4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
27 * DVA-based Adjustable Replacement Cache
29 * While much of the theory of operation used here is
30 * based on the self-tuning, low overhead replacement cache
31 * presented by Megiddo and Modha at FAST 2003, there are some
32 * significant differences:
34 * 1. The Megiddo and Modha model assumes any page is evictable.
35 * Pages in its cache cannot be "locked" into memory. This makes
36 * the eviction algorithm simple: evict the last page in the list.
37 * This also make the performance characteristics easy to reason
38 * about. Our cache is not so simple. At any given moment, some
39 * subset of the blocks in the cache are un-evictable because we
40 * have handed out a reference to them. Blocks are only evictable
41 * when there are no external references active. This makes
42 * eviction far more problematic: we choose to evict the evictable
43 * blocks that are the "lowest" in the list.
45 * There are times when it is not possible to evict the requested
46 * space. In these circumstances we are unable to adjust the cache
47 * size. To prevent the cache growing unbounded at these times we
48 * implement a "cache throttle" that slows the flow of new data
49 * into the cache until we can make space available.
51 * 2. The Megiddo and Modha model assumes a fixed cache size.
52 * Pages are evicted when the cache is full and there is a cache
53 * miss. Our model has a variable sized cache. It grows with
54 * high use, but also tries to react to memory pressure from the
55 * operating system: decreasing its size when system memory is
58 * 3. The Megiddo and Modha model assumes a fixed page size. All
59 * elements of the cache are therefor exactly the same size. So
60 * when adjusting the cache size following a cache miss, its simply
61 * a matter of choosing a single page to evict. In our model, we
62 * have variable sized cache blocks (rangeing from 512 bytes to
63 * 128K bytes). We therefor choose a set of blocks to evict to make
64 * space for a cache miss that approximates as closely as possible
65 * the space used by the new block.
67 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache"
68 * by N. Megiddo & D. Modha, FAST 2003
74 * A new reference to a cache buffer can be obtained in two
75 * ways: 1) via a hash table lookup using the DVA as a key,
76 * or 2) via one of the ARC lists. The arc_read() interface
77 * uses method 1, while the internal arc algorithms for
78 * adjusting the cache use method 2. We therefor provide two
79 * types of locks: 1) the hash table lock array, and 2) the
82 * Buffers do not have their own mutexs, rather they rely on the
83 * hash table mutexs for the bulk of their protection (i.e. most
84 * fields in the arc_buf_hdr_t are protected by these mutexs).
86 * buf_hash_find() returns the appropriate mutex (held) when it
87 * locates the requested buffer in the hash table. It returns
88 * NULL for the mutex if the buffer was not in the table.
90 * buf_hash_remove() expects the appropriate hash mutex to be
91 * already held before it is invoked.
93 * Each arc state also has a mutex which is used to protect the
94 * buffer list associated with the state. When attempting to
95 * obtain a hash table lock while holding an arc list lock you
96 * must use: mutex_tryenter() to avoid deadlock. Also note that
97 * the active state mutex must be held before the ghost state mutex.
99 * Arc buffers may have an associated eviction callback function.
100 * This function will be invoked prior to removing the buffer (e.g.
101 * in arc_do_user_evicts()). Note however that the data associated
102 * with the buffer may be evicted prior to the callback. The callback
103 * must be made with *no locks held* (to prevent deadlock). Additionally,
104 * the users of callbacks must ensure that their private data is
105 * protected from simultaneous callbacks from arc_buf_evict()
106 * and arc_do_user_evicts().
108 * Note that the majority of the performance stats are manipulated
109 * with atomic operations.
111 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following:
113 * - L2ARC buflist creation
114 * - L2ARC buflist eviction
115 * - L2ARC write completion, which walks L2ARC buflists
116 * - ARC header destruction, as it removes from L2ARC buflists
117 * - ARC header release, as it removes from L2ARC buflists
122 #include <sys/zio_checksum.h>
123 #include <sys/zfs_context.h>
125 #include <sys/refcount.h>
126 #include <sys/vdev.h>
128 #include <sys/dnlc.h>
130 #include <sys/callb.h>
131 #include <sys/kstat.h>
134 #include <vm/vm_pageout.h>
136 static kmutex_t arc_reclaim_thr_lock;
137 static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */
138 static uint8_t arc_thread_exit;
140 extern int zfs_write_limit_shift;
141 extern uint64_t zfs_write_limit_max;
142 extern kmutex_t zfs_write_limit_lock;
144 #define ARC_REDUCE_DNLC_PERCENT 3
145 uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT;
147 typedef enum arc_reclaim_strategy {
148 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */
149 ARC_RECLAIM_CONS /* Conservative reclaim strategy */
150 } arc_reclaim_strategy_t;
152 /* number of seconds before growing cache again */
153 static int arc_grow_retry = 60;
155 /* shift of arc_c for calculating both min and max arc_p */
156 static int arc_p_min_shift = 4;
158 /* log2(fraction of arc to reclaim) */
159 static int arc_shrink_shift = 5;
162 * minimum lifespan of a prefetch block in clock ticks
163 * (initialized in arc_init())
165 static int arc_min_prefetch_lifespan;
168 extern int zfs_prefetch_disable;
171 * The arc has filled available memory and has now warmed up.
173 static boolean_t arc_warm;
176 * These tunables are for performance analysis.
178 uint64_t zfs_arc_max;
179 uint64_t zfs_arc_min;
180 uint64_t zfs_arc_meta_limit = 0;
181 int zfs_mdcomp_disable = 0;
182 int zfs_arc_grow_retry = 0;
183 int zfs_arc_shrink_shift = 0;
184 int zfs_arc_p_min_shift = 0;
186 TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max);
187 TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min);
188 TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit);
189 TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable);
190 SYSCTL_DECL(_vfs_zfs);
191 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0,
193 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0,
195 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RDTUN,
196 &zfs_mdcomp_disable, 0, "Disable metadata compression");
199 * Note that buffers can be in one of 6 states:
200 * ARC_anon - anonymous (discussed below)
201 * ARC_mru - recently used, currently cached
202 * ARC_mru_ghost - recentely used, no longer in cache
203 * ARC_mfu - frequently used, currently cached
204 * ARC_mfu_ghost - frequently used, no longer in cache
205 * ARC_l2c_only - exists in L2ARC but not other states
206 * When there are no active references to the buffer, they are
207 * are linked onto a list in one of these arc states. These are
208 * the only buffers that can be evicted or deleted. Within each
209 * state there are multiple lists, one for meta-data and one for
210 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
211 * etc.) is tracked separately so that it can be managed more
212 * explicitly: favored over data, limited explicitly.
214 * Anonymous buffers are buffers that are not associated with
215 * a DVA. These are buffers that hold dirty block copies
216 * before they are written to stable storage. By definition,
217 * they are "ref'd" and are considered part of arc_mru
218 * that cannot be freed. Generally, they will aquire a DVA
219 * as they are written and migrate onto the arc_mru list.
221 * The ARC_l2c_only state is for buffers that are in the second
222 * level ARC but no longer in any of the ARC_m* lists. The second
223 * level ARC itself may also contain buffers that are in any of
224 * the ARC_m* states - meaning that a buffer can exist in two
225 * places. The reason for the ARC_l2c_only state is to keep the
226 * buffer header in the hash table, so that reads that hit the
227 * second level ARC benefit from these fast lookups.
230 #define ARCS_LOCK_PAD CACHE_LINE_SIZE
234 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))];
239 * must be power of two for mask use to work
242 #define ARC_BUFC_NUMDATALISTS 16
243 #define ARC_BUFC_NUMMETADATALISTS 16
244 #define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS)
246 typedef struct arc_state {
247 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */
248 uint64_t arcs_size; /* total amount of data in this state */
249 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */
250 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE);
253 #define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock))
256 static arc_state_t ARC_anon;
257 static arc_state_t ARC_mru;
258 static arc_state_t ARC_mru_ghost;
259 static arc_state_t ARC_mfu;
260 static arc_state_t ARC_mfu_ghost;
261 static arc_state_t ARC_l2c_only;
263 typedef struct arc_stats {
264 kstat_named_t arcstat_hits;
265 kstat_named_t arcstat_misses;
266 kstat_named_t arcstat_demand_data_hits;
267 kstat_named_t arcstat_demand_data_misses;
268 kstat_named_t arcstat_demand_metadata_hits;
269 kstat_named_t arcstat_demand_metadata_misses;
270 kstat_named_t arcstat_prefetch_data_hits;
271 kstat_named_t arcstat_prefetch_data_misses;
272 kstat_named_t arcstat_prefetch_metadata_hits;
273 kstat_named_t arcstat_prefetch_metadata_misses;
274 kstat_named_t arcstat_mru_hits;
275 kstat_named_t arcstat_mru_ghost_hits;
276 kstat_named_t arcstat_mfu_hits;
277 kstat_named_t arcstat_mfu_ghost_hits;
278 kstat_named_t arcstat_allocated;
279 kstat_named_t arcstat_deleted;
280 kstat_named_t arcstat_stolen;
281 kstat_named_t arcstat_recycle_miss;
282 kstat_named_t arcstat_mutex_miss;
283 kstat_named_t arcstat_evict_skip;
284 kstat_named_t arcstat_evict_l2_cached;
285 kstat_named_t arcstat_evict_l2_eligible;
286 kstat_named_t arcstat_evict_l2_ineligible;
287 kstat_named_t arcstat_hash_elements;
288 kstat_named_t arcstat_hash_elements_max;
289 kstat_named_t arcstat_hash_collisions;
290 kstat_named_t arcstat_hash_chains;
291 kstat_named_t arcstat_hash_chain_max;
292 kstat_named_t arcstat_p;
293 kstat_named_t arcstat_c;
294 kstat_named_t arcstat_c_min;
295 kstat_named_t arcstat_c_max;
296 kstat_named_t arcstat_size;
297 kstat_named_t arcstat_hdr_size;
298 kstat_named_t arcstat_data_size;
299 kstat_named_t arcstat_other_size;
300 kstat_named_t arcstat_l2_hits;
301 kstat_named_t arcstat_l2_misses;
302 kstat_named_t arcstat_l2_feeds;
303 kstat_named_t arcstat_l2_rw_clash;
304 kstat_named_t arcstat_l2_read_bytes;
305 kstat_named_t arcstat_l2_write_bytes;
306 kstat_named_t arcstat_l2_writes_sent;
307 kstat_named_t arcstat_l2_writes_done;
308 kstat_named_t arcstat_l2_writes_error;
309 kstat_named_t arcstat_l2_writes_hdr_miss;
310 kstat_named_t arcstat_l2_evict_lock_retry;
311 kstat_named_t arcstat_l2_evict_reading;
312 kstat_named_t arcstat_l2_free_on_write;
313 kstat_named_t arcstat_l2_abort_lowmem;
314 kstat_named_t arcstat_l2_cksum_bad;
315 kstat_named_t arcstat_l2_io_error;
316 kstat_named_t arcstat_l2_size;
317 kstat_named_t arcstat_l2_hdr_size;
318 kstat_named_t arcstat_memory_throttle_count;
319 kstat_named_t arcstat_l2_write_trylock_fail;
320 kstat_named_t arcstat_l2_write_passed_headroom;
321 kstat_named_t arcstat_l2_write_spa_mismatch;
322 kstat_named_t arcstat_l2_write_in_l2;
323 kstat_named_t arcstat_l2_write_hdr_io_in_progress;
324 kstat_named_t arcstat_l2_write_not_cacheable;
325 kstat_named_t arcstat_l2_write_full;
326 kstat_named_t arcstat_l2_write_buffer_iter;
327 kstat_named_t arcstat_l2_write_pios;
328 kstat_named_t arcstat_l2_write_bytes_written;
329 kstat_named_t arcstat_l2_write_buffer_bytes_scanned;
330 kstat_named_t arcstat_l2_write_buffer_list_iter;
331 kstat_named_t arcstat_l2_write_buffer_list_null_iter;
334 static arc_stats_t arc_stats = {
335 { "hits", KSTAT_DATA_UINT64 },
336 { "misses", KSTAT_DATA_UINT64 },
337 { "demand_data_hits", KSTAT_DATA_UINT64 },
338 { "demand_data_misses", KSTAT_DATA_UINT64 },
339 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
340 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
341 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
342 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
343 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
344 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
345 { "mru_hits", KSTAT_DATA_UINT64 },
346 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
347 { "mfu_hits", KSTAT_DATA_UINT64 },
348 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
349 { "allocated", KSTAT_DATA_UINT64 },
350 { "deleted", KSTAT_DATA_UINT64 },
351 { "stolen", KSTAT_DATA_UINT64 },
352 { "recycle_miss", KSTAT_DATA_UINT64 },
353 { "mutex_miss", KSTAT_DATA_UINT64 },
354 { "evict_skip", KSTAT_DATA_UINT64 },
355 { "evict_l2_cached", KSTAT_DATA_UINT64 },
356 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
357 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
358 { "hash_elements", KSTAT_DATA_UINT64 },
359 { "hash_elements_max", KSTAT_DATA_UINT64 },
360 { "hash_collisions", KSTAT_DATA_UINT64 },
361 { "hash_chains", KSTAT_DATA_UINT64 },
362 { "hash_chain_max", KSTAT_DATA_UINT64 },
363 { "p", KSTAT_DATA_UINT64 },
364 { "c", KSTAT_DATA_UINT64 },
365 { "c_min", KSTAT_DATA_UINT64 },
366 { "c_max", KSTAT_DATA_UINT64 },
367 { "size", KSTAT_DATA_UINT64 },
368 { "hdr_size", KSTAT_DATA_UINT64 },
369 { "data_size", KSTAT_DATA_UINT64 },
370 { "other_size", KSTAT_DATA_UINT64 },
371 { "l2_hits", KSTAT_DATA_UINT64 },
372 { "l2_misses", KSTAT_DATA_UINT64 },
373 { "l2_feeds", KSTAT_DATA_UINT64 },
374 { "l2_rw_clash", KSTAT_DATA_UINT64 },
375 { "l2_read_bytes", KSTAT_DATA_UINT64 },
376 { "l2_write_bytes", KSTAT_DATA_UINT64 },
377 { "l2_writes_sent", KSTAT_DATA_UINT64 },
378 { "l2_writes_done", KSTAT_DATA_UINT64 },
379 { "l2_writes_error", KSTAT_DATA_UINT64 },
380 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 },
381 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
382 { "l2_evict_reading", KSTAT_DATA_UINT64 },
383 { "l2_free_on_write", KSTAT_DATA_UINT64 },
384 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
385 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
386 { "l2_io_error", KSTAT_DATA_UINT64 },
387 { "l2_size", KSTAT_DATA_UINT64 },
388 { "l2_hdr_size", KSTAT_DATA_UINT64 },
389 { "memory_throttle_count", KSTAT_DATA_UINT64 },
390 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 },
391 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 },
392 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 },
393 { "l2_write_in_l2", KSTAT_DATA_UINT64 },
394 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 },
395 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 },
396 { "l2_write_full", KSTAT_DATA_UINT64 },
397 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 },
398 { "l2_write_pios", KSTAT_DATA_UINT64 },
399 { "l2_write_bytes_written", KSTAT_DATA_UINT64 },
400 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 },
401 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 },
402 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }
405 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
407 #define ARCSTAT_INCR(stat, val) \
408 atomic_add_64(&arc_stats.stat.value.ui64, (val));
410 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
411 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
413 #define ARCSTAT_MAX(stat, val) { \
415 while ((val) > (m = arc_stats.stat.value.ui64) && \
416 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
420 #define ARCSTAT_MAXSTAT(stat) \
421 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
424 * We define a macro to allow ARC hits/misses to be easily broken down by
425 * two separate conditions, giving a total of four different subtypes for
426 * each of hits and misses (so eight statistics total).
428 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
431 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
433 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
437 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
439 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
444 static arc_state_t *arc_anon;
445 static arc_state_t *arc_mru;
446 static arc_state_t *arc_mru_ghost;
447 static arc_state_t *arc_mfu;
448 static arc_state_t *arc_mfu_ghost;
449 static arc_state_t *arc_l2c_only;
452 * There are several ARC variables that are critical to export as kstats --
453 * but we don't want to have to grovel around in the kstat whenever we wish to
454 * manipulate them. For these variables, we therefore define them to be in
455 * terms of the statistic variable. This assures that we are not introducing
456 * the possibility of inconsistency by having shadow copies of the variables,
457 * while still allowing the code to be readable.
459 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
460 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
461 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
462 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
463 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
465 static int arc_no_grow; /* Don't try to grow cache size */
466 static uint64_t arc_tempreserve;
467 static uint64_t arc_meta_used;
468 static uint64_t arc_meta_limit;
469 static uint64_t arc_meta_max = 0;
470 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RDTUN,
471 &arc_meta_used, 0, "ARC metadata used");
472 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RDTUN,
473 &arc_meta_limit, 0, "ARC metadata limit");
475 typedef struct l2arc_buf_hdr l2arc_buf_hdr_t;
477 typedef struct arc_callback arc_callback_t;
479 struct arc_callback {
481 arc_done_func_t *acb_done;
483 zio_t *acb_zio_dummy;
484 arc_callback_t *acb_next;
487 typedef struct arc_write_callback arc_write_callback_t;
489 struct arc_write_callback {
491 arc_done_func_t *awcb_ready;
492 arc_done_func_t *awcb_done;
497 /* protected by hash lock */
502 kmutex_t b_freeze_lock;
503 zio_cksum_t *b_freeze_cksum;
505 arc_buf_hdr_t *b_hash_next;
510 arc_callback_t *b_acb;
514 arc_buf_contents_t b_type;
518 /* protected by arc state mutex */
519 arc_state_t *b_state;
520 list_node_t b_arc_node;
522 /* updated atomically */
523 clock_t b_arc_access;
525 /* self protecting */
528 l2arc_buf_hdr_t *b_l2hdr;
529 list_node_t b_l2node;
532 static arc_buf_t *arc_eviction_list;
533 static kmutex_t arc_eviction_mtx;
534 static arc_buf_hdr_t arc_eviction_hdr;
535 static void arc_get_data_buf(arc_buf_t *buf);
536 static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock);
537 static int arc_evict_needed(arc_buf_contents_t type);
538 static void arc_evict_ghost(arc_state_t *state, spa_t *spa, int64_t bytes);
540 static boolean_t l2arc_write_eligible(spa_t *spa, arc_buf_hdr_t *ab);
542 #define GHOST_STATE(state) \
543 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
544 (state) == arc_l2c_only)
547 * Private ARC flags. These flags are private ARC only flags that will show up
548 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can
549 * be passed in as arc_flags in things like arc_read. However, these flags
550 * should never be passed and should only be set by ARC code. When adding new
551 * public flags, make sure not to smash the private ones.
