } else {
2299 2306 if (remove_reference(hdr, NULL, tag) > 0)
2300 2307 arc_buf_destroy(buf, TRUE);
2301 2308 else
2302 2309 arc_hdr_destroy(hdr);
2303 2310 }
2304 2311 }
2305 2312
2306 2313 boolean_t
2307 2314 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
2308 2315 {
2309 2316 arc_buf_hdr_t *hdr = buf->b_hdr;
2310 2317 kmutex_t *hash_lock = HDR_LOCK(hdr);
2311 2318 boolean_t no_callback = (buf->b_efunc == NULL);
2312 2319
2313 2320 if (hdr->b_l1hdr.b_state == arc_anon) {
2314 2321 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
2315 2322 arc_buf_free(buf, tag);
2316 2323 return (no_callback);
2317 2324 }
2318 2325
2319 2326 mutex_enter(hash_lock);
2320 2327 hdr = buf->b_hdr;
2321 2328 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
2322 2329 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
2323 2330 ASSERT(hdr->b_l1hdr.b_state != arc_anon);
2324 2331 ASSERT(buf->b_data != NULL);
2325 2332
2326 2333 (void) remove_reference(hdr, hash_lock, tag);
2327 2334 if (hdr->b_l1hdr.b_datacnt > 1) {
2328 2335 if (no_callback)
2329 2336 arc_buf_destroy(buf, TRUE);
2330 2337 } else if (no_callback) {
2331 2338 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL);
2332 2339 ASSERT(buf->b_efunc == NULL);
2333 2340 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
2334 2341 }
2335 2342 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 ||
2336 2343 refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2337 2344 mutex_exit(hash_lock);
2338 2345 return (no_callback);
2339 2346 }
2340 2347
2341 2348 int32_t
2342 2349 arc_buf_size(arc_buf_t *buf)
2343 2350 {
2344 2351 return (buf->b_hdr->b_size);
2345 2352 }
2346 2353
2347 2354 /*
2348 2355 * Called from the DMU to determine if the current buffer should be
2349 2356 * evicted. In order to ensure proper locking, the eviction must be initiated
2350 2357 * from the DMU. Return true if the buffer is associated with user data and
2351 2358 * duplicate buffers still exist.
2352 2359 */
2353 2360 boolean_t
2354 2361 arc_buf_eviction_needed(arc_buf_t *buf)
2355 2362 {
2356 2363 arc_buf_hdr_t *hdr;
2357 2364 boolean_t evict_needed = B_FALSE;
2358 2365
2359 2366 if (zfs_disable_dup_eviction)
2360 2367 return (B_FALSE);
2361 2368
2362 2369 mutex_enter(&buf->b_evict_lock);
2363 2370 hdr = buf->b_hdr;
2364 2371 if (hdr == NULL) {
2365 2372 /*
2366 2373 * We are in arc_do_user_evicts(); let that function
2367 2374 * perform the eviction.
2368 2375 */
2369 2376 ASSERT(buf->b_data == NULL);
2370 2377 mutex_exit(&buf->b_evict_lock);
2371 2378 return (B_FALSE);
2372 2379 } else if (buf->b_data == NULL) {
2373 2380 /*
2374 2381 * We have already been added to the arc eviction list;
2375 2382 * recommend eviction.
2376 2383 */
2377 2384 ASSERT3P(hdr, ==, &arc_eviction_hdr);
2378 2385 mutex_exit(&buf->b_evict_lock);
2379 2386 return (B_TRUE);
2380 2387 }
2381 2388
2382 2389 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr))
2383 2390 evict_needed = B_TRUE;
2384 2391
2385 2392 mutex_exit(&buf->b_evict_lock);
2386 2393 return (evict_needed);
2387 2394 }
2388 2395
2389 2396 /*
2390 2397 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
2391 2398 * state of the header is dependent on it's state prior to entering this
2392 2399 * function. The following transitions are possible:
2393 2400 *
2394 2401 * - arc_mru -> arc_mru_ghost
2395 2402 * - arc_mfu -> arc_mfu_ghost
2396 2403 * - arc_mru_ghost -> arc_l2c_only
2397 2404 * - arc_mru_ghost -> deleted
2398 2405 * - arc_mfu_ghost -> arc_l2c_only
2399 2406 * - arc_mfu_ghost -> deleted
2400 2407 */
2401 2408 static int64_t
2402 2409 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
2403 2410 {
2404 2411 arc_state_t *evicted_state, *state;
2405 2412 int64_t bytes_evicted = 0;
2406 2413
2407 2414 ASSERT(MUTEX_HELD(hash_lock));
2408 2415 ASSERT(HDR_HAS_L1HDR(hdr));
2409 2416
2410 2417 state = hdr->b_l1hdr.b_state;
2411 2418 if (GHOST_STATE(state)) {
2412 2419 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2413 2420 ASSERT(hdr->b_l1hdr.b_buf == NULL);
2414 2421
2415 2422 /*
2416 2423 * l2arc_write_buffers() relies on a header's L1 portion
2417 2424 * (i.e. it's b_tmp_cdata field) during it's write phase.
2418 2425 * Thus, we cannot push a header onto the arc_l2c_only
2419 2426 * state (removing it's L1 piece) until the header is
2420 2427 * done being written to the l2arc.
2421 2428 */
2422 2429 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
2423 2430 ARCSTAT_BUMP(arcstat_evict_l2_skip);
2424 2431 return (bytes_evicted);
2425 2432 }
2426 2433
2427 2434 ARCSTAT_BUMP(arcstat_deleted);
2428 2435 bytes_evicted += hdr->b_size;
2429 2436
2430 2437 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr);
2431 2438
2432 2439 if (HDR_HAS_L2HDR(hdr)) {
2433 2440 /*
2434 2441 * This buffer is cached on the 2nd Level ARC;
2435 2442 * don't destroy the header.
2436 2443 */
2437 2444 arc_change_state(arc_l2c_only, hdr, hash_lock);
2438 2445 /*
2439 2446 * dropping from L1+L2 cached to L2-only,
2440 2447 * realloc to remove the L1 header.
2441 2448 */
2442 2449 hdr = arc_hdr_realloc(hdr, hdr_full_cache,
2443 2450 hdr_l2only_cache);
2444 2451 } else {
2445 2452 arc_change_state(arc_anon, hdr, hash_lock);
2446 2453 arc_hdr_destroy(hdr);
2447 2454 }
2448 2455 return (bytes_evicted);
2449 2456 }
2450 2457
2451 2458 ASSERT(state == arc_mru || state == arc_mfu);
2452 2459 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
2453 2460
2454 2461 /* prefetch buffers have a minimum lifespan */
2455 2462 if (HDR_IO_IN_PROGRESS(hdr) ||
2456 2463 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
2457 2464 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access <
2458 2465 arc_min_prefetch_lifespan)) {
2459 2466 ARCSTAT_BUMP(arcstat_evict_skip);
2460 2467 return (bytes_evicted);
2461 2468 }
2462 2469
2463 2470 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
2464 2471 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0);
2465 2472 while (hdr->b_l1hdr.b_buf) {
2466 2473 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
2467 2474 if (!mutex_tryenter(&buf->b_evict_lock)) {
2468 2475 ARCSTAT_BUMP(arcstat_mutex_miss);
2469 2476 break;
2470 2477 }
2471 2478 if (buf->b_data != NULL)
2472 2479 bytes_evicted += hdr->b_size;
2473 2480 if (buf->b_efunc != NULL) {
2474 2481 mutex_enter(&arc_user_evicts_lock);
2475 2482 arc_buf_destroy(buf, FALSE);
2476 2483 hdr->b_l1hdr.b_buf = buf->b_next;
2477 2484 buf->b_hdr = &arc_eviction_hdr;
2478 2485 buf->b_next = arc_eviction_list;
2479 2486 arc_eviction_list = buf;
2480 2487 cv_signal(&arc_user_evicts_cv);
2481 2488 mutex_exit(&arc_user_evicts_lock);
2482 2489 mutex_exit(&buf->b_evict_lock);
2483 2490 } else {
2484 2491 mutex_exit(&buf->b_evict_lock);
2485 2492 arc_buf_destroy(buf, TRUE);
2486 2493 }
2487 2494 }
2488 2495
2489 2496 if (HDR_HAS_L2HDR(hdr)) {
2490 2497 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size);
2491 2498 } else {
2492 2499 if (l2arc_write_eligible(hdr->b_spa, hdr))
2493 2500 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size);
2494 2501 else
2495 2502 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size);
2496 2503 }
2497 2504
2498 2505 if (hdr->b_l1hdr.b_datacnt == 0) {
2499 2506 arc_change_state(evicted_state, hdr, hash_lock);
2500 2507 ASSERT(HDR_IN_HASH_TABLE(hdr));
2501 2508 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE;
2502 2509 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
2503 2510 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
2504 2511 }
2505 2512
2506 2513 return (bytes_evicted);
2507 2514 }
2508 2515
2509 2516 static uint64_t
2510 2517 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
2511 2518 uint64_t spa, int64_t bytes)
2512 2519 {
2513 2520 multilist_sublist_t *mls;
2514 2521 uint64_t bytes_evicted = 0;
2515 2522 arc_buf_hdr_t *hdr;
2516 2523 kmutex_t *hash_lock;
2517 2524 int evict_count = 0;
2518 2525
2519 2526 ASSERT3P(marker, !=, NULL);
2520 2527 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2521 2528
2522 2529 mls = multilist_sublist_lock(ml, idx);
2523 2530
2524 2531 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL;
2525 2532 hdr = multilist_sublist_prev(mls, marker)) {
2526 2533 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) ||
2527 2534 (evict_count >= zfs_arc_evict_batch_limit))
2528 2535 break;
2529 2536
2530 2537 /*
2531 2538 * To keep our iteration location, move the marker
2532 2539 * forward. Since we're not holding hdr's hash lock, we
2533 2540 * must be very careful and not remove 'hdr' from the
2534 2541 * sublist. Otherwise, other consumers might mistake the
2535 2542 * 'hdr' as not being on a sublist when they call the
2536 2543 * multilist_link_active() function (they all rely on
2537 2544 * the hash lock protecting concurrent insertions and
2538 2545 * removals). multilist_sublist_move_forward() was
2539 2546 * specifically implemented to ensure this is the case
2540 2547 * (only 'marker' will be removed and re-inserted).
2541 2548 */
2542 2549 multilist_sublist_move_forward(mls, marker);
2543 2550
2544 2551 /*
2545 2552 * The only case where the b_spa field should ever be
2546 2553 * zero, is the marker headers inserted by
2547 2554 * arc_evict_state(). It's possible for multiple threads
2548 2555 * to be calling arc_evict_state() concurrently (e.g.
2549 2556 * dsl_pool_close() and zio_inject_fault()), so we must
2550 2557 * skip any markers we see from these other threads.
2551 2558 */
2552 2559 if (hdr->b_spa == 0)
2553 2560 continue;
2554 2561
2555 2562 /* we're only interested in evicting buffers of a certain spa */
2556 2563 if (spa != 0 && hdr->b_spa != spa) {
2557 2564 ARCSTAT_BUMP(arcstat_evict_skip);
2558 2565 continue;
2559 2566 }
2560 2567
2561 2568 hash_lock = HDR_LOCK(hdr);
2562 2569
2563 2570 /*
2564 2571 * We aren't calling this function from any code path
2565 2572 * that would already be holding a hash lock, so we're
2566 2573 * asserting on this assumption to be defensive in case
2567 2574 * this ever changes. Without this check, it would be
2568 2575 * possible to incorrectly increment arcstat_mutex_miss
2569 2576 * below (e.g. if the code changed such that we called
2570 2577 * this function with a hash lock held).
2571 2578 */
2572 2579 ASSERT(!MUTEX_HELD(hash_lock));
2573 2580
2574 2581 if (mutex_tryenter(hash_lock)) {
2575 2582 uint64_t evicted = arc_evict_hdr(hdr, hash_lock);
2576 2583 mutex_exit(hash_lock);
2577 2584
2578 2585 bytes_evicted += evicted;
2579 2586
2580 2587 /*
2581 2588 * If evicted is zero, arc_evict_hdr() must have
2582 2589 * decided to skip this header, don't increment
2583 2590 * evict_count in this case.
2584 2591 */
2585 2592 if (evicted != 0)
2586 2593 evict_count++;
2587 2594
2588 2595 /*
2589 2596 * If arc_size isn't overflowing, signal any
2590 2597 * threads that might happen to be waiting.
2591 2598 *
2592 2599 * For each header evicted, we wake up a single
2593 2600 * thread. If we used cv_broadcast, we could
2594 2601 * wake up "too many" threads causing arc_size
2595 2602 * to significantly overflow arc_c; since
2596 2603 * arc_get_data_buf() doesn't check for overflow
2597 2604 * when it's woken up (it doesn't because it's
2598 2605 * possible for the ARC to be overflowing while
2599 2606 * full of un-evictable buffers, and the
2600 2607 * function should proceed in this case).
2601 2608 *
2602 2609 * If threads are left sleeping, due to not
2603 2610 * using cv_broadcast, they will be woken up
2604 2611 * just before arc_reclaim_thread() sleeps.
2605 2612 */
2606 2613 mutex_enter(&arc_reclaim_lock);
2607 2614 if (!arc_is_overflowing())
2608 2615 cv_signal(&arc_reclaim_waiters_cv);
2609 2616 mutex_exit(&arc_reclaim_lock);
2610 2617 } else {
2611 2618 ARCSTAT_BUMP(arcstat_mutex_miss);
2612 2619 }
2613 2620 }
2614 2621
2615 2622 multilist_sublist_unlock(mls);
2616 2623
2617 2624 return (bytes_evicted);
2618 2625 }
2619 2626
2620 2627 /*
2621 2628 * Evict buffers from the given arc state, until we've removed the
2622 2629 * specified number of bytes. Move the removed buffers to the
2623 2630 * appropriate evict state.
2624 2631 *
2625 2632 * This function makes a "best effort". It skips over any buffers
2626 2633 * it can't get a hash_lock on, and so, may not catch all candidates.
2627 2634 * It may also return without evicting as much space as requested.
2628 2635 *
2629 2636 * If bytes is specified using the special value ARC_EVICT_ALL, this
2630 2637 * will evict all available (i.e. unlocked and evictable) buffers from
2631 2638 * the given arc state; which is used by arc_flush().
2632 2639 */
2633 2640 static uint64_t
2634 2641 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes,
2635 2642 arc_buf_contents_t type)
2636 2643 {
2637 2644 uint64_t total_evicted = 0;
2638 2645 multilist_t *ml = &state->arcs_list[type];
2639 2646 int num_sublists;
2640 2647 arc_buf_hdr_t **markers;
2641 2648
2642 2649 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL);
2643 2650
2644 2651 num_sublists = multilist_get_num_sublists(ml);
2645 2652
2646 2653 /*
2647 2654 * If we've tried to evict from each sublist, made some
2648 2655 * progress, but still have not hit the target number of bytes
2649 2656 * to evict, we want to keep trying. The markers allow us to
2650 2657 * pick up where we left off for each individual sublist, rather
2651 2658 * than starting from the tail each time.
2652 2659 */
2653 2660 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
2654 2661 for (int i = 0; i < num_sublists; i++) {
2655 2662 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
2656 2663
2657 2664 /*
2658 2665 * A b_spa of 0 is used to indicate that this header is
2659 2666 * a marker. This fact is used in arc_adjust_type() and
2660 2667 * arc_evict_state_impl().
2661 2668 */
2662 2669 markers[i]->b_spa = 0;
2663 2670
2664 2671 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2665 2672 multilist_sublist_insert_tail(mls, markers[i]);
2666 2673 multilist_sublist_unlock(mls);
2667 2674 }
2668 2675
2669 2676 /*
2670 2677 * While we haven't hit our target number of bytes to evict, or
2671 2678 * we're evicting all available buffers.
2672 2679 */
2673 2680 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) {
2674 2681 /*
2675 2682 * Start eviction using a randomly selected sublist,
2676 2683 * this is to try and evenly balance eviction across all
2677 2684 * sublists. Always starting at the same sublist
2678 2685 * (e.g. index 0) would cause evictions to favor certain
2679 2686 * sublists over others.
2680 2687 */
2681 2688 int sublist_idx = multilist_get_random_index(ml);
2682 2689 uint64_t scan_evicted = 0;
2683 2690
2684 2691 for (int i = 0; i < num_sublists; i++) {
2685 2692 uint64_t bytes_remaining;
2686 2693 uint64_t bytes_evicted;
2687 2694
2688 2695 if (bytes == ARC_EVICT_ALL)
2689 2696 bytes_remaining = ARC_EVICT_ALL;
2690 2697 else if (total_evicted < bytes)
2691 2698 bytes_remaining = bytes - total_evicted;
2692 2699 else
2693 2700 break;
2694 2701
2695 2702 bytes_evicted = arc_evict_state_impl(ml, sublist_idx,
2696 2703 markers[sublist_idx], spa, bytes_remaining);
2697 2704
2698 2705 scan_evicted += bytes_evicted;
2699 2706 total_evicted += bytes_evicted;
2700 2707
2701 2708 /* we've reached the end, wrap to the beginning */
2702 2709 if (++sublist_idx >= num_sublists)
2703 2710 sublist_idx = 0;
2704 2711 }
2705 2712
2706 2713 /*
2707 2714 * If we didn't evict anything during this scan, we have
2708 2715 * no reason to believe we'll evict more during another
2709 2716 * scan, so break the loop.
2710 2717 */
2711 2718 if (scan_evicted == 0) {
2712 2719 /* This isn't possible, let's make that obvious */
2713 2720 ASSERT3S(bytes, !=, 0);
2714 2721
2715 2722 /*
2716 2723 * When bytes is ARC_EVICT_ALL, the only way to
2717 2724 * break the loop is when scan_evicted is zero.
2718 2725 * In that case, we actually have evicted enough,
2719 2726 * so we don't want to increment the kstat.
2720 2727 */
2721 2728 if (bytes != ARC_EVICT_ALL) {
2722 2729 ASSERT3S(total_evicted, <, bytes);
2723 2730 ARCSTAT_BUMP(arcstat_evict_not_enough);
2724 2731 }
2725 2732
2726 2733 break;
2727 2734 }
2728 2735 }
2729 2736
2730 2737 for (int i = 0; i < num_sublists; i++) {
2731 2738 multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
2732 2739 multilist_sublist_remove(mls, markers[i]);
2733 2740 multilist_sublist_unlock(mls);
2734 2741
2735 2742 kmem_cache_free(hdr_full_cache, markers[i]);
2736 2743 }
2737 2744 kmem_free(markers, sizeof (*markers) * num_sublists);
2738 2745
2739 2746 return (total_evicted);
2740 2747 }
2741 2748
2742 2749 /*
2743 2750 * Flush all "evictable" data of the given type from the arc state
2744 2751 * specified. This will not evict any "active" buffers (i.e. referenced).
2745 2752 *
2746 2753 * When 'retry' is set to FALSE, the function will make a single pass
2747 2754 * over the state and evict any buffers that it can. Since it doesn't
2748 2755 * continually retry the eviction, it might end up leaving some buffers
2749 2756 * in the ARC due to lock misses.
2750 2757 *
2751 2758 * When 'retry' is set to TRUE, the function will continually retry the
2752 2759 * eviction until *all* evictable buffers have been removed from the
2753 2760 * state. As a result, if concurrent insertions into the state are
2754 2761 * allowed (e.g. if the ARC isn't shutting down), this function might
2755 2762 * wind up in an infinite loop, continually trying to evict buffers.
2756 2763 */
2757 2764 static uint64_t
2758 2765 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type,
2759 2766 boolean_t retry)
2760 2767 {
2761 2768 uint64_t evicted = 0;
2762 2769
2763 2770 while (state->arcs_lsize[type] != 0) {
2764 2771 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type);
2765 2772
2766 2773 if (!retry)
2767 2774 break;
2768 2775 }
2769 2776
2770 2777 return (evicted);
2771 2778 }
2772 2779
2773 2780 /*
2774 2781 * Evict the specified number of bytes from the state specified,
2775 2782 * restricting eviction to the spa and type given. This function
2776 2783 * prevents us from trying to evict more from a state's list than
2777 2784 * is "evictable", and to skip evicting altogether when passed a
2778 2785 * negative value for "bytes". In contrast, arc_evict_state() will
2779 2786 * evict everything it can, when passed a negative value for "bytes".
2780 2787 */
2781 2788 static uint64_t
2782 2789 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
2783 2790 arc_buf_contents_t type)
2784 2791 {
2785 2792 int64_t delta;
2786 2793
2787 2794 if (bytes > 0 && state->arcs_lsize[type] > 0) {
2788 2795 delta = MIN(state->arcs_lsize[type], bytes);
2789 2796 return (arc_evict_state(state, spa, delta, type));
2790 2797 }
2791 2798
2792 2799 return (0);
2793 2800 }
2794 2801
2795 2802 /*
2796 2803 * Evict metadata buffers from the cache, such that arc_meta_used is
2797 2804 * capped by the arc_meta_limit tunable.
2798 2805 */
2799 2806 static uint64_t
2800 2807 arc_adjust_meta(void)
2801 2808 {
2802 2809 uint64_t total_evicted = 0;
2803 2810 int64_t target;
2804 2811
2805 2812 /*
2806 2813 * If we're over the meta limit, we want to evict enough
2807 2814 * metadata to get back under the meta limit. We don't want to
2808 2815 * evict so much that we drop the MRU below arc_p, though. If
2809 2816 * we're over the meta limit more than we're over arc_p, we
2810 2817 * evict some from the MRU here, and some from the MFU below.