554 #define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */
555 #define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */
556 #define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */
557 #define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */
558 #define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */
559 #define ARC_INDIRECT (1 << 14) /* this is an indirect block */
560 #define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */
561 #define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */
562 #define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */
563 #define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */
564 #define ARC_STORED (1 << 19) /* has been store()d to */
566 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE)
567 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS)
568 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR)
569 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH)
570 #define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ)
571 #define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE)
572 #define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS)
573 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE)
574 #define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \
575 (hdr)->b_l2hdr != NULL)
576 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING)
577 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED)
578 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD)
584 #define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
585 #define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t))
588 * Hash table routines
591 #define HT_LOCK_PAD CACHE_LINE_SIZE
596 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
600 #define BUF_LOCKS 256
601 typedef struct buf_hash_table {
603 arc_buf_hdr_t **ht_table;
604 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE);
607 static buf_hash_table_t buf_hash_table;
609 #define BUF_HASH_INDEX(spa, dva, birth) \
610 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
611 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
612 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
613 #define HDR_LOCK(buf) \
614 (BUF_HASH_LOCK(BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth)))
616 uint64_t zfs_crc64_table[256];
622 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
623 #define L2ARC_HEADROOM 2 /* num of writes */
624 #define L2ARC_FEED_SECS 1 /* caching interval secs */
625 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
627 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
628 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
631 * L2ARC Performance Tunables
633 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
634 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
635 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
636 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
637 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
638 boolean_t l2arc_noprefetch = B_FALSE; /* don't cache prefetch bufs */
639 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
640 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
642 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW,
643 &l2arc_write_max, 0, "max write size");
644 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW,
645 &l2arc_write_boost, 0, "extra write during warmup");
646 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW,
647 &l2arc_headroom, 0, "number of dev writes");
648 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW,
649 &l2arc_feed_secs, 0, "interval seconds");
650 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW,
651 &l2arc_feed_min_ms, 0, "min interval milliseconds");
653 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW,
654 &l2arc_noprefetch, 0, "don't cache prefetch bufs");
655 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW,
656 &l2arc_feed_again, 0, "turbo warmup");
657 SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW,
658 &l2arc_norw, 0, "no reads during writes");
660 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD,
661 &ARC_anon.arcs_size, 0, "size of anonymous state");
662 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD,
663 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state");
664 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD,
665 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state");
667 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD,
668 &ARC_mru.arcs_size, 0, "size of mru state");
669 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD,
670 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state");
671 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD,
672 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state");
674 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD,
675 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state");
676 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD,
677 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
678 "size of metadata in mru ghost state");
679 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD,
680 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
681 "size of data in mru ghost state");
683 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD,
684 &ARC_mfu.arcs_size, 0, "size of mfu state");
685 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD,
686 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state");
687 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD,
688 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state");
690 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD,
691 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state");
692 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD,
693 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0,
694 "size of metadata in mfu ghost state");
695 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD,
696 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0,
697 "size of data in mfu ghost state");
699 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD,
700 &ARC_l2c_only.arcs_size, 0, "size of mru state");
705 typedef struct l2arc_dev {
706 vdev_t *l2ad_vdev; /* vdev */
707 spa_t *l2ad_spa; /* spa */
708 uint64_t l2ad_hand; /* next write location */
709 uint64_t l2ad_write; /* desired write size, bytes */
710 uint64_t l2ad_boost; /* warmup write boost, bytes */
711 uint64_t l2ad_start; /* first addr on device */
712 uint64_t l2ad_end; /* last addr on device */
713 uint64_t l2ad_evict; /* last addr eviction reached */
714 boolean_t l2ad_first; /* first sweep through */
715 boolean_t l2ad_writing; /* currently writing */
716 list_t *l2ad_buflist; /* buffer list */
717 list_node_t l2ad_node; /* device list node */
720 static list_t L2ARC_dev_list; /* device list */
721 static list_t *l2arc_dev_list; /* device list pointer */
722 static kmutex_t l2arc_dev_mtx; /* device list mutex */
723 static l2arc_dev_t *l2arc_dev_last; /* last device used */
724 static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */
725 static list_t L2ARC_free_on_write; /* free after write buf list */
726 static list_t *l2arc_free_on_write; /* free after write list ptr */
727 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
728 static uint64_t l2arc_ndev; /* number of devices */
730 typedef struct l2arc_read_callback {
731 arc_buf_t *l2rcb_buf; /* read buffer */
732 spa_t *l2rcb_spa; /* spa */
733 blkptr_t l2rcb_bp; /* original blkptr */
734 zbookmark_t l2rcb_zb; /* original bookmark */
735 int l2rcb_flags; /* original flags */
736 } l2arc_read_callback_t;
738 typedef struct l2arc_write_callback {
739 l2arc_dev_t *l2wcb_dev; /* device info */
740 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
741 } l2arc_write_callback_t;
743 struct l2arc_buf_hdr {
744 /* protected by arc_buf_hdr mutex */
745 l2arc_dev_t *b_dev; /* L2ARC device */
746 uint64_t b_daddr; /* disk address, offset byte */
749 typedef struct l2arc_data_free {
750 /* protected by l2arc_free_on_write_mtx */
753 void (*l2df_func)(void *, size_t);
754 list_node_t l2df_list_node;
757 static kmutex_t l2arc_feed_thr_lock;
758 static kcondvar_t l2arc_feed_thr_cv;
759 static uint8_t l2arc_thread_exit;
761 static void l2arc_read_done(zio_t *zio);
762 static void l2arc_hdr_stat_add(void);
763 static void l2arc_hdr_stat_remove(void);
766 buf_hash(spa_t *spa, const dva_t *dva, uint64_t birth)
768 uintptr_t spav = (uintptr_t)spa;
769 uint8_t *vdva = (uint8_t *)dva;
770 uint64_t crc = -1ULL;
773 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
775 for (i = 0; i < sizeof (dva_t); i++)
776 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
778 crc ^= (spav>>8) ^ birth;
783 #define BUF_EMPTY(buf) \
784 ((buf)->b_dva.dva_word[0] == 0 && \
785 (buf)->b_dva.dva_word[1] == 0 && \
788 #define BUF_EQUAL(spa, dva, birth, buf) \
789 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
790 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
791 ((buf)->b_birth == birth) && ((buf)->b_spa == spa)
793 static arc_buf_hdr_t *
794 buf_hash_find(spa_t *spa, const dva_t *dva, uint64_t birth, kmutex_t **lockp)
796 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
797 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
800 mutex_enter(hash_lock);
801 for (buf = buf_hash_table.ht_table[idx]; buf != NULL;
802 buf = buf->b_hash_next) {
803 if (BUF_EQUAL(spa, dva, birth, buf)) {
808 mutex_exit(hash_lock);
814 * Insert an entry into the hash table. If there is already an element
815 * equal to elem in the hash table, then the already existing element
816 * will be returned and the new element will not be inserted.
817 * Otherwise returns NULL.
819 static arc_buf_hdr_t *
820 buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp)
822 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
823 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
827 ASSERT(!HDR_IN_HASH_TABLE(buf));
829 mutex_enter(hash_lock);
830 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL;
831 fbuf = fbuf->b_hash_next, i++) {
832 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf))
836 buf->b_hash_next = buf_hash_table.ht_table[idx];
837 buf_hash_table.ht_table[idx] = buf;
838 buf->b_flags |= ARC_IN_HASH_TABLE;
840 /* collect some hash table performance data */
842 ARCSTAT_BUMP(arcstat_hash_collisions);
844 ARCSTAT_BUMP(arcstat_hash_chains);
846 ARCSTAT_MAX(arcstat_hash_chain_max, i);
849 ARCSTAT_BUMP(arcstat_hash_elements);
850 ARCSTAT_MAXSTAT(arcstat_hash_elements);
856 buf_hash_remove(arc_buf_hdr_t *buf)
858 arc_buf_hdr_t *fbuf, **bufp;
859 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth);
861 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
862 ASSERT(HDR_IN_HASH_TABLE(buf));
864 bufp = &buf_hash_table.ht_table[idx];
865 while ((fbuf = *bufp) != buf) {
866 ASSERT(fbuf != NULL);
867 bufp = &fbuf->b_hash_next;
869 *bufp = buf->b_hash_next;
870 buf->b_hash_next = NULL;
871 buf->b_flags &= ~ARC_IN_HASH_TABLE;
873 /* collect some hash table performance data */
874 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
876 if (buf_hash_table.ht_table[idx] &&
877 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
878 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
882 * Global data structures and functions for the buf kmem cache.
884 static kmem_cache_t *hdr_cache;
885 static kmem_cache_t *buf_cache;
892 kmem_free(buf_hash_table.ht_table,
893 (buf_hash_table.ht_mask + 1) * sizeof (void *));
894 for (i = 0; i < BUF_LOCKS; i++)
895 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
896 kmem_cache_destroy(hdr_cache);
897 kmem_cache_destroy(buf_cache);
901 * Constructor callback - called when the cache is empty
902 * and a new buf is requested.
906 hdr_cons(void *vbuf, void *unused, int kmflag)
908 arc_buf_hdr_t *buf = vbuf;
910 bzero(buf, sizeof (arc_buf_hdr_t));
911 refcount_create(&buf->b_refcnt);
912 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL);
913 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
914 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
921 buf_cons(void *vbuf, void *unused, int kmflag)
923 arc_buf_t *buf = vbuf;
925 bzero(buf, sizeof (arc_buf_t));
926 rw_init(&buf->b_lock, NULL, RW_DEFAULT, NULL);
927 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS);
933 * Destructor callback - called when a cached buf is
934 * no longer required.
938 hdr_dest(void *vbuf, void *unused)
940 arc_buf_hdr_t *buf = vbuf;
942 refcount_destroy(&buf->b_refcnt);
943 cv_destroy(&buf->b_cv);
944 mutex_destroy(&buf->b_freeze_lock);
945 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS);
950 buf_dest(void *vbuf, void *unused)
952 arc_buf_t *buf = vbuf;
954 rw_destroy(&buf->b_lock);
955 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS);
959 * Reclaim callback -- invoked when memory is low.
963 hdr_recl(void *unused)
965 dprintf("hdr_recl called\n");
967 * umem calls the reclaim func when we destroy the buf cache,
968 * which is after we do arc_fini().
971 cv_signal(&arc_reclaim_thr_cv);
978 uint64_t hsize = 1ULL << 12;
982 * The hash table is big enough to fill all of physical memory
983 * with an average 64K block size. The table will take up
984 * totalmem*sizeof(void*)/64K (eg. 128KB/GB with 8-byte pointers).
986 while (hsize * 65536 < (uint64_t)physmem * PAGESIZE)
989 buf_hash_table.ht_mask = hsize - 1;
990 buf_hash_table.ht_table =
991 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
992 if (buf_hash_table.ht_table == NULL) {
993 ASSERT(hsize > (1ULL << 8));
998 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t),
999 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0);
1000 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1001 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1003 for (i = 0; i < 256; i++)
1004 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1005 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1007 for (i = 0; i < BUF_LOCKS; i++) {
1008 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1009 NULL, MUTEX_DEFAULT, NULL);
1013 #define ARC_MINTIME (hz>>4) /* 62 ms */
1016 arc_cksum_verify(arc_buf_t *buf)
1020 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1023 mutex_enter(&buf->b_hdr->b_freeze_lock);
1024 if (buf->b_hdr->b_freeze_cksum == NULL ||
1025 (buf->b_hdr->b_flags & ARC_IO_ERROR)) {
1026 mutex_exit(&buf->b_hdr->b_freeze_lock);
1029 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1030 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc))
1031 panic("buffer modified while frozen!");
1032 mutex_exit(&buf->b_hdr->b_freeze_lock);
1036 arc_cksum_equal(arc_buf_t *buf)
1041 mutex_enter(&buf->b_hdr->b_freeze_lock);
1042 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc);
1043 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc);
1044 mutex_exit(&buf->b_hdr->b_freeze_lock);
1050 arc_cksum_compute(arc_buf_t *buf, boolean_t force)
1052 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY))
1055 mutex_enter(&buf->b_hdr->b_freeze_lock);
1056 if (buf->b_hdr->b_freeze_cksum != NULL) {
1057 mutex_exit(&buf->b_hdr->b_freeze_lock);
1060 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
1061 fletcher_2_native(buf->b_data, buf->b_hdr->b_size,
1062 buf->b_hdr->b_freeze_cksum);
1063 mutex_exit(&buf->b_hdr->b_freeze_lock);
1067 arc_buf_thaw(arc_buf_t *buf)
1069 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1070 if (buf->b_hdr->b_state != arc_anon)
1071 panic("modifying non-anon buffer!");
1072 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS)
1073 panic("modifying buffer while i/o in progress!");
1074 arc_cksum_verify(buf);
1077 mutex_enter(&buf->b_hdr->b_freeze_lock);
1078 if (buf->b_hdr->b_freeze_cksum != NULL) {
1079 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1080 buf->b_hdr->b_freeze_cksum = NULL;
1082 mutex_exit(&buf->b_hdr->b_freeze_lock);
1086 arc_buf_freeze(arc_buf_t *buf)
1088 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1091 ASSERT(buf->b_hdr->b_freeze_cksum != NULL ||
1092 buf->b_hdr->b_state == arc_anon);
1093 arc_cksum_compute(buf, B_FALSE);
1097 get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock)
1099 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth);
1101 if (ab->b_type == ARC_BUFC_METADATA)
1102 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1);
1104 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1);
1105 buf_hashid += ARC_BUFC_NUMMETADATALISTS;
1108 *list = &state->arcs_lists[buf_hashid];
1109 *lock = ARCS_LOCK(state, buf_hashid);
1114 add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1117 ASSERT(MUTEX_HELD(hash_lock));
1119 if ((refcount_add(&ab->b_refcnt, tag) == 1) &&
1120 (ab->b_state != arc_anon)) {
1121 uint64_t delta = ab->b_size * ab->b_datacnt;
1122 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type];
1126 get_buf_info(ab, ab->b_state, &list, &lock);
1127 ASSERT(!MUTEX_HELD(lock));
1129 ASSERT(list_link_active(&ab->b_arc_node));
1130 list_remove(list, ab);
1131 if (GHOST_STATE(ab->b_state)) {
1132 ASSERT3U(ab->b_datacnt, ==, 0);
1133 ASSERT3P(ab->b_buf, ==, NULL);
1137 ASSERT3U(*size, >=, delta);
1138 atomic_add_64(size, -delta);
1140 /* remove the prefetch flag if we get a reference */
1141 if (ab->b_flags & ARC_PREFETCH)
1142 ab->b_flags &= ~ARC_PREFETCH;
1147 remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag)
1150 arc_state_t *state = ab->b_state;
1152 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock));
1153 ASSERT(!GHOST_STATE(state));
1155 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) &&
1156 (state != arc_anon)) {
1157 uint64_t *size = &state->arcs_lsize[ab->b_type];
1161 get_buf_info(ab, state, &list, &lock);
1162 ASSERT(!MUTEX_HELD(lock));
1164 ASSERT(!list_link_active(&ab->b_arc_node));
1165 list_insert_head(list, ab);
1166 ASSERT(ab->b_datacnt > 0);
1167 atomic_add_64(size, ab->b_size * ab->b_datacnt);
1174 * Move the supplied buffer to the indicated state. The mutex
1175 * for the buffer must be held by the caller.
1178 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock)
1180 arc_state_t *old_state = ab->b_state;
1181 int64_t refcnt = refcount_count(&ab->b_refcnt);
1182 uint64_t from_delta, to_delta;
1186 ASSERT(MUTEX_HELD(hash_lock));
1187 ASSERT(new_state != old_state);
1188 ASSERT(refcnt == 0 || ab->b_datacnt > 0);
1189 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state));
1191 from_delta = to_delta = ab->b_datacnt * ab->b_size;
1194 * If this buffer is evictable, transfer it from the
1195 * old state list to the new state list.
1198 if (old_state != arc_anon) {
1200 uint64_t *size = &old_state->arcs_lsize[ab->b_type];
1202 get_buf_info(ab, old_state, &list, &lock);
1203 use_mutex = !MUTEX_HELD(lock);
1207 ASSERT(list_link_active(&ab->b_arc_node));
1208 list_remove(list, ab);
1211 * If prefetching out of the ghost cache,
1212 * we will have a non-null datacnt.
1214 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) {
1215 /* ghost elements have a ghost size */
1216 ASSERT(ab->b_buf == NULL);
1217 from_delta = ab->b_size;
1219 ASSERT3U(*size, >=, from_delta);
1220 atomic_add_64(size, -from_delta);
1225 if (new_state != arc_anon) {
1227 uint64_t *size = &new_state->arcs_lsize[ab->b_type];
1229 get_buf_info(ab, new_state, &list, &lock);
1230 use_mutex = !MUTEX_HELD(lock);
1234 list_insert_head(list, ab);
1236 /* ghost elements have a ghost size */
1237 if (GHOST_STATE(new_state)) {
1238 ASSERT(ab->b_datacnt == 0);
1239 ASSERT(ab->b_buf == NULL);
1240 to_delta = ab->b_size;
1242 atomic_add_64(size, to_delta);
1249 ASSERT(!BUF_EMPTY(ab));
1250 if (new_state == arc_anon) {
1251 buf_hash_remove(ab);
1254 /* adjust state sizes */
1256 atomic_add_64(&new_state->arcs_size, to_delta);
1258 ASSERT3U(old_state->arcs_size, >=, from_delta);
1259 atomic_add_64(&old_state->arcs_size, -from_delta);
1261 ab->b_state = new_state;
1263 /* adjust l2arc hdr stats */
1264 if (new_state == arc_l2c_only)
1265 l2arc_hdr_stat_add();
1266 else if (old_state == arc_l2c_only)
1267 l2arc_hdr_stat_remove();
1271 arc_space_consume(uint64_t space, arc_space_type_t type)
1273 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1276 case ARC_SPACE_DATA:
1277 ARCSTAT_INCR(arcstat_data_size, space);
1279 case ARC_SPACE_OTHER:
1280 ARCSTAT_INCR(arcstat_other_size, space);
1282 case ARC_SPACE_HDRS:
1283 ARCSTAT_INCR(arcstat_hdr_size, space);
1285 case ARC_SPACE_L2HDRS:
1286 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
1290 atomic_add_64(&arc_meta_used, space);
1291 atomic_add_64(&arc_size, space);
1295 arc_space_return(uint64_t space, arc_space_type_t type)
1297 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
1300 case ARC_SPACE_DATA:
1301 ARCSTAT_INCR(arcstat_data_size, -space);
1303 case ARC_SPACE_OTHER:
1304 ARCSTAT_INCR(arcstat_other_size, -space);
1306 case ARC_SPACE_HDRS:
1307 ARCSTAT_INCR(arcstat_hdr_size, -space);
1309 case ARC_SPACE_L2HDRS:
1310 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
1314 ASSERT(arc_meta_used >= space);
1315 if (arc_meta_max < arc_meta_used)
1316 arc_meta_max = arc_meta_used;
1317 atomic_add_64(&arc_meta_used, -space);
1318 ASSERT(arc_size >= space);
1319 atomic_add_64(&arc_size, -space);
1323 arc_data_buf_alloc(uint64_t size)
1325 if (arc_evict_needed(ARC_BUFC_DATA))
1326 cv_signal(&arc_reclaim_thr_cv);
1327 atomic_add_64(&arc_size, size);
1328 return (zio_data_buf_alloc(size));
1332 arc_data_buf_free(void *buf, uint64_t size)
1334 zio_data_buf_free(buf, size);
1335 ASSERT(arc_size >= size);
1336 atomic_add_64(&arc_size, -size);
1340 arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type)
1345 ASSERT3U(size, >, 0);
1346 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
1347 ASSERT(BUF_EMPTY(hdr));
1351 hdr->b_state = arc_anon;
1352 hdr->b_arc_access = 0;
1353 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1356 buf->b_efunc = NULL;
1357 buf->b_private = NULL;
1360 arc_get_data_buf(buf);
1363 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1364 (void) refcount_add(&hdr->b_refcnt, tag);
1370 arc_buf_clone(arc_buf_t *from)
1373 arc_buf_hdr_t *hdr = from->b_hdr;
1374 uint64_t size = hdr->b_size;
1376 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
1379 buf->b_efunc = NULL;
1380 buf->b_private = NULL;
1381 buf->b_next = hdr->b_buf;
1383 arc_get_data_buf(buf);
1384 bcopy(from->b_data, buf->b_data, size);
1385 hdr->b_datacnt += 1;
1390 arc_buf_add_ref(arc_buf_t *buf, void* tag)
1393 kmutex_t *hash_lock;
1396 * Check to see if this buffer is evicted. Callers
1397 * must verify b_data != NULL to know if the add_ref
1400 rw_enter(&buf->b_lock, RW_READER);
1401 if (buf->b_data == NULL) {
1402 rw_exit(&buf->b_lock);
1406 ASSERT(hdr != NULL);
1407 hash_lock = HDR_LOCK(hdr);
1408 mutex_enter(hash_lock);
1409 rw_exit(&buf->b_lock);
1411 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
1412 add_reference(hdr, hash_lock, tag);
1413 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
1414 arc_access(hdr, hash_lock);
1415 mutex_exit(hash_lock);
1416 ARCSTAT_BUMP(arcstat_hits);
1417 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
1418 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
1419 data, metadata, hits);
1423 * Free the arc data buffer. If it is an l2arc write in progress,
1424 * the buffer is placed on l2arc_free_on_write to be freed later.
1427 arc_buf_data_free(arc_buf_hdr_t *hdr, void (*free_func)(void *, size_t),
1428 void *data, size_t size)
1430 if (HDR_L2_WRITING(hdr)) {
1431 l2arc_data_free_t *df;
1432 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP);
1433 df->l2df_data = data;
1434 df->l2df_size = size;
1435 df->l2df_func = free_func;
1436 mutex_enter(&l2arc_free_on_write_mtx);
1437 list_insert_head(l2arc_free_on_write, df);
1438 mutex_exit(&l2arc_free_on_write_mtx);
1439 ARCSTAT_BUMP(arcstat_l2_free_on_write);
1441 free_func(data, size);
1446 arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t all)
1450 /* free up data associated with the buf */
1452 arc_state_t *state = buf->b_hdr->b_state;
1453 uint64_t size = buf->b_hdr->b_size;
1454 arc_buf_contents_t type = buf->b_hdr->b_type;
1456 arc_cksum_verify(buf);
1458 if (type == ARC_BUFC_METADATA) {
1459 arc_buf_data_free(buf->b_hdr, zio_buf_free,
1461 arc_space_return(size, ARC_SPACE_DATA);
1463 ASSERT(type == ARC_BUFC_DATA);
1464 arc_buf_data_free(buf->b_hdr,
1465 zio_data_buf_free, buf->b_data, size);
1466 ARCSTAT_INCR(arcstat_data_size, -size);
1467 atomic_add_64(&arc_size, -size);
1470 if (list_link_active(&buf->b_hdr->b_arc_node)) {
1471 uint64_t *cnt = &state->arcs_lsize[type];
1473 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt));
1474 ASSERT(state != arc_anon);
1476 ASSERT3U(*cnt, >=, size);
1477 atomic_add_64(cnt, -size);
1479 ASSERT3U(state->arcs_size, >=, size);
1480 atomic_add_64(&state->arcs_size, -size);
1482 ASSERT(buf->b_hdr->b_datacnt > 0);
1483 buf->b_hdr->b_datacnt -= 1;
1486 /* only remove the buf if requested */
1490 /* remove the buf from the hdr list */
1491 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next)
1493 *bufp = buf->b_next;
1495 ASSERT(buf->b_efunc == NULL);
1497 /* clean up the buf */
1499 kmem_cache_free(buf_cache, buf);
1503 arc_hdr_destroy(arc_buf_hdr_t *hdr)
1505 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1506 ASSERT3P(hdr->b_state, ==, arc_anon);
1507 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1508 ASSERT(!(hdr->b_flags & ARC_STORED));
1510 if (hdr->b_l2hdr != NULL) {
1511 if (!MUTEX_HELD(&l2arc_buflist_mtx)) {
1513 * To prevent arc_free() and l2arc_evict() from
1514 * attempting to free the same buffer at the same time,
1515 * a FREE_IN_PROGRESS flag is given to arc_free() to
1516 * give it priority. l2arc_evict() can't destroy this
1517 * header while we are waiting on l2arc_buflist_mtx.
1519 * The hdr may be removed from l2ad_buflist before we
1520 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked.
1522 mutex_enter(&l2arc_buflist_mtx);
1523 if (hdr->b_l2hdr != NULL) {
1524 list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist,
1527 mutex_exit(&l2arc_buflist_mtx);
1529 list_remove(hdr->b_l2hdr->b_dev->l2ad_buflist, hdr);
1531 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size);
1532 kmem_free(hdr->b_l2hdr, sizeof (l2arc_buf_hdr_t));
1533 if (hdr->b_state == arc_l2c_only)
1534 l2arc_hdr_stat_remove();
1535 hdr->b_l2hdr = NULL;
1538 if (!BUF_EMPTY(hdr)) {
1539 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1540 bzero(&hdr->b_dva, sizeof (dva_t));
1544 while (hdr->b_buf) {
1545 arc_buf_t *buf = hdr->b_buf;
1548 mutex_enter(&arc_eviction_mtx);
1549 rw_enter(&buf->b_lock, RW_WRITER);
1550 ASSERT(buf->b_hdr != NULL);
1551 arc_buf_destroy(hdr->b_buf, FALSE, FALSE);
1552 hdr->b_buf = buf->b_next;
1553 buf->b_hdr = &arc_eviction_hdr;
1554 buf->b_next = arc_eviction_list;
1555 arc_eviction_list = buf;
1556 rw_exit(&buf->b_lock);
1557 mutex_exit(&arc_eviction_mtx);
1559 arc_buf_destroy(hdr->b_buf, FALSE, TRUE);
1562 if (hdr->b_freeze_cksum != NULL) {
1563 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1564 hdr->b_freeze_cksum = NULL;
1567 ASSERT(!list_link_active(&hdr->b_arc_node));
1568 ASSERT3P(hdr->b_hash_next, ==, NULL);
1569 ASSERT3P(hdr->b_acb, ==, NULL);
1570 kmem_cache_free(hdr_cache, hdr);
1574 arc_buf_free(arc_buf_t *buf, void *tag)
1576 arc_buf_hdr_t *hdr = buf->b_hdr;
1577 int hashed = hdr->b_state != arc_anon;
1579 ASSERT(buf->b_efunc == NULL);
1580 ASSERT(buf->b_data != NULL);
1583 kmutex_t *hash_lock = HDR_LOCK(hdr);
1585 mutex_enter(hash_lock);
1586 (void) remove_reference(hdr, hash_lock, tag);
1587 if (hdr->b_datacnt > 1)
1588 arc_buf_destroy(buf, FALSE, TRUE);
1590 hdr->b_flags |= ARC_BUF_AVAILABLE;
1591 mutex_exit(hash_lock);
1592 } else if (HDR_IO_IN_PROGRESS(hdr)) {
1595 * We are in the middle of an async write. Don't destroy
1596 * this buffer unless the write completes before we finish
1597 * decrementing the reference count.
1599 mutex_enter(&arc_eviction_mtx);
1600 (void) remove_reference(hdr, NULL, tag);
1601 ASSERT(refcount_is_zero(&hdr->b_refcnt));
1602 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr);
1603 mutex_exit(&arc_eviction_mtx);
1605 arc_hdr_destroy(hdr);
1607 if (remove_reference(hdr, NULL, tag) > 0) {
1608 ASSERT(HDR_IO_ERROR(hdr));
1609 arc_buf_destroy(buf, FALSE, TRUE);
1611 arc_hdr_destroy(hdr);
1617 arc_buf_remove_ref(arc_buf_t *buf, void* tag)
1619 arc_buf_hdr_t *hdr = buf->b_hdr;
1620 kmutex_t *hash_lock = HDR_LOCK(hdr);
1621 int no_callback = (buf->b_efunc == NULL);
1623 if (hdr->b_state == arc_anon) {
1624 arc_buf_free(buf, tag);
1625 return (no_callback);
1628 mutex_enter(hash_lock);
1629 ASSERT(hdr->b_state != arc_anon);
1630 ASSERT(buf->b_data != NULL);
1632 (void) remove_reference(hdr, hash_lock, tag);
1633 if (hdr->b_datacnt > 1) {
1635 arc_buf_destroy(buf, FALSE, TRUE);
1636 } else if (no_callback) {
1637 ASSERT(hdr->b_buf == buf && buf->b_next == NULL);
1638 hdr->b_flags |= ARC_BUF_AVAILABLE;
1640 ASSERT(no_callback || hdr->b_datacnt > 1 ||
1641 refcount_is_zero(&hdr->b_refcnt));
1642 mutex_exit(hash_lock);
1643 return (no_callback);
1647 arc_buf_size(arc_buf_t *buf)
1649 return (buf->b_hdr->b_size);
1653 * Evict buffers from list until we've removed the specified number of
1654 * bytes. Move the removed buffers to the appropriate evict state.
1655 * If the recycle flag is set, then attempt to "recycle" a buffer:
1656 * - look for a buffer to evict that is `bytes' long.
1657 * - return the data block from this buffer rather than freeing it.
1658 * This flag is used by callers that are trying to make space for a
1659 * new buffer in a full arc cache.
1661 * This function makes a "best effort". It skips over any buffers
1662 * it can't get a hash_lock on, and so may not catch all candidates.
1663 * It may also return without evicting as much space as requested.
1666 arc_evict(arc_state_t *state, spa_t *spa, int64_t bytes, boolean_t recycle,
1667 arc_buf_contents_t type)
1669 arc_state_t *evicted_state;
1670 uint64_t bytes_evicted = 0, skipped = 0, missed = 0;
1671 int64_t bytes_remaining;
1672 arc_buf_hdr_t *ab, *ab_prev = NULL;
1673 list_t *evicted_list, *list, *evicted_list_start, *list_start;
1674 kmutex_t *lock, *evicted_lock;
1675 kmutex_t *hash_lock;
1676 boolean_t have_lock;
1677 void *stolen = NULL;
1678 static int evict_metadata_offset, evict_data_offset;
1679 int i, idx, offset, list_count, count;
1681 ASSERT(state == arc_mru || state == arc_mfu);
1683 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
1685 if (type == ARC_BUFC_METADATA) {
1687 list_count = ARC_BUFC_NUMMETADATALISTS;
1688 list_start = &state->arcs_lists[0];
1689 evicted_list_start = &evicted_state->arcs_lists[0];
1690 idx = evict_metadata_offset;
1692 offset = ARC_BUFC_NUMMETADATALISTS;
1693 list_start = &state->arcs_lists[offset];
1694 evicted_list_start = &evicted_state->arcs_lists[offset];
1695 list_count = ARC_BUFC_NUMDATALISTS;
1696 idx = evict_data_offset;
1698 bytes_remaining = evicted_state->arcs_lsize[type];
1702 list = &list_start[idx];
1703 evicted_list = &evicted_list_start[idx];
1704 lock = ARCS_LOCK(state, (offset + idx));
1705 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx));
1708 mutex_enter(evicted_lock);
1710 for (ab = list_tail(list); ab; ab = ab_prev) {
1711 ab_prev = list_prev(list, ab);
1712 bytes_remaining -= (ab->b_size * ab->b_datacnt);
1713 /* prefetch buffers have a minimum lifespan */
1714 if (HDR_IO_IN_PROGRESS(ab) ||
1715 (spa && ab->b_spa != spa) ||
1716 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) &&
1717 LBOLT - ab->b_arc_access < arc_min_prefetch_lifespan)) {
1721 /* "lookahead" for better eviction candidate */
1722 if (recycle && ab->b_size != bytes &&
1723 ab_prev && ab_prev->b_size == bytes)
1725 hash_lock = HDR_LOCK(ab);
1726 have_lock = MUTEX_HELD(hash_lock);
1727 if (have_lock || mutex_tryenter(hash_lock)) {
1728 ASSERT3U(refcount_count(&ab->b_refcnt), ==, 0);
1729 ASSERT(ab->b_datacnt > 0);
1731 arc_buf_t *buf = ab->b_buf;
1732 if (!rw_tryenter(&buf->b_lock, RW_WRITER)) {
1737 bytes_evicted += ab->b_size;
1738 if (recycle && ab->b_type == type &&
1739 ab->b_size == bytes &&
1740 !HDR_L2_WRITING(ab)) {
1741 stolen = buf->b_data;
1746 mutex_enter(&arc_eviction_mtx);
1747 arc_buf_destroy(buf,
1748 buf->b_data == stolen, FALSE);
1749 ab->b_buf = buf->b_next;
1750 buf->b_hdr = &arc_eviction_hdr;
1751 buf->b_next = arc_eviction_list;
1752 arc_eviction_list = buf;
1753 mutex_exit(&arc_eviction_mtx);
1754 rw_exit(&buf->b_lock);
1756 rw_exit(&buf->b_lock);
1757 arc_buf_destroy(buf,
1758 buf->b_data == stolen, TRUE);
1763 ARCSTAT_INCR(arcstat_evict_l2_cached,
1766 if (l2arc_write_eligible(ab->b_spa, ab)) {
1767 ARCSTAT_INCR(arcstat_evict_l2_eligible,
1771 arcstat_evict_l2_ineligible,
1776 if (ab->b_datacnt == 0) {
1777 arc_change_state(evicted_state, ab, hash_lock);
1778 ASSERT(HDR_IN_HASH_TABLE(ab));
1779 ab->b_flags |= ARC_IN_HASH_TABLE;
1780 ab->b_flags &= ~ARC_BUF_AVAILABLE;
1781 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab);
1784 mutex_exit(hash_lock);
1785 if (bytes >= 0 && bytes_evicted >= bytes)
1787 if (bytes_remaining > 0) {
1788 mutex_exit(evicted_lock);
1790 idx = ((idx + 1) & (list_count - 1));
1799 mutex_exit(evicted_lock);
1802 idx = ((idx + 1) & (list_count - 1));
1805 if (bytes_evicted < bytes) {
1806 if (count < list_count)
1809 dprintf("only evicted %lld bytes from %x",
1810 (longlong_t)bytes_evicted, state);
1812 if (type == ARC_BUFC_METADATA)
1813 evict_metadata_offset = idx;
1815 evict_data_offset = idx;
1818 ARCSTAT_INCR(arcstat_evict_skip, skipped);
1821 ARCSTAT_INCR(arcstat_mutex_miss, missed);
1824 * We have just evicted some date into the ghost state, make
1825 * sure we also adjust the ghost state size if necessary.