2811 2818 */
2812 2819 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
2813 2820 (int64_t)(refcount_count(&arc_anon->arcs_size) +
2814 2821 refcount_count(&arc_mru->arcs_size) - arc_p));
2815 2822
2816 2823 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2817 2824
2818 2825 /*
2819 2826 * Similar to the above, we want to evict enough bytes to get us
2820 2827 * below the meta limit, but not so much as to drop us below the
2821 2828 * space alloted to the MFU (which is defined as arc_c - arc_p).
2822 2829 */
2823 2830 target = MIN((int64_t)(arc_meta_used - arc_meta_limit),
2824 2831 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
2825 2832
2826 2833 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
2827 2834
2828 2835 return (total_evicted);
2829 2836 }
2830 2837
2831 2838 /*
2832 2839 * Return the type of the oldest buffer in the given arc state
2833 2840 *
2834 2841 * This function will select a random sublist of type ARC_BUFC_DATA and
2835 2842 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
2836 2843 * is compared, and the type which contains the "older" buffer will be
2837 2844 * returned.
2838 2845 */
2839 2846 static arc_buf_contents_t
2840 2847 arc_adjust_type(arc_state_t *state)
2841 2848 {
2842 2849 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA];
2843 2850 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA];
2844 2851 int data_idx = multilist_get_random_index(data_ml);
2845 2852 int meta_idx = multilist_get_random_index(meta_ml);
2846 2853 multilist_sublist_t *data_mls;
2847 2854 multilist_sublist_t *meta_mls;
2848 2855 arc_buf_contents_t type;
2849 2856 arc_buf_hdr_t *data_hdr;
2850 2857 arc_buf_hdr_t *meta_hdr;
2851 2858
2852 2859 /*
2853 2860 * We keep the sublist lock until we're finished, to prevent
2854 2861 * the headers from being destroyed via arc_evict_state().
2855 2862 */
2856 2863 data_mls = multilist_sublist_lock(data_ml, data_idx);
2857 2864 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
2858 2865
2859 2866 /*
2860 2867 * These two loops are to ensure we skip any markers that
2861 2868 * might be at the tail of the lists due to arc_evict_state().
2862 2869 */
2863 2870
2864 2871 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
2865 2872 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
2866 2873 if (data_hdr->b_spa != 0)
2867 2874 break;
2868 2875 }
2869 2876
2870 2877 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
2871 2878 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
2872 2879 if (meta_hdr->b_spa != 0)
2873 2880 break;
2874 2881 }
2875 2882
2876 2883 if (data_hdr == NULL && meta_hdr == NULL) {
2877 2884 type = ARC_BUFC_DATA;
2878 2885 } else if (data_hdr == NULL) {
2879 2886 ASSERT3P(meta_hdr, !=, NULL);
2880 2887 type = ARC_BUFC_METADATA;
2881 2888 } else if (meta_hdr == NULL) {
2882 2889 ASSERT3P(data_hdr, !=, NULL);
2883 2890 type = ARC_BUFC_DATA;
2884 2891 } else {
2885 2892 ASSERT3P(data_hdr, !=, NULL);
2886 2893 ASSERT3P(meta_hdr, !=, NULL);
2887 2894
2888 2895 /* The headers can't be on the sublist without an L1 header */
2889 2896 ASSERT(HDR_HAS_L1HDR(data_hdr));
2890 2897 ASSERT(HDR_HAS_L1HDR(meta_hdr));
2891 2898
2892 2899 if (data_hdr->b_l1hdr.b_arc_access <
2893 2900 meta_hdr->b_l1hdr.b_arc_access) {
2894 2901 type = ARC_BUFC_DATA;
2895 2902 } else {
2896 2903 type = ARC_BUFC_METADATA;
2897 2904 }
2898 2905 }
2899 2906
2900 2907 multilist_sublist_unlock(meta_mls);
2901 2908 multilist_sublist_unlock(data_mls);
2902 2909
2903 2910 return (type);
2904 2911 }
2905 2912
2906 2913 /*
2907 2914 * Evict buffers from the cache, such that arc_size is capped by arc_c.
2908 2915 */
2909 2916 static uint64_t
2910 2917 arc_adjust(void)
2911 2918 {
2912 2919 uint64_t total_evicted = 0;
2913 2920 uint64_t bytes;
2914 2921 int64_t target;
2915 2922
2916 2923 /*
2917 2924 * If we're over arc_meta_limit, we want to correct that before
2918 2925 * potentially evicting data buffers below.
2919 2926 */
2920 2927 total_evicted += arc_adjust_meta();
2921 2928
2922 2929 /*
2923 2930 * Adjust MRU size
2924 2931 *
2925 2932 * If we're over the target cache size, we want to evict enough
2926 2933 * from the list to get back to our target size. We don't want
2927 2934 * to evict too much from the MRU, such that it drops below
2928 2935 * arc_p. So, if we're over our target cache size more than
2929 2936 * the MRU is over arc_p, we'll evict enough to get back to
2930 2937 * arc_p here, and then evict more from the MFU below.
2931 2938 */
2932 2939 target = MIN((int64_t)(arc_size - arc_c),
2933 2940 (int64_t)(refcount_count(&arc_anon->arcs_size) +
2934 2941 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
2935 2942
2936 2943 /*
2937 2944 * If we're below arc_meta_min, always prefer to evict data.
2938 2945 * Otherwise, try to satisfy the requested number of bytes to
2939 2946 * evict from the type which contains older buffers; in an
2940 2947 * effort to keep newer buffers in the cache regardless of their
2941 2948 * type. If we cannot satisfy the number of bytes from this
2942 2949 * type, spill over into the next type.
2943 2950 */
2944 2951 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
2945 2952 arc_meta_used > arc_meta_min) {
2946 2953 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2947 2954 total_evicted += bytes;
2948 2955
2949 2956 /*
2950 2957 * If we couldn't evict our target number of bytes from
2951 2958 * metadata, we try to get the rest from data.
2952 2959 */
2953 2960 target -= bytes;
2954 2961
2955 2962 total_evicted +=
2956 2963 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
2957 2964 } else {
2958 2965 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
2959 2966 total_evicted += bytes;
2960 2967
2961 2968 /*
2962 2969 * If we couldn't evict our target number of bytes from
2963 2970 * data, we try to get the rest from metadata.
2964 2971 */
2965 2972 target -= bytes;
2966 2973
2967 2974 total_evicted +=
2968 2975 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
2969 2976 }
2970 2977
2971 2978 /*
2972 2979 * Adjust MFU size
2973 2980 *
2974 2981 * Now that we've tried to evict enough from the MRU to get its
2975 2982 * size back to arc_p, if we're still above the target cache
2976 2983 * size, we evict the rest from the MFU.
2977 2984 */
2978 2985 target = arc_size - arc_c;
2979 2986
2980 2987 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
2981 2988 arc_meta_used > arc_meta_min) {
2982 2989 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
2983 2990 total_evicted += bytes;
2984 2991
2985 2992 /*
2986 2993 * If we couldn't evict our target number of bytes from
2987 2994 * metadata, we try to get the rest from data.
2988 2995 */
2989 2996 target -= bytes;
2990 2997
2991 2998 total_evicted +=
2992 2999 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
2993 3000 } else {
2994 3001 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
2995 3002 total_evicted += bytes;
2996 3003
2997 3004 /*
2998 3005 * If we couldn't evict our target number of bytes from
2999 3006 * data, we try to get the rest from data.
3000 3007 */
3001 3008 target -= bytes;
3002 3009
3003 3010 total_evicted +=
3004 3011 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3005 3012 }
3006 3013
3007 3014 /*
3008 3015 * Adjust ghost lists
3009 3016 *
3010 3017 * In addition to the above, the ARC also defines target values
3011 3018 * for the ghost lists. The sum of the mru list and mru ghost
3012 3019 * list should never exceed the target size of the cache, and
3013 3020 * the sum of the mru list, mfu list, mru ghost list, and mfu
3014 3021 * ghost list should never exceed twice the target size of the
3015 3022 * cache. The following logic enforces these limits on the ghost
3016 3023 * caches, and evicts from them as needed.
3017 3024 */
3018 3025 target = refcount_count(&arc_mru->arcs_size) +
3019 3026 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3020 3027
3021 3028 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3022 3029 total_evicted += bytes;
3023 3030
3024 3031 target -= bytes;
3025 3032
3026 3033 total_evicted +=
3027 3034 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3028 3035
3029 3036 /*
3030 3037 * We assume the sum of the mru list and mfu list is less than
3031 3038 * or equal to arc_c (we enforced this above), which means we
3032 3039 * can use the simpler of the two equations below:
3033 3040 *
3034 3041 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3035 3042 * mru ghost + mfu ghost <= arc_c
3036 3043 */
3037 3044 target = refcount_count(&arc_mru_ghost->arcs_size) +
3038 3045 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3039 3046
3040 3047 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3041 3048 total_evicted += bytes;
3042 3049
3043 3050 target -= bytes;
3044 3051
3045 3052 total_evicted +=
3046 3053 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3047 3054
3048 3055 return (total_evicted);
3049 3056 }
3050 3057
3051 3058 static void
3052 3059 arc_do_user_evicts(void)
3053 3060 {
3054 3061 mutex_enter(&arc_user_evicts_lock);
3055 3062 while (arc_eviction_list != NULL) {
3056 3063 arc_buf_t *buf = arc_eviction_list;
3057 3064 arc_eviction_list = buf->b_next;
3058 3065 mutex_enter(&buf->b_evict_lock);
3059 3066 buf->b_hdr = NULL;
3060 3067 mutex_exit(&buf->b_evict_lock);
3061 3068 mutex_exit(&arc_user_evicts_lock);
3062 3069
3063 3070 if (buf->b_efunc != NULL)
3064 3071 VERIFY0(buf->b_efunc(buf->b_private));
3065 3072
3066 3073 buf->b_efunc = NULL;
3067 3074 buf->b_private = NULL;
3068 3075 kmem_cache_free(buf_cache, buf);
3069 3076 mutex_enter(&arc_user_evicts_lock);
3070 3077 }
3071 3078 mutex_exit(&arc_user_evicts_lock);
3072 3079 }
3073 3080
3074 3081 void
3075 3082 arc_flush(spa_t *spa, boolean_t retry)
3076 3083 {
3077 3084 uint64_t guid = 0;
3078 3085
3079 3086 /*
3080 3087 * If retry is TRUE, a spa must not be specified since we have
3081 3088 * no good way to determine if all of a spa's buffers have been
3082 3089 * evicted from an arc state.
3083 3090 */
3084 3091 ASSERT(!retry || spa == 0);
3085 3092
3086 3093 if (spa != NULL)
3087 3094 guid = spa_load_guid(spa);
3088 3095
3089 3096 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3090 3097 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3091 3098
3092 3099 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3093 3100 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3094 3101
3095 3102 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3096 3103 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3097 3104
3098 3105 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3099 3106 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3100 3107
3101 3108 arc_do_user_evicts();
3102 3109 ASSERT(spa || arc_eviction_list == NULL);
3103 3110 }
3104 3111
3105 3112 void
3106 3113 arc_shrink(int64_t to_free)
3107 3114 {
3108 3115 if (arc_c > arc_c_min) {
3109 3116
3110 3117 if (arc_c > arc_c_min + to_free)
3111 3118 atomic_add_64(&arc_c, -to_free);
3112 3119 else
3113 3120 arc_c = arc_c_min;
3114 3121
3115 3122 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3116 3123 if (arc_c > arc_size)
3117 3124 arc_c = MAX(arc_size, arc_c_min);
3118 3125 if (arc_p > arc_c)
3119 3126 arc_p = (arc_c >> 1);
3120 3127 ASSERT(arc_c >= arc_c_min);
3121 3128 ASSERT((int64_t)arc_p >= 0);
3122 3129 }
3123 3130
3124 3131 if (arc_size > arc_c)
3125 3132 (void) arc_adjust();
3126 3133 }
3127 3134
3128 3135 typedef enum free_memory_reason_t {
3129 3136 FMR_UNKNOWN,
3130 3137 FMR_NEEDFREE,
3131 3138 FMR_LOTSFREE,
3132 3139 FMR_SWAPFS_MINFREE,
3133 3140 FMR_PAGES_PP_MAXIMUM,
3134 3141 FMR_HEAP_ARENA,
3135 3142 FMR_ZIO_ARENA,
3136 3143 } free_memory_reason_t;
3137 3144
3138 3145 int64_t last_free_memory;
3139 3146 free_memory_reason_t last_free_reason;
3140 3147
3141 3148 /*
3142 3149 * Additional reserve of pages for pp_reserve.
3143 3150 */
3144 3151 int64_t arc_pages_pp_reserve = 64;
3145 3152
3146 3153 /*
3147 3154 * Additional reserve of pages for swapfs.
3148 3155 */
3149 3156 int64_t arc_swapfs_reserve = 64;
3150 3157
3151 3158 /*
3152 3159 * Return the amount of memory that can be consumed before reclaim will be
3153 3160 * needed. Positive if there is sufficient free memory, negative indicates
3154 3161 * the amount of memory that needs to be freed up.
3155 3162 */
3156 3163 static int64_t
3157 3164 arc_available_memory(void)
3158 3165 {
3159 3166 int64_t lowest = INT64_MAX;
3160 3167 int64_t n;
3161 3168 free_memory_reason_t r = FMR_UNKNOWN;
3162 3169
3163 3170 #ifdef _KERNEL
3164 3171 if (needfree > 0) {
3165 3172 n = PAGESIZE * (-needfree);
3166 3173 if (n < lowest) {
3167 3174 lowest = n;
3168 3175 r = FMR_NEEDFREE;
3169 3176 }
3170 3177 }
3171 3178
3172 3179 /*
3173 3180 * check that we're out of range of the pageout scanner. It starts to
3174 3181 * schedule paging if freemem is less than lotsfree and needfree.
3175 3182 * lotsfree is the high-water mark for pageout, and needfree is the
3176 3183 * number of needed free pages. We add extra pages here to make sure
3177 3184 * the scanner doesn't start up while we're freeing memory.
3178 3185 */
3179 3186 n = PAGESIZE * (freemem - lotsfree - needfree - desfree);
3180 3187 if (n < lowest) {
3181 3188 lowest = n;
3182 3189 r = FMR_LOTSFREE;
3183 3190 }
3184 3191
3185 3192 /*
3186 3193 * check to make sure that swapfs has enough space so that anon
3187 3194 * reservations can still succeed. anon_resvmem() checks that the
3188 3195 * availrmem is greater than swapfs_minfree, and the number of reserved
3189 3196 * swap pages. We also add a bit of extra here just to prevent
3190 3197 * circumstances from getting really dire.
3191 3198 */
3192 3199 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve -
3193 3200 desfree - arc_swapfs_reserve);
3194 3201 if (n < lowest) {
3195 3202 lowest = n;
3196 3203 r = FMR_SWAPFS_MINFREE;
3197 3204 }
3198 3205
3199 3206
3200 3207 /*
3201 3208 * Check that we have enough availrmem that memory locking (e.g., via
3202 3209 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum
3203 3210 * stores the number of pages that cannot be locked; when availrmem
3204 3211 * drops below pages_pp_maximum, page locking mechanisms such as
3205 3212 * page_pp_lock() will fail.)
3206 3213 */
3207 3214 n = PAGESIZE * (availrmem - pages_pp_maximum -
3208 3215 arc_pages_pp_reserve);
3209 3216 if (n < lowest) {
3210 3217 lowest = n;
3211 3218 r = FMR_PAGES_PP_MAXIMUM;
3212 3219 }
3213 3220
3214 3221 #if defined(__i386)
3215 3222 /*
3216 3223 * If we're on an i386 platform, it's possible that we'll exhaust the
3217 3224 * kernel heap space before we ever run out of available physical
3218 3225 * memory. Most checks of the size of the heap_area compare against
3219 3226 * tune.t_minarmem, which is the minimum available real memory that we
3220 3227 * can have in the system. However, this is generally fixed at 25 pages
3221 3228 * which is so low that it's useless. In this comparison, we seek to
3222 3229 * calculate the total heap-size, and reclaim if more than 3/4ths of the
3223 3230 * heap is allocated. (Or, in the calculation, if less than 1/4th is
3224 3231 * free)
3225 3232 */
3226 3233 n = vmem_size(heap_arena, VMEM_FREE) -
3227 3234 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2);
3228 3235 if (n < lowest) {
3229 3236 lowest = n;
3230 3237 r = FMR_HEAP_ARENA;
3231 3238 }
3232 3239 #endif
3233 3240
3234 3241 /*
3235 3242 * If zio data pages are being allocated out of a separate heap segment,
3236 3243 * then enforce that the size of available vmem for this arena remains
3237 3244 * above about 1/16th free.
3238 3245 *
3239 3246 * Note: The 1/16th arena free requirement was put in place
3240 3247 * to aggressively evict memory from the arc in order to avoid
3241 3248 * memory fragmentation issues.
3242 3249 */
3243 3250 if (zio_arena != NULL) {
3244 3251 n = vmem_size(zio_arena, VMEM_FREE) -
3245 3252 (vmem_size(zio_arena, VMEM_ALLOC) >> 4);
3246 3253 if (n < lowest) {
3247 3254 lowest = n;
3248 3255 r = FMR_ZIO_ARENA;
3249 3256 }
3250 3257 }
3251 3258 #else
3252 3259 /* Every 100 calls, free a small amount */
3253 3260 if (spa_get_random(100) == 0)
3254 3261 lowest = -1024;
3255 3262 #endif
3256 3263
3257 3264 last_free_memory = lowest;
3258 3265 last_free_reason = r;
3259 3266
3260 3267 return (lowest);
3261 3268 }
3262 3269
3263 3270
3264 3271 /*
3265 3272 * Determine if the system is under memory pressure and is asking
3266 3273 * to reclaim memory. A return value of TRUE indicates that the system
3267 3274 * is under memory pressure and that the arc should adjust accordingly.
3268 3275 */
3269 3276 static boolean_t
3270 3277 arc_reclaim_needed(void)
3271 3278 {
3272 3279 return (arc_available_memory() < 0);
3273 3280 }
3274 3281
3275 3282 static void
3276 3283 arc_kmem_reap_now(void)
3277 3284 {
3278 3285 size_t i;
3279 3286 kmem_cache_t *prev_cache = NULL;
3280 3287 kmem_cache_t *prev_data_cache = NULL;
3281 3288 extern kmem_cache_t *zio_buf_cache[];
3282 3289 extern kmem_cache_t *zio_data_buf_cache[];
3283 3290 extern kmem_cache_t *range_seg_cache;
3284 3291
3285 3292 #ifdef _KERNEL
3286 3293 if (arc_meta_used >= arc_meta_limit) {
3287 3294 /*
3288 3295 * We are exceeding our meta-data cache limit.
3289 3296 * Purge some DNLC entries to release holds on meta-data.
3290 3297 */
3291 3298 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
3292 3299 }
3293 3300 #if defined(__i386)
3294 3301 /*
3295 3302 * Reclaim unused memory from all kmem caches.
3296 3303 */
3297 3304 kmem_reap();
3298 3305 #endif
3299 3306 #endif
3300 3307
3301 3308 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
3302 3309 if (zio_buf_cache[i] != prev_cache) {
3303 3310 prev_cache = zio_buf_cache[i];
3304 3311 kmem_cache_reap_now(zio_buf_cache[i]);
3305 3312 }
3306 3313 if (zio_data_buf_cache[i] != prev_data_cache) {
3307 3314 prev_data_cache = zio_data_buf_cache[i];
3308 3315 kmem_cache_reap_now(zio_data_buf_cache[i]);
3309 3316 }
3310 3317 }
3311 3318 kmem_cache_reap_now(buf_cache);
3312 3319 kmem_cache_reap_now(hdr_full_cache);
3313 3320 kmem_cache_reap_now(hdr_l2only_cache);
3314 3321 kmem_cache_reap_now(range_seg_cache);
3315 3322
3316 3323 if (zio_arena != NULL) {
3317 3324 /*
3318 3325 * Ask the vmem arena to reclaim unused memory from its
3319 3326 * quantum caches.
3320 3327 */
3321 3328 vmem_qcache_reap(zio_arena);
3322 3329 }
3323 3330 }
3324 3331
3325 3332 /*
3326 3333 * Threads can block in arc_get_data_buf() waiting for this thread to evict
3327 3334 * enough data and signal them to proceed. When this happens, the threads in
3328 3335 * arc_get_data_buf() are sleeping while holding the hash lock for their
3329 3336 * particular arc header. Thus, we must be careful to never sleep on a
3330 3337 * hash lock in this thread. This is to prevent the following deadlock:
3331 3338 *
3332 3339 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L",
3333 3340 * waiting for the reclaim thread to signal it.
3334 3341 *
3335 3342 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter,
3336 3343 * fails, and goes to sleep forever.
3337 3344 *
3338 3345 * This possible deadlock is avoided by always acquiring a hash lock
3339 3346 * using mutex_tryenter() from arc_reclaim_thread().
3340 3347 */
3341 3348 static void
3342 3349 arc_reclaim_thread(void)
3343 3350 {
3344 3351 clock_t growtime = 0;
3345 3352 callb_cpr_t cpr;
3346 3353
3347 3354 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG);
3348 3355
3349 3356 mutex_enter(&arc_reclaim_lock);
3350 3357 while (!arc_reclaim_thread_exit) {
3351 3358 int64_t free_memory = arc_available_memory();
3352 3359 uint64_t evicted = 0;
3353 3360
3354 3361 mutex_exit(&arc_reclaim_lock);
3355 3362
3356 3363 if (free_memory < 0) {
3357 3364
3358 3365 arc_no_grow = B_TRUE;
3359 3366 arc_warm = B_TRUE;
3360 3367
3361 3368 /*
3362 3369 * Wait at least zfs_grow_retry (default 60) seconds
3363 3370 * before considering growing.
3364 3371 */
3365 3372 growtime = ddi_get_lbolt() + (arc_grow_retry * hz);
3366 3373
3367 3374 arc_kmem_reap_now();
3368 3375
3369 3376 /*
3370 3377 * If we are still low on memory, shrink the ARC
3371 3378 * so that we have arc_shrink_min free space.