1828 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size > arc_c) {
1829 int64_t mru_over = arc_anon->arcs_size + arc_mru->arcs_size +
1830 arc_mru_ghost->arcs_size - arc_c;
1832 if (mru_over > 0 && arc_mru_ghost->arcs_lsize[type] > 0) {
1834 MIN(arc_mru_ghost->arcs_lsize[type], mru_over);
1835 arc_evict_ghost(arc_mru_ghost, NULL, todelete);
1836 } else if (arc_mfu_ghost->arcs_lsize[type] > 0) {
1837 int64_t todelete = MIN(arc_mfu_ghost->arcs_lsize[type],
1838 arc_mru_ghost->arcs_size +
1839 arc_mfu_ghost->arcs_size - arc_c);
1840 arc_evict_ghost(arc_mfu_ghost, NULL, todelete);
1844 ARCSTAT_BUMP(arcstat_stolen);
1850 * Remove buffers from list until we've removed the specified number of
1851 * bytes. Destroy the buffers that are removed.
1854 arc_evict_ghost(arc_state_t *state, spa_t *spa, int64_t bytes)
1856 arc_buf_hdr_t *ab, *ab_prev;
1857 list_t *list, *list_start;
1858 kmutex_t *hash_lock, *lock;
1859 uint64_t bytes_deleted = 0;
1860 uint64_t bufs_skipped = 0;
1861 static int evict_offset;
1862 int list_count, idx = evict_offset;
1863 int offset, count = 0;
1865 ASSERT(GHOST_STATE(state));
1868 * data lists come after metadata lists
1870 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS];
1871 list_count = ARC_BUFC_NUMDATALISTS;
1872 offset = ARC_BUFC_NUMMETADATALISTS;
1875 list = &list_start[idx];
1876 lock = ARCS_LOCK(state, idx + offset);
1879 for (ab = list_tail(list); ab; ab = ab_prev) {
1880 ab_prev = list_prev(list, ab);
1881 if (spa && ab->b_spa != spa)
1883 hash_lock = HDR_LOCK(ab);
1884 if (mutex_tryenter(hash_lock)) {
1885 ASSERT(!HDR_IO_IN_PROGRESS(ab));
1886 ASSERT(ab->b_buf == NULL);
1887 ARCSTAT_BUMP(arcstat_deleted);
1888 bytes_deleted += ab->b_size;
1890 if (ab->b_l2hdr != NULL) {
1892 * This buffer is cached on the 2nd Level ARC;
1893 * don't destroy the header.
1895 arc_change_state(arc_l2c_only, ab, hash_lock);
1896 mutex_exit(hash_lock);
1898 arc_change_state(arc_anon, ab, hash_lock);
1899 mutex_exit(hash_lock);
1900 arc_hdr_destroy(ab);
1903 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab);
1904 if (bytes >= 0 && bytes_deleted >= bytes)
1909 * we're draining the ARC, retry
1912 mutex_enter(hash_lock);
1913 mutex_exit(hash_lock);
1920 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1));
1923 if (count < list_count)
1927 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] &&
1928 (bytes < 0 || bytes_deleted < bytes)) {
1929 list_start = &state->arcs_lists[0];
1930 list_count = ARC_BUFC_NUMMETADATALISTS;
1936 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped);
1940 if (bytes_deleted < bytes)
1941 dprintf("only deleted %lld bytes from %p",
1942 (longlong_t)bytes_deleted, state);
1948 int64_t adjustment, delta;
1954 adjustment = MIN(arc_size - arc_c,
1955 arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - arc_p);
1957 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) {
1958 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment);
1959 (void) arc_evict(arc_mru, NULL, delta, FALSE, ARC_BUFC_DATA);
1960 adjustment -= delta;
1963 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1964 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment);
1965 (void) arc_evict(arc_mru, NULL, delta, FALSE,
1973 adjustment = arc_size - arc_c;
1975 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) {
1976 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]);
1977 (void) arc_evict(arc_mfu, NULL, delta, FALSE, ARC_BUFC_DATA);
1978 adjustment -= delta;
1981 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) {
1982 int64_t delta = MIN(adjustment,
1983 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]);
1984 (void) arc_evict(arc_mfu, NULL, delta, FALSE,
1989 * Adjust ghost lists
1992 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c;
1994 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) {
1995 delta = MIN(arc_mru_ghost->arcs_size, adjustment);
1996 arc_evict_ghost(arc_mru_ghost, NULL, delta);
2000 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c;
2002 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) {
2003 delta = MIN(arc_mfu_ghost->arcs_size, adjustment);
2004 arc_evict_ghost(arc_mfu_ghost, NULL, delta);
2009 arc_do_user_evicts(void)
2011 static arc_buf_t *tmp_arc_eviction_list;
2014 * Move list over to avoid LOR
2017 mutex_enter(&arc_eviction_mtx);
2018 tmp_arc_eviction_list = arc_eviction_list;
2019 arc_eviction_list = NULL;
2020 mutex_exit(&arc_eviction_mtx);
2022 while (tmp_arc_eviction_list != NULL) {
2023 arc_buf_t *buf = tmp_arc_eviction_list;
2024 tmp_arc_eviction_list = buf->b_next;
2025 rw_enter(&buf->b_lock, RW_WRITER);
2027 rw_exit(&buf->b_lock);
2029 if (buf->b_efunc != NULL)
2030 VERIFY(buf->b_efunc(buf) == 0);
2032 buf->b_efunc = NULL;
2033 buf->b_private = NULL;
2034 kmem_cache_free(buf_cache, buf);
2037 if (arc_eviction_list != NULL)
2042 * Flush all *evictable* data from the cache for the given spa.
2043 * NOTE: this will not touch "active" (i.e. referenced) data.
2046 arc_flush(spa_t *spa)
2048 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) {
2049 (void) arc_evict(arc_mru, spa, -1, FALSE, ARC_BUFC_DATA);
2053 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) {
2054 (void) arc_evict(arc_mru, spa, -1, FALSE, ARC_BUFC_METADATA);
2058 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) {
2059 (void) arc_evict(arc_mfu, spa, -1, FALSE, ARC_BUFC_DATA);
2063 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) {
2064 (void) arc_evict(arc_mfu, spa, -1, FALSE, ARC_BUFC_METADATA);
2069 arc_evict_ghost(arc_mru_ghost, spa, -1);
2070 arc_evict_ghost(arc_mfu_ghost, spa, -1);
2072 mutex_enter(&arc_reclaim_thr_lock);
2073 arc_do_user_evicts();
2074 mutex_exit(&arc_reclaim_thr_lock);
2075 ASSERT(spa || arc_eviction_list == NULL);
2081 if (arc_c > arc_c_min) {
2085 to_free = arc_c >> arc_shrink_shift;
2087 to_free = arc_c >> arc_shrink_shift;
2089 if (arc_c > arc_c_min + to_free)
2090 atomic_add_64(&arc_c, -to_free);
2094 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
2095 if (arc_c > arc_size)
2096 arc_c = MAX(arc_size, arc_c_min);
2098 arc_p = (arc_c >> 1);
2099 ASSERT(arc_c >= arc_c_min);
2100 ASSERT((int64_t)arc_p >= 0);
2103 if (arc_size > arc_c)
2107 static int needfree = 0;
2110 arc_reclaim_needed(void)
2119 if (arc_size > arc_c_max)
2121 if (arc_size <= arc_c_min)
2125 * If pages are needed or we're within 2048 pages
2126 * of needing to page need to reclaim
2128 if (vm_pages_needed || (vm_paging_target() > -2048))
2133 * take 'desfree' extra pages, so we reclaim sooner, rather than later
2138 * check that we're out of range of the pageout scanner. It starts to
2139 * schedule paging if freemem is less than lotsfree and needfree.
2140 * lotsfree is the high-water mark for pageout, and needfree is the
2141 * number of needed free pages. We add extra pages here to make sure
2142 * the scanner doesn't start up while we're freeing memory.
2144 if (freemem < lotsfree + needfree + extra)
2148 * check to make sure that swapfs has enough space so that anon
2149 * reservations can still succeed. anon_resvmem() checks that the
2150 * availrmem is greater than swapfs_minfree, and the number of reserved
2151 * swap pages. We also add a bit of extra here just to prevent
2152 * circumstances from getting really dire.
2154 if (availrmem < swapfs_minfree + swapfs_reserve + extra)
2159 * If we're on an i386 platform, it's possible that we'll exhaust the
2160 * kernel heap space before we ever run out of available physical
2161 * memory. Most checks of the size of the heap_area compare against
2162 * tune.t_minarmem, which is the minimum available real memory that we
2163 * can have in the system. However, this is generally fixed at 25 pages
2164 * which is so low that it's useless. In this comparison, we seek to
2165 * calculate the total heap-size, and reclaim if more than 3/4ths of the
2166 * heap is allocated. (Or, in the calculation, if less than 1/4th is
2169 if (btop(vmem_size(heap_arena, VMEM_FREE)) <
2170 (btop(vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2))
2174 if (kmem_used() > (kmem_size() * 3) / 4)
2179 if (spa_get_random(100) == 0)
2185 extern kmem_cache_t *zio_buf_cache[];
2186 extern kmem_cache_t *zio_data_buf_cache[];
2189 arc_kmem_reap_now(arc_reclaim_strategy_t strat)
2192 kmem_cache_t *prev_cache = NULL;
2193 kmem_cache_t *prev_data_cache = NULL;
2196 if (arc_meta_used >= arc_meta_limit) {
2198 * We are exceeding our meta-data cache limit.
2199 * Purge some DNLC entries to release holds on meta-data.
2201 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
2205 * Reclaim unused memory from all kmem caches.
2212 * An aggressive reclamation will shrink the cache size as well as
2213 * reap free buffers from the arc kmem caches.
2215 if (strat == ARC_RECLAIM_AGGR)
2218 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) {
2219 if (zio_buf_cache[i] != prev_cache) {
2220 prev_cache = zio_buf_cache[i];
2221 kmem_cache_reap_now(zio_buf_cache[i]);
2223 if (zio_data_buf_cache[i] != prev_data_cache) {
2224 prev_data_cache = zio_data_buf_cache[i];
2225 kmem_cache_reap_now(zio_data_buf_cache[i]);
2228 kmem_cache_reap_now(buf_cache);
2229 kmem_cache_reap_now(hdr_cache);
2233 arc_reclaim_thread(void *dummy __unused)
2235 clock_t growtime = 0;
2236 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS;
2239 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG);
2241 mutex_enter(&arc_reclaim_thr_lock);
2242 while (arc_thread_exit == 0) {
2243 if (arc_reclaim_needed()) {
2246 if (last_reclaim == ARC_RECLAIM_CONS) {
2247 last_reclaim = ARC_RECLAIM_AGGR;
2249 last_reclaim = ARC_RECLAIM_CONS;
2253 last_reclaim = ARC_RECLAIM_AGGR;
2257 /* reset the growth delay for every reclaim */
2258 growtime = LBOLT + (arc_grow_retry * hz);
2260 if (needfree && last_reclaim == ARC_RECLAIM_CONS) {
2262 * If needfree is TRUE our vm_lowmem hook
2263 * was called and in that case we must free some
2264 * memory, so switch to aggressive mode.
2267 last_reclaim = ARC_RECLAIM_AGGR;
2269 arc_kmem_reap_now(last_reclaim);
2272 } else if (arc_no_grow && LBOLT >= growtime) {
2273 arc_no_grow = FALSE;
2277 (2 * arc_c < arc_size +
2278 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size))
2281 if (arc_eviction_list != NULL)
2282 arc_do_user_evicts();
2284 if (arc_reclaim_needed()) {
2291 /* block until needed, or one second, whichever is shorter */
2292 CALLB_CPR_SAFE_BEGIN(&cpr);
2293 (void) cv_timedwait(&arc_reclaim_thr_cv,
2294 &arc_reclaim_thr_lock, hz);
2295 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock);
2298 arc_thread_exit = 0;
2299 cv_broadcast(&arc_reclaim_thr_cv);
2300 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */
2305 * Adapt arc info given the number of bytes we are trying to add and
2306 * the state that we are comming from. This function is only called
2307 * when we are adding new content to the cache.
2310 arc_adapt(int bytes, arc_state_t *state)
2313 uint64_t arc_p_min = (arc_c >> arc_p_min_shift);
2315 if (state == arc_l2c_only)
2320 * Adapt the target size of the MRU list:
2321 * - if we just hit in the MRU ghost list, then increase
2322 * the target size of the MRU list.
2323 * - if we just hit in the MFU ghost list, then increase
2324 * the target size of the MFU list by decreasing the
2325 * target size of the MRU list.
2327 if (state == arc_mru_ghost) {
2328 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ?
2329 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size));
2331 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult);
2332 } else if (state == arc_mfu_ghost) {
2335 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ?
2336 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size));
2338 delta = MIN(bytes * mult, arc_p);
2339 arc_p = MAX(arc_p_min, arc_p - delta);
2341 ASSERT((int64_t)arc_p >= 0);
2343 if (arc_reclaim_needed()) {
2344 cv_signal(&arc_reclaim_thr_cv);
2351 if (arc_c >= arc_c_max)
2355 * If we're within (2 * maxblocksize) bytes of the target
2356 * cache size, increment the target cache size
2358 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
2359 atomic_add_64(&arc_c, (int64_t)bytes);
2360 if (arc_c > arc_c_max)
2362 else if (state == arc_anon)
2363 atomic_add_64(&arc_p, (int64_t)bytes);
2367 ASSERT((int64_t)arc_p >= 0);
2371 * Check if the cache has reached its limits and eviction is required
2375 arc_evict_needed(arc_buf_contents_t type)
2377 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit)
2383 * If zio data pages are being allocated out of a separate heap segment,
2384 * then enforce that the size of available vmem for this area remains
2385 * above about 1/32nd free.
2387 if (type == ARC_BUFC_DATA && zio_arena != NULL &&
2388 vmem_size(zio_arena, VMEM_FREE) <
2389 (vmem_size(zio_arena, VMEM_ALLOC) >> 5))
2394 if (arc_reclaim_needed())
2397 return (arc_size > arc_c);
2401 * The buffer, supplied as the first argument, needs a data block.
2402 * So, if we are at cache max, determine which cache should be victimized.
2403 * We have the following cases:
2405 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) ->
2406 * In this situation if we're out of space, but the resident size of the MFU is
2407 * under the limit, victimize the MFU cache to satisfy this insertion request.
2409 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) ->
2410 * Here, we've used up all of the available space for the MRU, so we need to
2411 * evict from our own cache instead. Evict from the set of resident MRU
2414 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) ->
2415 * c minus p represents the MFU space in the cache, since p is the size of the
2416 * cache that is dedicated to the MRU. In this situation there's still space on
2417 * the MFU side, so the MRU side needs to be victimized.
2419 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) ->
2420 * MFU's resident set is consuming more space than it has been allotted. In
2421 * this situation, we must victimize our own cache, the MFU, for this insertion.
2424 arc_get_data_buf(arc_buf_t *buf)
2426 arc_state_t *state = buf->b_hdr->b_state;
2427 uint64_t size = buf->b_hdr->b_size;
2428 arc_buf_contents_t type = buf->b_hdr->b_type;
2430 arc_adapt(size, state);
2433 * We have not yet reached cache maximum size,
2434 * just allocate a new buffer.
2436 if (!arc_evict_needed(type)) {
2437 if (type == ARC_BUFC_METADATA) {
2438 buf->b_data = zio_buf_alloc(size);
2439 arc_space_consume(size, ARC_SPACE_DATA);
2441 ASSERT(type == ARC_BUFC_DATA);
2442 buf->b_data = zio_data_buf_alloc(size);
2443 ARCSTAT_INCR(arcstat_data_size, size);
2444 atomic_add_64(&arc_size, size);
2450 * If we are prefetching from the mfu ghost list, this buffer
2451 * will end up on the mru list; so steal space from there.
2453 if (state == arc_mfu_ghost)
2454 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu;
2455 else if (state == arc_mru_ghost)
2458 if (state == arc_mru || state == arc_anon) {
2459 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size;
2460 state = (arc_mfu->arcs_lsize[type] >= size &&
2461 arc_p > mru_used) ? arc_mfu : arc_mru;
2464 uint64_t mfu_space = arc_c - arc_p;
2465 state = (arc_mru->arcs_lsize[type] >= size &&
2466 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu;
2468 if ((buf->b_data = arc_evict(state, NULL, size, TRUE, type)) == NULL) {
2469 if (type == ARC_BUFC_METADATA) {
2470 buf->b_data = zio_buf_alloc(size);
2471 arc_space_consume(size, ARC_SPACE_DATA);
2473 ASSERT(type == ARC_BUFC_DATA);
2474 buf->b_data = zio_data_buf_alloc(size);
2475 ARCSTAT_INCR(arcstat_data_size, size);
2476 atomic_add_64(&arc_size, size);
2478 ARCSTAT_BUMP(arcstat_recycle_miss);
2480 ASSERT(buf->b_data != NULL);
2483 * Update the state size. Note that ghost states have a
2484 * "ghost size" and so don't need to be updated.
2486 if (!GHOST_STATE(buf->b_hdr->b_state)) {
2487 arc_buf_hdr_t *hdr = buf->b_hdr;
2489 atomic_add_64(&hdr->b_state->arcs_size, size);
2490 if (list_link_active(&hdr->b_arc_node)) {
2491 ASSERT(refcount_is_zero(&hdr->b_refcnt));
2492 atomic_add_64(&hdr->b_state->arcs_lsize[type], size);
2495 * If we are growing the cache, and we are adding anonymous
2496 * data, and we have outgrown arc_p, update arc_p
2498 if (arc_size < arc_c && hdr->b_state == arc_anon &&
2499 arc_anon->arcs_size + arc_mru->arcs_size > arc_p)
2500 arc_p = MIN(arc_c, arc_p + size);
2502 ARCSTAT_BUMP(arcstat_allocated);
2506 * This routine is called whenever a buffer is accessed.