3372 3379 */
3373 3380 free_memory = arc_available_memory();
3374 3381
3375 3382 int64_t to_free =
3376 3383 (arc_c >> arc_shrink_shift) - free_memory;
3377 3384 if (to_free > 0) {
3378 3385 #ifdef _KERNEL
3379 3386 to_free = MAX(to_free, ptob(needfree));
3380 3387 #endif
3381 3388 arc_shrink(to_free);
3382 3389 }
3383 3390 } else if (free_memory < arc_c >> arc_no_grow_shift) {
3384 3391 arc_no_grow = B_TRUE;
3385 3392 } else if (ddi_get_lbolt() >= growtime) {
3386 3393 arc_no_grow = B_FALSE;
3387 3394 }
3388 3395
3389 3396 evicted = arc_adjust();
3390 3397
3391 3398 mutex_enter(&arc_reclaim_lock);
3392 3399
3393 3400 /*
3394 3401 * If evicted is zero, we couldn't evict anything via
3395 3402 * arc_adjust(). This could be due to hash lock
3396 3403 * collisions, but more likely due to the majority of
3397 3404 * arc buffers being unevictable. Therefore, even if
3398 3405 * arc_size is above arc_c, another pass is unlikely to
3399 3406 * be helpful and could potentially cause us to enter an
3400 3407 * infinite loop.
3401 3408 */
3402 3409 if (arc_size <= arc_c || evicted == 0) {
3403 3410 /*
3404 3411 * We're either no longer overflowing, or we
3405 3412 * can't evict anything more, so we should wake
3406 3413 * up any threads before we go to sleep.
3407 3414 */
3408 3415 cv_broadcast(&arc_reclaim_waiters_cv);
3409 3416
3410 3417 /*
3411 3418 * Block until signaled, or after one second (we
3412 3419 * might need to perform arc_kmem_reap_now()
3413 3420 * even if we aren't being signalled)
3414 3421 */
3415 3422 CALLB_CPR_SAFE_BEGIN(&cpr);
3416 3423 (void) cv_timedwait(&arc_reclaim_thread_cv,
3417 3424 &arc_reclaim_lock, ddi_get_lbolt() + hz);
3418 3425 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
3419 3426 }
3420 3427 }
3421 3428
3422 3429 arc_reclaim_thread_exit = FALSE;
3423 3430 cv_broadcast(&arc_reclaim_thread_cv);
3424 3431 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */
3425 3432 thread_exit();
3426 3433 }
3427 3434
3428 3435 static void
3429 3436 arc_user_evicts_thread(void)
3430 3437 {
3431 3438 callb_cpr_t cpr;
3432 3439
3433 3440 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG);
3434 3441
3435 3442 mutex_enter(&arc_user_evicts_lock);
3436 3443 while (!arc_user_evicts_thread_exit) {
3437 3444 mutex_exit(&arc_user_evicts_lock);
3438 3445
3439 3446 arc_do_user_evicts();
3440 3447
3441 3448 /*
3442 3449 * This is necessary in order for the mdb ::arc dcmd to
3443 3450 * show up to date information. Since the ::arc command
3444 3451 * does not call the kstat's update function, without
3445 3452 * this call, the command may show stale stats for the
3446 3453 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even
3447 3454 * with this change, the data might be up to 1 second
3448 3455 * out of date; but that should suffice. The arc_state_t
3449 3456 * structures can be queried directly if more accurate
3450 3457 * information is needed.
3451 3458 */
3452 3459 if (arc_ksp != NULL)
3453 3460 arc_ksp->ks_update(arc_ksp, KSTAT_READ);
3454 3461
3455 3462 mutex_enter(&arc_user_evicts_lock);
3456 3463
3457 3464 /*
3458 3465 * Block until signaled, or after one second (we need to
3459 3466 * call the arc's kstat update function regularly).
3460 3467 */
3461 3468 CALLB_CPR_SAFE_BEGIN(&cpr);
3462 3469 (void) cv_timedwait(&arc_user_evicts_cv,
3463 3470 &arc_user_evicts_lock, ddi_get_lbolt() + hz);
3464 3471 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock);
3465 3472 }
3466 3473
3467 3474 arc_user_evicts_thread_exit = FALSE;
3468 3475 cv_broadcast(&arc_user_evicts_cv);
3469 3476 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */
3470 3477 thread_exit();
3471 3478 }
3472 3479
3473 3480 /*
3474 3481 * Adapt arc info given the number of bytes we are trying to add and
3475 3482 * the state that we are comming from. This function is only called
3476 3483 * when we are adding new content to the cache.
3477 3484 */
3478 3485 static void
3479 3486 arc_adapt(int bytes, arc_state_t *state)
3480 3487 {
3481 3488 int mult;
3482 3489 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
3483 3490 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size);
3484 3491 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size);
3485 3492
3486 3493 if (state == arc_l2c_only)
3487 3494 return;
3488 3495
3489 3496 ASSERT(bytes > 0);
3490 3497 /*
3491 3498 * Adapt the target size of the MRU list:
3492 3499 * - if we just hit in the MRU ghost list, then increase
3493 3500 * the target size of the MRU list.
3494 3501 * - if we just hit in the MFU ghost list, then increase
3495 3502 * the target size of the MFU list by decreasing the
3496 3503 * target size of the MRU list.
3497 3504 */
3498 3505 if (state == arc_mru_ghost) {
3499 3506 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size);
3500 3507 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */
3501 3508
3502 3509 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
3503 3510 } else if (state == arc_mfu_ghost) {
3504 3511 uint64_t delta;
3505 3512
3506 3513 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size);
3507 3514 mult = MIN(mult, 10);
3508 3515
3509 3516 delta = MIN(bytes * mult, arc_p);
3510 3517 arc_p = MAX(arc_p_min, arc_p - delta);
3511 3518 }
3512 3519 ASSERT((int64_t)arc_p >= 0);
3513 3520
3514 3521 if (arc_reclaim_needed()) {
3515 3522 cv_signal(&arc_reclaim_thread_cv);
3516 3523 return;
3517 3524 }
3518 3525
3519 3526 if (arc_no_grow)
3520 3527 return;
3521 3528
3522 3529 if (arc_c >= arc_c_max)
3523 3530 return;
3524 3531
3525 3532 /*
3526 3533 * If we're within (2 * maxblocksize) bytes of the target
3527 3534 * cache size, increment the target cache size
3528 3535 */
3529 3536 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
3530 3537 atomic_add_64(&arc_c, (int64_t)bytes);
3531 3538 if (arc_c > arc_c_max)
3532 3539 arc_c = arc_c_max;
3533 3540 else if (state == arc_anon)
3534 3541 atomic_add_64(&arc_p, (int64_t)bytes);
3535 3542 if (arc_p > arc_c)
3536 3543 arc_p = arc_c;
3537 3544 }
3538 3545 ASSERT((int64_t)arc_p >= 0);
3539 3546 }
3540 3547
3541 3548 /*
3542 3549 * Check if arc_size has grown past our upper threshold, determined by
3543 3550 * zfs_arc_overflow_shift.
3544 3551 */
3545 3552 static boolean_t
3546 3553 arc_is_overflowing(void)
3547 3554 {
3548 3555 /* Always allow at least one block of overflow */
3549 3556 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
3550 3557 arc_c >> zfs_arc_overflow_shift);
3551 3558
3552 3559 return (arc_size >= arc_c + overflow);
3553 3560 }
3554 3561
3555 3562 /*
3556 3563 * The buffer, supplied as the first argument, needs a data block. If we
3557 3564 * are hitting the hard limit for the cache size, we must sleep, waiting
3558 3565 * for the eviction thread to catch up. If we're past the target size
3559 3566 * but below the hard limit, we'll only signal the reclaim thread and
3560 3567 * continue on.
3561 3568 */
3562 3569 static void
3563 3570 arc_get_data_buf(arc_buf_t *buf)
3564 3571 {
3565 3572 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state;
3566 3573 uint64_t size = buf->b_hdr->b_size;
3567 3574 arc_buf_contents_t type = arc_buf_type(buf->b_hdr);
3568 3575
3569 3576 arc_adapt(size, state);
3570 3577
3571 3578 /*
3572 3579 * If arc_size is currently overflowing, and has grown past our
3573 3580 * upper limit, we must be adding data faster than the evict
3574 3581 * thread can evict. Thus, to ensure we don't compound the
3575 3582 * problem by adding more data and forcing arc_size to grow even
3576 3583 * further past it's target size, we halt and wait for the
3577 3584 * eviction thread to catch up.
3578 3585 *
3579 3586 * It's also possible that the reclaim thread is unable to evict
3580 3587 * enough buffers to get arc_size below the overflow limit (e.g.
3581 3588 * due to buffers being un-evictable, or hash lock collisions).
3582 3589 * In this case, we want to proceed regardless if we're
3583 3590 * overflowing; thus we don't use a while loop here.
3584 3591 */
3585 3592 if (arc_is_overflowing()) {
3586 3593 mutex_enter(&arc_reclaim_lock);
3587 3594
3588 3595 /*
3589 3596 * Now that we've acquired the lock, we may no longer be
3590 3597 * over the overflow limit, lets check.
3591 3598 *
3592 3599 * We're ignoring the case of spurious wake ups. If that
3593 3600 * were to happen, it'd let this thread consume an ARC
3594 3601 * buffer before it should have (i.e. before we're under
3595 3602 * the overflow limit and were signalled by the reclaim
3596 3603 * thread). As long as that is a rare occurrence, it
3597 3604 * shouldn't cause any harm.
3598 3605 */
3599 3606 if (arc_is_overflowing()) {
3600 3607 cv_signal(&arc_reclaim_thread_cv);
3601 3608 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
3602 3609 }
3603 3610
3604 3611 mutex_exit(&arc_reclaim_lock);
3605 3612 }
3606 3613
3607 3614 if (type == ARC_BUFC_METADATA) {
3608 3615 buf->b_data = zio_buf_alloc(size);
3609 3616 arc_space_consume(size, ARC_SPACE_META);
3610 3617 } else {
3611 3618 ASSERT(type == ARC_BUFC_DATA);
3612 3619 buf->b_data = zio_data_buf_alloc(size);
3613 3620 arc_space_consume(size, ARC_SPACE_DATA);
3614 3621 }
3615 3622
3616 3623 /*
3617 3624 * Update the state size. Note that ghost states have a
3618 3625 * "ghost size" and so don't need to be updated.
3619 3626 */
3620 3627 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) {
3621 3628 arc_buf_hdr_t *hdr = buf->b_hdr;
3622 3629 arc_state_t *state = hdr->b_l1hdr.b_state;
3623 3630
3624 3631 (void) refcount_add_many(&state->arcs_size, size, buf);
3625 3632
3626 3633 /*
3627 3634 * If this is reached via arc_read, the link is
3628 3635 * protected by the hash lock. If reached via
3629 3636 * arc_buf_alloc, the header should not be accessed by
3630 3637 * any other thread. And, if reached via arc_read_done,
3631 3638 * the hash lock will protect it if it's found in the
3632 3639 * hash table; otherwise no other thread should be
3633 3640 * trying to [add|remove]_reference it.
3634 3641 */
3635 3642 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3636 3643 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3637 3644 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type],
3638 3645 size);
3639 3646 }
3640 3647 /*
3641 3648 * If we are growing the cache, and we are adding anonymous
3642 3649 * data, and we have outgrown arc_p, update arc_p
3643 3650 */
3644 3651 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
3645 3652 (refcount_count(&arc_anon->arcs_size) +
3646 3653 refcount_count(&arc_mru->arcs_size) > arc_p))
3647 3654 arc_p = MIN(arc_c, arc_p + size);
3648 3655 }
3649 3656 }
3650 3657
3651 3658 /*
3652 3659 * This routine is called whenever a buffer is accessed.
3653 3660 * NOTE: the hash lock is dropped in this function.
3654 3661 */
3655 3662 static void
3656 3663 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3657 3664 {
3658 3665 clock_t now;
3659 3666
3660 3667 ASSERT(MUTEX_HELD(hash_lock));
3661 3668 ASSERT(HDR_HAS_L1HDR(hdr));
3662 3669
3663 3670 if (hdr->b_l1hdr.b_state == arc_anon) {
3664 3671 /*
3665 3672 * This buffer is not in the cache, and does not
3666 3673 * appear in our "ghost" list. Add the new buffer
3667 3674 * to the MRU state.
3668 3675 */
3669 3676
3670 3677 ASSERT0(hdr->b_l1hdr.b_arc_access);
3671 3678 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3672 3679 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3673 3680 arc_change_state(arc_mru, hdr, hash_lock);
3674 3681
3675 3682 } else if (hdr->b_l1hdr.b_state == arc_mru) {
3676 3683 now = ddi_get_lbolt();
3677 3684
3678 3685 /*
3679 3686 * If this buffer is here because of a prefetch, then either:
3680 3687 * - clear the flag if this is a "referencing" read
3681 3688 * (any subsequent access will bump this into the MFU state).
3682 3689 * or
3683 3690 * - move the buffer to the head of the list if this is
3684 3691 * another prefetch (to make it less likely to be evicted).
3685 3692 */
3686 3693 if (HDR_PREFETCH(hdr)) {
3687 3694 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
3688 3695 /* link protected by hash lock */
3689 3696 ASSERT(multilist_link_active(
3690 3697 &hdr->b_l1hdr.b_arc_node));
3691 3698 } else {
3692 3699 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3693 3700 ARCSTAT_BUMP(arcstat_mru_hits);
3694 3701 }
3695 3702 hdr->b_l1hdr.b_arc_access = now;
3696 3703 return;
3697 3704 }
3698 3705
3699 3706 /*
3700 3707 * This buffer has been "accessed" only once so far,
3701 3708 * but it is still in the cache. Move it to the MFU
3702 3709 * state.
3703 3710 */
3704 3711 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) {
3705 3712 /*
3706 3713 * More than 125ms have passed since we
3707 3714 * instantiated this buffer. Move it to the
3708 3715 * most frequently used state.
3709 3716 */
3710 3717 hdr->b_l1hdr.b_arc_access = now;
3711 3718 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3712 3719 arc_change_state(arc_mfu, hdr, hash_lock);
3713 3720 }
3714 3721 ARCSTAT_BUMP(arcstat_mru_hits);
3715 3722 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) {
3716 3723 arc_state_t *new_state;
3717 3724 /*
3718 3725 * This buffer has been "accessed" recently, but
3719 3726 * was evicted from the cache. Move it to the
3720 3727 * MFU state.
3721 3728 */
3722 3729
3723 3730 if (HDR_PREFETCH(hdr)) {
3724 3731 new_state = arc_mru;
3725 3732 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0)
3726 3733 hdr->b_flags &= ~ARC_FLAG_PREFETCH;
3727 3734 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
3728 3735 } else {
3729 3736 new_state = arc_mfu;
3730 3737 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3731 3738 }
3732 3739
3733 3740 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3734 3741 arc_change_state(new_state, hdr, hash_lock);
3735 3742
3736 3743 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
3737 3744 } else if (hdr->b_l1hdr.b_state == arc_mfu) {
3738 3745 /*
3739 3746 * This buffer has been accessed more than once and is
3740 3747 * still in the cache. Keep it in the MFU state.
3741 3748 *
3742 3749 * NOTE: an add_reference() that occurred when we did
3743 3750 * the arc_read() will have kicked this off the list.
3744 3751 * If it was a prefetch, we will explicitly move it to
3745 3752 * the head of the list now.
3746 3753 */
3747 3754 if ((HDR_PREFETCH(hdr)) != 0) {
3748 3755 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3749 3756 /* link protected by hash_lock */
3750 3757 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3751 3758 }
3752 3759 ARCSTAT_BUMP(arcstat_mfu_hits);
3753 3760 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3754 3761 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) {
3755 3762 arc_state_t *new_state = arc_mfu;
3756 3763 /*
3757 3764 * This buffer has been accessed more than once but has
3758 3765 * been evicted from the cache. Move it back to the
3759 3766 * MFU state.
3760 3767 */
3761 3768
3762 3769 if (HDR_PREFETCH(hdr)) {
3763 3770 /*
3764 3771 * This is a prefetch access...
3765 3772 * move this block back to the MRU state.
3766 3773 */
3767 3774 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt));
3768 3775 new_state = arc_mru;
3769 3776 }
3770 3777
3771 3778 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3772 3779 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3773 3780 arc_change_state(new_state, hdr, hash_lock);
3774 3781
3775 3782 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
3776 3783 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
3777 3784 /*
3778 3785 * This buffer is on the 2nd Level ARC.
3779 3786 */
3780 3787
3781 3788 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
3782 3789 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
3783 3790 arc_change_state(arc_mfu, hdr, hash_lock);
3784 3791 } else {
3785 3792 ASSERT(!"invalid arc state");
3786 3793 }
3787 3794 }
3788 3795
3789 3796 /* a generic arc_done_func_t which you can use */
3790 3797 /* ARGSUSED */
3791 3798 void
3792 3799 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
3793 3800 {
3794 3801 if (zio == NULL || zio->io_error == 0)
3795 3802 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
3796 3803 VERIFY(arc_buf_remove_ref(buf, arg));
3797 3804 }
3798 3805
3799 3806 /* a generic arc_done_func_t */
3800 3807 void
3801 3808 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
3802 3809 {
3803 3810 arc_buf_t **bufp = arg;
3804 3811 if (zio && zio->io_error) {
3805 3812 VERIFY(arc_buf_remove_ref(buf, arg));
3806 3813 *bufp = NULL;
3807 3814 } else {
3808 3815 *bufp = buf;
3809 3816 ASSERT(buf->b_data);
3810 3817 }
3811 3818 }
3812 3819
3813 3820 static void
3814 3821 arc_read_done(zio_t *zio)
3815 3822 {
3816 3823 arc_buf_hdr_t *hdr;
3817 3824 arc_buf_t *buf;
3818 3825 arc_buf_t *abuf; /* buffer we're assigning to callback */
3819 3826 kmutex_t *hash_lock = NULL;
3820 3827 arc_callback_t *callback_list, *acb;
3821 3828 int freeable = FALSE;
3822 3829
3823 3830 buf = zio->io_private;
3824 3831 hdr = buf->b_hdr;
3825 3832
3826 3833 /*
3827 3834 * The hdr was inserted into hash-table and removed from lists
3828 3835 * prior to starting I/O. We should find this header, since
3829 3836 * it's in the hash table, and it should be legit since it's
3830 3837 * not possible to evict it during the I/O. The only possible
3831 3838 * reason for it not to be found is if we were freed during the
3832 3839 * read.
3833 3840 */
3834 3841 if (HDR_IN_HASH_TABLE(hdr)) {
3835 3842 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp));
3836 3843 ASSERT3U(hdr->b_dva.dva_word[0], ==,
3837 3844 BP_IDENTITY(zio->io_bp)->dva_word[0]);
3838 3845 ASSERT3U(hdr->b_dva.dva_word[1], ==,
3839 3846 BP_IDENTITY(zio->io_bp)->dva_word[1]);
3840 3847
3841 3848 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp,
3842 3849 &hash_lock);
3843 3850
3844 3851 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) &&
3845 3852 hash_lock == NULL) ||
3846 3853 (found == hdr &&
3847 3854 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
3848 3855 (found == hdr && HDR_L2_READING(hdr)));
3849 3856 }
3850 3857
3851 3858 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED;
3852 3859 if (l2arc_noprefetch && HDR_PREFETCH(hdr))
3853 3860 hdr->b_flags &= ~ARC_FLAG_L2CACHE;
3854 3861
3855 3862 /* byteswap if necessary */
3856 3863 callback_list = hdr->b_l1hdr.b_acb;
3857 3864 ASSERT(callback_list != NULL);
3858 3865 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) {
3859 3866 dmu_object_byteswap_t bswap =
3860 3867 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp));
3861 3868 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
3862 3869 byteswap_uint64_array :
3863 3870 dmu_ot_byteswap[bswap].ob_func;
3864 3871 func(buf->b_data, hdr->b_size);
3865 3872 }
3866 3873
3867 3874 arc_cksum_compute(buf, B_FALSE);
3868 3875 arc_buf_watch(buf);
3869 3876
3870 3877 if (hash_lock && zio->io_error == 0 &&
3871 3878 hdr->b_l1hdr.b_state == arc_anon) {
3872 3879 /*
3873 3880 * Only call arc_access on anonymous buffers. This is because
3874 3881 * if we've issued an I/O for an evicted buffer, we've already
3875 3882 * called arc_access (to prevent any simultaneous readers from
3876 3883 * getting confused).
3877 3884 */
3878 3885 arc_access(hdr, hash_lock);
3879 3886 }
3880 3887
3881 3888 /* create copies of the data buffer for the callers */
3882 3889 abuf = buf;
3883 3890 for (acb = callback_list; acb; acb = acb->acb_next) {
3884 3891 if (acb->acb_done) {
3885 3892 if (abuf == NULL) {
3886 3893 ARCSTAT_BUMP(arcstat_duplicate_reads);
3887 3894 abuf = arc_buf_clone(buf);
3888 3895 }
3889 3896 acb->acb_buf = abuf;
3890 3897 abuf = NULL;
3891 3898 }
3892 3899 }
3893 3900 hdr->b_l1hdr.b_acb = NULL;
3894 3901 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
3895 3902 ASSERT(!HDR_BUF_AVAILABLE(hdr));
3896 3903 if (abuf == buf) {
3897 3904 ASSERT(buf->b_efunc == NULL);
3898 3905 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
3899 3906 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
3900 3907 }
3901 3908
3902 3909 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
3903 3910 callback_list != NULL);
3904 3911
3905 3912 if (zio->io_error != 0) {
3906 3913 hdr->b_flags |= ARC_FLAG_IO_ERROR;
3907 3914 if (hdr->b_l1hdr.b_state != arc_anon)
3908 3915 arc_change_state(arc_anon, hdr, hash_lock);
3909 3916 if (HDR_IN_HASH_TABLE(hdr))
3910 3917 buf_hash_remove(hdr);
3911 3918 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
3912 3919 }
3913 3920
3914 3921 /*
3915 3922 * Broadcast before we drop the hash_lock to avoid the possibility
3916 3923 * that the hdr (and hence the cv) might be freed before we get to
3917 3924 * the cv_broadcast().