2507 * NOTE: the hash lock is dropped in this function.
2510 arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock)
2512 ASSERT(MUTEX_HELD(hash_lock));
2514 if (buf->b_state == arc_anon) {
2516 * This buffer is not in the cache, and does not
2517 * appear in our "ghost" list. Add the new buffer
2521 ASSERT(buf->b_arc_access == 0);
2522 buf->b_arc_access = LBOLT;
2523 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2524 arc_change_state(arc_mru, buf, hash_lock);
2526 } else if (buf->b_state == arc_mru) {
2528 * If this buffer is here because of a prefetch, then either:
2529 * - clear the flag if this is a "referencing" read
2530 * (any subsequent access will bump this into the MFU state).
2532 * - move the buffer to the head of the list if this is
2533 * another prefetch (to make it less likely to be evicted).
2535 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2536 if (refcount_count(&buf->b_refcnt) == 0) {
2537 ASSERT(list_link_active(&buf->b_arc_node));
2539 buf->b_flags &= ~ARC_PREFETCH;
2540 ARCSTAT_BUMP(arcstat_mru_hits);
2542 buf->b_arc_access = LBOLT;
2547 * This buffer has been "accessed" only once so far,
2548 * but it is still in the cache. Move it to the MFU
2551 if (LBOLT > buf->b_arc_access + ARC_MINTIME) {
2553 * More than 125ms have passed since we
2554 * instantiated this buffer. Move it to the
2555 * most frequently used state.
2557 buf->b_arc_access = LBOLT;
2558 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2559 arc_change_state(arc_mfu, buf, hash_lock);
2561 ARCSTAT_BUMP(arcstat_mru_hits);
2562 } else if (buf->b_state == arc_mru_ghost) {
2563 arc_state_t *new_state;
2565 * This buffer has been "accessed" recently, but
2566 * was evicted from the cache. Move it to the
2570 if (buf->b_flags & ARC_PREFETCH) {
2571 new_state = arc_mru;
2572 if (refcount_count(&buf->b_refcnt) > 0)
2573 buf->b_flags &= ~ARC_PREFETCH;
2574 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf);
2576 new_state = arc_mfu;
2577 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2580 buf->b_arc_access = LBOLT;
2581 arc_change_state(new_state, buf, hash_lock);
2583 ARCSTAT_BUMP(arcstat_mru_ghost_hits);
2584 } else if (buf->b_state == arc_mfu) {
2586 * This buffer has been accessed more than once and is
2587 * still in the cache. Keep it in the MFU state.
2589 * NOTE: an add_reference() that occurred when we did
2590 * the arc_read() will have kicked this off the list.
2591 * If it was a prefetch, we will explicitly move it to
2592 * the head of the list now.
2594 if ((buf->b_flags & ARC_PREFETCH) != 0) {
2595 ASSERT(refcount_count(&buf->b_refcnt) == 0);
2596 ASSERT(list_link_active(&buf->b_arc_node));
2598 ARCSTAT_BUMP(arcstat_mfu_hits);
2599 buf->b_arc_access = LBOLT;
2600 } else if (buf->b_state == arc_mfu_ghost) {
2601 arc_state_t *new_state = arc_mfu;
2603 * This buffer has been accessed more than once but has
2604 * been evicted from the cache. Move it back to the
2608 if (buf->b_flags & ARC_PREFETCH) {
2610 * This is a prefetch access...
2611 * move this block back to the MRU state.
2613 ASSERT3U(refcount_count(&buf->b_refcnt), ==, 0);
2614 new_state = arc_mru;
2617 buf->b_arc_access = LBOLT;
2618 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2619 arc_change_state(new_state, buf, hash_lock);
2621 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
2622 } else if (buf->b_state == arc_l2c_only) {
2624 * This buffer is on the 2nd Level ARC.
2627 buf->b_arc_access = LBOLT;
2628 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf);
2629 arc_change_state(arc_mfu, buf, hash_lock);
2631 ASSERT(!"invalid arc state");
2635 /* a generic arc_done_func_t which you can use */
2638 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
2640 bcopy(buf->b_data, arg, buf->b_hdr->b_size);
2641 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2644 /* a generic arc_done_func_t */
2646 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
2648 arc_buf_t **bufp = arg;
2649 if (zio && zio->io_error) {
2650 VERIFY(arc_buf_remove_ref(buf, arg) == 1);
2658 arc_read_done(zio_t *zio)
2660 arc_buf_hdr_t *hdr, *found;
2662 arc_buf_t *abuf; /* buffer we're assigning to callback */
2663 kmutex_t *hash_lock;
2664 arc_callback_t *callback_list, *acb;
2665 int freeable = FALSE;
2667 buf = zio->io_private;
2671 * The hdr was inserted into hash-table and removed from lists
2672 * prior to starting I/O. We should find this header, since
2673 * it's in the hash table, and it should be legit since it's
2674 * not possible to evict it during the I/O. The only possible
2675 * reason for it not to be found is if we were freed during the
2678 found = buf_hash_find(zio->io_spa, &hdr->b_dva, hdr->b_birth,
2681 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && hash_lock == NULL) ||
2682 (found == hdr && DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) ||
2683 (found == hdr && HDR_L2_READING(hdr)));
2685 hdr->b_flags &= ~ARC_L2_EVICTED;
2686 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH))
2687 hdr->b_flags &= ~ARC_L2CACHE;
2689 /* byteswap if necessary */
2690 callback_list = hdr->b_acb;
2691 ASSERT(callback_list != NULL);
2692 if (BP_SHOULD_BYTESWAP(zio->io_bp)) {
2693 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ?
2694 byteswap_uint64_array :
2695 dmu_ot[BP_GET_TYPE(zio->io_bp)].ot_byteswap;
2696 func(buf->b_data, hdr->b_size);
2699 arc_cksum_compute(buf, B_FALSE);
2701 /* create copies of the data buffer for the callers */
2703 for (acb = callback_list; acb; acb = acb->acb_next) {
2704 if (acb->acb_done) {
2706 abuf = arc_buf_clone(buf);
2707 acb->acb_buf = abuf;
2712 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
2713 ASSERT(!HDR_BUF_AVAILABLE(hdr));
2715 hdr->b_flags |= ARC_BUF_AVAILABLE;
2717 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL);
2719 if (zio->io_error != 0) {
2720 hdr->b_flags |= ARC_IO_ERROR;
2721 if (hdr->b_state != arc_anon)
2722 arc_change_state(arc_anon, hdr, hash_lock);
2723 if (HDR_IN_HASH_TABLE(hdr))
2724 buf_hash_remove(hdr);
2725 freeable = refcount_is_zero(&hdr->b_refcnt);
2729 * Broadcast before we drop the hash_lock to avoid the possibility
2730 * that the hdr (and hence the cv) might be freed before we get to
2731 * the cv_broadcast().
2733 cv_broadcast(&hdr->b_cv);
2737 * Only call arc_access on anonymous buffers. This is because
2738 * if we've issued an I/O for an evicted buffer, we've already
2739 * called arc_access (to prevent any simultaneous readers from
2740 * getting confused).
2742 if (zio->io_error == 0 && hdr->b_state == arc_anon)
2743 arc_access(hdr, hash_lock);
2744 mutex_exit(hash_lock);
2747 * This block was freed while we waited for the read to
2748 * complete. It has been removed from the hash table and
2749 * moved to the anonymous state (so that it won't show up
2752 ASSERT3P(hdr->b_state, ==, arc_anon);
2753 freeable = refcount_is_zero(&hdr->b_refcnt);
2756 /* execute each callback and free its structure */
2757 while ((acb = callback_list) != NULL) {
2759 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
2761 if (acb->acb_zio_dummy != NULL) {
2762 acb->acb_zio_dummy->io_error = zio->io_error;
2763 zio_nowait(acb->acb_zio_dummy);
2766 callback_list = acb->acb_next;
2767 kmem_free(acb, sizeof (arc_callback_t));
2771 arc_hdr_destroy(hdr);
2775 * "Read" the block block at the specified DVA (in bp) via the
2776 * cache. If the block is found in the cache, invoke the provided
2777 * callback immediately and return. Note that the `zio' parameter
2778 * in the callback will be NULL in this case, since no IO was
2779 * required. If the block is not in the cache pass the read request
2780 * on to the spa with a substitute callback function, so that the
2781 * requested block will be added to the cache.
2783 * If a read request arrives for a block that has a read in-progress,
2784 * either wait for the in-progress read to complete (and return the
2785 * results); or, if this is a read with a "done" func, add a record
2786 * to the read to invoke the "done" func when the read completes,
2787 * and return; or just return.
2789 * arc_read_done() will invoke all the requested "done" functions
2790 * for readers of this block.
2792 * Normal callers should use arc_read and pass the arc buffer and offset
2793 * for the bp. But if you know you don't need locking, you can use
2797 arc_read(zio_t *pio, spa_t *spa, blkptr_t *bp, arc_buf_t *pbuf,
2798 arc_done_func_t *done, void *private, int priority, int zio_flags,
2799 uint32_t *arc_flags, const zbookmark_t *zb)
2803 ASSERT(!refcount_is_zero(&pbuf->b_hdr->b_refcnt));
2804 ASSERT3U((char *)bp - (char *)pbuf->b_data, <, pbuf->b_hdr->b_size);
2805 rw_enter(&pbuf->b_lock, RW_READER);
2807 err = arc_read_nolock(pio, spa, bp, done, private, priority,
2808 zio_flags, arc_flags, zb);
2809 rw_exit(&pbuf->b_lock);
2814 arc_read_nolock(zio_t *pio, spa_t *spa, blkptr_t *bp,
2815 arc_done_func_t *done, void *private, int priority, int zio_flags,
2816 uint32_t *arc_flags, const zbookmark_t *zb)
2820 kmutex_t *hash_lock;
2824 hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
2825 if (hdr && hdr->b_datacnt > 0) {
2827 *arc_flags |= ARC_CACHED;
2829 if (HDR_IO_IN_PROGRESS(hdr)) {
2831 if (*arc_flags & ARC_WAIT) {
2832 cv_wait(&hdr->b_cv, hash_lock);
2833 mutex_exit(hash_lock);
2836 ASSERT(*arc_flags & ARC_NOWAIT);
2839 arc_callback_t *acb = NULL;
2841 acb = kmem_zalloc(sizeof (arc_callback_t),
2843 acb->acb_done = done;
2844 acb->acb_private = private;
2846 acb->acb_zio_dummy = zio_null(pio,
2847 spa, NULL, NULL, zio_flags);
2849 ASSERT(acb->acb_done != NULL);
2850 acb->acb_next = hdr->b_acb;
2852 add_reference(hdr, hash_lock, private);
2853 mutex_exit(hash_lock);
2856 mutex_exit(hash_lock);
2860 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
2863 add_reference(hdr, hash_lock, private);
2865 * If this block is already in use, create a new
2866 * copy of the data so that we will be guaranteed
2867 * that arc_release() will always succeed.
2871 ASSERT(buf->b_data);
2872 if (HDR_BUF_AVAILABLE(hdr)) {
2873 ASSERT(buf->b_efunc == NULL);
2874 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
2876 buf = arc_buf_clone(buf);
2878 } else if (*arc_flags & ARC_PREFETCH &&
2879 refcount_count(&hdr->b_refcnt) == 0) {
2880 hdr->b_flags |= ARC_PREFETCH;
2882 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
2883 arc_access(hdr, hash_lock);
2884 if (*arc_flags & ARC_L2CACHE)
2885 hdr->b_flags |= ARC_L2CACHE;
2886 mutex_exit(hash_lock);
2887 ARCSTAT_BUMP(arcstat_hits);
2888 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2889 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2890 data, metadata, hits);
2893 done(NULL, buf, private);
2895 uint64_t size = BP_GET_LSIZE(bp);
2896 arc_callback_t *acb;
2899 boolean_t devw = B_FALSE;
2902 /* this block is not in the cache */
2903 arc_buf_hdr_t *exists;
2904 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
2905 buf = arc_buf_alloc(spa, size, private, type);
2907 hdr->b_dva = *BP_IDENTITY(bp);
2908 hdr->b_birth = bp->blk_birth;
2909 hdr->b_cksum0 = bp->blk_cksum.zc_word[0];
2910 exists = buf_hash_insert(hdr, &hash_lock);
2912 /* somebody beat us to the hash insert */
2913 mutex_exit(hash_lock);
2914 bzero(&hdr->b_dva, sizeof (dva_t));
2917 (void) arc_buf_remove_ref(buf, private);
2918 goto top; /* restart the IO request */
2920 /* if this is a prefetch, we don't have a reference */
2921 if (*arc_flags & ARC_PREFETCH) {
2922 (void) remove_reference(hdr, hash_lock,
2924 hdr->b_flags |= ARC_PREFETCH;
2926 if (*arc_flags & ARC_L2CACHE)
2927 hdr->b_flags |= ARC_L2CACHE;
2928 if (BP_GET_LEVEL(bp) > 0)
2929 hdr->b_flags |= ARC_INDIRECT;
2931 /* this block is in the ghost cache */
2932 ASSERT(GHOST_STATE(hdr->b_state));
2933 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2934 ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 0);
2935 ASSERT(hdr->b_buf == NULL);
2937 /* if this is a prefetch, we don't have a reference */
2938 if (*arc_flags & ARC_PREFETCH)
2939 hdr->b_flags |= ARC_PREFETCH;
2941 add_reference(hdr, hash_lock, private);
2942 if (*arc_flags & ARC_L2CACHE)
2943 hdr->b_flags |= ARC_L2CACHE;
2944 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2947 buf->b_efunc = NULL;
2948 buf->b_private = NULL;
2951 arc_get_data_buf(buf);
2952 ASSERT(hdr->b_datacnt == 0);
2957 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
2958 acb->acb_done = done;
2959 acb->acb_private = private;
2961 ASSERT(hdr->b_acb == NULL);
2963 hdr->b_flags |= ARC_IO_IN_PROGRESS;
2966 * If the buffer has been evicted, migrate it to a present state
2967 * before issuing the I/O. Once we drop the hash-table lock,
2968 * the header will be marked as I/O in progress and have an
2969 * attached buffer. At this point, anybody who finds this
2970 * buffer ought to notice that it's legit but has a pending I/O.
2973 if (GHOST_STATE(hdr->b_state))
2974 arc_access(hdr, hash_lock);
2976 if (HDR_L2CACHE(hdr) && hdr->b_l2hdr != NULL &&
2977 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) {
2978 devw = hdr->b_l2hdr->b_dev->l2ad_writing;
2979 addr = hdr->b_l2hdr->b_daddr;
2981 * Lock out device removal.
2983 if (vdev_is_dead(vd) ||
2984 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
2988 mutex_exit(hash_lock);
2990 ASSERT3U(hdr->b_size, ==, size);
2991 DTRACE_PROBE3(arc__miss, blkptr_t *, bp, uint64_t, size,
2993 ARCSTAT_BUMP(arcstat_misses);
2994 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH),
2995 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA,
2996 data, metadata, misses);
2998 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
3000 * Read from the L2ARC if the following are true:
3001 * 1. The L2ARC vdev was previously cached.
3002 * 2. This buffer still has L2ARC metadata.
3003 * 3. This buffer isn't currently writing to the L2ARC.
3004 * 4. The L2ARC entry wasn't evicted, which may
3005 * also have invalidated the vdev.
3006 * 5. This isn't prefetch and l2arc_noprefetch is set.
3008 if (hdr->b_l2hdr != NULL &&
3009 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
3010 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
3011 l2arc_read_callback_t *cb;
3013 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
3014 ARCSTAT_BUMP(arcstat_l2_hits);
3016 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
3018 cb->l2rcb_buf = buf;
3019 cb->l2rcb_spa = spa;
3022 cb->l2rcb_flags = zio_flags;
3025 * l2arc read. The SCL_L2ARC lock will be
3026 * released by l2arc_read_done().
3028 rzio = zio_read_phys(pio, vd, addr, size,
3029 buf->b_data, ZIO_CHECKSUM_OFF,
3030 l2arc_read_done, cb, priority, zio_flags |
3031 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
3032 ZIO_FLAG_DONT_PROPAGATE |
3033 ZIO_FLAG_DONT_RETRY, B_FALSE);
3034 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
3036 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
3038 if (*arc_flags & ARC_NOWAIT) {
3043 ASSERT(*arc_flags & ARC_WAIT);
3044 if (zio_wait(rzio) == 0)
3047 /* l2arc read error; goto zio_read() */
3049 DTRACE_PROBE1(l2arc__miss,
3050 arc_buf_hdr_t *, hdr);
3051 ARCSTAT_BUMP(arcstat_l2_misses);
3052 if (HDR_L2_WRITING(hdr))
3053 ARCSTAT_BUMP(arcstat_l2_rw_clash);
3054 spa_config_exit(spa, SCL_L2ARC, vd);
3058 spa_config_exit(spa, SCL_L2ARC, vd);
3059 if (l2arc_ndev != 0) {
3060 DTRACE_PROBE1(l2arc__miss,
3061 arc_buf_hdr_t *, hdr);
3062 ARCSTAT_BUMP(arcstat_l2_misses);
3066 rzio = zio_read(pio, spa, bp, buf->b_data, size,
3067 arc_read_done, buf, priority, zio_flags, zb);
3069 if (*arc_flags & ARC_WAIT)
3070 return (zio_wait(rzio));
3072 ASSERT(*arc_flags & ARC_NOWAIT);
3079 * arc_read() variant to support pool traversal. If the block is already
3080 * in the ARC, make a copy of it; otherwise, the caller will do the I/O.
3081 * The idea is that we don't want pool traversal filling up memory, but
3082 * if the ARC already has the data anyway, we shouldn't pay for the I/O.