3918 3925 */
3919 3926 cv_broadcast(&hdr->b_l1hdr.b_cv);
3920 3927
3921 3928 if (hash_lock != NULL) {
3922 3929 mutex_exit(hash_lock);
3923 3930 } else {
3924 3931 /*
3925 3932 * This block was freed while we waited for the read to
3926 3933 * complete. It has been removed from the hash table and
3927 3934 * moved to the anonymous state (so that it won't show up
3928 3935 * in the cache).
3929 3936 */
3930 3937 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3931 3938 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
3932 3939 }
3933 3940
3934 3941 /* execute each callback and free its structure */
3935 3942 while ((acb = callback_list) != NULL) {
3936 3943 if (acb->acb_done)
3937 3944 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
3938 3945
3939 3946 if (acb->acb_zio_dummy != NULL) {
3940 3947 acb->acb_zio_dummy->io_error = zio->io_error;
3941 3948 zio_nowait(acb->acb_zio_dummy);
3942 3949 }
3943 3950
3944 3951 callback_list = acb->acb_next;
3945 3952 kmem_free(acb, sizeof (arc_callback_t));
3946 3953 }
3947 3954
3948 3955 if (freeable)
3949 3956 arc_hdr_destroy(hdr);
3950 3957 }
3951 3958
3952 3959 /*
3953 3960 * "Read" the block at the specified DVA (in bp) via the
3954 3961 * cache. If the block is found in the cache, invoke the provided
3955 3962 * callback immediately and return. Note that the `zio' parameter
3956 3963 * in the callback will be NULL in this case, since no IO was
3957 3964 * required. If the block is not in the cache pass the read request
3958 3965 * on to the spa with a substitute callback function, so that the
3959 3966 * requested block will be added to the cache.
3960 3967 *
3961 3968 * If a read request arrives for a block that has a read in-progress,
3962 3969 * either wait for the in-progress read to complete (and return the
3963 3970 * results); or, if this is a read with a "done" func, add a record
3964 3971 * to the read to invoke the "done" func when the read completes,
3965 3972 * and return; or just return.
3966 3973 *
3967 3974 * arc_read_done() will invoke all the requested "done" functions
3968 3975 * for readers of this block.
3969 3976 */
3970 3977 int
3971 3978 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
3972 3979 void *private, zio_priority_t priority, int zio_flags,
3973 3980 arc_flags_t *arc_flags, const zbookmark_phys_t *zb)
3974 3981 {
3975 3982 arc_buf_hdr_t *hdr = NULL;
3976 3983 arc_buf_t *buf = NULL;
3977 3984 kmutex_t *hash_lock = NULL;
3978 3985 zio_t *rzio;
3979 3986 uint64_t guid = spa_load_guid(spa);
3980 3987
3981 3988 ASSERT(!BP_IS_EMBEDDED(bp) ||
3982 3989 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA);
3983 3990
3984 3991 top:
3985 3992 if (!BP_IS_EMBEDDED(bp)) {
3986 3993 /*
3987 3994 * Embedded BP's have no DVA and require no I/O to "read".
3988 3995 * Create an anonymous arc buf to back it.
3989 3996 */
3990 3997 hdr = buf_hash_find(guid, bp, &hash_lock);
3991 3998 }
3992 3999
3993 4000 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) {
3994 4001
3995 4002 *arc_flags |= ARC_FLAG_CACHED;
3996 4003
3997 4004 if (HDR_IO_IN_PROGRESS(hdr)) {
3998 4005
3999 4006 if (*arc_flags & ARC_FLAG_WAIT) {
4000 4007 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
4001 4008 mutex_exit(hash_lock);
4002 4009 goto top;
4003 4010 }
4004 4011 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4005 4012
4006 4013 if (done) {
4007 4014 arc_callback_t *acb = NULL;
4008 4015
4009 4016 acb = kmem_zalloc(sizeof (arc_callback_t),
4010 4017 KM_SLEEP);
4011 4018 acb->acb_done = done;
4012 4019 acb->acb_private = private;
4013 4020 if (pio != NULL)
4014 4021 acb->acb_zio_dummy = zio_null(pio,
4015 4022 spa, NULL, NULL, NULL, zio_flags);
4016 4023
4017 4024 ASSERT(acb->acb_done != NULL);
4018 4025 acb->acb_next = hdr->b_l1hdr.b_acb;
4019 4026 hdr->b_l1hdr.b_acb = acb;
4020 4027 add_reference(hdr, hash_lock, private);
4021 4028 mutex_exit(hash_lock);
4022 4029 return (0);
4023 4030 }
4024 4031 mutex_exit(hash_lock);
4025 4032 return (0);
4026 4033 }
4027 4034
4028 4035 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4029 4036 hdr->b_l1hdr.b_state == arc_mfu);
4030 4037
4031 4038 if (done) {
4032 4039 add_reference(hdr, hash_lock, private);
4033 4040 /*
4034 4041 * If this block is already in use, create a new
4035 4042 * copy of the data so that we will be guaranteed
4036 4043 * that arc_release() will always succeed.
4037 4044 */
4038 4045 buf = hdr->b_l1hdr.b_buf;
4039 4046 ASSERT(buf);
4040 4047 ASSERT(buf->b_data);
4041 4048 if (HDR_BUF_AVAILABLE(hdr)) {
4042 4049 ASSERT(buf->b_efunc == NULL);
4043 4050 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4044 4051 } else {
4045 4052 buf = arc_buf_clone(buf);
4046 4053 }
4047 4054
4048 4055 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4049 4056 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4050 4057 hdr->b_flags |= ARC_FLAG_PREFETCH;
4051 4058 }
4052 4059 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4053 4060 arc_access(hdr, hash_lock);
4054 4061 if (*arc_flags & ARC_FLAG_L2CACHE)
4055 4062 hdr->b_flags |= ARC_FLAG_L2CACHE;
4056 4063 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4057 4064 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4058 4065 mutex_exit(hash_lock);
4059 4066 ARCSTAT_BUMP(arcstat_hits);
4060 4067 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4061 4068 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4062 4069 data, metadata, hits);
4063 4070
4064 4071 if (done)
4065 4072 done(NULL, buf, private);
4066 4073 } else {
4067 4074 uint64_t size = BP_GET_LSIZE(bp);
4068 4075 arc_callback_t *acb;
4069 4076 vdev_t *vd = NULL;
4070 4077 uint64_t addr = 0;
4071 4078 boolean_t devw = B_FALSE;
4072 4079 enum zio_compress b_compress = ZIO_COMPRESS_OFF;
4073 4080 int32_t b_asize = 0;
4074 4081
4075 4082 if (hdr == NULL) {
4076 4083 /* this block is not in the cache */
4077 4084 arc_buf_hdr_t *exists = NULL;
4078 4085 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
4079 4086 buf = arc_buf_alloc(spa, size, private, type);
4080 4087 hdr = buf->b_hdr;
4081 4088 if (!BP_IS_EMBEDDED(bp)) {
4082 4089 hdr->b_dva = *BP_IDENTITY(bp);
4083 4090 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
4084 4091 exists = buf_hash_insert(hdr, &hash_lock);
4085 4092 }
4086 4093 if (exists != NULL) {
4087 4094 /* somebody beat us to the hash insert */
4088 4095 mutex_exit(hash_lock);
4089 4096 buf_discard_identity(hdr);
4090 4097 (void) arc_buf_remove_ref(buf, private);
4091 4098 goto top; /* restart the IO request */
4092 4099 }
4093 4100
4094 4101 /* if this is a prefetch, we don't have a reference */
4095 4102 if (*arc_flags & ARC_FLAG_PREFETCH) {
4096 4103 (void) remove_reference(hdr, hash_lock,
4097 4104 private);
4098 4105 hdr->b_flags |= ARC_FLAG_PREFETCH;
4099 4106 }
4100 4107 if (*arc_flags & ARC_FLAG_L2CACHE)
4101 4108 hdr->b_flags |= ARC_FLAG_L2CACHE;
4102 4109 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4103 4110 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4104 4111 if (BP_GET_LEVEL(bp) > 0)
4105 4112 hdr->b_flags |= ARC_FLAG_INDIRECT;
4106 4113 } else {
4107 4114 /*
4108 4115 * This block is in the ghost cache. If it was L2-only
4109 4116 * (and thus didn't have an L1 hdr), we realloc the
4110 4117 * header to add an L1 hdr.
4111 4118 */
4112 4119 if (!HDR_HAS_L1HDR(hdr)) {
4113 4120 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
4114 4121 hdr_full_cache);
4115 4122 }
4116 4123
4117 4124 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
4118 4125 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4119 4126 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4120 4127 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
4121 4128
4122 4129 /* if this is a prefetch, we don't have a reference */
4123 4130 if (*arc_flags & ARC_FLAG_PREFETCH)
4124 4131 hdr->b_flags |= ARC_FLAG_PREFETCH;
4125 4132 else
4126 4133 add_reference(hdr, hash_lock, private);
4127 4134 if (*arc_flags & ARC_FLAG_L2CACHE)
4128 4135 hdr->b_flags |= ARC_FLAG_L2CACHE;
4129 4136 if (*arc_flags & ARC_FLAG_L2COMPRESS)
4130 4137 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4131 4138 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
4132 4139 buf->b_hdr = hdr;
4133 4140 buf->b_data = NULL;
4134 4141 buf->b_efunc = NULL;
4135 4142 buf->b_private = NULL;
4136 4143 buf->b_next = NULL;
4137 4144 hdr->b_l1hdr.b_buf = buf;
4138 4145 ASSERT0(hdr->b_l1hdr.b_datacnt);
4139 4146 hdr->b_l1hdr.b_datacnt = 1;
4140 4147 arc_get_data_buf(buf);
4141 4148 arc_access(hdr, hash_lock);
4142 4149 }
4143 4150
4144 4151 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
4145 4152
4146 4153 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
4147 4154 acb->acb_done = done;
4148 4155 acb->acb_private = private;
4149 4156
4150 4157 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4151 4158 hdr->b_l1hdr.b_acb = acb;
4152 4159 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4153 4160
4154 4161 if (HDR_HAS_L2HDR(hdr) &&
4155 4162 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
4156 4163 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
4157 4164 addr = hdr->b_l2hdr.b_daddr;
4158 4165 b_compress = hdr->b_l2hdr.b_compress;
4159 4166 b_asize = hdr->b_l2hdr.b_asize;
4160 4167 /*
4161 4168 * Lock out device removal.
4162 4169 */
4163 4170 if (vdev_is_dead(vd) ||
4164 4171 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
4165 4172 vd = NULL;
4166 4173 }
4167 4174
4168 4175 if (hash_lock != NULL)
4169 4176 mutex_exit(hash_lock);
4170 4177
4171 4178 /*
4172 4179 * At this point, we have a level 1 cache miss. Try again in
4173 4180 * L2ARC if possible.
4174 4181 */
4175 4182 ASSERT3U(hdr->b_size, ==, size);
4176 4183 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
4177 4184 uint64_t, size, zbookmark_phys_t *, zb);
4178 4185 ARCSTAT_BUMP(arcstat_misses);
4179 4186 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4180 4187 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4181 4188 data, metadata, misses);
4182 4189
4183 4190 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
4184 4191 /*
4185 4192 * Read from the L2ARC if the following are true:
4186 4193 * 1. The L2ARC vdev was previously cached.
4187 4194 * 2. This buffer still has L2ARC metadata.
4188 4195 * 3. This buffer isn't currently writing to the L2ARC.
4189 4196 * 4. The L2ARC entry wasn't evicted, which may
4190 4197 * also have invalidated the vdev.
4191 4198 * 5. This isn't prefetch and l2arc_noprefetch is set.
4192 4199 */
4193 4200 if (HDR_HAS_L2HDR(hdr) &&
4194 4201 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
4195 4202 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
4196 4203 l2arc_read_callback_t *cb;
4197 4204
4198 4205 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
4199 4206 ARCSTAT_BUMP(arcstat_l2_hits);
4200 4207
4201 4208 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
4202 4209 KM_SLEEP);
4203 4210 cb->l2rcb_buf = buf;
4204 4211 cb->l2rcb_spa = spa;
4205 4212 cb->l2rcb_bp = *bp;
4206 4213 cb->l2rcb_zb = *zb;
4207 4214 cb->l2rcb_flags = zio_flags;
4208 4215 cb->l2rcb_compress = b_compress;
4209 4216
4210 4217 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
4211 4218 addr + size < vd->vdev_psize -
4212 4219 VDEV_LABEL_END_SIZE);
4213 4220
4214 4221 /*
4215 4222 * l2arc read. The SCL_L2ARC lock will be
4216 4223 * released by l2arc_read_done().
4217 4224 * Issue a null zio if the underlying buffer
4218 4225 * was squashed to zero size by compression.
4219 4226 */
4220 4227 if (b_compress == ZIO_COMPRESS_EMPTY) {
4221 4228 rzio = zio_null(pio, spa, vd,
4222 4229 l2arc_read_done, cb,
4223 4230 zio_flags | ZIO_FLAG_DONT_CACHE |
4224 4231 ZIO_FLAG_CANFAIL |
4225 4232 ZIO_FLAG_DONT_PROPAGATE |
4226 4233 ZIO_FLAG_DONT_RETRY);
4227 4234 } else {
4228 4235 rzio = zio_read_phys(pio, vd, addr,
4229 4236 b_asize, buf->b_data,
4230 4237 ZIO_CHECKSUM_OFF,
4231 4238 l2arc_read_done, cb, priority,
4232 4239 zio_flags | ZIO_FLAG_DONT_CACHE |
4233 4240 ZIO_FLAG_CANFAIL |
4234 4241 ZIO_FLAG_DONT_PROPAGATE |
4235 4242 ZIO_FLAG_DONT_RETRY, B_FALSE);
4236 4243 }
4237 4244 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
4238 4245 zio_t *, rzio);
4239 4246 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize);
4240 4247
4241 4248 if (*arc_flags & ARC_FLAG_NOWAIT) {
4242 4249 zio_nowait(rzio);
4243 4250 return (0);
4244 4251 }
4245 4252
4246 4253 ASSERT(*arc_flags & ARC_FLAG_WAIT);
4247 4254 if (zio_wait(rzio) == 0)
4248 4255 return (0);
4249 4256
4250 4257 /* l2arc read error; goto zio_read() */
4251 4258 } else {
4252 4259 DTRACE_PROBE1(l2arc__miss,
4253 4260 arc_buf_hdr_t *, hdr);
4254 4261 ARCSTAT_BUMP(arcstat_l2_misses);
4255 4262 if (HDR_L2_WRITING(hdr))
4256 4263 ARCSTAT_BUMP(arcstat_l2_rw_clash);
4257 4264 spa_config_exit(spa, SCL_L2ARC, vd);
4258 4265 }
4259 4266 } else {
4260 4267 if (vd != NULL)
4261 4268 spa_config_exit(spa, SCL_L2ARC, vd);
4262 4269 if (l2arc_ndev != 0) {
4263 4270 DTRACE_PROBE1(l2arc__miss,
4264 4271 arc_buf_hdr_t *, hdr);
4265 4272 ARCSTAT_BUMP(arcstat_l2_misses);
4266 4273 }
4267 4274 }
4268 4275
4269 4276 rzio = zio_read(pio, spa, bp, buf->b_data, size,
4270 4277 arc_read_done, buf, priority, zio_flags, zb);
4271 4278
4272 4279 if (*arc_flags & ARC_FLAG_WAIT)
4273 4280 return (zio_wait(rzio));
4274 4281
4275 4282 ASSERT(*arc_flags & ARC_FLAG_NOWAIT);
4276 4283 zio_nowait(rzio);
4277 4284 }
4278 4285 return (0);
4279 4286 }
4280 4287
4281 4288 void
4282 4289 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
4283 4290 {
4284 4291 ASSERT(buf->b_hdr != NULL);
4285 4292 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon);
4286 4293 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) ||
4287 4294 func == NULL);
4288 4295 ASSERT(buf->b_efunc == NULL);
4289 4296 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr));
4290 4297
4291 4298 buf->b_efunc = func;
4292 4299 buf->b_private = private;
4293 4300 }
4294 4301
4295 4302 /*
4296 4303 * Notify the arc that a block was freed, and thus will never be used again.
4297 4304 */
4298 4305 void
4299 4306 arc_freed(spa_t *spa, const blkptr_t *bp)
4300 4307 {
4301 4308 arc_buf_hdr_t *hdr;
4302 4309 kmutex_t *hash_lock;
4303 4310 uint64_t guid = spa_load_guid(spa);
4304 4311
4305 4312 ASSERT(!BP_IS_EMBEDDED(bp));
4306 4313
4307 4314 hdr = buf_hash_find(guid, bp, &hash_lock);
4308 4315 if (hdr == NULL)
4309 4316 return;
4310 4317 if (HDR_BUF_AVAILABLE(hdr)) {
4311 4318 arc_buf_t *buf = hdr->b_l1hdr.b_buf;
4312 4319 add_reference(hdr, hash_lock, FTAG);
4313 4320 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE;
4314 4321 mutex_exit(hash_lock);
4315 4322
4316 4323 arc_release(buf, FTAG);
4317 4324 (void) arc_buf_remove_ref(buf, FTAG);
4318 4325 } else {
4319 4326 mutex_exit(hash_lock);
4320 4327 }
4321 4328
4322 4329 }
4323 4330
4324 4331 /*
4325 4332 * Clear the user eviction callback set by arc_set_callback(), first calling
4326 4333 * it if it exists. Because the presence of a callback keeps an arc_buf cached
4327 4334 * clearing the callback may result in the arc_buf being destroyed. However,
4328 4335 * it will not result in the *last* arc_buf being destroyed, hence the data
4329 4336 * will remain cached in the ARC. We make a copy of the arc buffer here so
4330 4337 * that we can process the callback without holding any locks.
4331 4338 *
4332 4339 * It's possible that the callback is already in the process of being cleared
4333 4340 * by another thread. In this case we can not clear the callback.
4334 4341 *
4335 4342 * Returns B_TRUE if the callback was successfully called and cleared.
4336 4343 */
4337 4344 boolean_t
4338 4345 arc_clear_callback(arc_buf_t *buf)
4339 4346 {
4340 4347 arc_buf_hdr_t *hdr;
4341 4348 kmutex_t *hash_lock;
4342 4349 arc_evict_func_t *efunc = buf->b_efunc;
4343 4350 void *private = buf->b_private;
4344 4351
4345 4352 mutex_enter(&buf->b_evict_lock);
4346 4353 hdr = buf->b_hdr;
4347 4354 if (hdr == NULL) {
4348 4355 /*
4349 4356 * We are in arc_do_user_evicts().
4350 4357 */
4351 4358 ASSERT(buf->b_data == NULL);
4352 4359 mutex_exit(&buf->b_evict_lock);
4353 4360 return (B_FALSE);
4354 4361 } else if (buf->b_data == NULL) {
4355 4362 /*
4356 4363 * We are on the eviction list; process this buffer now
4357 4364 * but let arc_do_user_evicts() do the reaping.
4358 4365 */
4359 4366 buf->b_efunc = NULL;
4360 4367 mutex_exit(&buf->b_evict_lock);
4361 4368 VERIFY0(efunc(private));
4362 4369 return (B_TRUE);
4363 4370 }
4364 4371 hash_lock = HDR_LOCK(hdr);
4365 4372 mutex_enter(hash_lock);
4366 4373 hdr = buf->b_hdr;
4367 4374 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4368 4375
4369 4376 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <,
4370 4377 hdr->b_l1hdr.b_datacnt);
4371 4378 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
4372 4379 hdr->b_l1hdr.b_state == arc_mfu);
4373 4380
4374 4381 buf->b_efunc = NULL;
4375 4382 buf->b_private = NULL;
4376 4383
4377 4384 if (hdr->b_l1hdr.b_datacnt > 1) {
4378 4385 mutex_exit(&buf->b_evict_lock);
4379 4386 arc_buf_destroy(buf, TRUE);
4380 4387 } else {
4381 4388 ASSERT(buf == hdr->b_l1hdr.b_buf);
4382 4389 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE;
4383 4390 mutex_exit(&buf->b_evict_lock);
4384 4391 }
4385 4392
4386 4393 mutex_exit(hash_lock);
4387 4394 VERIFY0(efunc(private));
4388 4395 return (B_TRUE);
4389 4396 }
4390 4397
4391 4398 /*
4392 4399 * Release this buffer from the cache, making it an anonymous buffer. This
4393 4400 * must be done after a read and prior to modifying the buffer contents.
4394 4401 * If the buffer has more than one reference, we must make
4395 4402 * a new hdr for the buffer.
4396 4403 */
4397 4404 void
4398 4405 arc_release(arc_buf_t *buf, void *tag)
4399 4406 {
4400 4407 arc_buf_hdr_t *hdr = buf->b_hdr;
4401 4408
4402 4409 /*
4403 4410 * It would be nice to assert that if it's DMU metadata (level >
4404 4411 * 0 || it's the dnode file), then it must be syncing context.
4405 4412 * But we don't know that information at this level.
4406 4413 */
4407 4414
4408 4415 mutex_enter(&buf->b_evict_lock);
4409 4416
4410 4417 ASSERT(HDR_HAS_L1HDR(hdr));
4411 4418
4412 4419 /*
4413 4420 * We don't grab the hash lock prior to this check, because if
4414 4421 * the buffer's header is in the arc_anon state, it won't be
4415 4422 * linked into the hash table.
4416 4423 */
4417 4424 if (hdr->b_l1hdr.b_state == arc_anon) {
4418 4425 mutex_exit(&buf->b_evict_lock);
4419 4426 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4420 4427 ASSERT(!HDR_IN_HASH_TABLE(hdr));
4421 4428 ASSERT(!HDR_HAS_L2HDR(hdr));
4422 4429 ASSERT(BUF_EMPTY(hdr));
4423 4430
4424 4431 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1);
4425 4432 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1);
4426 4433 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node));
4427 4434
4428 4435 ASSERT3P(buf->b_efunc, ==, NULL);
4429 4436 ASSERT3P(buf->b_private, ==, NULL);
4430 4437
4431 4438 hdr->b_l1hdr.b_arc_access = 0;
4432 4439 arc_buf_thaw(buf);
4433 4440
4434 4441 return;
4435 4442 }
4436 4443
4437 4444 kmutex_t *hash_lock = HDR_LOCK(hdr);
4438 4445 mutex_enter(hash_lock);
4439 4446
4440 4447 /*
4441 4448 * This assignment is only valid as long as the hash_lock is
4442 4449 * held, we must be careful not to reference state or the
4443 4450 * b_state field after dropping the lock.
4444 4451 */
4445 4452 arc_state_t *state = hdr->b_l1hdr.b_state;
4446 4453 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
4447 4454 ASSERT3P(state, !=, arc_anon);
4448 4455
4449 4456 /* this buffer is not on any list */
4450 4457 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0);
4451 4458
4452 4459 if (HDR_HAS_L2HDR(hdr)) {
4453 4460 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4454 4461
4455 4462 /*
4456 4463 * We have to recheck this conditional again now that
4457 4464 * we're holding the l2ad_mtx to prevent a race with
4458 4465 * another thread which might be concurrently calling
4459 4466 * l2arc_evict(). In that case, l2arc_evict() might have
4460 4467 * destroyed the header's L2 portion as we were waiting
4461 4468 * to acquire the l2ad_mtx.
4462 4469 */
4463 4470 if (HDR_HAS_L2HDR(hdr))
4464 4471 arc_hdr_l2hdr_destroy(hdr);
4465 4472
4466 4473 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx);
4467 4474 }
4468 4475
4469 4476 /*
4470 4477 * Do we have more than one buf?
4471 4478 */
4472 4479 if (hdr->b_l1hdr.b_datacnt > 1) {
4473 4480 arc_buf_hdr_t *nhdr;
4474 4481 arc_buf_t **bufp;
4475 4482 uint64_t blksz = hdr->b_size;
4476 4483 uint64_t spa = hdr->b_spa;
4477 4484 arc_buf_contents_t type = arc_buf_type(hdr);
4478 4485 uint32_t flags = hdr->b_flags;
4479 4486
4480 4487 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
4481 4488 /*
4482 4489 * Pull the data off of this hdr and attach it to
4483 4490 * a new anonymous hdr.
4484 4491 */
4485 4492 (void) remove_reference(hdr, hash_lock, tag);
4486 4493 bufp = &hdr->b_l1hdr.b_buf;
4487 4494 while (*bufp != buf)
4488 4495 bufp = &(*bufp)->b_next;
4489 4496 *bufp = buf->b_next;
4490 4497 buf->b_next = NULL;
4491 4498
4492 4499 ASSERT3P(state, !=, arc_l2c_only);
4493 4500
4494 4501 (void) refcount_remove_many(
4495 4502 &state->arcs_size, hdr->b_size, buf);
4496 4503
4497 4504 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) {
4498 4505 ASSERT3P(state, !=, arc_l2c_only);
4499 4506 uint64_t *size = &state->arcs_lsize[type];
4500 4507 ASSERT3U(*size, >=, hdr->b_size);
4501 4508 atomic_add_64(size, -hdr->b_size);
4502 4509 }
4503 4510
4504 4511 /*
4505 4512 * We're releasing a duplicate user data buffer, update
4506 4513 * our statistics accordingly.
4507 4514 */
4508 4515 if (HDR_ISTYPE_DATA(hdr)) {
4509 4516 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers);
4510 4517 ARCSTAT_INCR(arcstat_duplicate_buffers_size,
4511 4518 -hdr->b_size);
4512 4519 }
4513 4520 hdr->b_l1hdr.b_datacnt -= 1;
4514 4521 arc_cksum_verify(buf);
4515 4522 arc_buf_unwatch(buf);
4516 4523
4517 4524 mutex_exit(hash_lock);
4518 4525
4519 4526 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
4520 4527 nhdr->b_size = blksz;
4521 4528 nhdr->b_spa = spa;
4522 4529
4523 4530 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING;
4524 4531 nhdr->b_flags |= arc_bufc_to_flags(type);
4525 4532 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR;
4526 4533
4527 4534 nhdr->b_l1hdr.b_buf = buf;
4528 4535 nhdr->b_l1hdr.b_datacnt = 1;
4529 4536 nhdr->b_l1hdr.b_state = arc_anon;
4530 4537 nhdr->b_l1hdr.b_arc_access = 0;
4531 4538 nhdr->b_l1hdr.b_tmp_cdata = NULL;
4532 4539 nhdr->b_freeze_cksum = NULL;
4533 4540
4534 4541 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
4535 4542 buf->b_hdr = nhdr;
4536 4543 mutex_exit(&buf->b_evict_lock);
4537 4544 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf);
4538 4545 } else {
4539 4546 mutex_exit(&buf->b_evict_lock);
4540 4547 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
4541 4548 /* protected by hash lock, or hdr is on arc_anon */
4542 4549 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
4543 4550 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4544 4551 arc_change_state(arc_anon, hdr, hash_lock);
4545 4552 hdr->b_l1hdr.b_arc_access = 0;
4546 4553 mutex_exit(hash_lock);
4547 4554
4548 4555 buf_discard_identity(hdr);
4549 4556 arc_buf_thaw(buf);
4550 4557 }
4551 4558 buf->b_efunc = NULL;
4552 4559 buf->b_private = NULL;
4553 4560 }
4554 4561
4555 4562 int
4556 4563 arc_released(arc_buf_t *buf)
4557 4564 {
4558 4565 int released;
4559 4566
4560 4567 mutex_enter(&buf->b_evict_lock);
4561 4568 released = (buf->b_data != NULL &&
4562 4569 buf->b_hdr->b_l1hdr.b_state == arc_anon);
4563 4570 mutex_exit(&buf->b_evict_lock);
4564 4571 return (released);
4565 4572 }
4566 4573
4567 4574 #ifdef ZFS_DEBUG
4568 4575 int
4569 4576 arc_referenced(arc_buf_t *buf)
4570 4577 {
4571 4578 int referenced;
4572 4579
4573 4580 mutex_enter(&buf->b_evict_lock);
4574 4581 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt));
4575 4582 mutex_exit(&buf->b_evict_lock);
4576 4583 return (referenced);
4577 4584 }
4578 4585 #endif
4579 4586
4580 4587 static void
4581 4588 arc_write_ready(zio_t *zio)
4582 4589 {
4583 4590 arc_write_callback_t *callback = zio->io_private;
4584 4591 arc_buf_t *buf = callback->awcb_buf;
4585 4592 arc_buf_hdr_t *hdr = buf->b_hdr;
4586 4593
4587 4594 ASSERT(HDR_HAS_L1HDR(hdr));
4588 4595 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt));
4589 4596 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4590 4597 callback->awcb_ready(zio, buf, callback->awcb_private);
4591 4598
4592 4599 /*
4593 4600 * If the IO is already in progress, then this is a re-write
4594 4601 * attempt, so we need to thaw and re-compute the cksum.
4595 4602 * It is the responsibility of the callback to handle the
4596 4603 * accounting for any re-write attempt.
4597 4604 */
4598 4605 if (HDR_IO_IN_PROGRESS(hdr)) {
4599 4606 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
4600 4607 if (hdr->b_freeze_cksum != NULL) {
4601 4608 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
4602 4609 hdr->b_freeze_cksum = NULL;
4603 4610 }
4604 4611 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
4605 4612 }
4606 4613 arc_cksum_compute(buf, B_FALSE);
4607 4614 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS;
4608 4615 }
4609 4616
4610 4617 /*
4611 4618 * The SPA calls this callback for each physical write that happens on behalf
4612 4619 * of a logical write. See the comment in dbuf_write_physdone() for details.
4613 4620 */
4614 4621 static void
4615 4622 arc_write_physdone(zio_t *zio)
4616 4623 {
4617 4624 arc_write_callback_t *cb = zio->io_private;
4618 4625 if (cb->awcb_physdone != NULL)
4619 4626 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private);
4620 4627 }
4621 4628
4622 4629 static void
4623 4630 arc_write_done(zio_t *zio)
4624 4631 {
4625 4632 arc_write_callback_t *callback = zio->io_private;
4626 4633 arc_buf_t *buf = callback->awcb_buf;
4627 4634 arc_buf_hdr_t *hdr = buf->b_hdr;
4628 4635
4629 4636 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4630 4637
4631 4638 if (zio->io_error == 0) {
4632 4639 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) {
4633 4640 buf_discard_identity(hdr);
4634 4641 } else {
4635 4642 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
4636 4643 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp);
4637 4644 }
4638 4645 } else {
4639 4646 ASSERT(BUF_EMPTY(hdr));
4640 4647 }
4641 4648
4642 4649 /*
4643 4650 * If the block to be written was all-zero or compressed enough to be
4644 4651 * embedded in the BP, no write was performed so there will be no
4645 4652 * dva/birth/checksum. The buffer must therefore remain anonymous
4646 4653 * (and uncached).
4647 4654 */
4648 4655 if (!BUF_EMPTY(hdr)) {
4649 4656 arc_buf_hdr_t *exists;
4650 4657 kmutex_t *hash_lock;
4651 4658
4652 4659 ASSERT(zio->io_error == 0);
4653 4660
4654 4661 arc_cksum_verify(buf);
4655 4662
4656 4663 exists = buf_hash_insert(hdr, &hash_lock);
4657 4664 if (exists != NULL) {
4658 4665 /*
4659 4666 * This can only happen if we overwrite for
4660 4667 * sync-to-convergence, because we remove
4661 4668 * buffers from the hash table when we arc_free().
4662 4669 */
4663 4670 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
4664 4671 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4665 4672 panic("bad overwrite, hdr=%p exists=%p",
4666 4673 (void *)hdr, (void *)exists);
4667 4674 ASSERT(refcount_is_zero(
4668 4675 &exists->b_l1hdr.b_refcnt));
4669 4676 arc_change_state(arc_anon, exists, hash_lock);
4670 4677 mutex_exit(hash_lock);
4671 4678 arc_hdr_destroy(exists);
4672 4679 exists = buf_hash_insert(hdr, &hash_lock);
4673 4680 ASSERT3P(exists, ==, NULL);
4674 4681 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
4675 4682 /* nopwrite */
4676 4683 ASSERT(zio->io_prop.zp_nopwrite);
4677 4684 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
4678 4685 panic("bad nopwrite, hdr=%p exists=%p",
4679 4686 (void *)hdr, (void *)exists);
4680 4687 } else {
4681 4688 /* Dedup */
4682 4689 ASSERT(hdr->b_l1hdr.b_datacnt == 1);
4683 4690 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
4684 4691 ASSERT(BP_GET_DEDUP(zio->io_bp));
4685 4692 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
4686 4693 }
4687 4694 }
4688 4695 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4689 4696 /* if it's not anon, we are doing a scrub */
4690 4697 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
4691 4698 arc_access(hdr, hash_lock);
4692 4699 mutex_exit(hash_lock);
4693 4700 } else {
4694 4701 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS;
4695 4702 }
4696 4703
4697 4704 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4698 4705 callback->awcb_done(zio, buf, callback->awcb_private);
4699 4706
4700 4707 kmem_free(callback, sizeof (arc_write_callback_t));
4701 4708 }
4702 4709
4703 4710 zio_t *
4704 4711 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
4705 4712 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress,
4706 4713 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone,
4707 4714 arc_done_func_t *done, void *private, zio_priority_t priority,
4708 4715 int zio_flags, const zbookmark_phys_t *zb)
4709 4716 {
4710 4717 arc_buf_hdr_t *hdr = buf->b_hdr;
4711 4718 arc_write_callback_t *callback;
4712 4719 zio_t *zio;
4713 4720
4714 4721 ASSERT(ready != NULL);
4715 4722 ASSERT(done != NULL);
4716 4723 ASSERT(!HDR_IO_ERROR(hdr));
4717 4724 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
4718 4725 ASSERT(hdr->b_l1hdr.b_acb == NULL);
4719 4726 ASSERT(hdr->b_l1hdr.b_datacnt > 0);
4720 4727 if (l2arc)
4721 4728 hdr->b_flags |= ARC_FLAG_L2CACHE;
4722 4729 if (l2arc_compress)
4723 4730 hdr->b_flags |= ARC_FLAG_L2COMPRESS;
4724 4731 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
4725 4732 callback->awcb_ready = ready;
4726 4733 callback->awcb_physdone = physdone;
4727 4734 callback->awcb_done = done;
4728 4735 callback->awcb_private = private;
4729 4736 callback->awcb_buf = buf;
4730 4737
4731 4738 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp,
4732 4739 arc_write_ready, arc_write_physdone, arc_write_done, callback,
4733 4740 priority, zio_flags, zb);
4734 4741
4735 4742 return (zio);
4736 4743 }
4737 4744
4738 4745 static int
4739 4746 arc_memory_throttle(uint64_t reserve, uint64_t txg)
4740 4747 {
4741 4748 #ifdef _KERNEL
4742 4749 uint64_t available_memory = ptob(freemem);
4743 4750 static uint64_t page_load = 0;
4744 4751 static uint64_t last_txg = 0;
4745 4752
4746 4753 #if defined(__i386)
4747 4754 available_memory =
4748 4755 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
4749 4756 #endif
4750 4757
4751 4758 if (freemem > physmem * arc_lotsfree_percent / 100)
4752 4759 return (0);
4753 4760
4754 4761 if (txg > last_txg) {
4755 4762 last_txg = txg;
4756 4763 page_load = 0;
4757 4764 }
4758 4765 /*
4759 4766 * If we are in pageout, we know that memory is already tight,
4760 4767 * the arc is already going to be evicting, so we just want to
4761 4768 * continue to let page writes occur as quickly as possible.
4762 4769 */
4763 4770 if (curproc == proc_pageout) {
4764 4771 if (page_load > MAX(ptob(minfree), available_memory) / 4)
4765 4772 return (SET_ERROR(ERESTART));
4766 4773 /* Note: reserve is inflated, so we deflate */
4767 4774 page_load += reserve / 8;
4768 4775 return (0);
4769 4776 } else if (page_load > 0 && arc_reclaim_needed()) {
4770 4777 /* memory is low, delay before restarting */
4771 4778 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
4772 4779 return (SET_ERROR(EAGAIN));
4773 4780 }
4774 4781 page_load = 0;
4775 4782 #endif
4776 4783 return (0);
4777 4784 }
4778 4785
4779 4786 void
4780 4787 arc_tempreserve_clear(uint64_t reserve)
4781 4788 {
4782 4789 atomic_add_64(&arc_tempreserve, -reserve);
4783 4790 ASSERT((int64_t)arc_tempreserve >= 0);
4784 4791 }
4785 4792
4786 4793 int
4787 4794 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
4788 4795 {
4789 4796 int error;
4790 4797 uint64_t anon_size;
4791 4798
4792 4799 if (reserve > arc_c/4 && !arc_no_grow)
4793 4800 arc_c = MIN(arc_c_max, reserve * 4);
4794 4801 if (reserve > arc_c)
4795 4802 return (SET_ERROR(ENOMEM));
4796 4803
4797 4804 /*
4798 4805 * Don't count loaned bufs as in flight dirty data to prevent long
4799 4806 * network delays from blocking transactions that are ready to be
4800 4807 * assigned to a txg.
4801 4808 */
4802 4809 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
4803 4810 arc_loaned_bytes), 0);
4804 4811
4805 4812 /*
4806 4813 * Writes will, almost always, require additional memory allocations
4807 4814 * in order to compress/encrypt/etc the data. We therefore need to
4808 4815 * make sure that there is sufficient available memory for this.
4809 4816 */
4810 4817 error = arc_memory_throttle(reserve, txg);
4811 4818 if (error != 0)
4812 4819 return (error);
4813 4820
4814 4821 /*
4815 4822 * Throttle writes when the amount of dirty data in the cache
4816 4823 * gets too large. We try to keep the cache less than half full
4817 4824 * of dirty blocks so that our sync times don't grow too large.
4818 4825 * Note: if two requests come in concurrently, we might let them
4819 4826 * both succeed, when one of them should fail. Not a huge deal.
4820 4827 */
4821 4828
4822 4829 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
4823 4830 anon_size > arc_c / 4) {
4824 4831 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
4825 4832 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
4826 4833 arc_tempreserve>>10,
4827 4834 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
4828 4835 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
4829 4836 reserve>>10, arc_c>>10);
4830 4837 return (SET_ERROR(ERESTART));
4831 4838 }
4832 4839 atomic_add_64(&arc_tempreserve, reserve);
4833 4840 return (0);
4834 4841 }
4835 4842
4836 4843 static void
4837 4844 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
4838 4845 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
4839 4846 {
4840 4847 size->value.ui64 = refcount_count(&state->arcs_size);
4841 4848 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA];
4842 4849 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA];
4843 4850 }
4844 4851
4845 4852 static int
4846 4853 arc_kstat_update(kstat_t *ksp, int rw)
4847 4854 {
4848 4855 arc_stats_t *as = ksp->ks_data;
4849 4856
4850 4857 if (rw == KSTAT_WRITE) {
4851 4858 return (EACCES);
4852 4859 } else {
4853 4860 arc_kstat_update_state(arc_anon,
4854 4861 &as->arcstat_anon_size,
4855 4862 &as->arcstat_anon_evictable_data,
4856 4863 &as->arcstat_anon_evictable_metadata);
4857 4864 arc_kstat_update_state(arc_mru,
4858 4865 &as->arcstat_mru_size,
4859 4866 &as->arcstat_mru_evictable_data,
4860 4867 &as->arcstat_mru_evictable_metadata);
4861 4868 arc_kstat_update_state(arc_mru_ghost,
4862 4869 &as->arcstat_mru_ghost_size,
4863 4870 &as->arcstat_mru_ghost_evictable_data,
4864 4871 &as->arcstat_mru_ghost_evictable_metadata);
4865 4872 arc_kstat_update_state(arc_mfu,
4866 4873 &as->arcstat_mfu_size,
4867 4874 &as->arcstat_mfu_evictable_data,
4868 4875 &as->arcstat_mfu_evictable_metadata);
4869 4876 arc_kstat_update_state(arc_mfu_ghost,
4870 4877 &as->arcstat_mfu_ghost_size,
4871 4878 &as->arcstat_mfu_ghost_evictable_data,
4872 4879 &as->arcstat_mfu_ghost_evictable_metadata);
4873 4880 }
4874 4881
4875 4882 return (0);
4876 4883 }
4877 4884
4878 4885 /*
4879 4886 * This function *must* return indices evenly distributed between all
4880 4887 * sublists of the multilist. This is needed due to how the ARC eviction
4881 4888 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
4882 4889 * distributed between all sublists and uses this assumption when
4883 4890 * deciding which sublist to evict from and how much to evict from it.
4884 4891 */
4885 4892 unsigned int
4886 4893 arc_state_multilist_index_func(multilist_t *ml, void *obj)
4887 4894 {
4888 4895 arc_buf_hdr_t *hdr = obj;
4889 4896
4890 4897 /*
4891 4898 * We rely on b_dva to generate evenly distributed index
4892 4899 * numbers using buf_hash below. So, as an added precaution,
4893 4900 * let's make sure we never add empty buffers to the arc lists.
4894 4901 */
4895 4902 ASSERT(!BUF_EMPTY(hdr));
4896 4903
4897 4904 /*
4898 4905 * The assumption here, is the hash value for a given
4899 4906 * arc_buf_hdr_t will remain constant throughout it's lifetime
4900 4907 * (i.e. it's b_spa, b_dva, and b_birth fields don't change).
4901 4908 * Thus, we don't need to store the header's sublist index
4902 4909 * on insertion, as this index can be recalculated on removal.
4903 4910 *
4904 4911 * Also, the low order bits of the hash value are thought to be
4905 4912 * distributed evenly. Otherwise, in the case that the multilist
4906 4913 * has a power of two number of sublists, each sublists' usage
4907 4914 * would not be evenly distributed.
4908 4915 */
4909 4916 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
4910 4917 multilist_get_num_sublists(ml));
4911 4918 }
4912 4919
4913 4920 void
4914 4921 arc_init(void)
4915 4922 {
4916 4923 /*
4917 4924 * allmem is "all memory that we could possibly use".
4918 4925 */
4919 4926 #ifdef _KERNEL
4920 4927 uint64_t allmem = ptob(physmem - swapfs_minfree);
4921 4928 #else
4922 4929 uint64_t allmem = (physmem * PAGESIZE) / 2;
4923 4930 #endif
4924 4931
4925 4932 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
4926 4933 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL);
4927 4934 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL);
4928 4935
4929 4936 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL);
4930 4937 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL);
4931 4938
4932 4939 /* Convert seconds to clock ticks */
4933 4940 arc_min_prefetch_lifespan = 1 * hz;
4934 4941
4935 4942 /* Start out with 1/8 of all memory */
4936 4943 arc_c = allmem / 8;
4937 4944
4938 4945 #ifdef _KERNEL
4939 4946 /*
4940 4947 * On architectures where the physical memory can be larger
4941 4948 * than the addressable space (intel in 32-bit mode), we may
4942 4949 * need to limit the cache to 1/8 of VM size.