3085 arc_tryread(spa_t *spa, blkptr_t *bp, void *data)
3091 hdr = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_mtx);
3093 if (hdr && hdr->b_datacnt > 0 && !HDR_IO_IN_PROGRESS(hdr)) {
3094 arc_buf_t *buf = hdr->b_buf;
3097 while (buf->b_data == NULL) {
3101 bcopy(buf->b_data, data, hdr->b_size);
3107 mutex_exit(hash_mtx);
3113 arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private)
3115 ASSERT(buf->b_hdr != NULL);
3116 ASSERT(buf->b_hdr->b_state != arc_anon);
3117 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL);
3118 buf->b_efunc = func;
3119 buf->b_private = private;
3123 * This is used by the DMU to let the ARC know that a buffer is
3124 * being evicted, so the ARC should clean up. If this arc buf
3125 * is not yet in the evicted state, it will be put there.
3128 arc_buf_evict(arc_buf_t *buf)
3131 kmutex_t *hash_lock;
3133 list_t *list, *evicted_list;
3134 kmutex_t *lock, *evicted_lock;
3136 rw_enter(&buf->b_lock, RW_WRITER);
3140 * We are in arc_do_user_evicts().
3142 ASSERT(buf->b_data == NULL);
3143 rw_exit(&buf->b_lock);
3145 } else if (buf->b_data == NULL) {
3146 arc_buf_t copy = *buf; /* structure assignment */
3148 * We are on the eviction list; process this buffer now
3149 * but let arc_do_user_evicts() do the reaping.
3151 buf->b_efunc = NULL;
3152 rw_exit(&buf->b_lock);
3153 VERIFY(copy.b_efunc(©) == 0);
3156 hash_lock = HDR_LOCK(hdr);
3157 mutex_enter(hash_lock);
3159 ASSERT(buf->b_hdr == hdr);
3160 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt);
3161 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu);
3164 * Pull this buffer off of the hdr
3167 while (*bufp != buf)
3168 bufp = &(*bufp)->b_next;
3169 *bufp = buf->b_next;
3171 ASSERT(buf->b_data != NULL);
3172 arc_buf_destroy(buf, FALSE, FALSE);
3174 if (hdr->b_datacnt == 0) {
3175 arc_state_t *old_state = hdr->b_state;
3176 arc_state_t *evicted_state;
3178 ASSERT(refcount_is_zero(&hdr->b_refcnt));
3181 (old_state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
3183 get_buf_info(hdr, old_state, &list, &lock);
3184 get_buf_info(hdr, evicted_state, &evicted_list, &evicted_lock);
3186 mutex_enter(evicted_lock);
3188 arc_change_state(evicted_state, hdr, hash_lock);
3189 ASSERT(HDR_IN_HASH_TABLE(hdr));
3190 hdr->b_flags |= ARC_IN_HASH_TABLE;
3191 hdr->b_flags &= ~ARC_BUF_AVAILABLE;
3193 mutex_exit(evicted_lock);
3196 mutex_exit(hash_lock);
3197 rw_exit(&buf->b_lock);
3199 VERIFY(buf->b_efunc(buf) == 0);
3200 buf->b_efunc = NULL;
3201 buf->b_private = NULL;
3203 kmem_cache_free(buf_cache, buf);
3208 * Release this buffer from the cache. This must be done
3209 * after a read and prior to modifying the buffer contents.
3210 * If the buffer has more than one reference, we must make
3211 * a new hdr for the buffer.
3214 arc_release(arc_buf_t *buf, void *tag)
3217 kmutex_t *hash_lock;
3218 l2arc_buf_hdr_t *l2hdr;
3220 boolean_t released = B_FALSE;
3222 rw_enter(&buf->b_lock, RW_WRITER);
3225 /* this buffer is not on any list */
3226 ASSERT(refcount_count(&hdr->b_refcnt) > 0);
3227 ASSERT(!(hdr->b_flags & ARC_STORED));
3229 if (hdr->b_state == arc_anon) {
3230 /* this buffer is already released */
3231 ASSERT3U(refcount_count(&hdr->b_refcnt), ==, 1);
3232 ASSERT(BUF_EMPTY(hdr));
3233 ASSERT(buf->b_efunc == NULL);
3235 rw_exit(&buf->b_lock);
3238 hash_lock = HDR_LOCK(hdr);
3239 mutex_enter(hash_lock);
3242 l2hdr = hdr->b_l2hdr;
3244 mutex_enter(&l2arc_buflist_mtx);
3245 hdr->b_l2hdr = NULL;
3246 buf_size = hdr->b_size;
3253 * Do we have more than one buf?
3255 if (hdr->b_datacnt > 1) {
3256 arc_buf_hdr_t *nhdr;
3258 uint64_t blksz = hdr->b_size;
3259 spa_t *spa = hdr->b_spa;
3260 arc_buf_contents_t type = hdr->b_type;
3261 uint32_t flags = hdr->b_flags;
3263 ASSERT(hdr->b_buf != buf || buf->b_next != NULL);
3265 * Pull the data off of this buf and attach it to
3266 * a new anonymous buf.
3268 (void) remove_reference(hdr, hash_lock, tag);
3270 while (*bufp != buf)
3271 bufp = &(*bufp)->b_next;
3272 *bufp = (*bufp)->b_next;
3275 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size);
3276 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size);
3277 if (refcount_is_zero(&hdr->b_refcnt)) {
3278 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type];
3279 ASSERT3U(*size, >=, hdr->b_size);
3280 atomic_add_64(size, -hdr->b_size);
3282 hdr->b_datacnt -= 1;
3283 arc_cksum_verify(buf);
3285 mutex_exit(hash_lock);
3287 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
3288 nhdr->b_size = blksz;
3290 nhdr->b_type = type;
3292 nhdr->b_state = arc_anon;
3293 nhdr->b_arc_access = 0;
3294 nhdr->b_flags = flags & ARC_L2_WRITING;
3295 nhdr->b_l2hdr = NULL;
3296 nhdr->b_datacnt = 1;
3297 nhdr->b_freeze_cksum = NULL;
3298 (void) refcount_add(&nhdr->b_refcnt, tag);
3300 rw_exit(&buf->b_lock);
3301 atomic_add_64(&arc_anon->arcs_size, blksz);
3303 rw_exit(&buf->b_lock);
3304 ASSERT(refcount_count(&hdr->b_refcnt) == 1);
3305 ASSERT(!list_link_active(&hdr->b_arc_node));
3306 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3307 arc_change_state(arc_anon, hdr, hash_lock);
3308 hdr->b_arc_access = 0;
3309 mutex_exit(hash_lock);
3311 bzero(&hdr->b_dva, sizeof (dva_t));
3316 buf->b_efunc = NULL;
3317 buf->b_private = NULL;
3321 list_remove(l2hdr->b_dev->l2ad_buflist, hdr);
3322 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t));
3323 ARCSTAT_INCR(arcstat_l2_size, -buf_size);
3324 mutex_exit(&l2arc_buflist_mtx);
3329 arc_released(arc_buf_t *buf)
3333 rw_enter(&buf->b_lock, RW_READER);
3334 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon);
3335 rw_exit(&buf->b_lock);
3340 arc_has_callback(arc_buf_t *buf)
3344 rw_enter(&buf->b_lock, RW_READER);
3345 callback = (buf->b_efunc != NULL);
3346 rw_exit(&buf->b_lock);
3352 arc_referenced(arc_buf_t *buf)
3356 rw_enter(&buf->b_lock, RW_READER);
3357 referenced = (refcount_count(&buf->b_hdr->b_refcnt));
3358 rw_exit(&buf->b_lock);
3359 return (referenced);
3364 arc_write_ready(zio_t *zio)
3366 arc_write_callback_t *callback = zio->io_private;
3367 arc_buf_t *buf = callback->awcb_buf;
3368 arc_buf_hdr_t *hdr = buf->b_hdr;
3370 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt));
3371 callback->awcb_ready(zio, buf, callback->awcb_private);
3374 * If the IO is already in progress, then this is a re-write
3375 * attempt, so we need to thaw and re-compute the cksum.
3376 * It is the responsibility of the callback to handle the
3377 * accounting for any re-write attempt.
3379 if (HDR_IO_IN_PROGRESS(hdr)) {
3380 mutex_enter(&hdr->b_freeze_lock);
3381 if (hdr->b_freeze_cksum != NULL) {
3382 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
3383 hdr->b_freeze_cksum = NULL;
3385 mutex_exit(&hdr->b_freeze_lock);
3387 arc_cksum_compute(buf, B_FALSE);
3388 hdr->b_flags |= ARC_IO_IN_PROGRESS;
3392 arc_write_done(zio_t *zio)
3394 arc_write_callback_t *callback = zio->io_private;
3395 arc_buf_t *buf = callback->awcb_buf;
3396 arc_buf_hdr_t *hdr = buf->b_hdr;
3400 hdr->b_dva = *BP_IDENTITY(zio->io_bp);
3401 hdr->b_birth = zio->io_bp->blk_birth;
3402 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0];
3404 * If the block to be written was all-zero, we may have
3405 * compressed it away. In this case no write was performed
3406 * so there will be no dva/birth-date/checksum. The buffer
3407 * must therefor remain anonymous (and uncached).
3409 if (!BUF_EMPTY(hdr)) {
3410 arc_buf_hdr_t *exists;
3411 kmutex_t *hash_lock;
3413 arc_cksum_verify(buf);
3415 exists = buf_hash_insert(hdr, &hash_lock);
3418 * This can only happen if we overwrite for
3419 * sync-to-convergence, because we remove
3420 * buffers from the hash table when we arc_free().
3422 ASSERT(zio->io_flags & ZIO_FLAG_IO_REWRITE);
3423 ASSERT(DVA_EQUAL(BP_IDENTITY(&zio->io_bp_orig),
3424 BP_IDENTITY(zio->io_bp)));
3425 ASSERT3U(zio->io_bp_orig.blk_birth, ==,
3426 zio->io_bp->blk_birth);
3428 ASSERT(refcount_is_zero(&exists->b_refcnt));
3429 arc_change_state(arc_anon, exists, hash_lock);
3430 mutex_exit(hash_lock);
3431 arc_hdr_destroy(exists);
3432 exists = buf_hash_insert(hdr, &hash_lock);
3433 ASSERT3P(exists, ==, NULL);
3435 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3436 /* if it's not anon, we are doing a scrub */
3437 if (hdr->b_state == arc_anon)
3438 arc_access(hdr, hash_lock);
3439 mutex_exit(hash_lock);
3440 } else if (callback->awcb_done == NULL) {
3443 * This is an anonymous buffer with no user callback,
3444 * destroy it if there are no active references.
3446 mutex_enter(&arc_eviction_mtx);
3447 destroy_hdr = refcount_is_zero(&hdr->b_refcnt);
3448 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3449 mutex_exit(&arc_eviction_mtx);
3451 arc_hdr_destroy(hdr);
3453 hdr->b_flags &= ~ARC_IO_IN_PROGRESS;
3455 hdr->b_flags &= ~ARC_STORED;
3457 if (callback->awcb_done) {
3458 ASSERT(!refcount_is_zero(&hdr->b_refcnt));
3459 callback->awcb_done(zio, buf, callback->awcb_private);
3462 kmem_free(callback, sizeof (arc_write_callback_t));
3466 write_policy(spa_t *spa, const writeprops_t *wp, zio_prop_t *zp)
3468 boolean_t ismd = (wp->wp_level > 0 || dmu_ot[wp->wp_type].ot_metadata);
3470 /* Determine checksum setting */
3473 * Metadata always gets checksummed. If the data
3474 * checksum is multi-bit correctable, and it's not a
3475 * ZBT-style checksum, then it's suitable for metadata
3476 * as well. Otherwise, the metadata checksum defaults
3479 if (zio_checksum_table[wp->wp_oschecksum].ci_correctable &&
3480 !zio_checksum_table[wp->wp_oschecksum].ci_zbt)
3481 zp->zp_checksum = wp->wp_oschecksum;
3483 zp->zp_checksum = ZIO_CHECKSUM_FLETCHER_4;
3485 zp->zp_checksum = zio_checksum_select(wp->wp_dnchecksum,
3489 /* Determine compression setting */
3492 * XXX -- we should design a compression algorithm
3493 * that specializes in arrays of bps.
3495 zp->zp_compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
3498 zp->zp_compress = zio_compress_select(wp->wp_dncompress,
3502 zp->zp_type = wp->wp_type;
3503 zp->zp_level = wp->wp_level;
3504 zp->zp_ndvas = MIN(wp->wp_copies + ismd, spa_max_replication(spa));
3508 arc_write(zio_t *pio, spa_t *spa, const writeprops_t *wp,
3509 boolean_t l2arc, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
3510 arc_done_func_t *ready, arc_done_func_t *done, void *private, int priority,
3511 int zio_flags, const zbookmark_t *zb)
3513 arc_buf_hdr_t *hdr = buf->b_hdr;
3514 arc_write_callback_t *callback;
3518 ASSERT(ready != NULL);
3519 ASSERT(!HDR_IO_ERROR(hdr));
3520 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0);
3521 ASSERT(hdr->b_acb == 0);
3523 hdr->b_flags |= ARC_L2CACHE;
3524 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP);
3525 callback->awcb_ready = ready;
3526 callback->awcb_done = done;
3527 callback->awcb_private = private;
3528 callback->awcb_buf = buf;
3530 write_policy(spa, wp, &zp);
3531 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, &zp,
3532 arc_write_ready, arc_write_done, callback, priority, zio_flags, zb);
3538 arc_free(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
3539 zio_done_func_t *done, void *private, uint32_t arc_flags)
3542 kmutex_t *hash_lock;
3546 * If this buffer is in the cache, release it, so it
3549 ab = buf_hash_find(spa, BP_IDENTITY(bp), bp->blk_birth, &hash_lock);
3552 * The checksum of blocks to free is not always
3553 * preserved (eg. on the deadlist). However, if it is
3554 * nonzero, it should match what we have in the cache.
3556 ASSERT(bp->blk_cksum.zc_word[0] == 0 ||
3557 bp->blk_cksum.zc_word[0] == ab->b_cksum0 ||
3558 bp->blk_fill == BLK_FILL_ALREADY_FREED);
3560 if (ab->b_state != arc_anon)
3561 arc_change_state(arc_anon, ab, hash_lock);
3562 if (HDR_IO_IN_PROGRESS(ab)) {
3564 * This should only happen when we prefetch.
3566 ASSERT(ab->b_flags & ARC_PREFETCH);
3567 ASSERT3U(ab->b_datacnt, ==, 1);
3568 ab->b_flags |= ARC_FREED_IN_READ;
3569 if (HDR_IN_HASH_TABLE(ab))
3570 buf_hash_remove(ab);
3571 ab->b_arc_access = 0;
3572 bzero(&ab->b_dva, sizeof (dva_t));
3575 ab->b_buf->b_efunc = NULL;
3576 ab->b_buf->b_private = NULL;
3577 mutex_exit(hash_lock);
3578 } else if (refcount_is_zero(&ab->b_refcnt)) {
3579 ab->b_flags |= ARC_FREE_IN_PROGRESS;
3580 mutex_exit(hash_lock);
3581 arc_hdr_destroy(ab);
3582 ARCSTAT_BUMP(arcstat_deleted);
3585 * We still have an active reference on this
3586 * buffer. This can happen, e.g., from
3587 * dbuf_unoverride().
3589 ASSERT(!HDR_IN_HASH_TABLE(ab));
3590 ab->b_arc_access = 0;
3591 bzero(&ab->b_dva, sizeof (dva_t));
3594 ab->b_buf->b_efunc = NULL;
3595 ab->b_buf->b_private = NULL;
3596 mutex_exit(hash_lock);
3600 zio = zio_free(pio, spa, txg, bp, done, private, ZIO_FLAG_MUSTSUCCEED);
3602 if (arc_flags & ARC_WAIT)
3603 return (zio_wait(zio));
3605 ASSERT(arc_flags & ARC_NOWAIT);
3612 arc_memory_throttle(uint64_t reserve, uint64_t txg)
3615 uint64_t inflight_data = arc_anon->arcs_size;
3616 uint64_t available_memory = ptoa((uintmax_t)cnt.v_free_count);
3617 static uint64_t page_load = 0;
3618 static uint64_t last_txg = 0;
3623 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
3626 if (available_memory >= zfs_write_limit_max)
3629 if (txg > last_txg) {
3634 * If we are in pageout, we know that memory is already tight,
3635 * the arc is already going to be evicting, so we just want to
3636 * continue to let page writes occur as quickly as possible.
3638 if (curproc == pageproc) {
3639 if (page_load > available_memory / 4)
3641 /* Note: reserve is inflated, so we deflate */
3642 page_load += reserve / 8;
3644 } else if (page_load > 0 && arc_reclaim_needed()) {
3645 /* memory is low, delay before restarting */
3646 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3651 if (arc_size > arc_c_min) {
3652 uint64_t evictable_memory =
3653 arc_mru->arcs_lsize[ARC_BUFC_DATA] +
3654 arc_mru->arcs_lsize[ARC_BUFC_METADATA] +
3655 arc_mfu->arcs_lsize[ARC_BUFC_DATA] +
3656 arc_mfu->arcs_lsize[ARC_BUFC_METADATA];
3657 available_memory += MIN(evictable_memory, arc_size - arc_c_min);
3660 if (inflight_data > available_memory / 4) {
3661 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
3669 arc_tempreserve_clear(uint64_t reserve)
3671 atomic_add_64(&arc_tempreserve, -reserve);
3672 ASSERT((int64_t)arc_tempreserve >= 0);
3676 arc_tempreserve_space(uint64_t reserve, uint64_t txg)
3682 * Once in a while, fail for no reason. Everything should cope.
3684 if (spa_get_random(10000) == 0) {
3685 dprintf("forcing random failure\n");
3689 if (reserve > arc_c/4 && !arc_no_grow)
3690 arc_c = MIN(arc_c_max, reserve * 4);
3691 if (reserve > arc_c)
3695 * Writes will, almost always, require additional memory allocations
3696 * in order to compress/encrypt/etc the data. We therefor need to
3697 * make sure that there is sufficient available memory for this.