4943 4950 */
4944 4951 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
4945 4952 #endif
4946 4953
4947 4954 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */
4948 4955 arc_c_min = MAX(allmem / 32, 64 << 20);
4949 4956 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */
4950 4957 if (allmem >= 1 << 30)
4951 4958 arc_c_max = allmem - (1 << 30);
4952 4959 else
4953 4960 arc_c_max = arc_c_min;
4954 4961 arc_c_max = MAX(allmem * 3 / 4, arc_c_max);
4955 4962
4956 4963 /*
4957 4964 * Allow the tunables to override our calculations if they are
4958 4965 * reasonable (ie. over 64MB)
4959 4966 */
4960 4967 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem)
4961 4968 arc_c_max = zfs_arc_max;
4962 4969 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
4963 4970 arc_c_min = zfs_arc_min;
4964 4971
4965 4972 arc_c = arc_c_max;
4966 4973 arc_p = (arc_c >> 1);
4967 4974
4968 4975 /* limit meta-data to 1/4 of the arc capacity */
4969 4976 arc_meta_limit = arc_c_max / 4;
4970 4977
4971 4978 /* Allow the tunable to override if it is reasonable */
4972 4979 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
4973 4980 arc_meta_limit = zfs_arc_meta_limit;
4974 4981
4975 4982 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
4976 4983 arc_c_min = arc_meta_limit / 2;
4977 4984
4978 4985 if (zfs_arc_meta_min > 0) {
4979 4986 arc_meta_min = zfs_arc_meta_min;
4980 4987 } else {
4981 4988 arc_meta_min = arc_c_min / 2;
4982 4989 }
4983 4990
4984 4991 if (zfs_arc_grow_retry > 0)
4985 4992 arc_grow_retry = zfs_arc_grow_retry;
4986 4993
4987 4994 if (zfs_arc_shrink_shift > 0)
4988 4995 arc_shrink_shift = zfs_arc_shrink_shift;
4989 4996
4990 4997 /*
4991 4998 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
4992 4999 */
4993 5000 if (arc_no_grow_shift >= arc_shrink_shift)
4994 5001 arc_no_grow_shift = arc_shrink_shift - 1;
4995 5002
4996 5003 if (zfs_arc_p_min_shift > 0)
4997 5004 arc_p_min_shift = zfs_arc_p_min_shift;
4998 5005
4999 5006 if (zfs_arc_num_sublists_per_state < 1)
5000 5007 zfs_arc_num_sublists_per_state = MAX(boot_ncpus, 1);
5001 5008
5002 5009 /* if kmem_flags are set, lets try to use less memory */
5003 5010 if (kmem_debugging())
5004 5011 arc_c = arc_c / 2;
5005 5012 if (arc_c < arc_c_min)
5006 5013 arc_c = arc_c_min;
5007 5014
5008 5015 arc_anon = &ARC_anon;
5009 5016 arc_mru = &ARC_mru;
5010 5017 arc_mru_ghost = &ARC_mru_ghost;
5011 5018 arc_mfu = &ARC_mfu;
5012 5019 arc_mfu_ghost = &ARC_mfu_ghost;
5013 5020 arc_l2c_only = &ARC_l2c_only;
5014 5021 arc_size = 0;
5015 5022
5016 5023 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
5017 5024 sizeof (arc_buf_hdr_t),
5018 5025 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5019 5026 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5020 5027 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
5021 5028 sizeof (arc_buf_hdr_t),
5022 5029 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5023 5030 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5024 5031 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
5025 5032 sizeof (arc_buf_hdr_t),
5026 5033 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5027 5034 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5028 5035 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
5029 5036 sizeof (arc_buf_hdr_t),
5030 5037 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5031 5038 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5032 5039 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
5033 5040 sizeof (arc_buf_hdr_t),
5034 5041 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5035 5042 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5036 5043 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
5037 5044 sizeof (arc_buf_hdr_t),
5038 5045 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5039 5046 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5040 5047 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
5041 5048 sizeof (arc_buf_hdr_t),
5042 5049 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5043 5050 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5044 5051 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
5045 5052 sizeof (arc_buf_hdr_t),
5046 5053 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5047 5054 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5048 5055 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
5049 5056 sizeof (arc_buf_hdr_t),
5050 5057 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5051 5058 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5052 5059 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
5053 5060 sizeof (arc_buf_hdr_t),
5054 5061 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5055 5062 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func);
5056 5063
5057 5064 refcount_create(&arc_anon->arcs_size);
5058 5065 refcount_create(&arc_mru->arcs_size);
5059 5066 refcount_create(&arc_mru_ghost->arcs_size);
5060 5067 refcount_create(&arc_mfu->arcs_size);
5061 5068 refcount_create(&arc_mfu_ghost->arcs_size);
5062 5069 refcount_create(&arc_l2c_only->arcs_size);
5063 5070
5064 5071 buf_init();
5065 5072
5066 5073 arc_reclaim_thread_exit = FALSE;
5067 5074 arc_user_evicts_thread_exit = FALSE;
5068 5075 arc_eviction_list = NULL;
5069 5076 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
5070 5077
5071 5078 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
5072 5079 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
5073 5080
5074 5081 if (arc_ksp != NULL) {
5075 5082 arc_ksp->ks_data = &arc_stats;
5076 5083 arc_ksp->ks_update = arc_kstat_update;
5077 5084 kstat_install(arc_ksp);
5078 5085 }
5079 5086
5080 5087 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
5081 5088 TS_RUN, minclsyspri);
5082 5089
5083 5090 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0,
5084 5091 TS_RUN, minclsyspri);
5085 5092
5086 5093 arc_dead = FALSE;
5087 5094 arc_warm = B_FALSE;
5088 5095
5089 5096 /*
5090 5097 * Calculate maximum amount of dirty data per pool.
5091 5098 *
5092 5099 * If it has been set by /etc/system, take that.
5093 5100 * Otherwise, use a percentage of physical memory defined by
5094 5101 * zfs_dirty_data_max_percent (default 10%) with a cap at
5095 5102 * zfs_dirty_data_max_max (default 4GB).
5096 5103 */
5097 5104 if (zfs_dirty_data_max == 0) {
5098 5105 zfs_dirty_data_max = physmem * PAGESIZE *
5099 5106 zfs_dirty_data_max_percent / 100;
5100 5107 zfs_dirty_data_max = MIN(zfs_dirty_data_max,
5101 5108 zfs_dirty_data_max_max);
5102 5109 }
5103 5110 }
5104 5111
5105 5112 void
5106 5113 arc_fini(void)
5107 5114 {
5108 5115 mutex_enter(&arc_reclaim_lock);
5109 5116 arc_reclaim_thread_exit = TRUE;
5110 5117 /*
5111 5118 * The reclaim thread will set arc_reclaim_thread_exit back to
5112 5119 * FALSE when it is finished exiting; we're waiting for that.
5113 5120 */
5114 5121 while (arc_reclaim_thread_exit) {
5115 5122 cv_signal(&arc_reclaim_thread_cv);
5116 5123 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
5117 5124 }
5118 5125 mutex_exit(&arc_reclaim_lock);
5119 5126
5120 5127 mutex_enter(&arc_user_evicts_lock);
5121 5128 arc_user_evicts_thread_exit = TRUE;
5122 5129 /*
5123 5130 * The user evicts thread will set arc_user_evicts_thread_exit
5124 5131 * to FALSE when it is finished exiting; we're waiting for that.
5125 5132 */
5126 5133 while (arc_user_evicts_thread_exit) {
5127 5134 cv_signal(&arc_user_evicts_cv);
5128 5135 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock);
5129 5136 }
5130 5137 mutex_exit(&arc_user_evicts_lock);
5131 5138
5132 5139 /* Use TRUE to ensure *all* buffers are evicted */
5133 5140 arc_flush(NULL, TRUE);
5134 5141
5135 5142 arc_dead = TRUE;
5136 5143
5137 5144 if (arc_ksp != NULL) {
5138 5145 kstat_delete(arc_ksp);
5139 5146 arc_ksp = NULL;
5140 5147 }
5141 5148
5142 5149 mutex_destroy(&arc_reclaim_lock);
5143 5150 cv_destroy(&arc_reclaim_thread_cv);
5144 5151 cv_destroy(&arc_reclaim_waiters_cv);
5145 5152
5146 5153 mutex_destroy(&arc_user_evicts_lock);
5147 5154 cv_destroy(&arc_user_evicts_cv);
5148 5155
5149 5156 refcount_destroy(&arc_anon->arcs_size);
5150 5157 refcount_destroy(&arc_mru->arcs_size);
5151 5158 refcount_destroy(&arc_mru_ghost->arcs_size);
5152 5159 refcount_destroy(&arc_mfu->arcs_size);
5153 5160 refcount_destroy(&arc_mfu_ghost->arcs_size);
5154 5161 refcount_destroy(&arc_l2c_only->arcs_size);
5155 5162
5156 5163 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
5157 5164 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
5158 5165 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]);
5159 5166 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
5160 5167 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]);
5161 5168 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
5162 5169 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]);
5163 5170 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
5164 5171
5165 5172 buf_fini();
5166 5173
5167 5174 ASSERT0(arc_loaned_bytes);
5168 5175 }
5169 5176
5170 5177 /*
5171 5178 * Level 2 ARC
5172 5179 *
5173 5180 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
5174 5181 * It uses dedicated storage devices to hold cached data, which are populated
5175 5182 * using large infrequent writes. The main role of this cache is to boost
5176 5183 * the performance of random read workloads. The intended L2ARC devices
5177 5184 * include short-stroked disks, solid state disks, and other media with
5178 5185 * substantially faster read latency than disk.
5179 5186 *
5180 5187 * +-----------------------+
5181 5188 * | ARC |
5182 5189 * +-----------------------+
5183 5190 * | ^ ^
5184 5191 * | | |
5185 5192 * l2arc_feed_thread() arc_read()
5186 5193 * | | |
5187 5194 * | l2arc read |
5188 5195 * V | |
5189 5196 * +---------------+ |
5190 5197 * | L2ARC | |
5191 5198 * +---------------+ |
5192 5199 * | ^ |
5193 5200 * l2arc_write() | |
5194 5201 * | | |
5195 5202 * V | |
5196 5203 * +-------+ +-------+
5197 5204 * | vdev | | vdev |
5198 5205 * | cache | | cache |
5199 5206 * +-------+ +-------+
5200 5207 * +=========+ .-----.
5201 5208 * : L2ARC : |-_____-|
5202 5209 * : devices : | Disks |
5203 5210 * +=========+ `-_____-'
5204 5211 *
5205 5212 * Read requests are satisfied from the following sources, in order:
5206 5213 *
5207 5214 * 1) ARC
5208 5215 * 2) vdev cache of L2ARC devices
5209 5216 * 3) L2ARC devices
5210 5217 * 4) vdev cache of disks
5211 5218 * 5) disks
5212 5219 *
5213 5220 * Some L2ARC device types exhibit extremely slow write performance.
5214 5221 * To accommodate for this there are some significant differences between
5215 5222 * the L2ARC and traditional cache design:
5216 5223 *
5217 5224 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
5218 5225 * the ARC behave as usual, freeing buffers and placing headers on ghost
5219 5226 * lists. The ARC does not send buffers to the L2ARC during eviction as
5220 5227 * this would add inflated write latencies for all ARC memory pressure.
5221 5228 *
5222 5229 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
5223 5230 * It does this by periodically scanning buffers from the eviction-end of
5224 5231 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
5225 5232 * not already there. It scans until a headroom of buffers is satisfied,
5226 5233 * which itself is a buffer for ARC eviction. If a compressible buffer is
5227 5234 * found during scanning and selected for writing to an L2ARC device, we
5228 5235 * temporarily boost scanning headroom during the next scan cycle to make
5229 5236 * sure we adapt to compression effects (which might significantly reduce
5230 5237 * the data volume we write to L2ARC). The thread that does this is
5231 5238 * l2arc_feed_thread(), illustrated below; example sizes are included to
5232 5239 * provide a better sense of ratio than this diagram:
5233 5240 *
5234 5241 * head --> tail
5235 5242 * +---------------------+----------+
5236 5243 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
5237 5244 * +---------------------+----------+ | o L2ARC eligible
5238 5245 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
5239 5246 * +---------------------+----------+ |
5240 5247 * 15.9 Gbytes ^ 32 Mbytes |
5241 5248 * headroom |
5242 5249 * l2arc_feed_thread()
5243 5250 * |
5244 5251 * l2arc write hand <--[oooo]--'
5245 5252 * | 8 Mbyte
5246 5253 * | write max
5247 5254 * V
5248 5255 * +==============================+
5249 5256 * L2ARC dev |####|#|###|###| |####| ... |
5250 5257 * +==============================+
5251 5258 * 32 Gbytes
5252 5259 *
5253 5260 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
5254 5261 * evicted, then the L2ARC has cached a buffer much sooner than it probably
5255 5262 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
5256 5263 * safe to say that this is an uncommon case, since buffers at the end of
5257 5264 * the ARC lists have moved there due to inactivity.
5258 5265 *
5259 5266 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
5260 5267 * then the L2ARC simply misses copying some buffers. This serves as a
5261 5268 * pressure valve to prevent heavy read workloads from both stalling the ARC
5262 5269 * with waits and clogging the L2ARC with writes. This also helps prevent
5263 5270 * the potential for the L2ARC to churn if it attempts to cache content too
5264 5271 * quickly, such as during backups of the entire pool.
5265 5272 *
5266 5273 * 5. After system boot and before the ARC has filled main memory, there are
5267 5274 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
5268 5275 * lists can remain mostly static. Instead of searching from tail of these
5269 5276 * lists as pictured, the l2arc_feed_thread() will search from the list heads
5270 5277 * for eligible buffers, greatly increasing its chance of finding them.
5271 5278 *
5272 5279 * The L2ARC device write speed is also boosted during this time so that
5273 5280 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
5274 5281 * there are no L2ARC reads, and no fear of degrading read performance
5275 5282 * through increased writes.
5276 5283 *
5277 5284 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
5278 5285 * the vdev queue can aggregate them into larger and fewer writes. Each
5279 5286 * device is written to in a rotor fashion, sweeping writes through
5280 5287 * available space then repeating.
5281 5288 *
5282 5289 * 7. The L2ARC does not store dirty content. It never needs to flush
5283 5290 * write buffers back to disk based storage.
5284 5291 *
5285 5292 * 8. If an ARC buffer is written (and dirtied) which also exists in the
5286 5293 * L2ARC, the now stale L2ARC buffer is immediately dropped.
5287 5294 *
5288 5295 * The performance of the L2ARC can be tweaked by a number of tunables, which
5289 5296 * may be necessary for different workloads:
5290 5297 *
5291 5298 * l2arc_write_max max write bytes per interval
5292 5299 * l2arc_write_boost extra write bytes during device warmup
5293 5300 * l2arc_noprefetch skip caching prefetched buffers
5294 5301 * l2arc_headroom number of max device writes to precache
5295 5302 * l2arc_headroom_boost when we find compressed buffers during ARC
5296 5303 * scanning, we multiply headroom by this
5297 5304 * percentage factor for the next scan cycle,
5298 5305 * since more compressed buffers are likely to
5299 5306 * be present
5300 5307 * l2arc_feed_secs seconds between L2ARC writing
5301 5308 *
5302 5309 * Tunables may be removed or added as future performance improvements are
5303 5310 * integrated, and also may become zpool properties.
5304 5311 *
5305 5312 * There are three key functions that control how the L2ARC warms up:
5306 5313 *
5307 5314 * l2arc_write_eligible() check if a buffer is eligible to cache
5308 5315 * l2arc_write_size() calculate how much to write
5309 5316 * l2arc_write_interval() calculate sleep delay between writes
5310 5317 *
5311 5318 * These three functions determine what to write, how much, and how quickly
5312 5319 * to send writes.
5313 5320 */
5314 5321
5315 5322 static boolean_t
5316 5323 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
5317 5324 {
5318 5325 /*
5319 5326 * A buffer is *not* eligible for the L2ARC if it:
5320 5327 * 1. belongs to a different spa.
5321 5328 * 2. is already cached on the L2ARC.
5322 5329 * 3. has an I/O in progress (it may be an incomplete read).
5323 5330 * 4. is flagged not eligible (zfs property).
5324 5331 */
5325 5332 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
5326 5333 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
5327 5334 return (B_FALSE);
5328 5335
5329 5336 return (B_TRUE);
5330 5337 }
5331 5338
5332 5339 static uint64_t
5333 5340 l2arc_write_size(void)
5334 5341 {
5335 5342 uint64_t size;
5336 5343
5337 5344 /*
5338 5345 * Make sure our globals have meaningful values in case the user
5339 5346 * altered them.
5340 5347 */
5341 5348 size = l2arc_write_max;
5342 5349 if (size == 0) {
5343 5350 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must "
5344 5351 "be greater than zero, resetting it to the default (%d)",
5345 5352 L2ARC_WRITE_SIZE);
5346 5353 size = l2arc_write_max = L2ARC_WRITE_SIZE;
5347 5354 }
5348 5355
5349 5356 if (arc_warm == B_FALSE)
5350 5357 size += l2arc_write_boost;
5351 5358
5352 5359 return (size);
5353 5360
5354 5361 }
5355 5362
5356 5363 static clock_t
5357 5364 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
5358 5365 {
5359 5366 clock_t interval, next, now;
5360 5367
5361 5368 /*
5362 5369 * If the ARC lists are busy, increase our write rate; if the
5363 5370 * lists are stale, idle back. This is achieved by checking
5364 5371 * how much we previously wrote - if it was more than half of
5365 5372 * what we wanted, schedule the next write much sooner.
5366 5373 */
5367 5374 if (l2arc_feed_again && wrote > (wanted / 2))
5368 5375 interval = (hz * l2arc_feed_min_ms) / 1000;
5369 5376 else
5370 5377 interval = hz * l2arc_feed_secs;
5371 5378
5372 5379 now = ddi_get_lbolt();
5373 5380 next = MAX(now, MIN(now + interval, began + interval));
5374 5381
5375 5382 return (next);
5376 5383 }
5377 5384
5378 5385 /*
5379 5386 * Cycle through L2ARC devices. This is how L2ARC load balances.
5380 5387 * If a device is returned, this also returns holding the spa config lock.
5381 5388 */
5382 5389 static l2arc_dev_t *
5383 5390 l2arc_dev_get_next(void)
5384 5391 {
5385 5392 l2arc_dev_t *first, *next = NULL;
5386 5393
5387 5394 /*
5388 5395 * Lock out the removal of spas (spa_namespace_lock), then removal
5389 5396 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
5390 5397 * both locks will be dropped and a spa config lock held instead.
5391 5398 */
5392 5399 mutex_enter(&spa_namespace_lock);
5393 5400 mutex_enter(&l2arc_dev_mtx);
5394 5401
5395 5402 /* if there are no vdevs, there is nothing to do */
5396 5403 if (l2arc_ndev == 0)
5397 5404 goto out;
5398 5405
5399 5406 first = NULL;
5400 5407 next = l2arc_dev_last;
5401 5408 do {
5402 5409 /* loop around the list looking for a non-faulted vdev */
5403 5410 if (next == NULL) {
5404 5411 next = list_head(l2arc_dev_list);
5405 5412 } else {
5406 5413 next = list_next(l2arc_dev_list, next);
5407 5414 if (next == NULL)
5408 5415 next = list_head(l2arc_dev_list);
5409 5416 }
5410 5417
5411 5418 /* if we have come back to the start, bail out */
5412 5419 if (first == NULL)
5413 5420 first = next;
5414 5421 else if (next == first)
5415 5422 break;
5416 5423
5417 5424 } while (vdev_is_dead(next->l2ad_vdev));
5418 5425
5419 5426 /* if we were unable to find any usable vdevs, return NULL */
5420 5427 if (vdev_is_dead(next->l2ad_vdev))
5421 5428 next = NULL;
5422 5429
5423 5430 l2arc_dev_last = next;
5424 5431
5425 5432 out:
5426 5433 mutex_exit(&l2arc_dev_mtx);
5427 5434
5428 5435 /*
5429 5436 * Grab the config lock to prevent the 'next' device from being
5430 5437 * removed while we are writing to it.
5431 5438 */
5432 5439 if (next != NULL)
5433 5440 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
5434 5441 mutex_exit(&spa_namespace_lock);
5435 5442
5436 5443 return (next);
5437 5444 }
5438 5445
5439 5446 /*
5440 5447 * Free buffers that were tagged for destruction.
5441 5448 */
5442 5449 static void
5443 5450 l2arc_do_free_on_write()
5444 5451 {
5445 5452 list_t *buflist;
5446 5453 l2arc_data_free_t *df, *df_prev;
5447 5454
5448 5455 mutex_enter(&l2arc_free_on_write_mtx);
5449 5456 buflist = l2arc_free_on_write;
5450 5457
5451 5458 for (df = list_tail(buflist); df; df = df_prev) {
5452 5459 df_prev = list_prev(buflist, df);
5453 5460 ASSERT(df->l2df_data != NULL);
5454 5461 ASSERT(df->l2df_func != NULL);
5455 5462 df->l2df_func(df->l2df_data, df->l2df_size);
5456 5463 list_remove(buflist, df);
5457 5464 kmem_free(df, sizeof (l2arc_data_free_t));
5458 5465 }
5459 5466
5460 5467 mutex_exit(&l2arc_free_on_write_mtx);
5461 5468 }
5462 5469
5463 5470 /*
5464 5471 * A write to a cache device has completed. Update all headers to allow
5465 5472 * reads from these buffers to begin.