3699 if (error = arc_memory_throttle(reserve, txg))
3703 * Throttle writes when the amount of dirty data in the cache
3704 * gets too large. We try to keep the cache less than half full
3705 * of dirty blocks so that our sync times don't grow too large.
3706 * Note: if two requests come in concurrently, we might let them
3707 * both succeed, when one of them should fail. Not a huge deal.
3709 if (reserve + arc_tempreserve + arc_anon->arcs_size > arc_c / 2 &&
3710 arc_anon->arcs_size > arc_c / 4) {
3711 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
3712 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
3713 arc_tempreserve>>10,
3714 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10,
3715 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10,
3716 reserve>>10, arc_c>>10);
3719 atomic_add_64(&arc_tempreserve, reserve);
3723 static kmutex_t arc_lowmem_lock;
3725 static eventhandler_tag arc_event_lowmem = NULL;
3728 arc_lowmem(void *arg __unused, int howto __unused)
3731 /* Serialize access via arc_lowmem_lock. */
3732 mutex_enter(&arc_lowmem_lock);
3734 cv_signal(&arc_reclaim_thr_cv);
3736 tsleep(&needfree, 0, "zfs:lowmem", hz / 5);
3737 mutex_exit(&arc_lowmem_lock);
3744 int prefetch_tunable_set = 0;
3747 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL);
3748 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL);
3749 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL);
3751 /* Convert seconds to clock ticks */
3752 arc_min_prefetch_lifespan = 1 * hz;
3754 /* Start out with 1/8 of all memory */
3755 arc_c = kmem_size() / 8;
3759 * On architectures where the physical memory can be larger
3760 * than the addressable space (intel in 32-bit mode), we may
3761 * need to limit the cache to 1/8 of VM size.
3763 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8);
3766 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */
3767 arc_c_min = MAX(arc_c / 4, 64<<18);
3768 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */
3769 if (arc_c * 8 >= 1<<30)
3770 arc_c_max = (arc_c * 8) - (1<<30);
3772 arc_c_max = arc_c_min;
3773 arc_c_max = MAX(arc_c * 5, arc_c_max);
3776 * Allow the tunables to override our calculations if they are
3777 * reasonable (ie. over 16MB)
3779 if (zfs_arc_max >= 64<<18 && zfs_arc_max < kmem_size())
3780 arc_c_max = zfs_arc_max;
3781 if (zfs_arc_min >= 64<<18 && zfs_arc_min <= arc_c_max)
3782 arc_c_min = zfs_arc_min;
3785 arc_p = (arc_c >> 1);
3787 /* limit meta-data to 1/4 of the arc capacity */
3788 arc_meta_limit = arc_c_max / 4;
3790 /* Allow the tunable to override if it is reasonable */
3791 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
3792 arc_meta_limit = zfs_arc_meta_limit;
3794 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
3795 arc_c_min = arc_meta_limit / 2;
3797 if (zfs_arc_grow_retry > 0)
3798 arc_grow_retry = zfs_arc_grow_retry;
3800 if (zfs_arc_shrink_shift > 0)
3801 arc_shrink_shift = zfs_arc_shrink_shift;
3803 if (zfs_arc_p_min_shift > 0)
3804 arc_p_min_shift = zfs_arc_p_min_shift;
3806 /* if kmem_flags are set, lets try to use less memory */
3807 if (kmem_debugging())
3809 if (arc_c < arc_c_min)
3812 zfs_arc_min = arc_c_min;
3813 zfs_arc_max = arc_c_max;
3815 arc_anon = &ARC_anon;
3817 arc_mru_ghost = &ARC_mru_ghost;
3819 arc_mfu_ghost = &ARC_mfu_ghost;
3820 arc_l2c_only = &ARC_l2c_only;
3823 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
3824 mutex_init(&arc_anon->arcs_locks[i].arcs_lock,
3825 NULL, MUTEX_DEFAULT, NULL);
3826 mutex_init(&arc_mru->arcs_locks[i].arcs_lock,
3827 NULL, MUTEX_DEFAULT, NULL);
3828 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock,
3829 NULL, MUTEX_DEFAULT, NULL);
3830 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock,
3831 NULL, MUTEX_DEFAULT, NULL);
3832 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock,
3833 NULL, MUTEX_DEFAULT, NULL);
3834 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock,
3835 NULL, MUTEX_DEFAULT, NULL);
3837 list_create(&arc_mru->arcs_lists[i],
3838 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3839 list_create(&arc_mru_ghost->arcs_lists[i],
3840 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3841 list_create(&arc_mfu->arcs_lists[i],
3842 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3843 list_create(&arc_mfu_ghost->arcs_lists[i],
3844 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3845 list_create(&arc_mfu_ghost->arcs_lists[i],
3846 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3847 list_create(&arc_l2c_only->arcs_lists[i],
3848 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node));
3853 arc_thread_exit = 0;
3854 arc_eviction_list = NULL;
3855 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL);
3856 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t));
3858 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED,
3859 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
3861 if (arc_ksp != NULL) {
3862 arc_ksp->ks_data = &arc_stats;
3863 kstat_install(arc_ksp);
3866 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0,
3867 TS_RUN, minclsyspri);
3870 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL,
3871 EVENTHANDLER_PRI_FIRST);
3877 if (zfs_write_limit_max == 0)
3878 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
3880 zfs_write_limit_shift = 0;
3881 mutex_init(&zfs_write_limit_lock, NULL, MUTEX_DEFAULT, NULL);
3884 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable))
3885 prefetch_tunable_set = 1;
3888 if (prefetch_tunable_set == 0) {
3889 printf("ZFS NOTICE: Prefetch is disabled by default on i386 "
3891 printf(" add \"vfs.zfs.prefetch_disable=0\" "
3892 "to /boot/loader.conf.\n");
3893 zfs_prefetch_disable=1;
3896 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) &&
3897 prefetch_tunable_set == 0) {
3898 printf("ZFS NOTICE: Prefetch is disabled by default if less "
3899 "than 4GB of RAM is present;\n"
3900 " to enable, add \"vfs.zfs.prefetch_disable=0\" "
3901 "to /boot/loader.conf.\n");
3902 zfs_prefetch_disable=1;
3905 /* Warn about ZFS memory and address space requirements. */
3906 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) {
3907 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; "
3908 "expect unstable behavior.\n");
3910 if (kmem_size() < 512 * (1 << 20)) {
3911 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; "
3912 "expect unstable behavior.\n");
3913 printf(" Consider tuning vm.kmem_size and "
3914 "vm.kmem_size_max\n");
3915 printf(" in /boot/loader.conf.\n");
3925 mutex_enter(&arc_reclaim_thr_lock);
3926 arc_thread_exit = 1;
3927 cv_signal(&arc_reclaim_thr_cv);
3928 while (arc_thread_exit != 0)
3929 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock);
3930 mutex_exit(&arc_reclaim_thr_lock);
3936 if (arc_ksp != NULL) {
3937 kstat_delete(arc_ksp);
3941 mutex_destroy(&arc_eviction_mtx);
3942 mutex_destroy(&arc_reclaim_thr_lock);
3943 cv_destroy(&arc_reclaim_thr_cv);
3945 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) {
3946 list_destroy(&arc_mru->arcs_lists[i]);
3947 list_destroy(&arc_mru_ghost->arcs_lists[i]);
3948 list_destroy(&arc_mfu->arcs_lists[i]);
3949 list_destroy(&arc_mfu_ghost->arcs_lists[i]);
3950 list_destroy(&arc_l2c_only->arcs_lists[i]);
3952 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock);
3953 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock);
3954 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock);
3955 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock);
3956 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock);
3957 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock);
3960 mutex_destroy(&zfs_write_limit_lock);
3964 mutex_destroy(&arc_lowmem_lock);
3966 if (arc_event_lowmem != NULL)
3967 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem);
3974 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
3975 * It uses dedicated storage devices to hold cached data, which are populated
3976 * using large infrequent writes. The main role of this cache is to boost
3977 * the performance of random read workloads. The intended L2ARC devices
3978 * include short-stroked disks, solid state disks, and other media with
3979 * substantially faster read latency than disk.
3981 * +-----------------------+
3983 * +-----------------------+
3986 * l2arc_feed_thread() arc_read()
3990 * +---------------+ |
3992 * +---------------+ |
3997 * +-------+ +-------+
3999 * | cache | | cache |
4000 * +-------+ +-------+
4001 * +=========+ .-----.
4002 * : L2ARC : |-_____-|
4003 * : devices : | Disks |
4004 * +=========+ `-_____-'
4006 * Read requests are satisfied from the following sources, in order:
4009 * 2) vdev cache of L2ARC devices
4011 * 4) vdev cache of disks
4014 * Some L2ARC device types exhibit extremely slow write performance.
4015 * To accommodate for this there are some significant differences between
4016 * the L2ARC and traditional cache design:
4018 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from
4019 * the ARC behave as usual, freeing buffers and placing headers on ghost
4020 * lists. The ARC does not send buffers to the L2ARC during eviction as
4021 * this would add inflated write latencies for all ARC memory pressure.
4023 * 2. The L2ARC attempts to cache data from the ARC before it is evicted.
4024 * It does this by periodically scanning buffers from the eviction-end of
4025 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are
4026 * not already there. It scans until a headroom of buffers is satisfied,
4027 * which itself is a buffer for ARC eviction. The thread that does this is
4028 * l2arc_feed_thread(), illustrated below; example sizes are included to
4029 * provide a better sense of ratio than this diagram:
4032 * +---------------------+----------+
4033 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC
4034 * +---------------------+----------+ | o L2ARC eligible
4035 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer
4036 * +---------------------+----------+ |
4037 * 15.9 Gbytes ^ 32 Mbytes |
4039 * l2arc_feed_thread()
4041 * l2arc write hand <--[oooo]--'
4045 * +==============================+
4046 * L2ARC dev |####|#|###|###| |####| ... |
4047 * +==============================+
4050 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of
4051 * evicted, then the L2ARC has cached a buffer much sooner than it probably
4052 * needed to, potentially wasting L2ARC device bandwidth and storage. It is
4053 * safe to say that this is an uncommon case, since buffers at the end of
4054 * the ARC lists have moved there due to inactivity.
4056 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom,
4057 * then the L2ARC simply misses copying some buffers. This serves as a
4058 * pressure valve to prevent heavy read workloads from both stalling the ARC
4059 * with waits and clogging the L2ARC with writes. This also helps prevent
4060 * the potential for the L2ARC to churn if it attempts to cache content too
4061 * quickly, such as during backups of the entire pool.
4063 * 5. After system boot and before the ARC has filled main memory, there are
4064 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru
4065 * lists can remain mostly static. Instead of searching from tail of these
4066 * lists as pictured, the l2arc_feed_thread() will search from the list heads
4067 * for eligible buffers, greatly increasing its chance of finding them.
4069 * The L2ARC device write speed is also boosted during this time so that
4070 * the L2ARC warms up faster. Since there have been no ARC evictions yet,
4071 * there are no L2ARC reads, and no fear of degrading read performance
4072 * through increased writes.
4074 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that
4075 * the vdev queue can aggregate them into larger and fewer writes. Each
4076 * device is written to in a rotor fashion, sweeping writes through
4077 * available space then repeating.
4079 * 7. The L2ARC does not store dirty content. It never needs to flush
4080 * write buffers back to disk based storage.
4082 * 8. If an ARC buffer is written (and dirtied) which also exists in the
4083 * L2ARC, the now stale L2ARC buffer is immediately dropped.
4085 * The performance of the L2ARC can be tweaked by a number of tunables, which
4086 * may be necessary for different workloads:
4088 * l2arc_write_max max write bytes per interval
4089 * l2arc_write_boost extra write bytes during device warmup
4090 * l2arc_noprefetch skip caching prefetched buffers
4091 * l2arc_headroom number of max device writes to precache
4092 * l2arc_feed_secs seconds between L2ARC writing
4094 * Tunables may be removed or added as future performance improvements are
4095 * integrated, and also may become zpool properties.
4097 * There are three key functions that control how the L2ARC warms up:
4099 * l2arc_write_eligible() check if a buffer is eligible to cache
4100 * l2arc_write_size() calculate how much to write
4101 * l2arc_write_interval() calculate sleep delay between writes
4103 * These three functions determine what to write, how much, and how quickly
4108 l2arc_write_eligible(spa_t *spa, arc_buf_hdr_t *ab)
4111 * A buffer is *not* eligible for the L2ARC if it:
4112 * 1. belongs to a different spa.
4113 * 2. is already cached on the L2ARC.
4114 * 3. has an I/O in progress (it may be an incomplete read).
4115 * 4. is flagged not eligible (zfs property).
4117 if (ab->b_spa != spa) {
4118 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch);
4121 if (ab->b_l2hdr != NULL) {
4122 ARCSTAT_BUMP(arcstat_l2_write_in_l2);
4125 if (HDR_IO_IN_PROGRESS(ab)) {
4126 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress);
4129 if (!HDR_L2CACHE(ab)) {
4130 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable);
4138 l2arc_write_size(l2arc_dev_t *dev)
4142 size = dev->l2ad_write;
4144 if (arc_warm == B_FALSE)
4145 size += dev->l2ad_boost;
4152 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote)
4154 clock_t interval, next;
4157 * If the ARC lists are busy, increase our write rate; if the
4158 * lists are stale, idle back. This is achieved by checking
4159 * how much we previously wrote - if it was more than half of
4160 * what we wanted, schedule the next write much sooner.
4162 if (l2arc_feed_again && wrote > (wanted / 2))
4163 interval = (hz * l2arc_feed_min_ms) / 1000;
4165 interval = hz * l2arc_feed_secs;
4167 next = MAX(LBOLT, MIN(LBOLT + interval, began + interval));
4173 l2arc_hdr_stat_add(void)
4175 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE);
4176 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE);
4180 l2arc_hdr_stat_remove(void)
4182 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE));
4183 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE);
4187 * Cycle through L2ARC devices. This is how L2ARC load balances.
4188 * If a device is returned, this also returns holding the spa config lock.
4190 static l2arc_dev_t *
4191 l2arc_dev_get_next(void)
4193 l2arc_dev_t *first, *next = NULL;
4196 * Lock out the removal of spas (spa_namespace_lock), then removal
4197 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
4198 * both locks will be dropped and a spa config lock held instead.
4200 mutex_enter(&spa_namespace_lock);
4201 mutex_enter(&l2arc_dev_mtx);
4203 /* if there are no vdevs, there is nothing to do */
4204 if (l2arc_ndev == 0)
4208 next = l2arc_dev_last;
4210 /* loop around the list looking for a non-faulted vdev */
4212 next = list_head(l2arc_dev_list);
4214 next = list_next(l2arc_dev_list, next);
4216 next = list_head(l2arc_dev_list);
4219 /* if we have come back to the start, bail out */
4222 else if (next == first)
4225 } while (vdev_is_dead(next->l2ad_vdev));
4227 /* if we were unable to find any usable vdevs, return NULL */
4228 if (vdev_is_dead(next->l2ad_vdev))
4231 l2arc_dev_last = next;
4234 mutex_exit(&l2arc_dev_mtx);
4237 * Grab the config lock to prevent the 'next' device from being
4238 * removed while we are writing to it.
4241 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
4242 mutex_exit(&spa_namespace_lock);
4248 * Free buffers that were tagged for destruction.
4251 l2arc_do_free_on_write()
4254 l2arc_data_free_t *df, *df_prev;
4256 mutex_enter(&l2arc_free_on_write_mtx);
4257 buflist = l2arc_free_on_write;
4259 for (df = list_tail(buflist); df; df = df_prev) {
4260 df_prev = list_prev(buflist, df);
4261 ASSERT(df->l2df_data != NULL);
4262 ASSERT(df->l2df_func != NULL);
4263 df->l2df_func(df->l2df_data, df->l2df_size);
4264 list_remove(buflist, df);
4265 kmem_free(df, sizeof (l2arc_data_free_t));
4268 mutex_exit(&l2arc_free_on_write_mtx);
4272 * A write to a cache device has completed. Update all headers to allow
4273 * reads from these buffers to begin.
4276 l2arc_write_done(zio_t *zio)
4278 l2arc_write_callback_t *cb;
4281 arc_buf_hdr_t *head, *ab, *ab_prev;
4282 l2arc_buf_hdr_t *abl2;
4283 kmutex_t *hash_lock;
4285 cb = zio->io_private;
4287 dev = cb->l2wcb_dev;
4288 ASSERT(dev != NULL);
4289 head = cb->l2wcb_head;
4290 ASSERT(head != NULL);
4291 buflist = dev->l2ad_buflist;
4292 ASSERT(buflist != NULL);
4293 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
4294 l2arc_write_callback_t *, cb);
4296 if (zio->io_error != 0)
4297 ARCSTAT_BUMP(arcstat_l2_writes_error);
4299 mutex_enter(&l2arc_buflist_mtx);
4302 * All writes completed, or an error was hit.
4304 for (ab = list_prev(buflist, head); ab; ab = ab_prev) {
4305 ab_prev = list_prev(buflist, ab);
4307 hash_lock = HDR_LOCK(ab);
4308 if (!mutex_tryenter(hash_lock)) {
4310 * This buffer misses out. It may be in a stage
4311 * of eviction. Its ARC_L2_WRITING flag will be
4312 * left set, denying reads to this buffer.
4314 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss);
4318 if (zio->io_error != 0) {
4320 * Error - drop L2ARC entry.
4322 list_remove(buflist, ab);
4325 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4326 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4330 * Allow ARC to begin reads to this L2ARC entry.