5466 5473 */
5467 5474 static void
5468 5475 l2arc_write_done(zio_t *zio)
5469 5476 {
5470 5477 l2arc_write_callback_t *cb;
5471 5478 l2arc_dev_t *dev;
5472 5479 list_t *buflist;
5473 5480 arc_buf_hdr_t *head, *hdr, *hdr_prev;
5474 5481 kmutex_t *hash_lock;
5475 5482 int64_t bytes_dropped = 0;
5476 5483
5477 5484 cb = zio->io_private;
5478 5485 ASSERT(cb != NULL);
5479 5486 dev = cb->l2wcb_dev;
5480 5487 ASSERT(dev != NULL);
5481 5488 head = cb->l2wcb_head;
5482 5489 ASSERT(head != NULL);
5483 5490 buflist = &dev->l2ad_buflist;
5484 5491 ASSERT(buflist != NULL);
5485 5492 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
5486 5493 l2arc_write_callback_t *, cb);
5487 5494
5488 5495 if (zio->io_error != 0)
5489 5496 ARCSTAT_BUMP(arcstat_l2_writes_error);
5490 5497
5491 5498 /*
5492 5499 * All writes completed, or an error was hit.
5493 5500 */
5494 5501 top:
5495 5502 mutex_enter(&dev->l2ad_mtx);
5496 5503 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) {
5497 5504 hdr_prev = list_prev(buflist, hdr);
5498 5505
5499 5506 hash_lock = HDR_LOCK(hdr);
5500 5507
5501 5508 /*
5502 5509 * We cannot use mutex_enter or else we can deadlock
5503 5510 * with l2arc_write_buffers (due to swapping the order
5504 5511 * the hash lock and l2ad_mtx are taken).
5505 5512 */
5506 5513 if (!mutex_tryenter(hash_lock)) {
5507 5514 /*
5508 5515 * Missed the hash lock. We must retry so we
5509 5516 * don't leave the ARC_FLAG_L2_WRITING bit set.
5510 5517 */
5511 5518 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry);
5512 5519
5513 5520 /*
5514 5521 * We don't want to rescan the headers we've
5515 5522 * already marked as having been written out, so
5516 5523 * we reinsert the head node so we can pick up
5517 5524 * where we left off.
5518 5525 */
5519 5526 list_remove(buflist, head);
5520 5527 list_insert_after(buflist, hdr, head);
5521 5528
5522 5529 mutex_exit(&dev->l2ad_mtx);
5523 5530
5524 5531 /*
5525 5532 * We wait for the hash lock to become available
5526 5533 * to try and prevent busy waiting, and increase
5527 5534 * the chance we'll be able to acquire the lock
5528 5535 * the next time around.
5529 5536 */
5530 5537 mutex_enter(hash_lock);
5531 5538 mutex_exit(hash_lock);
5532 5539 goto top;
5533 5540 }
5534 5541
5535 5542 /*
5536 5543 * We could not have been moved into the arc_l2c_only
5537 5544 * state while in-flight due to our ARC_FLAG_L2_WRITING
5538 5545 * bit being set. Let's just ensure that's being enforced.
5539 5546 */
5540 5547 ASSERT(HDR_HAS_L1HDR(hdr));
5541 5548
5542 5549 /*
5543 5550 * We may have allocated a buffer for L2ARC compression,
5544 5551 * we must release it to avoid leaking this data.
5545 5552 */
5546 5553 l2arc_release_cdata_buf(hdr);
5547 5554
5548 5555 if (zio->io_error != 0) {
5549 5556 /*
5550 5557 * Error - drop L2ARC entry.
5551 5558 */
5552 5559 list_remove(buflist, hdr);
5553 5560 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR;
5554 5561
5555 5562 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize);
5556 5563 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
5557 5564
5558 5565 bytes_dropped += hdr->b_l2hdr.b_asize;
5559 5566 (void) refcount_remove_many(&dev->l2ad_alloc,
5560 5567 hdr->b_l2hdr.b_asize, hdr);
5561 5568 }
5562 5569
5563 5570 /*
5564 5571 * Allow ARC to begin reads and ghost list evictions to
5565 5572 * this L2ARC entry.
5566 5573 */
5567 5574 hdr->b_flags &= ~ARC_FLAG_L2_WRITING;
5568 5575
5569 5576 mutex_exit(hash_lock);
5570 5577 }
5571 5578
5572 5579 atomic_inc_64(&l2arc_writes_done);
5573 5580 list_remove(buflist, head);
5574 5581 ASSERT(!HDR_HAS_L1HDR(head));
5575 5582 kmem_cache_free(hdr_l2only_cache, head);
5576 5583 mutex_exit(&dev->l2ad_mtx);
5577 5584
5578 5585 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
5579 5586
5580 5587 l2arc_do_free_on_write();
5581 5588
5582 5589 kmem_free(cb, sizeof (l2arc_write_callback_t));
5583 5590 }
5584 5591
5585 5592 /*
5586 5593 * A read to a cache device completed. Validate buffer contents before
5587 5594 * handing over to the regular ARC routines.
5588 5595 */
5589 5596 static void
5590 5597 l2arc_read_done(zio_t *zio)
5591 5598 {
5592 5599 l2arc_read_callback_t *cb;
5593 5600 arc_buf_hdr_t *hdr;
5594 5601 arc_buf_t *buf;
5595 5602 kmutex_t *hash_lock;
5596 5603 int equal;
5597 5604
5598 5605 ASSERT(zio->io_vd != NULL);
5599 5606 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
5600 5607
5601 5608 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
5602 5609
5603 5610 cb = zio->io_private;
5604 5611 ASSERT(cb != NULL);
5605 5612 buf = cb->l2rcb_buf;
5606 5613 ASSERT(buf != NULL);
5607 5614
5608 5615 hash_lock = HDR_LOCK(buf->b_hdr);
5609 5616 mutex_enter(hash_lock);
5610 5617 hdr = buf->b_hdr;
5611 5618 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr));
5612 5619
5613 5620 /*
5614 5621 * If the buffer was compressed, decompress it first.
5615 5622 */
5616 5623 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF)
5617 5624 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress);
5618 5625 ASSERT(zio->io_data != NULL);
5619 5626 ASSERT3U(zio->io_size, ==, hdr->b_size);
5620 5627 ASSERT3U(BP_GET_LSIZE(&cb->l2rcb_bp), ==, hdr->b_size);
5621 5628
5622 5629 /*
5623 5630 * Check this survived the L2ARC journey.
5624 5631 */
5625 5632 equal = arc_cksum_equal(buf);
5626 5633 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
5627 5634 mutex_exit(hash_lock);
5628 5635 zio->io_private = buf;
5629 5636 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
5630 5637 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
5631 5638 arc_read_done(zio);
5632 5639 } else {
5633 5640 mutex_exit(hash_lock);
5634 5641 /*
5635 5642 * Buffer didn't survive caching. Increment stats and
5636 5643 * reissue to the original storage device.
5637 5644 */
5638 5645 if (zio->io_error != 0) {
5639 5646 ARCSTAT_BUMP(arcstat_l2_io_error);
5640 5647 } else {
5641 5648 zio->io_error = SET_ERROR(EIO);
5642 5649 }
5643 5650 if (!equal)
5644 5651 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
5645 5652
5646 5653 /*
5647 5654 * If there's no waiter, issue an async i/o to the primary
5648 5655 * storage now. If there *is* a waiter, the caller must
5649 5656 * issue the i/o in a context where it's OK to block.
5650 5657 */
5651 5658 if (zio->io_waiter == NULL) {
5652 5659 zio_t *pio = zio_unique_parent(zio);
5653 5660
5654 5661 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
5655 5662
5656 5663 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp,
5657 5664 buf->b_data, hdr->b_size, arc_read_done, buf,
5658 5665 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
5659 5666 }
5660 5667 }
5661 5668
5662 5669 kmem_free(cb, sizeof (l2arc_read_callback_t));
5663 5670 }
5664 5671
5665 5672 /*
5666 5673 * This is the list priority from which the L2ARC will search for pages to
5667 5674 * cache. This is used within loops (0..3) to cycle through lists in the
5668 5675 * desired order. This order can have a significant effect on cache
5669 5676 * performance.
5670 5677 *
5671 5678 * Currently the metadata lists are hit first, MFU then MRU, followed by
5672 5679 * the data lists. This function returns a locked list, and also returns
5673 5680 * the lock pointer.
5674 5681 */
5675 5682 static multilist_sublist_t *
5676 5683 l2arc_sublist_lock(int list_num)
5677 5684 {
5678 5685 multilist_t *ml = NULL;
5679 5686 unsigned int idx;
5680 5687
5681 5688 ASSERT(list_num >= 0 && list_num <= 3);
5682 5689
5683 5690 switch (list_num) {
5684 5691 case 0:
5685 5692 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA];
5686 5693 break;
5687 5694 case 1:
5688 5695 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA];
5689 5696 break;
5690 5697 case 2:
5691 5698 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA];
5692 5699 break;
5693 5700 case 3:
5694 5701 ml = &arc_mru->arcs_list[ARC_BUFC_DATA];
5695 5702 break;
5696 5703 }
5697 5704
5698 5705 /*
5699 5706 * Return a randomly-selected sublist. This is acceptable
5700 5707 * because the caller feeds only a little bit of data for each
5701 5708 * call (8MB). Subsequent calls will result in different
5702 5709 * sublists being selected.
5703 5710 */
5704 5711 idx = multilist_get_random_index(ml);
5705 5712 return (multilist_sublist_lock(ml, idx));
5706 5713 }
5707 5714
5708 5715 /*
5709 5716 * Evict buffers from the device write hand to the distance specified in
5710 5717 * bytes. This distance may span populated buffers, it may span nothing.
5711 5718 * This is clearing a region on the L2ARC device ready for writing.
5712 5719 * If the 'all' boolean is set, every buffer is evicted.
5713 5720 */
5714 5721 static void
5715 5722 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
5716 5723 {
5717 5724 list_t *buflist;
5718 5725 arc_buf_hdr_t *hdr, *hdr_prev;
5719 5726 kmutex_t *hash_lock;
5720 5727 uint64_t taddr;
5721 5728
5722 5729 buflist = &dev->l2ad_buflist;
5723 5730
5724 5731 if (!all && dev->l2ad_first) {
5725 5732 /*
5726 5733 * This is the first sweep through the device. There is
5727 5734 * nothing to evict.
5728 5735 */
5729 5736 return;
5730 5737 }
5731 5738
5732 5739 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
5733 5740 /*
5734 5741 * When nearing the end of the device, evict to the end
5735 5742 * before the device write hand jumps to the start.
5736 5743 */
5737 5744 taddr = dev->l2ad_end;
5738 5745 } else {
5739 5746 taddr = dev->l2ad_hand + distance;
5740 5747 }
5741 5748 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
5742 5749 uint64_t, taddr, boolean_t, all);
5743 5750
5744 5751 top:
5745 5752 mutex_enter(&dev->l2ad_mtx);
5746 5753 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
5747 5754 hdr_prev = list_prev(buflist, hdr);
5748 5755
5749 5756 hash_lock = HDR_LOCK(hdr);
5750 5757
5751 5758 /*
5752 5759 * We cannot use mutex_enter or else we can deadlock
5753 5760 * with l2arc_write_buffers (due to swapping the order
5754 5761 * the hash lock and l2ad_mtx are taken).
5755 5762 */
5756 5763 if (!mutex_tryenter(hash_lock)) {
5757 5764 /*
5758 5765 * Missed the hash lock. Retry.
5759 5766 */
5760 5767 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
5761 5768 mutex_exit(&dev->l2ad_mtx);
5762 5769 mutex_enter(hash_lock);
5763 5770 mutex_exit(hash_lock);
5764 5771 goto top;
5765 5772 }
5766 5773
5767 5774 if (HDR_L2_WRITE_HEAD(hdr)) {
5768 5775 /*
5769 5776 * We hit a write head node. Leave it for
5770 5777 * l2arc_write_done().
5771 5778 */
5772 5779 list_remove(buflist, hdr);
5773 5780 mutex_exit(hash_lock);
5774 5781 continue;
5775 5782 }
5776 5783
5777 5784 if (!all && HDR_HAS_L2HDR(hdr) &&
5778 5785 (hdr->b_l2hdr.b_daddr > taddr ||
5779 5786 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) {
5780 5787 /*
5781 5788 * We've evicted to the target address,
5782 5789 * or the end of the device.
5783 5790 */
5784 5791 mutex_exit(hash_lock);
5785 5792 break;
5786 5793 }
5787 5794
5788 5795 ASSERT(HDR_HAS_L2HDR(hdr));
5789 5796 if (!HDR_HAS_L1HDR(hdr)) {
5790 5797 ASSERT(!HDR_L2_READING(hdr));
5791 5798 /*
5792 5799 * This doesn't exist in the ARC. Destroy.
5793 5800 * arc_hdr_destroy() will call list_remove()
5794 5801 * and decrement arcstat_l2_size.
5795 5802 */
5796 5803 arc_change_state(arc_anon, hdr, hash_lock);
5797 5804 arc_hdr_destroy(hdr);
5798 5805 } else {
5799 5806 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
5800 5807 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
5801 5808 /*
5802 5809 * Invalidate issued or about to be issued
5803 5810 * reads, since we may be about to write
5804 5811 * over this location.
5805 5812 */
5806 5813 if (HDR_L2_READING(hdr)) {
5807 5814 ARCSTAT_BUMP(arcstat_l2_evict_reading);
5808 5815 hdr->b_flags |= ARC_FLAG_L2_EVICTED;
5809 5816 }
5810 5817
5811 5818 /* Ensure this header has finished being written */
5812 5819 ASSERT(!HDR_L2_WRITING(hdr));
5813 5820 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
5814 5821
5815 5822 arc_hdr_l2hdr_destroy(hdr);
5816 5823 }
5817 5824 mutex_exit(hash_lock);
5818 5825 }
5819 5826 mutex_exit(&dev->l2ad_mtx);
5820 5827 }
5821 5828
5822 5829 /*
5823 5830 * Find and write ARC buffers to the L2ARC device.
5824 5831 *
5825 5832 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
5826 5833 * for reading until they have completed writing.
5827 5834 * The headroom_boost is an in-out parameter used to maintain headroom boost
5828 5835 * state between calls to this function.
5829 5836 *
5830 5837 * Returns the number of bytes actually written (which may be smaller than
5831 5838 * the delta by which the device hand has changed due to alignment).
5832 5839 */
5833 5840 static uint64_t
5834 5841 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
5835 5842 boolean_t *headroom_boost)
5836 5843 {
5837 5844 arc_buf_hdr_t *hdr, *hdr_prev, *head;
5838 5845 uint64_t write_asize, write_psize, write_sz, headroom,
5839 5846 buf_compress_minsz;
5840 5847 void *buf_data;
5841 5848 boolean_t full;
5842 5849 l2arc_write_callback_t *cb;
5843 5850 zio_t *pio, *wzio;
5844 5851 uint64_t guid = spa_load_guid(spa);
5845 5852 const boolean_t do_headroom_boost = *headroom_boost;
5846 5853
5847 5854 ASSERT(dev->l2ad_vdev != NULL);
5848 5855
5849 5856 /* Lower the flag now, we might want to raise it again later. */
5850 5857 *headroom_boost = B_FALSE;
5851 5858
5852 5859 pio = NULL;
5853 5860 write_sz = write_asize = write_psize = 0;
5854 5861 full = B_FALSE;
5855 5862 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
5856 5863 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD;
5857 5864 head->b_flags |= ARC_FLAG_HAS_L2HDR;
5858 5865
5859 5866 /*
5860 5867 * We will want to try to compress buffers that are at least 2x the
5861 5868 * device sector size.
5862 5869 */
5863 5870 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift;
5864 5871
5865 5872 /*
5866 5873 * Copy buffers for L2ARC writing.
5867 5874 */
5868 5875 for (int try = 0; try <= 3; try++) {
5869 5876 multilist_sublist_t *mls = l2arc_sublist_lock(try);
5870 5877 uint64_t passed_sz = 0;
5871 5878
5872 5879 /*
5873 5880 * L2ARC fast warmup.
5874 5881 *
5875 5882 * Until the ARC is warm and starts to evict, read from the
5876 5883 * head of the ARC lists rather than the tail.
5877 5884 */
5878 5885 if (arc_warm == B_FALSE)
5879 5886 hdr = multilist_sublist_head(mls);
5880 5887 else
5881 5888 hdr = multilist_sublist_tail(mls);
5882 5889
5883 5890 headroom = target_sz * l2arc_headroom;
5884 5891 if (do_headroom_boost)
5885 5892 headroom = (headroom * l2arc_headroom_boost) / 100;
5886 5893
5887 5894 for (; hdr; hdr = hdr_prev) {
5888 5895 kmutex_t *hash_lock;
5889 5896 uint64_t buf_sz;
5890 5897
5891 5898 if (arc_warm == B_FALSE)
5892 5899 hdr_prev = multilist_sublist_next(mls, hdr);
5893 5900 else
5894 5901 hdr_prev = multilist_sublist_prev(mls, hdr);
5895 5902
5896 5903 hash_lock = HDR_LOCK(hdr);
5897 5904 if (!mutex_tryenter(hash_lock)) {
5898 5905 /*
5899 5906 * Skip this buffer rather than waiting.
5900 5907 */
5901 5908 continue;
5902 5909 }
5903 5910
5904 5911 passed_sz += hdr->b_size;
5905 5912 if (passed_sz > headroom) {
5906 5913 /*
5907 5914 * Searched too far.
5908 5915 */
5909 5916 mutex_exit(hash_lock);
5910 5917 break;
5911 5918 }
5912 5919
5913 5920 if (!l2arc_write_eligible(guid, hdr)) {
5914 5921 mutex_exit(hash_lock);
5915 5922 continue;
5916 5923 }
5917 5924
5918 5925 if ((write_sz + hdr->b_size) > target_sz) {
5919 5926 full = B_TRUE;
5920 5927 mutex_exit(hash_lock);
5921 5928 break;
5922 5929 }
5923 5930
5924 5931 if (pio == NULL) {
5925 5932 /*
5926 5933 * Insert a dummy header on the buflist so
5927 5934 * l2arc_write_done() can find where the
5928 5935 * write buffers begin without searching.
5929 5936 */
5930 5937 mutex_enter(&dev->l2ad_mtx);
5931 5938 list_insert_head(&dev->l2ad_buflist, head);
5932 5939 mutex_exit(&dev->l2ad_mtx);
5933 5940
5934 5941 cb = kmem_alloc(
5935 5942 sizeof (l2arc_write_callback_t), KM_SLEEP);
5936 5943 cb->l2wcb_dev = dev;
5937 5944 cb->l2wcb_head = head;
5938 5945 pio = zio_root(spa, l2arc_write_done, cb,
5939 5946 ZIO_FLAG_CANFAIL);
5940 5947 }
5941 5948
5942 5949 /*
5943 5950 * Create and add a new L2ARC header.
5944 5951 */
5945 5952 hdr->b_l2hdr.b_dev = dev;
5946 5953 hdr->b_flags |= ARC_FLAG_L2_WRITING;
5947 5954 /*
5948 5955 * Temporarily stash the data buffer in b_tmp_cdata.
5949 5956 * The subsequent write step will pick it up from
5950 5957 * there. This is because can't access b_l1hdr.b_buf
5951 5958 * without holding the hash_lock, which we in turn
5952 5959 * can't access without holding the ARC list locks
5953 5960 * (which we want to avoid during compression/writing).
5954 5961 */
5955 5962 hdr->b_l2hdr.b_compress = ZIO_COMPRESS_OFF;
5956 5963 hdr->b_l2hdr.b_asize = hdr->b_size;
5957 5964 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data;
5958 5965
5959 5966 /*
5960 5967 * Explicitly set the b_daddr field to a known
5961 5968 * value which means "invalid address". This
5962 5969 * enables us to differentiate which stage of
5963 5970 * l2arc_write_buffers() the particular header
5964 5971 * is in (e.g. this loop, or the one below).
5965 5972 * ARC_FLAG_L2_WRITING is not enough to make
5966 5973 * this distinction, and we need to know in
5967 5974 * order to do proper l2arc vdev accounting in
5968 5975 * arc_release() and arc_hdr_destroy().
5969 5976 *
5970 5977 * Note, we can't use a new flag to distinguish
5971 5978 * the two stages because we don't hold the
5972 5979 * header's hash_lock below, in the second stage
5973 5980 * of this function. Thus, we can't simply
5974 5981 * change the b_flags field to denote that the
5975 5982 * IO has been sent. We can change the b_daddr
5976 5983 * field of the L2 portion, though, since we'll
5977 5984 * be holding the l2ad_mtx; which is why we're
5978 5985 * using it to denote the header's state change.
5979 5986 */
5980 5987 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET;
5981 5988
5982 5989 buf_sz = hdr->b_size;
5983 5990 hdr->b_flags |= ARC_FLAG_HAS_L2HDR;
5984 5991
5985 5992 mutex_enter(&dev->l2ad_mtx);
5986 5993 list_insert_head(&dev->l2ad_buflist, hdr);
5987 5994 mutex_exit(&dev->l2ad_mtx);
5988 5995
5989 5996 /*
5990 5997 * Compute and store the buffer cksum before
5991 5998 * writing. On debug the cksum is verified first.
5992 5999 */
5993 6000 arc_cksum_verify(hdr->b_l1hdr.b_buf);
5994 6001 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE);
5995 6002
5996 6003 mutex_exit(hash_lock);
5997 6004
5998 6005 write_sz += buf_sz;
5999 6006 }
6000 6007
6001 6008 multilist_sublist_unlock(mls);
6002 6009
6003 6010 if (full == B_TRUE)
6004 6011 break;
6005 6012 }
6006 6013
6007 6014 /* No buffers selected for writing? */
6008 6015 if (pio == NULL) {
6009 6016 ASSERT0(write_sz);
6010 6017 ASSERT(!HDR_HAS_L1HDR(head));
6011 6018 kmem_cache_free(hdr_l2only_cache, head);
6012 6019 return (0);
6013 6020 }
6014 6021
6015 6022 mutex_enter(&dev->l2ad_mtx);
6016 6023
6017 6024 /*
6018 6025 * Now start writing the buffers. We're starting at the write head
6019 6026 * and work backwards, retracing the course of the buffer selector
6020 6027 * loop above.