4332 ab->b_flags &= ~ARC_L2_WRITING;
4334 mutex_exit(hash_lock);
4337 atomic_inc_64(&l2arc_writes_done);
4338 list_remove(buflist, head);
4339 kmem_cache_free(hdr_cache, head);
4340 mutex_exit(&l2arc_buflist_mtx);
4342 l2arc_do_free_on_write();
4344 kmem_free(cb, sizeof (l2arc_write_callback_t));
4348 * A read to a cache device completed. Validate buffer contents before
4349 * handing over to the regular ARC routines.
4352 l2arc_read_done(zio_t *zio)
4354 l2arc_read_callback_t *cb;
4357 kmutex_t *hash_lock;
4360 ASSERT(zio->io_vd != NULL);
4361 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
4363 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
4365 cb = zio->io_private;
4367 buf = cb->l2rcb_buf;
4368 ASSERT(buf != NULL);
4370 ASSERT(hdr != NULL);
4372 hash_lock = HDR_LOCK(hdr);
4373 mutex_enter(hash_lock);
4376 * Check this survived the L2ARC journey.
4378 equal = arc_cksum_equal(buf);
4379 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) {
4380 mutex_exit(hash_lock);
4381 zio->io_private = buf;
4382 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */
4383 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */
4386 mutex_exit(hash_lock);
4388 * Buffer didn't survive caching. Increment stats and
4389 * reissue to the original storage device.
4391 if (zio->io_error != 0) {
4392 ARCSTAT_BUMP(arcstat_l2_io_error);
4394 zio->io_error = EIO;
4397 ARCSTAT_BUMP(arcstat_l2_cksum_bad);
4400 * If there's no waiter, issue an async i/o to the primary
4401 * storage now. If there *is* a waiter, the caller must
4402 * issue the i/o in a context where it's OK to block.
4404 if (zio->io_waiter == NULL)
4405 zio_nowait(zio_read(zio->io_parent,
4406 cb->l2rcb_spa, &cb->l2rcb_bp,
4407 buf->b_data, zio->io_size, arc_read_done, buf,
4408 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb));
4411 kmem_free(cb, sizeof (l2arc_read_callback_t));
4415 * This is the list priority from which the L2ARC will search for pages to
4416 * cache. This is used within loops (0..3) to cycle through lists in the
4417 * desired order. This order can have a significant effect on cache
4420 * Currently the metadata lists are hit first, MFU then MRU, followed by
4421 * the data lists. This function returns a locked list, and also returns
4425 l2arc_list_locked(int list_num, kmutex_t **lock)
4430 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS);
4432 if (list_num < ARC_BUFC_NUMMETADATALISTS) {
4434 list = &arc_mfu->arcs_lists[idx];
4435 *lock = ARCS_LOCK(arc_mfu, idx);
4436 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) {
4437 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4438 list = &arc_mru->arcs_lists[idx];
4439 *lock = ARCS_LOCK(arc_mru, idx);
4440 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 +
4441 ARC_BUFC_NUMDATALISTS)) {
4442 idx = list_num - ARC_BUFC_NUMMETADATALISTS;
4443 list = &arc_mfu->arcs_lists[idx];
4444 *lock = ARCS_LOCK(arc_mfu, idx);
4446 idx = list_num - ARC_BUFC_NUMLISTS;
4447 list = &arc_mru->arcs_lists[idx];
4448 *lock = ARCS_LOCK(arc_mru, idx);
4451 ASSERT(!(MUTEX_HELD(*lock)));
4457 * Evict buffers from the device write hand to the distance specified in
4458 * bytes. This distance may span populated buffers, it may span nothing.
4459 * This is clearing a region on the L2ARC device ready for writing.
4460 * If the 'all' boolean is set, every buffer is evicted.
4463 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
4466 l2arc_buf_hdr_t *abl2;
4467 arc_buf_hdr_t *ab, *ab_prev;
4468 kmutex_t *hash_lock;
4471 buflist = dev->l2ad_buflist;
4473 if (buflist == NULL)
4476 if (!all && dev->l2ad_first) {
4478 * This is the first sweep through the device. There is
4484 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
4486 * When nearing the end of the device, evict to the end
4487 * before the device write hand jumps to the start.
4489 taddr = dev->l2ad_end;
4491 taddr = dev->l2ad_hand + distance;
4493 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
4494 uint64_t, taddr, boolean_t, all);
4497 mutex_enter(&l2arc_buflist_mtx);
4498 for (ab = list_tail(buflist); ab; ab = ab_prev) {
4499 ab_prev = list_prev(buflist, ab);
4501 hash_lock = HDR_LOCK(ab);
4502 if (!mutex_tryenter(hash_lock)) {
4504 * Missed the hash lock. Retry.
4506 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry);
4507 mutex_exit(&l2arc_buflist_mtx);
4508 mutex_enter(hash_lock);
4509 mutex_exit(hash_lock);
4513 if (HDR_L2_WRITE_HEAD(ab)) {
4515 * We hit a write head node. Leave it for
4516 * l2arc_write_done().
4518 list_remove(buflist, ab);
4519 mutex_exit(hash_lock);
4523 if (!all && ab->b_l2hdr != NULL &&
4524 (ab->b_l2hdr->b_daddr > taddr ||
4525 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) {
4527 * We've evicted to the target address,
4528 * or the end of the device.
4530 mutex_exit(hash_lock);
4534 if (HDR_FREE_IN_PROGRESS(ab)) {
4536 * Already on the path to destruction.
4538 mutex_exit(hash_lock);
4542 if (ab->b_state == arc_l2c_only) {
4543 ASSERT(!HDR_L2_READING(ab));
4545 * This doesn't exist in the ARC. Destroy.
4546 * arc_hdr_destroy() will call list_remove()
4547 * and decrement arcstat_l2_size.
4549 arc_change_state(arc_anon, ab, hash_lock);
4550 arc_hdr_destroy(ab);
4553 * Invalidate issued or about to be issued
4554 * reads, since we may be about to write
4555 * over this location.
4557 if (HDR_L2_READING(ab)) {
4558 ARCSTAT_BUMP(arcstat_l2_evict_reading);
4559 ab->b_flags |= ARC_L2_EVICTED;
4563 * Tell ARC this no longer exists in L2ARC.
4565 if (ab->b_l2hdr != NULL) {
4568 kmem_free(abl2, sizeof (l2arc_buf_hdr_t));
4569 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size);
4571 list_remove(buflist, ab);
4574 * This may have been leftover after a
4577 ab->b_flags &= ~ARC_L2_WRITING;
4579 mutex_exit(hash_lock);
4581 mutex_exit(&l2arc_buflist_mtx);
4583 spa_l2cache_space_update(dev->l2ad_vdev, 0, -(taddr - dev->l2ad_evict));
4584 dev->l2ad_evict = taddr;
4588 * Find and write ARC buffers to the L2ARC device.
4590 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid
4591 * for reading until they have completed writing.
4594 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
4596 arc_buf_hdr_t *ab, *ab_prev, *head;
4597 l2arc_buf_hdr_t *hdrl2;
4599 uint64_t passed_sz, write_sz, buf_sz, headroom;
4601 kmutex_t *hash_lock, *list_lock;
4602 boolean_t have_lock, full;
4603 l2arc_write_callback_t *cb;
4607 ASSERT(dev->l2ad_vdev != NULL);
4612 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE);
4613 head->b_flags |= ARC_L2_WRITE_HEAD;
4615 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter);
4617 * Copy buffers for L2ARC writing.
4619 mutex_enter(&l2arc_buflist_mtx);
4620 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) {
4621 list = l2arc_list_locked(try, &list_lock);
4623 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter);
4626 * L2ARC fast warmup.
4628 * Until the ARC is warm and starts to evict, read from the
4629 * head of the ARC lists rather than the tail.
4631 headroom = target_sz * l2arc_headroom;
4632 if (arc_warm == B_FALSE)
4633 ab = list_head(list);
4635 ab = list_tail(list);
4637 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter);
4639 for (; ab; ab = ab_prev) {
4640 if (arc_warm == B_FALSE)
4641 ab_prev = list_next(list, ab);
4643 ab_prev = list_prev(list, ab);
4644 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size);
4646 hash_lock = HDR_LOCK(ab);
4647 have_lock = MUTEX_HELD(hash_lock);
4648 if (!have_lock && !mutex_tryenter(hash_lock)) {
4649 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail);
4651 * Skip this buffer rather than waiting.
4656 passed_sz += ab->b_size;
4657 if (passed_sz > headroom) {
4661 mutex_exit(hash_lock);
4662 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom);
4666 if (!l2arc_write_eligible(spa, ab)) {
4667 mutex_exit(hash_lock);
4671 if ((write_sz + ab->b_size) > target_sz) {
4673 mutex_exit(hash_lock);
4674 ARCSTAT_BUMP(arcstat_l2_write_full);
4680 * Insert a dummy header on the buflist so
4681 * l2arc_write_done() can find where the
4682 * write buffers begin without searching.
4684 list_insert_head(dev->l2ad_buflist, head);
4687 sizeof (l2arc_write_callback_t), KM_SLEEP);
4688 cb->l2wcb_dev = dev;
4689 cb->l2wcb_head = head;
4690 pio = zio_root(spa, l2arc_write_done, cb,
4692 ARCSTAT_BUMP(arcstat_l2_write_pios);
4695 ARCSTAT_INCR(arcstat_l2_write_bytes_written, ab->b_size);
4697 * Create and add a new L2ARC header.
4699 hdrl2 = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP);
4701 hdrl2->b_daddr = dev->l2ad_hand;
4703 ab->b_flags |= ARC_L2_WRITING;
4704 ab->b_l2hdr = hdrl2;
4705 list_insert_head(dev->l2ad_buflist, ab);
4706 buf_data = ab->b_buf->b_data;
4707 buf_sz = ab->b_size;
4710 * Compute and store the buffer cksum before
4711 * writing. On debug the cksum is verified first.
4713 arc_cksum_verify(ab->b_buf);
4714 arc_cksum_compute(ab->b_buf, B_TRUE);
4716 mutex_exit(hash_lock);
4718 wzio = zio_write_phys(pio, dev->l2ad_vdev,
4719 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF,
4720 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE,
4721 ZIO_FLAG_CANFAIL, B_FALSE);
4723 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
4725 (void) zio_nowait(wzio);
4728 * Keep the clock hand suitably device-aligned.
4730 buf_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz);
4733 dev->l2ad_hand += buf_sz;
4736 mutex_exit(list_lock);
4741 mutex_exit(&l2arc_buflist_mtx);
4744 ASSERT3U(write_sz, ==, 0);
4745 kmem_cache_free(hdr_cache, head);
4749 ASSERT3U(write_sz, <=, target_sz);
4750 ARCSTAT_BUMP(arcstat_l2_writes_sent);
4751 ARCSTAT_INCR(arcstat_l2_write_bytes, write_sz);
4752 ARCSTAT_INCR(arcstat_l2_size, write_sz);
4753 spa_l2cache_space_update(dev->l2ad_vdev, 0, write_sz);
4756 * Bump device hand to the device start if it is approaching the end.
4757 * l2arc_evict() will already have evicted ahead for this case.
4759 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
4760 spa_l2cache_space_update(dev->l2ad_vdev, 0,
4761 dev->l2ad_end - dev->l2ad_hand);
4762 dev->l2ad_hand = dev->l2ad_start;
4763 dev->l2ad_evict = dev->l2ad_start;
4764 dev->l2ad_first = B_FALSE;
4767 dev->l2ad_writing = B_TRUE;
4768 (void) zio_wait(pio);
4769 dev->l2ad_writing = B_FALSE;
4775 * This thread feeds the L2ARC at regular intervals. This is the beating
4776 * heart of the L2ARC.
4779 l2arc_feed_thread(void *dummy __unused)
4784 uint64_t size, wrote;
4785 clock_t begin, next = LBOLT;
4787 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
4789 mutex_enter(&l2arc_feed_thr_lock);
4791 while (l2arc_thread_exit == 0) {
4792 CALLB_CPR_SAFE_BEGIN(&cpr);
4793 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
4795 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
4799 * Quick check for L2ARC devices.
4801 mutex_enter(&l2arc_dev_mtx);
4802 if (l2arc_ndev == 0) {
4803 mutex_exit(&l2arc_dev_mtx);
4806 mutex_exit(&l2arc_dev_mtx);
4810 * This selects the next l2arc device to write to, and in
4811 * doing so the next spa to feed from: dev->l2ad_spa. This
4812 * will return NULL if there are now no l2arc devices or if
4813 * they are all faulted.
4815 * If a device is returned, its spa's config lock is also
4816 * held to prevent device removal. l2arc_dev_get_next()
4817 * will grab and release l2arc_dev_mtx.
4819 if ((dev = l2arc_dev_get_next()) == NULL)
4822 spa = dev->l2ad_spa;
4823 ASSERT(spa != NULL);
4826 * Avoid contributing to memory pressure.
4828 if (arc_reclaim_needed()) {
4829 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
4830 spa_config_exit(spa, SCL_L2ARC, dev);
4834 ARCSTAT_BUMP(arcstat_l2_feeds);
4836 size = l2arc_write_size(dev);
4839 * Evict L2ARC buffers that will be overwritten.
4841 l2arc_evict(dev, size, B_FALSE);
4844 * Write ARC buffers.
4846 wrote = l2arc_write_buffers(spa, dev, size);
4849 * Calculate interval between writes.
4851 next = l2arc_write_interval(begin, size, wrote);
4852 spa_config_exit(spa, SCL_L2ARC, dev);
4855 l2arc_thread_exit = 0;
4856 cv_broadcast(&l2arc_feed_thr_cv);
4857 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
4862 l2arc_vdev_present(vdev_t *vd)
4866 mutex_enter(&l2arc_dev_mtx);
4867 for (dev = list_head(l2arc_dev_list); dev != NULL;
4868 dev = list_next(l2arc_dev_list, dev)) {
4869 if (dev->l2ad_vdev == vd)
4872 mutex_exit(&l2arc_dev_mtx);
4874 return (dev != NULL);
4878 * Add a vdev for use by the L2ARC. By this point the spa has already
4879 * validated the vdev and opened it.
4882 l2arc_add_vdev(spa_t *spa, vdev_t *vd, uint64_t start, uint64_t end)
4884 l2arc_dev_t *adddev;
4886 ASSERT(!l2arc_vdev_present(vd));
4889 * Create a new l2arc device entry.
4891 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
4892 adddev->l2ad_spa = spa;
4893 adddev->l2ad_vdev = vd;
4894 adddev->l2ad_write = l2arc_write_max;
4895 adddev->l2ad_boost = l2arc_write_boost;
4896 adddev->l2ad_start = start;
4897 adddev->l2ad_end = end;
4898 adddev->l2ad_hand = adddev->l2ad_start;
4899 adddev->l2ad_evict = adddev->l2ad_start;
4900 adddev->l2ad_first = B_TRUE;
4901 adddev->l2ad_writing = B_FALSE;
4902 ASSERT3U(adddev->l2ad_write, >, 0);
4905 * This is a list of all ARC buffers that are still valid on the
4908 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP);
4909 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
4910 offsetof(arc_buf_hdr_t, b_l2node));
4912 spa_l2cache_space_update(vd, adddev->l2ad_end - adddev->l2ad_hand, 0);
4915 * Add device to global list
4917 mutex_enter(&l2arc_dev_mtx);
4918 list_insert_head(l2arc_dev_list, adddev);
4919 atomic_inc_64(&l2arc_ndev);
4920 mutex_exit(&l2arc_dev_mtx);
4924 * Remove a vdev from the L2ARC.
4927 l2arc_remove_vdev(vdev_t *vd)
4929 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
4932 * Find the device by vdev
4934 mutex_enter(&l2arc_dev_mtx);
4935 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
4936 nextdev = list_next(l2arc_dev_list, dev);
4937 if (vd == dev->l2ad_vdev) {
4942 ASSERT(remdev != NULL);
4945 * Remove device from global list
4947 list_remove(l2arc_dev_list, remdev);
4948 l2arc_dev_last = NULL; /* may have been invalidated */
4949 atomic_dec_64(&l2arc_ndev);
4950 mutex_exit(&l2arc_dev_mtx);
4953 * Clear all buflists and ARC references. L2ARC device flush.
4955 l2arc_evict(remdev, 0, B_TRUE);
4956 list_destroy(remdev->l2ad_buflist);
4957 kmem_free(remdev->l2ad_buflist, sizeof (list_t));
4958 kmem_free(remdev, sizeof (l2arc_dev_t));
4964 l2arc_thread_exit = 0;
4966 l2arc_writes_sent = 0;
4967 l2arc_writes_done = 0;
4969 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
4970 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
4971 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
4972 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL);
4973 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
4975 l2arc_dev_list = &L2ARC_dev_list;
4976 l2arc_free_on_write = &L2ARC_free_on_write;
4977 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
4978 offsetof(l2arc_dev_t, l2ad_node));
4979 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
4980 offsetof(l2arc_data_free_t, l2df_list_node));
4987 * This is called from dmu_fini(), which is called from spa_fini();
4988 * Because of this, we can assume that all l2arc devices have
4989 * already been removed when the pools themselves were removed.
4992 l2arc_do_free_on_write();
4994 mutex_destroy(&l2arc_feed_thr_lock);
4995 cv_destroy(&l2arc_feed_thr_cv);
4996 mutex_destroy(&l2arc_dev_mtx);
4997 mutex_destroy(&l2arc_buflist_mtx);
4998 mutex_destroy(&l2arc_free_on_write_mtx);
5000 list_destroy(l2arc_dev_list);
5001 list_destroy(l2arc_free_on_write);
5007 if (!(spa_mode & FWRITE))
5010 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
5011 TS_RUN, minclsyspri);
5017 if (!(spa_mode & FWRITE))
5020 mutex_enter(&l2arc_feed_thr_lock);
5021 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
5022 l2arc_thread_exit = 1;
5023 while (l2arc_thread_exit != 0)
5024 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
5025 mutex_exit(&l2arc_feed_thr_lock);