6021 6028 */
6022 6029 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr;
6023 6030 hdr = list_prev(&dev->l2ad_buflist, hdr)) {
6024 6031 uint64_t buf_sz;
6025 6032
6026 6033 /*
6027 6034 * We rely on the L1 portion of the header below, so
6028 6035 * it's invalid for this header to have been evicted out
6029 6036 * of the ghost cache, prior to being written out. The
6030 6037 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6031 6038 */
6032 6039 ASSERT(HDR_HAS_L1HDR(hdr));
6033 6040
6034 6041 /*
6035 6042 * We shouldn't need to lock the buffer here, since we flagged
6036 6043 * it as ARC_FLAG_L2_WRITING in the previous step, but we must
6037 6044 * take care to only access its L2 cache parameters. In
6038 6045 * particular, hdr->l1hdr.b_buf may be invalid by now due to
6039 6046 * ARC eviction.
6040 6047 */
6041 6048 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
6042 6049
6043 6050 if ((HDR_L2COMPRESS(hdr)) &&
6044 6051 hdr->b_l2hdr.b_asize >= buf_compress_minsz) {
6045 6052 if (l2arc_compress_buf(hdr)) {
6046 6053 /*
6047 6054 * If compression succeeded, enable headroom
6048 6055 * boost on the next scan cycle.
6049 6056 */
6050 6057 *headroom_boost = B_TRUE;
6051 6058 }
6052 6059 }
6053 6060
6054 6061 /*
6055 6062 * Pick up the buffer data we had previously stashed away
6056 6063 * (and now potentially also compressed).
6057 6064 */
6058 6065 buf_data = hdr->b_l1hdr.b_tmp_cdata;
6059 6066 buf_sz = hdr->b_l2hdr.b_asize;
6060 6067
6061 6068 /*
6062 6069 * We need to do this regardless if buf_sz is zero or
6063 6070 * not, otherwise, when this l2hdr is evicted we'll
6064 6071 * remove a reference that was never added.
6065 6072 */
6066 6073 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr);
6067 6074
6068 6075 /* Compression may have squashed the buffer to zero length. */
6069 6076 if (buf_sz != 0) {
6070 6077 uint64_t buf_p_sz;
6071 6078
6072 6079 wzio = zio_write_phys(pio, dev->l2ad_vdev,
6073 6080 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
6074 6081 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
6075 6082 ZIO_FLAG_CANFAIL, B_FALSE);
6076 6083
6077 6084 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
6078 6085 zio_t *, wzio);
6079 6086 (void) zio_nowait(wzio);
6080 6087
6081 6088 write_asize += buf_sz;
6082 6089
6083 6090 /*
6084 6091 * Keep the clock hand suitably device-aligned.
6085 6092 */
6086 6093 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
6087 6094 write_psize += buf_p_sz;
6088 6095 dev->l2ad_hand += buf_p_sz;
6089 6096 }
6090 6097 }
6091 6098
6092 6099 mutex_exit(&dev->l2ad_mtx);
6093 6100
6094 6101 ASSERT3U(write_asize, <=, target_sz);
6095 6102 ARCSTAT_BUMP(arcstat_l2_writes_sent);
6096 6103 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize);
6097 6104 ARCSTAT_INCR(arcstat_l2_size, write_sz);
6098 6105 ARCSTAT_INCR(arcstat_l2_asize, write_asize);
6099 6106 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0);
6100 6107
6101 6108 /*
6102 6109 * Bump device hand to the device start if it is approaching the end.
6103 6110 * l2arc_evict() will already have evicted ahead for this case.
6104 6111 */
6105 6112 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
6106 6113 dev->l2ad_hand = dev->l2ad_start;
6107 6114 dev->l2ad_first = B_FALSE;
6108 6115 }
6109 6116
6110 6117 dev->l2ad_writing = B_TRUE;
6111 6118 (void) zio_wait(pio);
6112 6119 dev->l2ad_writing = B_FALSE;
6113 6120
6114 6121 return (write_asize);
6115 6122 }
6116 6123
6117 6124 /*
6118 6125 * Compresses an L2ARC buffer.
6119 6126 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its
6120 6127 * size in l2hdr->b_asize. This routine tries to compress the data and
6121 6128 * depending on the compression result there are three possible outcomes:
6122 6129 * *) The buffer was incompressible. The original l2hdr contents were left
6123 6130 * untouched and are ready for writing to an L2 device.
6124 6131 * *) The buffer was all-zeros, so there is no need to write it to an L2
6125 6132 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is
6126 6133 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY.
6127 6134 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary
6128 6135 * data buffer which holds the compressed data to be written, and b_asize
6129 6136 * tells us how much data there is. b_compress is set to the appropriate
6130 6137 * compression algorithm. Once writing is done, invoke
6131 6138 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer.
6132 6139 *
6133 6140 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the
6134 6141 * buffer was incompressible).
6135 6142 */
6136 6143 static boolean_t
6137 6144 l2arc_compress_buf(arc_buf_hdr_t *hdr)
6138 6145 {
6139 6146 void *cdata;
6140 6147 size_t csize, len, rounded;
6141 6148 ASSERT(HDR_HAS_L2HDR(hdr));
6142 6149 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
6143 6150
6144 6151 ASSERT(HDR_HAS_L1HDR(hdr));
6145 6152 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF);
6146 6153 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6147 6154
6148 6155 len = l2hdr->b_asize;
6149 6156 cdata = zio_data_buf_alloc(len);
6150 6157 ASSERT3P(cdata, !=, NULL);
6151 6158 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata,
6152 6159 cdata, l2hdr->b_asize);
6153 6160
6154 6161 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE);
6155 6162 if (rounded > csize) {
6156 6163 bzero((char *)cdata + csize, rounded - csize);
6157 6164 csize = rounded;
6158 6165 }
6159 6166
6160 6167 if (csize == 0) {
6161 6168 /* zero block, indicate that there's nothing to write */
6162 6169 zio_data_buf_free(cdata, len);
6163 6170 l2hdr->b_compress = ZIO_COMPRESS_EMPTY;
6164 6171 l2hdr->b_asize = 0;
6165 6172 hdr->b_l1hdr.b_tmp_cdata = NULL;
6166 6173 ARCSTAT_BUMP(arcstat_l2_compress_zeros);
6167 6174 return (B_TRUE);
6168 6175 } else if (csize > 0 && csize < len) {
6169 6176 /*
6170 6177 * Compression succeeded, we'll keep the cdata around for
6171 6178 * writing and release it afterwards.
6172 6179 */
6173 6180 l2hdr->b_compress = ZIO_COMPRESS_LZ4;
6174 6181 l2hdr->b_asize = csize;
6175 6182 hdr->b_l1hdr.b_tmp_cdata = cdata;
6176 6183 ARCSTAT_BUMP(arcstat_l2_compress_successes);
6177 6184 return (B_TRUE);
6178 6185 } else {
6179 6186 /*
6180 6187 * Compression failed, release the compressed buffer.
6181 6188 * l2hdr will be left unmodified.
6182 6189 */
6183 6190 zio_data_buf_free(cdata, len);
6184 6191 ARCSTAT_BUMP(arcstat_l2_compress_failures);
6185 6192 return (B_FALSE);
6186 6193 }
6187 6194 }
6188 6195
6189 6196 /*
6190 6197 * Decompresses a zio read back from an l2arc device. On success, the
6191 6198 * underlying zio's io_data buffer is overwritten by the uncompressed
6192 6199 * version. On decompression error (corrupt compressed stream), the
6193 6200 * zio->io_error value is set to signal an I/O error.
6194 6201 *
6195 6202 * Please note that the compressed data stream is not checksummed, so
6196 6203 * if the underlying device is experiencing data corruption, we may feed
6197 6204 * corrupt data to the decompressor, so the decompressor needs to be
6198 6205 * able to handle this situation (LZ4 does).
6199 6206 */
6200 6207 static void
6201 6208 l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c)
6202 6209 {
6203 6210 ASSERT(L2ARC_IS_VALID_COMPRESS(c));
6204 6211
6205 6212 if (zio->io_error != 0) {
6206 6213 /*
6207 6214 * An io error has occured, just restore the original io
6208 6215 * size in preparation for a main pool read.
6209 6216 */
6210 6217 zio->io_orig_size = zio->io_size = hdr->b_size;
6211 6218 return;
6212 6219 }
6213 6220
6214 6221 if (c == ZIO_COMPRESS_EMPTY) {
6215 6222 /*
6216 6223 * An empty buffer results in a null zio, which means we
6217 6224 * need to fill its io_data after we're done restoring the
6218 6225 * buffer's contents.
6219 6226 */
6220 6227 ASSERT(hdr->b_l1hdr.b_buf != NULL);
6221 6228 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size);
6222 6229 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data;
6223 6230 } else {
6224 6231 ASSERT(zio->io_data != NULL);
6225 6232 /*
6226 6233 * We copy the compressed data from the start of the arc buffer
6227 6234 * (the zio_read will have pulled in only what we need, the
6228 6235 * rest is garbage which we will overwrite at decompression)
6229 6236 * and then decompress back to the ARC data buffer. This way we
6230 6237 * can minimize copying by simply decompressing back over the
6231 6238 * original compressed data (rather than decompressing to an
6232 6239 * aux buffer and then copying back the uncompressed buffer,
6233 6240 * which is likely to be much larger).
6234 6241 */
6235 6242 uint64_t csize;
6236 6243 void *cdata;
6237 6244
6238 6245 csize = zio->io_size;
6239 6246 cdata = zio_data_buf_alloc(csize);
6240 6247 bcopy(zio->io_data, cdata, csize);
6241 6248 if (zio_decompress_data(c, cdata, zio->io_data, csize,
6242 6249 hdr->b_size) != 0)
6243 6250 zio->io_error = EIO;
6244 6251 zio_data_buf_free(cdata, csize);
6245 6252 }
6246 6253
6247 6254 /* Restore the expected uncompressed IO size. */
6248 6255 zio->io_orig_size = zio->io_size = hdr->b_size;
6249 6256 }
6250 6257
6251 6258 /*
6252 6259 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure.
6253 6260 * This buffer serves as a temporary holder of compressed data while
6254 6261 * the buffer entry is being written to an l2arc device. Once that is
6255 6262 * done, we can dispose of it.
6256 6263 */
6257 6264 static void
6258 6265 l2arc_release_cdata_buf(arc_buf_hdr_t *hdr)
6259 6266 {
6260 6267 ASSERT(HDR_HAS_L2HDR(hdr));
6261 6268 enum zio_compress comp = hdr->b_l2hdr.b_compress;
6262 6269
6263 6270 ASSERT(HDR_HAS_L1HDR(hdr));
6264 6271 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp));
6265 6272
6266 6273 if (comp == ZIO_COMPRESS_OFF) {
6267 6274 /*
6268 6275 * In this case, b_tmp_cdata points to the same buffer
6269 6276 * as the arc_buf_t's b_data field. We don't want to
6270 6277 * free it, since the arc_buf_t will handle that.
6271 6278 */
6272 6279 hdr->b_l1hdr.b_tmp_cdata = NULL;
6273 6280 } else if (comp == ZIO_COMPRESS_EMPTY) {
6274 6281 /*
6275 6282 * In this case, b_tmp_cdata was compressed to an empty
6276 6283 * buffer, thus there's nothing to free and b_tmp_cdata
6277 6284 * should have been set to NULL in l2arc_write_buffers().
6278 6285 */
6279 6286 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL);
6280 6287 } else {
6281 6288 /*
6282 6289 * If the data was compressed, then we've allocated a
6283 6290 * temporary buffer for it, so now we need to release it.
6284 6291 */
6285 6292 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL);
6286 6293 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata,
6287 6294 hdr->b_size);
6288 6295 hdr->b_l1hdr.b_tmp_cdata = NULL;
6289 6296 }
6290 6297
6291 6298 }
6292 6299
6293 6300 /*
6294 6301 * This thread feeds the L2ARC at regular intervals. This is the beating
6295 6302 * heart of the L2ARC.
6296 6303 */
6297 6304 static void
6298 6305 l2arc_feed_thread(void)
6299 6306 {
6300 6307 callb_cpr_t cpr;
6301 6308 l2arc_dev_t *dev;
6302 6309 spa_t *spa;
6303 6310 uint64_t size, wrote;
6304 6311 clock_t begin, next = ddi_get_lbolt();
6305 6312 boolean_t headroom_boost = B_FALSE;
6306 6313
6307 6314 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
6308 6315
6309 6316 mutex_enter(&l2arc_feed_thr_lock);
6310 6317
6311 6318 while (l2arc_thread_exit == 0) {
6312 6319 CALLB_CPR_SAFE_BEGIN(&cpr);
6313 6320 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
6314 6321 next);
6315 6322 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
6316 6323 next = ddi_get_lbolt() + hz;
6317 6324
6318 6325 /*
6319 6326 * Quick check for L2ARC devices.
6320 6327 */
6321 6328 mutex_enter(&l2arc_dev_mtx);
6322 6329 if (l2arc_ndev == 0) {
6323 6330 mutex_exit(&l2arc_dev_mtx);
6324 6331 continue;
6325 6332 }
6326 6333 mutex_exit(&l2arc_dev_mtx);
6327 6334 begin = ddi_get_lbolt();
6328 6335
6329 6336 /*
6330 6337 * This selects the next l2arc device to write to, and in
6331 6338 * doing so the next spa to feed from: dev->l2ad_spa. This
6332 6339 * will return NULL if there are now no l2arc devices or if
6333 6340 * they are all faulted.
6334 6341 *
6335 6342 * If a device is returned, its spa's config lock is also
6336 6343 * held to prevent device removal. l2arc_dev_get_next()
6337 6344 * will grab and release l2arc_dev_mtx.
6338 6345 */
6339 6346 if ((dev = l2arc_dev_get_next()) == NULL)
6340 6347 continue;
6341 6348
6342 6349 spa = dev->l2ad_spa;
6343 6350 ASSERT(spa != NULL);
6344 6351
6345 6352 /*
6346 6353 * If the pool is read-only then force the feed thread to
6347 6354 * sleep a little longer.
6348 6355 */
6349 6356 if (!spa_writeable(spa)) {
6350 6357 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
6351 6358 spa_config_exit(spa, SCL_L2ARC, dev);
6352 6359 continue;
6353 6360 }
6354 6361
6355 6362 /*
6356 6363 * Avoid contributing to memory pressure.
6357 6364 */
6358 6365 if (arc_reclaim_needed()) {
6359 6366 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
6360 6367 spa_config_exit(spa, SCL_L2ARC, dev);
6361 6368 continue;
6362 6369 }
6363 6370
6364 6371 ARCSTAT_BUMP(arcstat_l2_feeds);
6365 6372
6366 6373 size = l2arc_write_size();
6367 6374
6368 6375 /*
6369 6376 * Evict L2ARC buffers that will be overwritten.
6370 6377 */
6371 6378 l2arc_evict(dev, size, B_FALSE);
6372 6379
6373 6380 /*
6374 6381 * Write ARC buffers.
6375 6382 */
6376 6383 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost);
6377 6384
6378 6385 /*
6379 6386 * Calculate interval between writes.
6380 6387 */
6381 6388 next = l2arc_write_interval(begin, size, wrote);
6382 6389 spa_config_exit(spa, SCL_L2ARC, dev);
6383 6390 }
6384 6391
6385 6392 l2arc_thread_exit = 0;
6386 6393 cv_broadcast(&l2arc_feed_thr_cv);
6387 6394 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
6388 6395 thread_exit();
6389 6396 }
6390 6397
6391 6398 boolean_t
6392 6399 l2arc_vdev_present(vdev_t *vd)
6393 6400 {
6394 6401 l2arc_dev_t *dev;
6395 6402
6396 6403 mutex_enter(&l2arc_dev_mtx);
6397 6404 for (dev = list_head(l2arc_dev_list); dev != NULL;
6398 6405 dev = list_next(l2arc_dev_list, dev)) {
6399 6406 if (dev->l2ad_vdev == vd)
6400 6407 break;
6401 6408 }
6402 6409 mutex_exit(&l2arc_dev_mtx);
6403 6410
6404 6411 return (dev != NULL);
6405 6412 }
6406 6413
6407 6414 /*
6408 6415 * Add a vdev for use by the L2ARC. By this point the spa has already
6409 6416 * validated the vdev and opened it.
6410 6417 */
6411 6418 void
6412 6419 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
6413 6420 {
6414 6421 l2arc_dev_t *adddev;
6415 6422
6416 6423 ASSERT(!l2arc_vdev_present(vd));
6417 6424
6418 6425 /*
6419 6426 * Create a new l2arc device entry.
6420 6427 */
6421 6428 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
6422 6429 adddev->l2ad_spa = spa;
6423 6430 adddev->l2ad_vdev = vd;
6424 6431 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
6425 6432 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
6426 6433 adddev->l2ad_hand = adddev->l2ad_start;
6427 6434 adddev->l2ad_first = B_TRUE;
6428 6435 adddev->l2ad_writing = B_FALSE;
6429 6436
6430 6437 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
6431 6438 /*
6432 6439 * This is a list of all ARC buffers that are still valid on the
6433 6440 * device.
6434 6441 */
6435 6442 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
6436 6443 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
6437 6444
6438 6445 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
6439 6446 refcount_create(&adddev->l2ad_alloc);
6440 6447
6441 6448 /*
6442 6449 * Add device to global list
6443 6450 */
6444 6451 mutex_enter(&l2arc_dev_mtx);
6445 6452 list_insert_head(l2arc_dev_list, adddev);
6446 6453 atomic_inc_64(&l2arc_ndev);
6447 6454 mutex_exit(&l2arc_dev_mtx);
6448 6455 }
6449 6456
6450 6457 /*
6451 6458 * Remove a vdev from the L2ARC.
6452 6459 */
6453 6460 void
6454 6461 l2arc_remove_vdev(vdev_t *vd)
6455 6462 {
6456 6463 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
6457 6464
6458 6465 /*
6459 6466 * Find the device by vdev
6460 6467 */
6461 6468 mutex_enter(&l2arc_dev_mtx);
6462 6469 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
6463 6470 nextdev = list_next(l2arc_dev_list, dev);
6464 6471 if (vd == dev->l2ad_vdev) {
6465 6472 remdev = dev;
6466 6473 break;
6467 6474 }
6468 6475 }
6469 6476 ASSERT(remdev != NULL);
6470 6477
6471 6478 /*
6472 6479 * Remove device from global list
6473 6480 */
6474 6481 list_remove(l2arc_dev_list, remdev);
6475 6482 l2arc_dev_last = NULL; /* may have been invalidated */
6476 6483 atomic_dec_64(&l2arc_ndev);
6477 6484 mutex_exit(&l2arc_dev_mtx);
6478 6485
6479 6486 /*
6480 6487 * Clear all buflists and ARC references. L2ARC device flush.
6481 6488 */
6482 6489 l2arc_evict(remdev, 0, B_TRUE);
6483 6490 list_destroy(&remdev->l2ad_buflist);
6484 6491 mutex_destroy(&remdev->l2ad_mtx);
6485 6492 refcount_destroy(&remdev->l2ad_alloc);
6486 6493 kmem_free(remdev, sizeof (l2arc_dev_t));
6487 6494 }
6488 6495
6489 6496 void
6490 6497 l2arc_init(void)
6491 6498 {
6492 6499 l2arc_thread_exit = 0;
6493 6500 l2arc_ndev = 0;
6494 6501 l2arc_writes_sent = 0;
6495 6502 l2arc_writes_done = 0;
6496 6503
6497 6504 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
6498 6505 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
6499 6506 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
6500 6507 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
6501 6508
6502 6509 l2arc_dev_list = &L2ARC_dev_list;
6503 6510 l2arc_free_on_write = &L2ARC_free_on_write;
6504 6511 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
6505 6512 offsetof(l2arc_dev_t, l2ad_node));
6506 6513 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
6507 6514 offsetof(l2arc_data_free_t, l2df_list_node));
6508 6515 }
6509 6516
6510 6517 void
6511 6518 l2arc_fini(void)
6512 6519 {
6513 6520 /*
6514 6521 * This is called from dmu_fini(), which is called from spa_fini();
6515 6522 * Because of this, we can assume that all l2arc devices have
6516 6523 * already been removed when the pools themselves were removed.
6517 6524 */
6518 6525
6519 6526 l2arc_do_free_on_write();
6520 6527
6521 6528 mutex_destroy(&l2arc_feed_thr_lock);
6522 6529 cv_destroy(&l2arc_feed_thr_cv);
6523 6530 mutex_destroy(&l2arc_dev_mtx);
6524 6531 mutex_destroy(&l2arc_free_on_write_mtx);
6525 6532
6526 6533 list_destroy(l2arc_dev_list);
6527 6534 list_destroy(l2arc_free_on_write);
6528 6535 }
6529 6536
6530 6537 void
6531 6538 l2arc_start(void)
6532 6539 {
6533 6540 if (!(spa_mode_global & FWRITE))
6534 6541 return;
6535 6542
6536 6543 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
6537 6544 TS_RUN, minclsyspri);
6538 6545 }
6539 6546
6540 6547 void
6541 6548 l2arc_stop(void)
6542 6549 {
6543 6550 if (!(spa_mode_global & FWRITE))
6544 6551 return;
6545 6552
6546 6553 mutex_enter(&l2arc_feed_thr_lock);
6547 6554 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
6548 6555 l2arc_thread_exit = 1;
6549 6556 while (l2arc_thread_exit != 0)
6550 6557 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
6551 6558 mutex_exit(&l2arc_feed_thr_lock);
6552 6559 }
|
↓ open down ↓ |
5239 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX